Adaptive Control

Adaptive Control
Adaptive Control
Edited by
Kwanho You
In-Tech
intechweb.org
IV
Published by In-Tech
In-Tech
Kirchengasse 43/3, A-1070 Vienna, Austria
Hosti 80b, 51000 Rijeka, Croatia
Abstracting and non-profit use of the material is permitted with credit to the source. Statements and
opinions expressed in the chapters are these of the individual contributors and not necessarily those of
the editors or publisher. No responsibility is accepted for the accuracy of information contained in the
published articles. Publisher assumes no responsibility liability for any damage or injury to persons or
property arising out of the use of any materials, instructions, methods or ideas contained inside. After
this work has been published by the In-Teh, authors have the right to republish it, in whole or part, in
any publication of which they are an author or editor, and the make other personal use of the work.
2009 In-tech
www.intechweb.org
Additional copies can be obtained from:
publication@intechweb.org
First published January 2009
Printed in Croatia
Adaptive Control, Edited by Kwanho You

p. cm.
ISBN 978-953-7619-47-3
1. Adaptive Control I. Kwanho You
Preface
Adaptive control has been a remarkable field for industrial and academic research since
1950s. Since more and more adaptive algorithms are applied in various control applications,
it is considered as important for practical implementation. As it can be confirmed from the
increasing number of conferences and journals on adaptive control topics, it is certain that
the adaptive control is a significant guidance for technology development.
Also adaptive control has been believed as a breakthrough for realization of intelligent
control systems. Even with the parametric and model uncertainties, adaptive control enables
the control system to monitor the time varying changes and manipulate the controller for
desired performance. Therefore adaptive control has been considered to be essential for time
varying multivariable systems. Moreover, now with the advent of high-speed microprocessors, it is possible to implement the innovative adaptive algorithms even in real time situation.
With the efforts of many control researchers, the adaptive control field is abundant in
mathematical analysis, programming tools, and implementational algorithms. The authors
of each chapter in this book are the professionals in their areas. The results in the book
introduce their recent research results and provide new idea for improved performance in
various control application problems.
The book is organized in the following way. There are 16 chapters discussing the issues
of adaptive control application to model generation, adaptive estimation, output regulation
and feedback, electrical drives, optical communication, neural estimator, simulation and
implementation:
Chapter One: Automatic 3D Model Generation based on a Matching of Adaptive
Control Points, by N. Lee, J. Lee, G. Kim, and H. Choi
Chapter Two: Adaptive Estimation and Control for Systems with Parametric and
Nonparametric Uncertainties, by H. Ma and K. Lum
Chapter Three: Adaptive Output Regulation of Unknown Linear Systems with
Unknown Exosystems, by I. Mizumoto and Z. Iwai
Chapter Four: Output Feedback Direct Adaptive Control for a Two-Link Flexible
Robot Subject to Parameter Changes, by S. Ozcelik and E. Miranda
Chapter Five: Discrete Model Matching Adaptive Control for Potentially Inversely Non-Stable Continuous-Time Plants by Using Multirate Sampling, by S.
Alonso-Quesada and M. De la Sen
Chapter Six: Hybrid Schemes
Queiroz
for Adaptive Control Strategies, by R. Ribeiro and K.
VI
Chapter Seven: Adaptive Control for Systems with Randomly Missing Measurements in a Network Environment, by Y. Shi and H. Fang
Chapter Eight: Adaptive Control based on Neural Network, by S. Wei, Z. Lujin, Z.
Jinhai, and M. Siyi
Chapter Nine: Adaptive Control of the Electrical Drives with the Elastic Coupling
using Kalman Filter, by K. Szabat and T. Orlowska-Kowalska
Chapter Ten: Adaptive Control of Dynamic Systems with Sandwiched Hysteresis
based on Neural Estimator, by Y. Tan, R. Dong, and X. Zhao
Chapter Eleven: High-Speed Adaptive Control Technique based on Steepest De-
scent Method for Adaptive Chromatic Dispersion Compensation in Optical Communications, by K. Tanizawa and A. Hirose
Chapter Twelve: Adaptive Control of Piezoelectric Actuators with Unknown Hysteresis, by W. Xie, J. Fu, H. Yao, and C. Su
Chapter Thirteen: On the Adaptive Tracking Control of 3-D Overhead Crane Systems
Chapter Fourteen: Adaptive Inverse Optimal Control of a Magnetic Levitation
System, by Y. Satoh, H. Nakamura, H. Katayama, and H. Nishitani
Chapter Fifteen: Adaptive Precision Geolocation Algorithm with Multiple Model
Uncertainties, by W. Sung and K. You
Chapter Sixteen: Adaptive Control for a Class of Non-affine Nonlinear Systems
via Neural Networks, by Z. Tong
We expect that the readers have taken a basic course in automatic control, linear systems,
and sampled data systems. This book is tried to be written in a self-contained way for better
understanding. Since this book introduces the development and recent progress of the
theory and application of adaptive control research, it is useful as a reference especially for
industrial engineers, graduate students in advanced study, and the researchers who are related in adaptive control field such as electrical, aeronautical, and mechanical engineering.
Kwanho You
Sungkyunkwan University, Korea
VII
Contents
1.
Preface
Automatic 3D Model Generation based on a Matching of Adaptive Control

Points
001
Na-Young Lee, Joong-Jae Lee, Gye-Young Kim and Hyung-Il Choi
2.
Adaptive Estimation and Control for Systems with Parametric and

Nonparametric Uncertainties
015
Hongbin Ma and Kai-Yew Lum
3.
Adaptive output regulation of unknown linear systems with unknown

exosystems
065
Ikuro Mizumoto and Zenta Iwai
4.
Output Feedback Direct Adaptive Control for a Two-Link Flexible Robot

Subject to Parameter Changes
087
Selahattin Ozcelik and Elroy Miranda
5.
Discrete Model Matching Adaptive Control for Potentially Inversely NonStable Continuous-Time Plants by Using Multirate Sampling
113
S. Alonso-Quesada and M. De la Sen
6.
Hybrid Schemes for Adaptive Control Strategies
137
Ricardo Ribeiro and Kurios Queiroz
7.
Adaptive Control for Systems with Randomly Missing Measurements in a

Network Environment
161
Yang Shi and Huazhen Fang
8.
Adaptive Control Based On Neural Network
181
Sun Wei, Zhang Lujin, Zou Jinhai and Miao Siyi
9.
Adaptive control of the electrical drives with the elastic coupling using Kalman filter
205
Krzysztof Szabat and Teresa Orlowska-Kowalska
10. Adaptive Control of Dynamic Systems with Sandwiched Hysteresis Based

on Neural Estimator
Yonghong Tan, Ruili Dong and Xinlong Zhao
227
VIII
11. High-Speed Adaptive Control Technique Based on Steepest Descent

Method for Adaptive Chromatic Dispersion Compensation in Optical Communications
243
Ken Tanizawa and Akira Hirose
12. Adaptive Control of Piezoelectric Actuators with Unknown Hysteresis
259
Wen-Fang Xie, Jun Fu, Han Yao and C.-Y. Su
13. On the Adaptive Tracking Control of 3-D Overhead Crane Systems
277
Yang, Jung Hua
14.
Adaptive Inverse Optimal Control of a Magnetic Levitation System
307
Yasuyuki Satoh, Hisakazu Nakamura, Hitoshi Katayama and Hirokazu Nishitani
15. Adaptive Precision Geolocation Algorithm with Multiple Model Uncertainties 323
Wookjin Sung and Kwanho You
16. Adaptive Control for a Class of Non-affine Nonlinear Systems via Neural
Networks
Zhao Tong
337
1
Automatic 3D Model Generation based on a
Matching of Adaptive Control Points
Na-Young Lee1, Joong-Jae Lee2, Gye-Young Kim3 and Hyung-Il Choi4
2Center
1Radioisotope
Research Division, Korea Atomic Energy Research Institute

for Cognitive Robotics Research, Korea Institute of Sciene and Technology
3School of Computing, Soongsil University
4School of Media, Soongsil University
Republic of Korea
Abstract
The use of a 3D model helps to diagnosis and accurately locate a disease where it is neither
available, nor can be exactly measured in a 2D image. Therefore, highly accurate software
for a 3D model of vessel is required for an accurate diagnosis of patients. We have generated
standard vessel because the shape of the arterial is different for each individual vessel,
where the standard vessel can be adjusted to suit individual vessel. In this paper, we
propose a new approach for an automatic 3D model generation based on a matching of
adaptive control points. The proposed method is carried out in three steps. First, standard
and individual vessels are acquired. The standard vessel is acquired by a 3D model
projection, while the individual vessel of the first segmented vessel bifurcation is obtained.
Second is matching the corresponding control points between the standard and individual
vessels, where a set of control and corner points are automatically extracted using the Harris
corner detector. If control points exist between corner points in an individual vessel, it is
adaptively interpolated in the corresponding standard vessel which is proportional to the
distance ratio. And then, the control points of corresponding individual vessel match with
those control points of standard vessel. Finally, we apply warping on the standard vessel to
suit the individual vessel using the TPS (Thin Plate Spline) interpolation function. For
experiments, we used angiograms of various patients from a coronary angiography in
Sanggye Paik Hospital.
Keywords: Coronary angiography, adaptive control point, standard vessel, individual
vessel, vessel warping.
1. Introduction
X-ray angiography is the most frequently used imaging modality to diagnose coronary
artery diseases and to assess their severity. Traditionally, this assessment is performed
directly from the angiograms, and thus, can suffer from viewpoint orientation dependence
and lack of precision of quantitative measures due to magnification factor uncertainty
Adaptive Control
(Messenger et al., 2000), (Lee et al., 2006) and (Lee et al., 2007). 3D model is provided to
display the morphology of vessel malformations such as stenoses, arteriovenous
malformations and aneurysms (Holger et al., 2005). Consequently, accurate software for a
3D model of a coronary tree is required for an accurate diagnosis of patients. It could lead to
a fast diagnosis and make it more accurate in an ambiguous condition.
In this paper, we present an automatic 3D model generation based on a matching of
adaptive control points. Fig. 1 shows the overall flow of the proposed method for the 3D
modelling of the individual vessel. The proposed method is composed as the following
three steps: image acquisition, matching of the adaptive control points and the vessel
warping. In Section 2, the acquisitions of the input image in standard and individual vessels
are described. Section 3 presents the matching of the corresponding control points between
the standard and individual vessels. Section 4 describes the 3D modelling of the individual
vessel which is performed through a vessel warping with the corresponding control points.
Experimental results of the vessel transformation are given in Section 5. Finally, we present
the conclusion in Section 6.
Fig. 1. Overall flow of the system configuration
Automatic 3D Model Generation based on a Matching of Adaptive Control Points
2. Image Acquisition
We have generated a standard vessel because the shape of the arterial is different for each
individual vessel, where the standard vessel can be adjusted to suit the individual vessel
(Chalopin et al., 2001), (Lee et al., 2006) and (Lee et al., 2007). The proposed approach is
based on a 3D model of standard vessel which is built from a database that implemented a
Korean vascular system (Lee et al., 2006).
We have limited the scope of the main arteries for the 3D model of the standard vessel as
depicted in Fig. 2.
Fig. 2. Vessel scope of the database for the 3D model of the standard vessel
Table 1 shows the database of the coronary artery of Lt. main (Left Main Coronary Artery),
LAD (Left Anterior Descending) and LCX (Left Circumflex artery) information. This
database consists of 40 people with mixed gender information.
age
below 60 years of
old (male)
above 60 years of
old (male)
below 60 years of
old (female)
above 60 years of
old (female)
Lt. main
Os distal length Os
LAD
LCX
distal length Os distal length
48.45.9 4.30.4 4.10.5 9.94.2 3.80.4 3.60.4 17.05.2 3.50.4 3.30.3 19.26.1
67.55.4 4.50.5 4.40.4 8.43.8 3.90.3 3.60.3 17.25.8 3.60.4 3.40.4 24.68.9
44.919.9 3.71.8 3.41.6 10.66.2 3.31.5 3.11.4 14.15.5 2.91.3 2.81.2 21.39.2
70.7 4.4 4.30.7 4.10.6 12.57.9 3.50.6 3.40.5 22.37.3 3.30.4 3.10.3 27.53.7
Table 1. Database of the coronary artery
Adaptive Control
To quantify the 3D model of the coronary artery, the angles of the vessel bifurcation are
measured with references to LCX, Lt. main and LAD, as in Table 2. Ten individuals
regardless of their gender and age were selected randomly for measuring the angles of the
vessel bifurcation from six angiograms. The measurement results, and the average and
standard deviations of each individual measurement are shown in Table 2.
RAO30
CAUD30
RAO30
CRA30
AP0 LAO60 LAO60

AP0
CRA30 CRA30 CAUD30 CAUD30
69.17
123.31
38.64
61.32
84.01
50.98
53.58
72.02
23.80
51.75
99.73
73.92
77.28
97.70
21.20
57.72
100.71
71.33
94.12
24.67
22.38
81.99
75.6
69.57
64.12
33.25
31.24
40.97
135.00
61.87
55.34
51.27
41.8
80.89
119.84
57.14
71.93
79.32
50.92
87.72
114.71
58.22
67.70
59.14
31.84
58.93
92.36
70.16
85.98
60.85
35.77
54.45
118.80
78.93
10
47.39
60.26
34.50
47.39
67.52
34.79
Average
68.67
66.18
33.21
62.31
100.83
62.69
Standard
deviation
14.56
29.07
9.32
15.86
21.46
13.06
Table 2. Measured angles of the vessel bifurcation from six angiographies

Fig. 3 illustrates the results of the 3D model generation of the standard vessel from six
angiographies: RAO (Right Anterior Oblique)30 CAUD (Caudal)30, RAO30 CRA (Cranial
Anterior)30, AP (Anterior Posterior)0 CRA (Cranial Anterior)30, LAO (Left Anterior
Oblique)60 CRA30, LAO60 CAUD30, AP0 CAUD30.
View
RAO30
CAUD30
RAO30
CRA30
AP0
CRA30
LAO60
CRA30
LAO60
CAUD30
AP0
CAUD30
Angiogram
3D Model
View
Angiogram
3D Model
Fig. 3. 3D model generation of the standard vessel from six angiographies

Evaluating the angles of the vessel bifurcation from six angiographies can reduce the
possible measurement error which occurs when the angle from a single view is measured.
Adaptive Control
It is difficult to transform the standard vessel into individual vessel in a 3D space (Lee et al.,
2006) and (Lee et al., 2007). Therefore, we projected the 3D model of the standard vessel into
2D projection. Fig. 4 shows the projected images of the standard vessel on a 2D plane
through the projection. The projection result can be view as vertices or polygons based.
Fig. 4. Projection result for 2D image of standard vessel
3. Matching of the Adaptive Control Points

To transform a standard vessel into an individual vessel, it is important to match
corresponding control points (Lee et al., 2006) and (Lee et al., 2007). In this paper, we
extracted feature points of the vessel automatically and defined as control points (Lee et al.,
2006) and (Lee et al., 2007). Feature points mean is referred to the corner points of an object
or points with higher variance brightness compared to the surrounding pixels in an image,
which are differentiated from other points in an image. Such feature points can be defined in
many different ways in (Parker, 1996) and (Pitas, 2000). They are sometimes defined as
points that have a high gradient in different directions, or as points that have properties that
do not change in spite of specific transformations. Generally feature points can be divided
into three categories (Cizek et al., 2004). The first one uses a non-linear filter, such as the
SUSAN corner detector proposed by Smith (Woods et al., 1993) which relates each pixel to
an area centered by a pixel. In this area, it is called the SUSAN area; all the pixels have
similar intensities as the center pixel. If the center pixel is a feature point (some times a
feature point is also referred to as a "corner"), SUSAN area is the smallest one among the
pixels around it. A SUSAN corner detector can suppress a noise effectively without
derivating an image. The second one is based on a curvature, such as the Kitchen and
Rosenfeld's method (Maes et al., 1997). This kind of method needs to extract edges in
advance, and then elucidate the feature points using the information on the curvature of the
edges. The disadvantage of this method is required more needs a complicated computation,
e.g. curve on fitting, thus its processing speed is relatively slow. The third method is exploits
a change of the pixel intensity. A typical one is the Harris and Stephens' method (Pluim et
al., 2003). It produces a corner response through an eigenvalues analysis. Since it does not
need to use a slide window explicitly, its processing speed is very fast. Accordingly, this
paper used the Harris corner detector to find the control points of standard and individual
vessels (Lee et al., 2006) and (Lee, 2007).
3.1 Extraction of the Control Points
The Harris corner detector is a popular interest point detector due to its strong invariance
such as rotation, scale, illumination variation and image noise (Schmid et al., 2000) and
(Derpanis, 2004). It is based on the local auto-correlation function of a signal. The local autocorrelation function measures the local changes of the signal with patches shifted by a small
amount in different directions (Derpanis, 2004). However, the Harris corner detector has a
problem where it can mistake those non-corner points.
Fig. 5 shows extracted 9 control points in individual vessel by using the Harris corner
detector. We noticed that some of the extracted control points are non-corner points. To
solve this problem of the Harris corner detector, we extracted more control points of
individual vessel than standard vessel. Fig. 6 shows the extraction of control points from
individual and standard vessels.
Fig. 5. Extracted 9 control points in individual vessel

3.2 Extraction of Corner Points
We performed thinning by using the structural characteristics of vessel to find the corner
points among the control points of individual vessel which is extracted with the Harris
corner detector (Lee, 2007). Fig. 7 shows the thinning process for detection of corner points
in individual vessel.
(a) Segmented vessel

(b) Thinned vessel
Fig. 6. Thinning process for detection of corner points in individual vessel
A vascular tree can be divided into a set of elementary components, or primitives, which are
the vascular segments, and bifurcation (Wahle et al., 1994). Using this intuitive
Adaptive Control
representation, it is natural to describe the coronary tree by a graph structure (Chalopin et

al., 2001) and (Lee, 2007).
A vascular tree of thinned vessel consists of three vertices ( v po int ) and one bifurcation ( bif )
as the following equation (1). Here, vertices ( v po int ) are comprised a start point ( vstart _ po int )
and two end points ( vend _ po int 1 , vend _ po int 2 ).
I thin = { v po int , bif }

(1)
v po int = { vstart _ po int , vend _ po int 1, vend _ po int 2 }
If the reference point is a vertex, the closest two control points to the vertex are defined as
the corner points. If the reference point is a bifurcation, the three control points that are
closest to it after comparing the distances between the bifurcation and all control points are
defined as the corner points. As shown in Fig. 7, if the reference point is the vertex
( vstart _ po int ), v1 and v2 become the corner points; if the reference point is the bifurcation
( bif ), v6, v11 and v15 become the corner points (Lee, 2007).
v8
vend_ point2
v1
v7
v9
bif
vstart_ point
v6
v2
v3
v5
v4
vend _ point1
Fig. 7. Primitives of a vascular net

3.3 Adaptive Interpolation of the Control Points between Corner Points
Once the control points and corner points are extracted from an individual vessel, an
interpolation for a standard vessel is applied. For an accurate matching, the control points
are adaptively interpolated into the corresponding standard vessel in proportion to the
distance ratio if there are control points between the corner points in an individual vessel
(Lee, 2007).
Fig. 8 shows the process of an interpolation of the control points. Control points of a
standard vessel are adaptively interpolated by the distance rate between control point ( v3 )
and two corner points ( v2 , v4 ) of an individual vessel. Fig. 8 (a) shows the extracted control
points from an individual vessel, and (b) shows an example of control point interpolated
between a standard vessel and the corresponding corner points from (a) image.
(a) Individual vessel

(b) Standard vessel
Fig. 8. Interpolation of the control points for a standard vessel
Fig. 9 shows the result of extracting the control points by using the Harris corner detector to
the segmented vessel in the individual vessel and an adaptive interpolation of the
corresponding the control points in the standard vessel.
Fig. 9. Result of an adaptive interpolation of the corresponding control points
4. Vessel Warping
We have warped the standard vessel with respect to the individual vessel. Given the two
sets of corresponding control points, S = {s1, s2 ,K sm } and I = {i1,i2 ,K im } , the warping is applied
the standard vessel to suit the individual vessel. Here, S is a set of control points in the
standard vessel and I is a set of one in the individual vessel (Lee et al., 2006) and (Lee et al.,
2007).
Adaptive Control
10
Standard vessel warping was performed using the TPS (Thin-Plate-Spline) algorithm
(Bentoutou et al., 2002) from the two sets of control points.
The TPS is the interpolation functions that exactly represent a distortion at each feature
point, and for defining a minimum curvature surface between control points. A TPS
function is a flexible transformation that allows for a rotation, translation, scaling, and
skewing. It also allows for lines to bend according by the TPS model (Bentoutou et al., 2002).
Therefore, a large number of deformations can be characterized by the TPS model.
The TPS interpolation function can be written as equation (2).
m
h( x) = Ax + t + Wi K (|| x xi ||)
(2)
i =1
The variables A and t are the affine transformation parameters matrices, Wi are the weights
of the non-linear radial interpolation function K , and xi are the control points. The function
K (r ) is the solution of the biharmonic equation (2 K = 0) that satisfies the condition of a

bending energy minimization, namely K (r ) = r 2 log (r 2 ) .
The complete set of parameters, the interpolating registration transformation is defined, and
then it is used to transform the standard vessel. It should be noted that in order to be able to
carry out the warping of the standard vessel with respect to the individual vessel, it is
required to have a complete description of the TPS interpolation function (Lee et al., 2006)
and (Lee et al., 2007).
Fig. 10 shows the results of modifying the standard vessel to suit the individual vessel.
(a) Individual vessel

(b) Standard vessel
Fig. 10. Results of the warped vessel in standard vessel
(c) Warped vessel
5. Results of the Vessel Transformation

We simulated the system environment that is Microsoft Windows XP on a Pentium 3GHz,
Intel Corp. and the compiler VC++ 6.0 is used. The image of 512 512 is used for the
experimentation. Each image has a gray-value resolution of 8 bits, i.e., 256 gray levels.
Fig. 11 shows the 3D model of the standard vessel from six different angiographic views.
The results of the standard vessel warping using TPS algorithm to suit the individual vessel
is shown in Fig. 13.
Fig. 11. 3D model of the standard vessel in angiographic of six different views
Fig. 12. Result of standard vessel warping
11
12
Adaptive Control
Fig. 13 shows the result for an automatically 3D model generation of individual vessel.
Fig. 13. Result of 3D model generation for the individual vessel in six views
6. Conclusion
We proposed a fully automatic and effective algorithm to perform a 3D modelling of
individual vessel from angiograms in six views. This approach can be used to recover the
geometry of the main arteries. The 3D model of the vessel enables patients to visualize their
progress and improvement for a disease. Such a model should not only enhance the level of
reliability but also provide a fast and accurate identification. In order words, this method
can be expected to reduce the number of misdiagnosed cases (Lee et al., 2006) and (Lee et al.,
2007).
7. Acknowledgement
This Work was supported by Soongsil University and Korea Research Foundation Grant
(KRF-2006-005-J03801) Funded by Korean Government.
8. References
A, Venot.; J.F, Lebruchec. & J.C, Roucayrol. (1984). A new class of similarity measures for
robust image registration, Comput. Vision Graph. Image Process, pp. 176-184.
13
A, Wahle.; E, Wellnhofer.; I, Mugaragu.;H.U, Sauer.; H, Oswald. & E, Fleck. (1994). Accurate

3-D reconstruction and statistics for assessment of diffuse coronary artery disease,
Computers in Cardiology, IEEE Computer Society, Los Alamitos, CA, pp.669-672
B.G, Brown.; E, Bolson.; M, Frimer. & H, Dodge. (1977). Quantitative coronary arteriography
estimation of dimensions, hemodynamic resistance, and atheroma mass of
coronary artery lesions using the arteriogram and digital computation, Circulation,
vol. 55, pp. 329-337
C, Blondel.; R, Vaillant.; F, Devernary. ; G, Malandain. & N, Ayache. (2002). Automatic
trinocular 3D reconstruction of coronary artery centerlines from rotational X-ray
angiography, Computer Assisted Radiology and Surgery 2002 Proceedings, Paris, June
2002, Springer Publishers, Heidelberg
Claire, Chalopin.; Gerard, Finet. & Isabelle E, Magnin. (2001). Modeling the 3D coronary tree
for labeling purposes. Medical Image Analysis, pp.301-315
Chris, Harris. & Mike, Stephens. (1988). A combined corner and edge detector, Proceedings of
the Fourth Alvey Vision Conference, pp.147-151, Manchester
C, Lorenz.; S, Renisch.; S, Schlatholter. & T, Bulow. (2003). Simultaneous Segmentation and
Tree Reconstruction of the Coronary Arteries in MSCT Images, Int. Symposium
Medical Imaging, San Diego, Proc. SPIE Vol. 5032, pp. 167-177
C, Schmid.; R, Mohr. & C, Bauckhage. (2000). Evaluation of interest point detectors.
International Journal of Computer Vision, pp. 151-172
F.L, Bookstein. (1989). Principal warps: thin-plate splines and the decomposition of
deformations, IEEE-PAMI 11, pp. 567-585
F, Maes.; A, Collignon.; G, Marchal. & P, Suetens. (1997). Multimodality image registration
by maximization of mutual information, IEEE Transaction on Medical Imaging, Vol.
16, No. 2, pp.187-198
Holger, Schmitt.; Michael, Grass.; Rolf, Suurmond.; Thomas Kohler.; Volker, Rasche.; Stefan,
Hahnel. & Sabine, Heiland. (2005). Reconstruction of blood propagation in threedimensional rotational X-ray angiography. Computerized Medical Imaging and
Graphics, 29, pp. 507-520
I, Pitas. (2000). Digital image processing algorithms and applications, 1st Ed., John Wiley & Sons,
Inc., ISBN 0471377392, New York
J.C, Messenger.; S.Y,Chen.; J.D, Carroll.; J.E, Burchenal.; K, Kioussopoulos. & B.M, Groves.
(2000). 3D coronary reconstruction from routine single-place coronary angiograms:
clinical validation and quantitative analysis of the right coronary artery in 100
patients, The International Journal of Cardiac Imaging 16(6), pp.413-427
J, Cizek.; K, Herholz.; S, Vollmar.; R, Schrader.; J, Klein. & W.D, Heiss. (2004). Fast and
robust registration of PET and MR image of human brin, Neuroimage, Vol. 22, Iss.
1, pp.434-442
J, Flusser. & T,Suk. (1998). Degraded image analysis: an invariant approach, IEEE Trans.
Pattern Anal. Mach. Intell, pp. 590-603
Jianbo, Shi. & Carlo, Tomasi. (1994). Good features to track, IEEE Conference on CVPR Seattle,
pp. 593-600
J.P.W, Pluim.; J.B.A, Maintz. & M.A, Viergever. (2003). Mutual information based
registration of medical images: a survey, IEEE Transactions on Medical Imaging,
Vol. 22, No. 8, pp. 986-1004
14
Adaptive Control
J.R, Parker. (1996). Algorithms for image processing and computer vision, 1st Ed., John Wiley &
Sons, Inc., IBSN 0471140562, USA
J, Ross. et al. (1987). Guidelines for coronary angiography, Circulation, vol.76
K, Derpanis. (2004). The Harris corner detector
M, Grass.; R, Koppe.; E, Klotz.; R, Proksa.; MH, Kuhn.; H, Aerts. & etc (1999). 3D
reconstruction of high contrast objects using C-arm image intensifier projection
data, Computer Med Imaging Graphics, 23(6):311-321
M, Grass.; R, Koppe.; E, Klotz.; Op de Beek J. & R, Kemkers. (1999). 3D reconstruction and
imaging based on C-arm systems, Med Biol Eng Comput, 37(2):1520-1
M, Grass.; R, Guillemaud.; R, Koppe.; E, Klotz.; V, Rasche. & Op de Beek J. (2002). La
radiologie tridimensionelle, La tomography medicale-imagerie morphologique et imagerie
fonctionelle, Hermes Science Publications, Paris
Na-Young, Lee.; Gye-Young, Kim. & Hyung-Il, Choi. (2006). Automatic generation
technique of three-dimensional model corresponding to individual vessels,
Computational Science and Its Applications-ICCSA 2006, LNCS 3984, PP.441-449,
Glasgow, UK, May 2006, Springer-Verlag Berlin Heidelberg
Na-Young, Lee. (2006). An automatic generating technique of individual vessels based on
general vessels, The 7th International Workshop on Image Analysis for Multimedia
Interactive Services, pp.241-244, Korea, April, 2006, Incheon
NaYoung, Lee.; JeongHee, Cha.; JinWook, On.; GyeYoung, Kim. & HyungIl, Choi. (2006).
An automatic generation technique of 3D vessels model form angiograms,
Proceedings of The 2006 International Conference on Image Processing, Computer Vision,
& Pattern Recognition, pp.371-376, USA, June, 2006, CSREA Press, Las Vegas
Na-Young, Lee. (2007). Automatic generation of 3D vessels model using vessels image
matching based on adaptive control points, Sixth International Conference on
Advanced Language Processing and Web Information Technology, pp.265-270, ISBN 07695-2930-5, China, August 2007, IEEE Computer Society, Luoyang, Henan
Na-Young, Lee.; Gye-Young, Kim. & Hyung-Il, Choi. (2007). 3D modelling of the vessels
from X-ray angiography, Digital Human Modeling, HCII 2007, LNCS 4561, pp.646654, IBSN 978-3-540-73318-8 , Springer Berlin Heidelberg
Na-Young, Lee.; Gye-Young, Kim. & Hyung-Il, Choi. (2007). 3D model of vessels from
angiograms, Proceedings of the 13th Japan-Korea join workshop on frontiers of computer
vision, pp.235-240, Korea, January 2007, Busan
Na-Young, Lee. (2007). Three Dimensional Modeling of Individual Vessels Based on
Matching of Adaptive Control Points, MICAI 2007: Advances in Artificial Intelligence,
LNAI 4827, pp. 11431150, Springer-Verlag Berlin Heidelberg
P, de Feyter.; J, Vos.; J, Reiber. & P, Serruys. (1993). Value and limitations of quantitative
coronary angiography to assess progression and regression of coronary
atherosclerosis. In Advances in Quantitative Coronary Arteriography, pp. 255-271
R.P, Woods.; J.C, Mazziotta. & S.R, Cherry. (1993). MRI-PET registration with automated
algorithm, Journal of computer assisted tomography. Vol. 17:44, pp.536-546
Y, Bentoutou.; N, Taleb.; M, Chikr El Mezouar, M, Taleb. & L, Jetto. (2002). An invariant
approach for image registration in digital subtraction angiography. Pattern
Recognition, Vol. 35, Iss. 12, December 2002, pp. 2853-2865
2
Adaptive Estimation and Control for Systems
with Parametric and
Nonparametric Uncertainties
Hongbin Ma* and Kai-Yew Lum
Temasek Laboratories, National University of Singapore

tslmh@nus.edu.sg*
tsllumky@nus.edu.sg
Abstract
Adaptive control has been developed for decades, and now it has become a rigorous and
mature discipline which mainly focuses on dealing parametric uncertainties in control
systems, especially linear parametric systems. Nonparametric uncertainties were seldom
studied or addressed in the literature of adaptive control until new areas on exploring
limitations and capability of feedback control emerged in recent years. Comparing with the
approach of robust control to deal with parametric or nonparametric uncertainties, the
approach of adaptive control can deal with relatively larger uncertainties and gain more
flexibility to fit the unknown plant because adaptive control usually involves adaptive
estimation algorithms which play role of learning in some sense.
This chapter will introduce a new challenging topic on dealing with both parametric and
nonparametric internal uncertainties in the same system. The existence of both two kinds of
uncertainties makes it very difficult or even impossible to apply the traditional recursive
identification algorithms which are designed for parametric systems. We will discuss by
examples why conventional adaptive estimation and hence conventional adaptive control
cannot be applied directly to deal with combination of parametric and nonparametric
uncertainties. And we will also introduce basic ideas to handle the difficulties involved in
the adaptive estimation problem for the system with combination of parametric and
nonparametric uncertainties. Especially, we will propose and discuss a novel class of
adaptive estimators, i.e. information-concentration (IC) estimators. This area is still in its infant
stage, and more efforts are expected in the future for gainning comprehensive
understanding to resolve challenging difficulties.
Furthermore, we will give two concrete examples of semi-parametric adaptive control to
demonstrate the ideas and the principles to deal with both parametric and nonparametric
uncertainties in the plant. (1) In the first example, a simple first-order discrete-time nonlinear
system with both kinds of internal uncertainties is investigated, where the uncertainty of
non-parametric part is characterized by a Lipschitz constant L, and the nonlinearity of
parametric part is characterized by an exponent index b. In this example, based on the idea
of the IC estimator, we construct a unified adaptive controller in both cases of b = 1 and
Adaptive Control
16
b > 1, and its closed-loop stability is established under some conditions. When the
parametric part is bilinear (b = 1), the conditions given reveal the magic number
3
+ 2 which
2
appeared in previous study on capability and limitations of the feedback
mechanism. (2) In the second example with both parametric uncertainties and nonparametric uncertainties, the controller gain is also supposed to be unknown besides the
unknown parameter in the parametric part, and we only consider the noise-free case. For this
model, according to some a priori knowledge on the non-parametric part and the unknown
controller gain, we design another type of adaptive controller based on a gradient-like
adaptation law with time-varying deadzone so as to deal with both kinds of uncertainties.
And in this example we can establish the asymptotic convergence of tracking error under
some mild conditions, althouth these conditions required are not as perfect as in the first
example in sense that L < 0.5 is far away from the best possible bound
3
+ 2.
2
These two examples illustrate different methods of designing adaptive estimation and
control algorithms. However, their essential ideas and principles are all based on the a
priori knowledge on the system model, especially on the parametric part and the nonparametric part. From these examples, we can see that the closed-loop stability analysis is
rather nontrivial. These examples demonstrate new adaptive control ideas to deal with two
kinds of internal uncertainties simultaneously and illustrates our elementary theoretical
attempts in establishing closed-loop stability.
1. Introduction
This chapter will focus on a special topic on adaptive estimation and control for systems with
parametric and nonparametric uncertainties. Our discussion on this topic starts with a very
brief introduction to adaptive control.
1.1 Adaptive Control
As stated in [SB89], Research in adaptive control has a long and vigorous history since
the initial study in 1950s on adaptive control which was motivated by the problem of
designing autopilots for air-craft operating at a wide range of speeds and altitudes. With
decades of efforts, adaptive control has become a rigorous and mature discipline which
mainly focuses on dealing parametric uncertainties in control systems, especially linear
parametric systems.
From the initial stage of adaptive control, this area has been aiming at study how to deal
with large uncertainties in control systems. This goal of adaptive control essentially means
that one adaptive control law cannot be a fixed controller with fixed structure and fixed
parameters because any fixed controller usually can only deal with small uncertainties in
control systems. The fact that most fixed controllers with certain structure (e.g. linear
feedback control) designed for an exact system model (called nominal model) can also work
for a small range of changes in the system parameter is often referred to as robustness,
which is the kernel concept of another area, robust control. While robust control focuses on
studying the stability margin of fixed controllers (mainly linear feedback controller), whose
Adaptive Estimation and Control for Systems with Parametric and Nonparametric Uncertainties
17
design essentially relies on priori knowledge on exact nominal system model and bounds
of uncertain parameters, adaptive control generally does not need a priori information
about the bounds on the uncertain or (slow) time-varying parameters. Briefly speaking,
comparing with the approach of robust control to deal with parametric or nonparametric
uncertainties, the approach of adaptive control can deal with relatively larger uncertainties
and gain more flexibility to fit the unknown plant because adaptive control usually
involves adaptive estimation algorithms which play role of learning in some sense.
The advantages of adaptive control come from the fact that adaptive controllers can adapt
themselves to modify the control law based on estimation of unknown parameters by
recursive identification algorithms. Hence the area of adaptive control has close connections
with system identification, which is an area aiming at providing and investigating
mathematical tools and algorithms that build dynamical models from measured data.
Typically, in system identification, a certain model structure is chosen by the user which
contains unknown parameters and then some recursive algorithms are put forward based
on the structural features of the model and statistical properties of the data or noise. The
methods or algorithms developed in system identification are borrowed in adaptive control
in order to estimate the unknown parameters in the closed loop. For convenience, the
parameter estimation methods or algorithms adopted in adaptive control are often
referred to as adaptive estimation methods. Adaptive estimation and system identification
share many similar characteristics, for example, both of them originate and benefit from
the development of statistics. One typical example is the frequently used least-squares (LS)
algorithm, which gives parameter estimation by minimizing the sum of squared errors (or
residuals), and we know that LS algorithm plays important role in many areas including
statistics, system identification and adaptive control. We shall also remark that, in spite of
the significant similarities and the same origin, adaptive estimation is different from
system identification in sense that adaptive estimation serves for adaptive control and
deals with dynamic data generated in the closed loop of adaptive controller, which means
that statistical properties generally cannot be guaranteed or verified in the analysis of
adaptive estimation. This unique feature of adaptive estimation and control brings many
difficulties in mathematical analysis, and we will show such difficulties in later examples
given in this paper.
1.2 Linear Regression Model and Least Square Algorithm
Major parts in existing study on regression analysis (a branch of statistics) [DS98, Ber04,
Wik08j], time series analysis [BJR08, Tsa05], system identification [Lju98, VV07] and
adaptive control [GS84, AW89, SB89, CG91, FL99] center on the following linear regression
model
z k = k + vk
where
(1)
{z k } , k , vk represent observation data, regression vector and noise disturbance (or
external uncertainties), respectively. Here is the unknown parameter to be estimated.

Linear regression models have many applications in many disciplines of science and
engineering [Wik08g, web08, DS98, Hel63, Wei05, MPV07, Fox97, BDB95]. For example, as
Adaptive Control
18
stated in [web08], Linear regression is probably the most widely used, and useful, statistical
technique for solving environmental problems. Linear regression models are extremely powerful, and
have the power to empirically tease out very complicated relationships between variables. Due to the
importance of model (1.1), we list several simple examples for illustration:
Assume that a series of (stationary) data (xk , yk ) (k = 1, 2, , N ) are generated from the
following model
Y = 0 + 1 X +
where 0 , 1 are unknown parameters, {xk } are i. i. d. taken from a certain probability
distribution, and
k N (0, 2 )
is random noise independent of X . For this model, let
= [0 , 1 ] , k = [1, xk ] , then we have
y k = k + k .
This example is a classic
topic in statistics to study the statistical properties of parameter estimates N as the data size
N grows to infinity. The statistical properties of interests may include
and so on.
Unlike the above example, in this example we assume that
xk
E( ), Var() ,
and
xk +1
have close
relationship modeled by
xk +1 = 0 + 1 xk + k
where 0, 1 are unknown parameters, and
k N (0, 2 )
are i. i. d. random noise
independent of {x1, x2, , xk}.

This model is an example of linear time series analysis, which aims to study asymptotic
statistical properties of parameter estimates
under certain assumptions on statistical
k . Note that for this example, it is possible to deduce an explicit expression

of xk in terms of j ( j = 0,1,L , k 1 ).
properties of
In this example, we consider a simple control system
xk +1 = 0 + 1 xk + buk + k
where b 0 is the controller gain,
is the noise disturbance at time step k. For this model,
= [ 0 , 1 ] , k = [1, xk 1 ] ,
= [ 0 , 1 , b] , k = [1, xk 1 ] , zk = xk buk 1 .
in case where b is known a priori, we can take;
z k = xk bu k 1 ;otherwise, we can take
In both cases, the system can be rewritten as
z k = k + k
19
which implies that intuitively, can be estimated by using the identification algorithm since
both data zk and
k are available at time step k. Let k denote the parameter estimates at
time step k , then we can design the control signal u k by regarding as the real parameter
:
where { rk } is the known reference signal to be tracked, and
0 , 1 ,
b , 0 , 1
are estimates of
b,
respectively. Note that for this example, the closed-loop system will be very
complex because the data generated in the closed loop essentially depend on all history
signals. In the closed-loop system of an adaptive controller, generally it is difficult to
analyze or verify statistical properties of signals, and this fact makes that adaptive
estimation and control cannot directly employ techniques or results from system
identification. Now we briefly introduce the frequently-used LS algorithm for model (1.1)
due to its importance and wide applications [LH74, Gio85, Wik08e, Wik08f, Wik08d]. The
idea of LS algorithm is simply to minimize the sum of squared errors, that is to say,
(1.2)
This idea has a long history rooted from great mathematician Carl Friedrich Gauss in 1795
and published first by Legendre in 1805. In 1809, Gauss published this method in volume
two of his classical work on celestial mechanics, heoria Motus Corporum Coelestium in
sectionibus conicis solem ambientium[Gau09], and later in 1829, Gauss was able to state that the
LS estimator is optimal in the sense that in a linear model where the errors have a mean of
zero, are uncorrelated, and have equal variances, the best linear unbiased estimators of the
coefficients is the least-squares estimators. This result is known as the Gauss-Markov
theorem [Wik08a].
By Eq. (1.2), at every time step, we need to minimize the sum of squared errors, which
requires much computation cost. To improve the computational efficiency, in practice we
often use the recursive form of LS algorithm, often referred to as recursive LS algorithm,
which will be derived in the following. First, introducing the following notations
(1.3)
and using Eq. (1.1), we obtain that
Adaptive Control
20
Noting that
where the last equation is derived from properties of Moore-Penrose pseudoinverse

[Wik08h]
we know that the minimum of
[ Z n n ] [ Z n n ]
can be achieved at
(1.4)
which is the LS estimate of . Let
and then, by Eq. (1.3), with the help of matrix inverse identity
we can obtain that

1
Pn = ( Pn 1 + nn ) 1
= [ A1 + A1 B C 1 BA1 ]1
= Pn 1 ( Pn 1n )[1 + (n Pn 1 ) Pn11 ( Pn 1n )]1 (n Pn 1 )
= Pn 1 an Pn 1nn Pn 1
where
Further,
21
Thus, we can obtain the following recursive LS algorithm
where Pn1 and n1 reflect only information up to step n 1, while an,
n and z n n n 1
reflect information up to step n.

In statistics, besides linear parametric regression, there also exist generalized linear models
[Wik08b] and non-parametric regression methods [Wik08i], such as kernel regression
[Wik08c]. Interested readers can refer to the wiki pages mentioned above and the references
therein.
1.3 Uncertainties and Feedback Mechanism
By the discussions above, we shall emphasize that, in a certain sense, linear regression
models are kernel of classical (discrete-time) adaptive control theory, which focuses to cope
with the parametric uncertainties in linear plants. In recent years, parametric uncertainties
in nonlinear plants have also gained much attention in the literature[MT95, Bos95, Guo97,
ASL98, GHZ99, LQF03]. Reviewing the development of adaptive control, we find that
parametric uncertainties were of primary interests in the study of adaptive control, no
matter whether the considered plants are linear or nonlinear. Nonparametric uncertainties
were seldom studied or addressed in the literature of adaptive control until some new areas
on understanding limitations and capability of feedback control emerged in recent years.
Here we mainly introduce the work initiated by Guo, who also motivated the authors
exploration in the direction which will be discussed in later parts.
Guos work started from trying to understand fundamental relationship between the
uncertainties and the feedback control. Unlike traditional adaptive theory, which focuses on
investigating closed-loop stability of certain types of adaptive controllers, Guo began to
think over a general set of adaptive controllers, called feedback mechanism, i.e., all possible
feedback control laws. Here the feedback control laws need not be restricted in a certain
class of controllers, and any series of mappings from the space of history data to the space of
control signals is regarded as a feedback control law. With this concept in mind, since the
most fundamental concept in automatic control, feedback, aims to reduce the effects of the
Adaptive Control
22
plant uncertainty on the desired control performance, by introducing the set F of internal
uncertainties in the plant and the whole feedback mechanism U, we wonder the following
basic problems:
1. Given an uncertainty set F, does there exist any feedback control law in U which can
stabilize the plant? This question leads to the problem of how to characterize the maximum
capability of feedback mechanism.
2. If the uncertainty set F is too large, is it possible that any feedback control law in U cannot
stabilize the plant? This question leads to the problem of how to characterize the limitations
of feedback mechanism.
The philosophical thoughts to these problems result in fruitful study [Guo97, XG00, ZG02,
XG01, LX06, Ma08a, Ma08b].
The first step towards this direction was made in [Guo97], where Guo attempted to answer
the following question for a nontrivial example of discrete-time nonlinear polynomial plant
model with parametric uncertainty: What is the largest nonlinearity that can be dealt with
by feedback? More specifically, in [Guo97], for the following nonlinear uncertain system
(1.5)
where is the unknown parameter, b characterizes the nonlinear growth rate of the
system, and { w t } is the Gaussian noise sequence, a critical stability result is found system
(1.5) is not a.s. globally stabilizable if and only if b 4. This result indicates that there exist
limitations of the feedback mechanism in controlling the discrete-time nonlinear adaptive
systems, which is not seen in the corresponding continuous-time nonlinear systems (see
[Guo97, Kan94]). The impossibility result has been extended to some classes of uncertain
nonlinear systems with unknown vector parameters in [XG99, Ma08a] and a similar result
for system (1.5) with bounded noise is obtained in [LX06].
Stimulated by the pioneering work in [Guo97], a series of efforts ([XG00, ZG02, XG01,
MG05]) have been made to explore the maximum capability and limitations of feedback
mechanism. Among these work, a breakthrough for non-parametric uncertain systems was
made by Xie and Guo in [XG00], where a class of first-order discrete-time dynamical control
systems
(1.6)
is studied and another interesting critical stability phenomenon is proved by using new
techniques which are totally different from those in [Guo97]. More specifically, in [XG00],
F(L) is a class of nonlinear functions satisfying Lipschitz condition, hence the Lipschitz
constant L can characterize the size of the uncertainty set F(L). Xie and Guo obtained the
following results: if L 3 +
2 , then there exists a feedback control law such that for any
F(L), the corresponding closed-loop control system is globally stable; and if
3
L< + 2
2
, then for any feedback control law and any
y 0 R1 ,
there always exists
some
23
f F (L) such that the corresponding closed-loop system is unstable. So for system
(1.6), the magic number 3 + 2 characterizes the capability and limits of the whole
2
feedback mechanism. The impossibility part of the above results has been generalized to
similar high-order discrete-time nonlinear systems with single Lipschitz constant [ZG02]
and multiple Lipschitz constants [Ma08a]. From the work mentioned above, we can see two
different threads: one is focused on parametric nonlinear systems and the other one is
focused on non-parametric nonlinear systems. By examining the techniques in these threads,
we find that different difficulties exist in the two threads, different controllers are designed
to deal with the uncertainties and completely different methods are used to explore the
capability and limitations of the feedback mechanism.
1.4 Motivation of Our Work
From the above introduction, we know that only parametric uncertainties were considered
in traditional adaptive control and non-parametric uncertainties were only addressed in
recent study on the whole feedback mechanism. This motivates us to explore the following
problems: When both parametric and non-parametric uncertainties are present in the
system, what is the maximum capability of feedback mechanism in dealing with these
uncertainties? And how to design feedback control laws to deal with both kinds of internal
uncertainties? Obviously, in most practical systems, there exist parametric uncertainties
(unknown model parameters) as well as non-parametric uncertainties (e.g. unmodeled
dynamics). Hence, it is valuable to explore answers to these fundamental yet novel
problems. Noting that parametric uncertainties and non-parametric uncertainties essentially
have different nature and require completely different techniques to deal with, generally it
is difficult to deal with them in the same loop. Therefore, adaptive estimation and control in
systems with parametric and non-parametric uncertainties is a new challenging direction. In
this chapter, as a preliminary study, we shall discuss some basic ideas and principles of
adaptive estimation in systems with both parametric and non-parametric uncertainties; as to
the most difficult adaptive control problem in systems with both parametric and nonparametric uncertainties, we shall discuss two concrete examples involving both kinds of
uncertainties, which will illustrate some proposed ideas of adaptive estimation and special
techniques to overcome the difficulties in the analysis closed-loop system. Because of
significant difficulties in this new direction, it is not possible to give systematic and
comprehensive discussions here for this topic, however, our study may shed light on the
aforementioned problems, which deserve further investigation.
The remainder of this chapter is organized as follows. In Section 2, a simple semi-parametric
model with parametric part and non-parametric part will be introduced first and then we
will discuss some basic ideas and principles of adaptive estimation for this model. Later in
Section 3 and Section 4, we will apply the proposed ideas of adaptive estimation and
investigate two concrete examples of discrete-time adaptive control: in the first example, a
discrete-time first-order nonlinear semi-parametric model with bounded external noise
disturbance is discussed with an adaptive controller based on information-contraction
estimator, and we give rigorous proof of closed-loop stability in case where the uncertain
parametric part is of linear growth rate, and our results reveal again the magic number
Adaptive Control
24
3
+
2
2 ; in the second example, another noise-free semi-parametric model with
parametric uncertainties and non-parametric uncertainties is discussed, where a new

adaptive controller based on a novel type of update law with deadzone will be adopted to
stabilize the system, which provides yet another view point for the adaptive estimation and
control problem for the semi-parametric model. Finally, we give some concluding remarks
in Section 5.
2. Semi-parametric Adaptive Estimation: Principles and Examples

2.1 One Semi-parametric System Model
Consider the following semi-parametric model
z k = k + f (k ) + k
where
denotes unknown parameter vector, f()
(2.1)
F denotes unknown function and
k k denote external noise disturbance. Here , F and k represent a priori knowledge

on possible , f (k ) and k , respectively. In this model, let
then Eq. (2.1) becomes Eq. (1.1). Because each term of right hand side of Eq. (2.1) involves
uncertainty, it is difficult to estimate
, f (k )
and
simultaneously.
Adaptive estimation problem can be formulated as follows: Given a priori knowledge on ,

f() and
k , how to estimate and f() according to a series of data { k , z k ; k = 1, 2 , L , n }
Or in other words, given a priori knowledge on and vk, how to estimate and vk according
to a series of data { k , z k ; k = 1, 2 , L , n }.
Now we list some examples of a priori knowledge to show various forms of adaptive
estimation problem.
Example 2.1 As to the unknown parameter , here are some commonly-seen examples of a priori
knowledge:
There is no any a priori knowledge on except for its dimension. This means that can be
arbitrary and we do not know its upper bound or lower bound.
The upper and lower bounds of are known, i.e. , where and are constant vector
and the relationship means element-wise less or equal.
The distance between and a nominal 0 is bounded by a known constant, i.e. || 0 || r,
where r 0 is a known constant and 0 is the center of set .
The unknown parameter lies in a known countable or finite set of values, that is to say, { 1, 2,
3, }.
Example 2.2 As to the unknown function f(), here are some possible examples of a priori knowledge:
f(x) = 0 for all x. This case means that there is no unmodeled dynamics.
Function f is bounded by other known functions, that is to say,
25
f ( x) f ( x) f ( x) for any x.
The distance between f and a nominal f0 is bounded by a known constant, i.e. ||f f0|| rf ,
where rf 0 is a known constant and f0 can be regarded as the center of a ball F in a metric functional
space with norm || ||.
The unknown function lies in a known countable or finite set of functions, that is to say, f {f1, f2,
f3, }.
Function f is Lipschitz, i.e. f ( x1 ) f ( x 2 ) L | x1 x 2 | for some constant L > 0.
Function f is monotone (increasing or decreasing) with respect to its arguments.
Function f is convex (or concave).
Function f is even (or odd).
Example 2.3 As to the unknown noise term
knowledge:
, here are some possible examples of a priori
Sequence
k = 0. This case means that no noise/disturbance exists.
Sequence
is
is bounded in a known range, that is to say,
for any k. One special case
= .
Sequence
k is bounded by a diminishing sequence, e.g, | k |
1
k
for any
k . This case means
that the noise disturbance converges to zero with a certain rate. Other typical rate sequences include
1
} , { k } ( 0 < < 1 ), and so on.
2
k
Sequence
k is bounded by other known sequences, that is to say,
for any k.
This case generalizes the above

cases.
Sequence
k is in a known finite set of values, that is to say, k {e1 , e2 ,L , eN } . This case
may happen in digital systems where all signals can only take values in a finite set.
k is oscillatory with specific patterns, e.g. k > 0 if k is even and k < 0 if k is odd.
2
2
Sequence k has some statistical properties, for example, Eek = 0 , Eek = ;; for another
example, sequence { k } is i.i.d. taken from a probability distribution e.g. k U (0,1) .
Sequence
Parameter estimation problems (without non-parametric part) involving statistical

properties of noise disturbance are studied extensively in statistics, system identification
and traditional adaptive control. However, we shall remark that other non-statistic
descriptions on a priori knowledge is more useful in practice yet seldom addressed in
existing literature. In fact, in practical problems, usually the probability distribution of the
noise/disturbance (if any) is not known and many cases cannot be described by any
probability distribution since noise/disturbance in practical systems may come from many
different types of sources. Without any a priori knowledge in mind, one frequently-used way
to handle the noise is to simply assume the noise is Gaussian white noise, which is
Adaptive Control
26
reasonable in a certain sense. But in practice, from the point of view of engineering, we can
usually conclude the noise/disturbance is bounded in a certain range. This chapter will
focus on uncertainties with non-statistical a priori knowledge. Without loss of generality, in
this section we often regard
knowledge on
vk = f (k ) + k
as a whole part, and correspondingly, a priori
vk , (e.g. v k vk v k ), should be provided for the study.
2.2 An Example Problem

Now we take a simple example to show that it may not be appropriate to apply traditional
identification algorithms blindly so as to get the estimate of unknown parameter.
Consider the following system
z k = k + f (k , k ) + k
where , f() and
(2.2)
are unknown parameter, unknown function and unmeasurable noise,
respectively. For this model, suppose that we have the following a priori knowledge on the
system:
No a priori knowledge on is known.
is of form
. Here
is an
At any step k, the term
1.
unknown sequence satisfying 0
Noise
is diminishing with
And in this example, our problem is how to use the data generated from model (2.2) so as to
get a good estimate of true value of parameter . In our experiment, the data is generated by
the following settings (k = 1, 2, , 50):
= 5 , k =
where
k
1
, f (k , k ) = exp(| sin k | k ) , k = ( k 0.5)
10
k
{ k } are i.i.d. taken from uniform distribution U(0, 1). Here we have N = 50 groups
of data
.
Since model (2.2) involves various uncertainties, we rewrite it into the following form of
linear regression
(2.3)
by letting
vk = f (k , k ) + k .
From the a priori knowledge for model (2.2), we can obtain the following a priori knowledge
for the term vk
27
where
Since model (2.3) has the form of linear regression, we can use try traditional identification
algorithms to estimate . Fig. 1 illustrates the parameter estimates for this problem by using
standard LS algorithm, which clearly show that LS algorithm cannot give good parameter
estimate in this example because the final parameter estimation error
k = 5 .68284
is very large.
Fig. 1. The dotted line illustrates the parameter estimates obtained by standard least-squares
algorithm. The straight line denotes the true parameter.
One may then argue that why LS algorithm fails here is just because the term
vk
is in fact
biased and we indeed do not utilize the a priori knowledge on vk. Therefore, we may try a
modified LS algorithm for this problem: let
Adaptive Control
28
then we can conclude that
yk = k + wk and wk [d k , d k ] ,
where
[ d k , d k ] is
symmetric interval for every k. Then, intuitively, we can apply LS algorithm to data
{ (k , z k ) , k = 1, 2, ,N}. The curve of parameter estimates obtained by this modified LS
algorithm is plotted in Fig. 2. Since the modified LS algorithm has removed the bias in the a
priori knowledge, one may expect the modified LS algorithm may give better parameter
estimates, which can be verified from Fig. 2 since the final parameter estimation error
N = N 1.83314 .
In this example, although the modified LS algorithm can
work better than the standard LS algorithm, the modified LS algorithm in fact does not help
much in solving our problem since the estimation error is still very large comparing with the
true value of the unknown parameter.
Fig. 2. The dotted line illustrates the parameter estimates obtained by modified least-squares
algorithm. The straight line denotes the true parameter.
29
From this example, we do not aim to conclude that traditional identification algorithms
developed in linear regression are not good, however, we want to emphasize the following
particular point: Although traditional identification algorithms (such as LS algorithm) are very
powerful and useful in practice, generally it is not wise to apply them blindly when the matching
conditions, which guarantee the convergence of those algorithms, cannot be verified or asserted a
priori. This particular point is in fact one main reason why the so-called minimum-variance
self tuning regulator, developed in the area of adaptive control based on the LS algorithm,
attracted several leading scholars to analyze its closed-loop stability throughout past
decades from the early stage of adaptive control.
To solve this example and many similar examples with a priori knowledge, we will propose
new ideas to estimate the parametric uncertainties and the non-parametric uncertainties.
2.3 Information-Concentration Estimator
We have seen that there exist various forms of a priori knowledge on system model. With the
a priori knowledge, how can we estimate the parametric part and the non-parametric part?
Now we introduce the so-called information-concentration estimator. The basic idea of this
estimator is, the a priori knowledge at each time step can be regarded as some constraints of
the unknown parameter or function, hence the growing data can provide more and more
information (constraints) on the true parameter or function, which enable us to reduce the
uncertainties step by step. We explain this general idea by the simple model
(2.4)
with a priori knowledge that
current data k,
R d , k V k . Then, at k-th step (k 1), with the
k , z k we can define the so-called information set Ik at step k:

(2.5)
For convenience, let I0 = . Then we can define the so-called concentrated information set Ck at
step k as follows
(2.6)
which can be recursively written as
(2.7)
with initial set C0 = . Eq. (2.7) with Eq. (2.5) is called information-concentration estimator
(short for IC estimator) throughout this chapter, and any value in the set
Ck
can be taken as
one possible estimate of unknown parameter at time step k . The IC estimator differs
from existing parameter identification in the sense that the IC estimator is in fact a set-
Adaptive Control
30
valued estimator rather than a real-valued estimator. In practical applications, generally
Ck is a domain in R d , and naturally we can take the center point of Ck
as
k .
Remark 2.1 The definition of information set varies with system model. In general cases, it can be
extended to the set of possible instances of (and/or f ) which do not contradict with the data at
step k. We will see an example involving unknown f in next section.
From the definition of the IC estimator, the following proposition can be obtained without
difficulty:
Proposition 2.1 Information-concentration estimator has the following properties:
(i) Monotonicity:
C0 C1 C2 L
(ii) Convergence: Sequence {Ck} has a limit set
C = C k ;
k =1
must be a non-empty set

(iii) If the system model and the a priori knowledge are correct, then
with property
and any element of
can match the data and the model;
(iv) If
C = , then the data {k , z k }
cannot be generated by the system model used by the IC
estimator under the specified a priori knowledge.

Proposition 2.1 tells us the following particular points of the IC estimator: property (i)
implies that the IC estimator will provide more and more exact estimation; property (ii)
means that the there exists a limitation in the accuracy of estimation; property (iii) means
that true parameter lies in every C k if the system model and a priori knowledge are correct;
and property (iv) means that the IC estimator provides also a method to validate the system
model and the a priori knowledge. Now we discuss the IC estimator for model (2.4) in more
details. In the following discussions, we only consider a typical a priori knowledge on
v k vk v k
are two known sequences of vectors (or scalars).
2.3.1 Scalar case: d = 1

By Eq. (2.5), we have
Solving the inequality in Ik, we obtain that
and consequently, if
31
k 0 , then we have
where
Here sign(x) denotes the sign of x: sign(x) = 1, 0,1 for positive number, zero, and negative
number, respectively. Then, by Eq. (2.7), we can explicitly obtain that
where
and
can be recursively obtained by
Fig. 3. The straight line may intersect the polygon V and split it into two sub-polygons, one
of which will become new polygon V'. The polygon V' can be efficiently calculated from the
polygon V.
Adaptive Control
32
2.3.2 Vector case: d > 1
In case of d > 1, since and
are vectors, we cannot directly obtain explicit solution of
inequality
(2.8)
Notice that Eq. (2.8) can be rewritten into two separate inequalities:
we need only study linear equalities of the form
T c . Generally speaking, the solution
to a system of inequalities represents a polyhedral (or polygonal) domain in Rd, hence we

need only determine the vertices of the polyhedral (or polygonal) domain. In case of d = 2, it
is easy to graph linear equalities since every inequality
T c represents a half-plane. In
v k = {/ i , i = 1,2,L , p k } denote the distinct vertices of the domain Ck

and pk denote the number of vertices of domain C k , then we discuss how to deduce Vk
from Vk 1 . The domain Ck has two more linear constraints than the domain C k 1
general case, let
with
We need only add these two constraints one by one, that is to say,
is an algorithm whose function is to add linear constraint
where
T c
to the polygon represented by vertex set V and to return the vertex set of the new
polygon with added constraint.

Now we discuss how to implement the algorithm AddLinearConstraint.
2D Case: In case of d = 2,
T c
represents a straight line which splits the plane into two
half-planes (see Fig. 3). In this case, we can use an efficient algorithm
AddLinearConstraint2D which is listed in Algorithm 1. Its basic idea is to simply test each
vertex of V to see whether to keep original vertex or generate new vertex. The time
33
complexity of Algorithm 1 is O(s), where s is the number of vertices of domain V. Note that
T c does not intersect with the polygon

Pi of polygon V does not satisfy T Pi > c . And the vertex number of
it is possible that V' = if the straight line L :

V and any vertex
polygon V ' can in fact vary within the range from 0 to s according to the geometric
relationship between the straight line L and the polygon V.
High-dimensional Case: In case of d > 2,
T c
represents a hyperplane which splits
the whole space into two half-hyperplanes.

Unlike in case of d = 2, the vertices in this case generally cannot be arranged in a certain
natural order (such as clock-wise order). In this case, we can use an algorithm
AddLinearConstraintND which is listed in Algorithm 2. The idea of this algorithm is to
classify the vertices of V first according to their relationship with the hyperplane determined
by hyperplane
T c .
Algorithm 2 AddLinearConstraintND(V, ", c): Add linear constraint
T c
(" % Rd) to a
polyhedron V
2.3.3 Implementation issues
In the IC estimator, the key problem is to calculate the information set Ik or the concentrated
information set Ck at every step. From the discussions above, we can see that it is easy to
solve this basic problem in case of d = 1. However, in case of d > 1, generally the vertex
Adaptive Control
34
number of domain
Ck
may grow as
. Therefore, it may be impractical to
implement the IC estimator in case of d > 1 since it may require growing memory as
k To overcome this problem, noticing the fact that the domain Ck will shrink
gradually as k in order to get a feasible IC estimate of the unknown parameter

vector, generally we need not use too many vertices to represent the exact concentrated
information set Ck. That is to say, in practical implementation of IC estimator in highdimensional case, we can use a domain k with only a small number (say up to M) of
vertices to approximate the exact concentrated information set Ck. With such an idea of
approximate IC estimator, the issue of computational complexity will not hinder the
applications of IC estimator.
We consider two typical cases of approximate IC estimator. One typical case is that
for any k, and the other case is that
for any k. Let
C = C k ,
k =1
then in the
former case (called loose IC estimator, see Fig. 4), we must have
which means that we will never mistakenly exclude the true parameter from the
concentrated approximate information sets; while in the latter case (called tight IC estimator,
see Fig. 5), we must have
which means that the true parameter may be outside of
C however any value in C can
be served as good estimate of true parameter.
Fig. 4. Idea of loose IC estimator: The polygon P1P2P3P4P5 can be approximated by a triangle
Q1P4Q2. Here M = 3.
35
Fig. 5. Idea of tight IC estimator: The polygon P1P2P3P4P5 can be approximated by a triangle
P3P4P5. Here M = 3.
Now we discuss implementation details of tight IC estimator and loose IC estimator. Without
loss of generality, we only explain the ideas in case of d = 2. Similar ideas can be applied in
cases of d > 2 without difficulty.
Tight IC estimator: To implement a tight IC estimator, one simple approach is to modify
Algorithm 1 so as it just keeps up to M vertices in the queue Q. To get good approximation,
in the loop of Algorithm 1, it is suggested to abandon the generated vertex P ' (in Line 12 of
Algorithm 1) which is very close to existing vertex Pj (let j = i if i < 0 and i1 > 0 or j = i 1
if i > 0 and i1 < 0). The closeness between P and existing vertex Pj can be measured by
checking the corresponding weight w .
Loose IC estimator: To implement a loose IC estimator, one simple approach is to modify
Algorithm 1 so as it can generate M vertices which surround all vertices in the queue Q. To
this end, in the loop of Algorithm 1, if the generated vertex P ' (in Line 12 of Algorithm 1) is
very close to existing vertex Pj (let j = i if i < 0 and i1 > 0 or j = i 1 if i > 0 and i1 < 0),
we can simply append vertex Pj instead of P to queue Q. In this way, we can avoid
increasing the vertex number by generating new vertices. The closeness between P and
existing vertex Pj can be measured by checking the corresponding weight w.
Besides the ideas of tight or loose IC estimator, to reduce the complexity of IC estimator, we
can also use other flexible approaches. For example, to avoid growth in the vertex number of
, we can approximate Vk by using a simple outline rectangle (see Fig. 6) every
Vk as
certain steps. For a polygon Vk with vertices P1, P2, , Ps, we can easily obtain its outline
rectangle by algorithm FindPolygonBounds listed in Algorithm 3. Here for convenience, the
operators max and min for vectors are defined element-wisely, i.e.
where
are two vectors in Rn.
Adaptive Control
36
Fig. 6. Idea of outline rectangle: The polygon

outline rectangle. In this case,
B1 , B1
P1 P2 P3 P4 P5
can be approximated by an
denote the lower bound and upper bound in the x-
axis (1st component of each vertex), and
B2 , B2
denote the lower bound and upper bound
in the y-axis (2nd component of each vertex)

2.4 IC Estimator vs. LS Estimator
2.4.1 Illustration of IC Estimator
Now we go back to the example problem discussed before. For this example,
and zk are
scalars, hence we need only apply the IC estimator introduced in Section 2.3.1. Since IC
estimator yields concentrated information set
Ck
at every step, we can take any value in
Ck
as parameter estimate of true parameter. In this example,
Ck
37
is an interval at every
step step. For comparison with other parameter estimation methods, we simply take
1
2
k = (b k + b k ) , i.e. the center of interval Ck , as the parameter estimate at step k.

In Fig. 7, we plot three curves
b k , b k and k . From this figure, we can see that, for this
particular example, with the help of a priori knowledge, the upper estimates
estimates
bk
consequently
bk
and lower
given by the IC estimator converge to true parameter = 5 quickly, and
also converges to true parameter.
Fig. 7. This figure illustrates the parameter estimates obtained by the proposed informationconcentration estimator. The upper curve and lower curve represent the upper bounds
and lower bounds
k =
1
bk + bk
2
bk
bk
for the parameter estimates. We use the center curve
) to yield the parameter estimates.
We should also remark that the parameter estimates given by the IC estimator are not
necessarily convergent as in this example. Whether the IC parameter estimates converge
Adaptive Control
38
largely depend on the accuracy of a priori knowledge and the richness of the practical data.
Note that the IC estimator generally does not require classical richness concepts (like
persistent excitation) which are useful in the analysis of traditional recursive identification
algorithms.
2.4.2 Advantages of IC Estimator
We have seen practical effects of IC estimator for the simple example given above. Why can
it perform better than the LS estimator? Roughly speaking, comparing with traditional
identification algorithm like LS algorithm, the proposed IC estimator has the following
advantages:
1. It can make full use of a priori information and posterior information. And in the ideal
case, no information is wasted in the iteration process of the IC estimator. This property is
not seen in traditional identification algorithms since only partial information and certain
stochastic a priori knowledge can be utilized in those algorithms.
2. It does not give single parameter estimate at every step; instead, it gives a (finite or
infinite) set of parameter estimates at every step. This property is also unique since
traditional identification algorithms always give parameter estimates directly.
3. It can gradually find out all (or most) possible values of true parameters; and this
property can even help people to check the consistence between the practical data and the
system model with a priori knowledge. This property distinguishes traditional identification
algorithms in sense that traditional identification algorithms generally have no mechanism
to validate the correctness of the system model.
4. The a priori knowledge can vary from case to case, not necessarily described in the
language of probability theory or statistics. This property enables the IC estimator to handle
various kinds of non-statistic a priori knowledge, which cannot be dealt with by traditional
identification algorithms.
5. It has great flexibilities in its implementation, and its design is largely determined by the
characteristics of a priori knowledge. The IC estimator has only one basic principleinformation
concentration! Any practical implementation approach using such a principle can be
regarded as an IC estimator. We have discussed some implementation details for a certain
type of IC estimator in last subsection, which have shown by examples how to design the IC
estimator according the known a priori knowledge and how to reduce computational
complexity in practical implementation.
6. Its accuracy will never degrade as time goes by. Generally speaking, the more steps
calculated, the more data involved, and the more accurate the estimates are. Generally
speaking, traditional identification algorithms can only have similar property (called strong
consistency) under certain matching conditions.
7. The IC estimator can not only provide reasonably good parameter estimates but also tell
people how accurate these estimates are. In our previous example, when we use
k =
1
b k + b k as the parameter estimate, we know also that the absolute parameter
2
estimation error ~ =
will not exceed
1
b k + b k . In some sense, such a property
2
may be conceptually similar to the so-called confidence level in statistics.
39
2.4.3 Disadvantages of IC Estimator

Although the IC estimator has many advantages over traditional identification algorithms, it
may have the following disadvantages:
1. The proposed IC estimator is relatively difficult to incorporate stochastic a priori
knowledge on noise term, especially unbounded random noise. In fact, in such cases
without non-parametric uncertainties, traditional identification algorithms like LS algorithm
may be more suitable and efficient to estimate the unknown parameter.
2. The efficiency of IC estimator largely depends on its implementation via the
characteristics of the a priori knowledge. Generally speaking, the IC estimator may involve a
little more computation operations than recursive identification algorithms like LS
algorithm. We shall remark also that this point is not always true since the numerical
operations involved in the IC estimator are relatively simple (see algorithms listed before),
while many traditional identification algorithms may involve costly numerical operations
like matrix product, matrix inversion, etc.
3. Although the IC estimator has simple and elegant properties such as monotonicity and
convergence, due to its nature of set-valued estimator, no explicit and recursive expressions can
be given directly for the IC parameter estimates, which may bring mathematical difficulties
in the applications of the IC estimator. However, generally speaking, we also know that
closed-loop analysis for adaptive control using traditional identification algorithms is not
easy, too.
Summarizing the above, we can conclude that the IC estimator provides a new approach or
principle to estimate parametric and even non-parametric uncertainties, and we have shown
that it is possible to design efficient IC estimator according to characteristics of a priori
knowledge.
3. Semi-parametric Adaptive Control: Example 1

In this section, we will give a first example of semi-parametric adaptive control, whose
design is essentially based on the IC estimator introduced in last section.
3.1 Problem Formulation
Consider the following system
(3.1)
where yt, ut and wt are the output, input and noise, respectively;
f () F ( L)
is an
unknown function (the set F(L) will be defined later) and is an unknown parameter. To
make further study, the following assumptions are used throughout this section:
Assumption 3.1 The unknown function
f : R R belongs to the following uncertainty set
(3.2)
Adaptive Control
40
where c is an arbitrary non-negative constant.
Assumption 3.2 The noise sequence
{wt } is bounded, i.e.
where w is an arbitrary positive constant.

Assumption 3.3 The tracking signal
{ yt*}
is bounded, i.e.
(3.3)
where S is a positive constant.

Assumption 3.4 In the parametric part
parameter , but
t , we have no any a priori information of the unknown
t = g ( yt ) is measurable and satisfies

(3.4)
for any
x1 x2 , where M' M are two positive constants and b 1 is a constant.
Remark 3.1 Assumption 3.4 implies that function g() has linear growth rate when b = 1. Especially
when g(x) = x, we can take M = M' = 1. Condition (3.4) need only hold for sufficiently large x1 and
x2, however we require it holds for all x1 x2 to simplify the proof. We shall also remark that Sokolov
[Sok03] has ever studied the adaptive estimation and control problem for a special case of model (3.1),
where
is simply taken as
ayt .
Remark 3.2 Assumption 3.4 excludes the case where g() is a bounded function, which can be
handled easily by previous research. In fact, in that case
w't +1 = t + wt +1
hence by the result of [XG00], system (3.1) is stabilizable if and only if L <
must be bounded,
3
+ 2.
2
3.2 Adaptive Controller Design

In the sequel, we shall construct a unified adaptive controller for both cases of b =1 and b >1.
For convenience, we introduce some notations which are used in later parts. Let I = [a, b] be
an interval, then
1
m( I ) = ( a + b)
2
(a+ b) denotes the center point of interval I, and
1
b a denotes the radius of interval I. And correspondingly, we
2
I ( x, ) = [x , x + ] denote a closed interval centered at x R with radius 0.
r (I ) =
let
Estimate of Parametric Part: At time t, we can use the following information: y0, y1, , yt,
u0, u1, , ut1 and
1 , 2 ,L , t . Define
41
(3.5)
and
(3.6)
where
(3.7)
then, we can take
(3.8)
as the estimate of parameter at time t and corresponding estimate error bound,
t = t t are the
estimates of the upper and lower bounds of the unknown parameter , respectively.
According to Eq. (3.6), obviously we can see that { t } is a non-increasing sequence and
{ t } is non-decreasing.
respectively. With
and t defined above,
t = t + t
and
Remark 3.3 Note that Eq. (3.6) makes use of a priori information on nonlinear function f(). This
estimator is another example of the IC estimator which demonstrates how to design the IC estimator
according to the Lipschitz property of function f(). With similar ideas, the IC estimator can be
designed based on other forms of a priori information of function f().
Estimate of Non-parametric Part: Since the non-parametric part
f ( yt ) may be unbounded
and the parametric part is also unknown, generally speaking it is not easy to estimate the
non-parametric part directly. To resolve this problem, we choose to estimate
as a whole part rather than to estimate f(yt) directly. In this way, consequently, we can
obtain the estimate of f(yt) by removing the estimate of parametric part from the estimate of
gt.
Define
(3.9)
then, we get
Adaptive Control
42
(3.10)
Thus, intuitively, we can take
(3.11)
as the estimate of
gt
at time
t.
Design of Control ut: Let
(3.12)
Under Assumptions 3.1-3.4, we can design the following control law
(3.13)
where D is an appropriately large constant, which will be addressed in the proof later.
Remark 3.4 The controller designed above is different from most traditional adaptive controllers in
its special form, information utilization and computational complexity. To reduce its computational
complexity, the interval It given by Eq. (3.6) can be calculated recursively based on the idea in Eq.
(3.12).
3.3 Stability of Closed-loop System
In this section, we shall investigate the closed-loop stability of system (3.1) using the
adaptive controller given above. We only discuss the case that the parametric part is of
linear growth rate, i.e. b = 1. For the case where the parametric part is of nonlinear growth
rate, i.e. b > 1, though simulations show that the constructed adaptive controller can stabilize
the system under some conditions, we have not rigorously established corresponding
theoretical results; further investigation is needed in the future to yield deeper
understanding.
3.3.1 Main Results
The adaptive controller constructed in last section has the following property:
Theorem 3.1 When
b = 1,
ML 3
< + 2 , the controller defined by Eqs. (3.5) (3.13) can
M' 2
guarantee that the output {yt} of the closed-loop system is bounded. More precisely, we have
43
(3.14)
Based on Theorem 3.1, we can classify the capability and limitations of feedback mechanism
for the system (3.1) in case of b = 1 as follows:
Corollary 3.1 For the system (3.1) with both parametric and non-parametric uncertainties, the
following results can be obtained in case of b = 1:
(i) If
b = 1,
ML 3
< + 2 , then there exists a feedback control law guaranteeing that the closedM' 2
loop system is stabilized.

(ii) When
t = yt
the critical value
(i.e.
g ( x) = x ), the presence of uncertain parametric part t
3
+ 2 of the feedback
2
does not reduce
mechanism which is determined by the uncertainties of
non-parametric part.
Proof of Corollary 3.1: (i) This result follows from Theorem 3.1 directly. (ii) When g(x) = x, we
can take M = M = 1. In this case, the sufficiency can be immediately obtained via Theorem
3.1; on the other hand, the necessity can be obtained by the impossibility part of Theorem
1 in [XG00]. In fact, if
3
+ 2 , for any given control law {ut}, we need only take the
2
parameter = 0, then by [XG00, Theorem 2.1], there exists a function f such that system (3.1)
cannot be stabilized by the given control law.
Remark 3.5 As we have mentioned in the introduction part, system (1.6), a special case of system
(3.1), has been studied in [XG00]. Comparing system (3.1) and system (1.6), we can see that system
(3.1) has also parametric uncertainty besides nonparametric uncertainty and noise disturbance.
Hence intuitively speaking, it will be more difficult for the feedback mechanism to deal with
uncertainties in system (3.1) than those in system (1.6). Noting that M' M, we know this fact has
been partially verified by Theorem 3.1. And Corollary 3.1 (ii) indicates that in the special case of
t = yt , since the structure of parametric part is completely determined, the uncertainty in non-
parametric part becomes the main difficulty in designing controller, and the parametric uncertainty
has no influence on the capability of the feedback mechanism, that is to say, the feedback mechanism
can still deal with the non-parametric uncertainty characterized by the set F(L) with
L<
3
+ 2.
2
Remark 3.6 Theorem 3.1 is also consistent with classic results on adaptive control for linear systems.
In fact, when L = 0, the non-parametric part f(yt) vanishes, consequently system (3.1) becomes a
linear-in-parameter system
yt +1 = t + ut + wt +1
(3.15)
Adaptive Control
44
where
is the unknown parameter, and
t = g ( yt )
can have arbitrary linear growth rate because
by Theorem 3.1, we can see that no restrictions are imposed on the values of M and M ' when L =
0. Based on the knowledge from existing adaptive control theory [CG91], system (3.15) can be always
stabilized by algorithms such as minimum-variance adaptive controller no matter how large the is.
Thus the special case of Theorem 3.1 reveals again the well-known result in a new way, where the
adaptive controller is defined by Eq. (3.13) together with Eqs. (3.5)(3.12).
Corollary 3.2 If b = 1,
ML 3
< + 2 , c = w = 0 , then the adaptive controller defined by Eqs.
M' 2
(3.5) (3.13) can asymptotically stabilize the corresponding noise-free system, i.e.
(3.16)
3.3.2 Preliminary Lemmas
To prove Theorem 3.1, we need the following Lemmas:
Lemma 3.1 Assume {xn} is a bounded sequence of real numbers, then we must have
(3.17)
Proof: It is a direct conclusion of [XG00, Lemma 3.4]. It can be proved by argument of
contradiction.
Lemma 3.2 Assume that L (0,
3
+ 2 ), d 0, n0 0 . If non-negative sequence {hn, n 0}
2
satisfies
(3.18)
where
x + = max( x,0), x R , then we must have
(3.19)
Proof: See [XG00, Lemma 3.3].
3.3.3 Proof of Theorem 3.1
Proof of Theorem 3.1: We divide the proof into four steps. In Step 1, we deduce the basic
relation between yt+1 and , and then a key inequality describing the upper bound of
| yt yit |
is established in Step 2. Consequently, in Step 3, we prove that
| yt yit | 0
45
as t if yt is not bounded, and hence the boundedness of output sequence {yt} can be
guaranteed. Finally, in the last step, the bound of tracking error can be further estimated
based on the stability result obtained in Step 3.
Step 1: Let
(3.20)
then, by definition of ut and Eq. (3.13), obviously we get
(3.21)
yt#+1 . By Eq. (3.11) and Eq. (3.1), we get
Now we discuss
(3.22)
In case of
t = i
, i.e. yt = yit , obviously we get
(3.23)
otherwise, we get
(3.24)
where
Obviously
Dij = D ji . In the latter case, i.e. when t it , for any (i, j ) J t , noting that
Adaptive Control
46
(3.25)
we obtain that
(3.26)
Therefore
(3.27)
where
(3.28)
Step 2: Since
3
ML 3
ML
< + 2 , there exists a constant > 0 such that
+ < + 2 .
2
M' 2
M'
Let
(3.29)
and consequently
(3.30)
By the definitions of
bt , bt
and
Bt , we obtain that
(3.31)
By the definition of
it , obviously we get
47
(3.32)
Step 3: Based on Assumption 3.4, for any fixed
> 0 , we can take constants D andD such
4 M (2 w + c ) when | y y |> D . Now we are ready to show that for

that | i j |> D >
t
it
'
any s > 0, there always exists t > s such that
| yt yit |> D
In fact, suppose that it is not true, then there must exist s > 0 such that
any t > s, correspondingly
t it
| yt yit |> D for
> D. Consequently, by the definition of D, for
sufficiently large t and j < t, we obtain that
(3.33)
together with the definition of
t , we know that for any s < i < j < t,
(3.34)
hence for
s < j < t , i = i j , we get
(3.35)
Now we consider
Let
d n = Dn ,in ,
Dt ,it D j ,i j .
then, by the definition of Di,j , noting that
| y j yi || y j yi j |> D for
any j > s, we obtain that
(3.36)
so we can conclude that {dn, n > s} is bounded. Then, by Lemma 3.1, we conclude that
(3.37)
48
Adaptive Control
Consequently there exists s > s such that for any t > s, we can always find a corresponding
j=j(t) satisfying
(3.38)
Summarizing the above, for any t > s, by taking j = j(t), we get
(3.39)
Therefore
(3.40)
Since |yt yit | > D, we know that
(3.41)
From Eq. (3.39) together with the result in Step 2, we obtain that
(3.42)
Thus noting (3.40), we obtain the following key inequality:
(3.43)
where
(3.44)
Considering the arbitrariness of t > s, together with Lemma 3.2, we obtain that
49
(3.45)
and consequently { | Bt | } must be bounded. By applying Lemma 3.1 again, we conclude
that
(3.46)
which contradicts the former assumption!
Step 4: According to the results in Step 3, for any s > 0, there always exists t > s such that
| yt yit | D
. Then, we can easily obtain that {
~
| t | }
is bounded, say
~
| t | L'
Considering that
(3.47)
we can conclude that
(3.48)
.
where
The proof below is similar to that in [XG00]. Let
(3.49)
Because of the result obtained above, we conclude that for any n 1, tn is well-defined and tn
< . Let
vn = ytn , then obviously {vn} is bounded. Then, by applying Lemma 3.1, we get
(3.50)
as
n . Thus for any > 0 , there exists an integer n0 such that for any n > n0,
(3.51)
So
(3.52)
Adaptive Control
50
By taking
sufficiently small, we obtain that

(3.53)
for any n > n0.

Thus based on definition of tn, we conclude that tn+1 = tn + 1! Therefore for any t t n ,
0
(3.54)
which means that the sequence {yt} is bounded.
Finally, by applying Lemma 3.1 again, for sufficiently large t,
| y t yit | consequently
(3.55)
Because of arbitrariness of
, Theorem 3.1 is true.
3.4 Simulation Study

In this section, two simulation examples will be given to illustrate the effects of the adaptive
controller designed above. In both simulations, the tracking signal is taken as
y t* = 10 sin
t and the noise sequence is i.i.d. randomly taken from uniform distribution
10
U(0, 1). The simulation results for two examples are depicted in Figure 8 and Figure 9,
respectively. In each figure, the output sequence
and the reference sequence
are
plotted in the top-left subfigure; the tracking error sequence

bottom-left subfigure; the control sequence
ut
et = yt yt*
is plotted in the
is plotted in the top-right subfigure; and the
parameter together with its upper and lower estimated bounds is plotted in the bottomright subfigure.
Simulation Example 1: This example is for case of b = 1, and the unknown plant is
(3.56)
with
L = 2 .9 <
3
+ 2 , g ( x) = x
2
(i.e.
b = 1, M = M ' = 1 )
and
(3.57)
For this example, we can verify that
51
(3.58)
consequently | f ( x ) f ( y ) |< L | x y | , i.e. f () F ( L) ..
Simulation Example 2: This example is for case of b > 1, and the unknown plant is
(3.59)
with
L = 2 .9 , g ( x ) = x 2
(i.e.
b = 2 , M = M ' = 1 ), and
(3.60)
For this example, we can verify that | f ( x ) f ( y ) |< L | x y | +2 , i.e. f ( ) F ( L ) .

From the simulation results, we can see that in both examples, the adaptive controller can
track the reference signal successfully. The simulation study verified our theoretical result
and indicate that under some conditions, the adaptive control law constructed in this paper
can deal with both parametric and non-parametric uncertainties, even in some cases when
the parametric part is of nonlinear growth rate. In case of b = 1, the stabilizability criteria
have been completely characterized by a simple algebraic condition; however, in case of b >
1, it is very difficult to give complete theoretical characterization. Note that usually more
accurate estimate of parameter can be obtained in case of b > 1 than in case of b = 1,
however, worse transient performance may be encountered.
Fig. 8. Simulation example 1: (g(x) = x, b = 1,M = M = 1)
Adaptive Control
52
Fig. 9. Simulation example 2: (g(x) = x2, b = 2,M = M = 1)
4. Semi-parametric Adaptive Control: Example 2

In this section, we shall give another example of adaptive estimation and control for a semiparametric model. Although the system considered in this section is similar to the model
considered in last section, there are several particular points in this example:
The controller gain in this model is also unknown with a priori knowledge on its sign and
its lower bound.
The system is noise-free, and correspondingly the asymptotic tracking is rigorously
established in this example.
The algorithm in this example has a form of gradient algorithm, however, it partially
makes use of a priori knowledge on the non-parametric part.
Due to the limitation of this algorithm and technical difficulties, unlike the algorithm in
last section, we can only establish stability of the closed-loop system under condition
0 < L < 0.5
0L<
for the parametric part, which is much stronger than the condition
3
+ 2
.
2
This example is given here only for the purpose of demonstrating that there exist other
possible ways to make use of a priori knowledge on the parametric uncertainties and nonparametric uncertainties. By comparing the examples in this section and last section, the
readers may get a deeper understanding to adaptive estimation and control problems for
semi-parametric models.
53
4.1 Problem Formulation

We consider the following system model
(4.1)
where
yk R1
and
u k R1
the unknown parameter,
are output and control signals, respectively. Here
R1 is
b R1 is the unknown controller gain, () is a known function,
and f() is the unknown function. We have the following a priori knowledge on the real
system:
Assumption
4.1
The
nonparametric
uncertain
function
f()
is
Lipschitz,
i.e.,
|| f ( x1 ) f ( x2 ) || L || x1 x2 ||, x1 , x2 R , where L < 0.5. The known function () is also a
Lipschitz function with Lipschitz constant L.

Assumption 4.2 The sign of unknown controller gain b is known. Without loss of generality, we
assume that
bb>0
where
b is a known constant.
yk* is a known bounded deterministic signal.

design the control law u k such that the output
Assumption 4.3 The reference signal

The control objective is to
asymptotically tracks a bounded reference trajectory
yk*
signal yk
and all the closed-loop signals are
guaranteed to be bounded.
4.2 Adaptive Control Design
To design the adaptive controller, the following notations will be used throughtout this
section:
(4.2)
Obviously, at time step k, with the history information {yj , j k} and the a priori knowledge,
the index
lk
and the tracking error
ek
are available. Later we will see important roles of
lk and ek in the controller design.

Estimation of parametric part: The estimates of the parameter
and the controller gain b at
and , respectively. We design the following adaptive

time step k are denoted by
update law to update the parameter estimates recursively:
Adaptive Control
54
where
0 < < 1 and the coefficient ak
is defined by a time-varying deadzone:
(4.4)
Estimation of non-parametric part: As in last section, we do not estimate the nonparametric part directly. Instead, we try to estimate the parametric part and nonparametric part as a whole part
(4.5)
Noticing of the system model (4.1), we know that
(4.6)
consequently, from Eqs. (4.5) and (4.6), it is easy to derive
(4.7)
Since function f() is unknown and parameters
#
k
and b are unknown, we simply estimate
by the following eqution
(4.8)
where
and
f ( y k ) f ( yl k )
are
regarded
as
true
parameters,
in Eq. (4.7) is simply dropped off.
and
the
unknown
term
55
Adaptive control law: By Eq. (4.6), according to the certainty equivallence principle, we can
design the following adaptive control law
(4.9)
Where
bk
and
y k# are given by Eqs. (4.3) and (4.7). The closed-loop stability will be given
later.
4.3 Asymptotic Tracking Performance
4.3.1 Main Results
Theorem 4.1 In the closed-loop system (4.1) with control law (4.9) and parameters adaptation law
(4.3), under Assumptions 4.14.3, all the signals in the closed-loop system are bounded and further
the tracking error
ek
will converge to zero.
4.3.2 Preliminaries
Definition 4.1 Let
We denote
xk
and
yk ( k 0 ) be two discrete-time scalar or vector signals.
xk = O[ yk ] ,
if there exist positive constants m1 and m2 such that || xk || m1
max j k || y j || + m2 , k > k0 and k0 is the initial time step.

We denote
xk = o[ yk ] , if there exists a sequence k
satisfying
lim k k 0
such that
|| xk || m1 max j k || y j || + m2 , k > k0 .
We denote
xk ~ y k
if they satisfy
xk = O[ yk ] and yk = O[ xk ] .
Lemma 4.1 Consider the following parameter update law
(4.10)
(4.11)
(4.12)
Adaptive Control
56
where
and k
where
is an unknown scalar,
R is any
sequence. Then,
k' = k'
and
is its estimate at time step
is guaranteed and the following properties hold:
k = k .
Proof: According to Eqs. (4.10) and (4.11), it is obvious that

(4.12), we see that
k always hold. From Eq.
| Proj ( k ) |=| k | , hence Proj2 ( k ) = k2 . Further, we have
From (4.10), we see that

that when
k , is the lower bound of ,
k' = k
if
k' >
such that
k'2 = k2
when
k' > . Noticing
k' , we have , so that
(4.13)
Therefore, we always have
k'2 k2 . This completes the proof.
Lemma 4.2 Given a bounded sequence
X k R m . Define
Then, we have
Proof: This lemma is an extension of Lemma 3.1. Its proof can be found in [Ma06].
Lemma 4.3 (Key Technical Lemma)Let
of vectors such that
{st } be a sequence of real numbers and { t } be a sequence
57
Assume that
where
1 > 0, 2 > 0 . Then || t || is bounded.
Proof: This lemma can be found in [AW89, GS84].

4.3.3 Proof of Theorem 4.1
Define parameter estimate errors
we have
and
. From Eqs. (4.7) and (4.8),
(4.14)
Then, we can derive the following error dynamics:
(4.15)
According to Assumption 4.1, we have

(4.16)
where
can be any constant satisfying
From the error dynamics Eq. (4.15), we have
(4.17)
Adaptive Control
58
Choose Lyapunov function candidate as
(4.18)
From the adaptation laws (4.3), we obtain that
(4.19)
(4.20)
(4.21)
Together with the error dynamics Eq. (4.17), we can derive that the difference of Vk
(4.22)
Noting that 0 < ( < 1 and taking summation on both hand sides of Eq. (4.22), we obtain
Which implies
(4.23)
and the boundedness of
and
. Considering
, we have
where and C2 are some constants. From the definition of deadzone in Eq. (4.4), we have
.
Therefore, we have
59
(4.24)
Therefore, we have
(4.25)
Note that < 0.5, we have
(4.26)
holds for all < *, where C3 is some finite number. Note that inequality Eq. (4.26) means
that
yk 1 = O[ak ek ] . Further we have
Therefore, we can apply the Key Technical Lemma (Lemma 4.3) to Eq. (4.23) and obtain that
(4.27)
which guarantees the boundedness of yk from Eq. (4.26) and thus, the boundedness of
output yk, tracking error ek. Therefore, applying Lemma 4.2 yields
(4.28)
lim k k ek 0 leads to lim k ek 0 . From

definition of deadzone in Eq. (4.4), we have ak [0,1) . Let us define the following sets:
Next, we will show that
the
(4.29)
which results in
Z1+ Z 2+ =
and
Z1+ Z 2+ = Z + . The following three cases need to
be considered. In every case, we only need to discuss the case where k belongs to an infinite
set.
Adaptive Control
60
Case i).
Z1+
is an infinite set and
Z 2+
is a finite set. Let us discuss
k Z1+
. From the
definition in Eq. (4.29), it follows that ak = 0. Hence it is clear from the definition of deadzone
(4.4) that 0 | ek | | yk 1 yl | which means lim k ek 0 according to Eq. (4.28).
k 1
Case ii).
+
1
is a finite set and
Z 2+
is an infinite set. Let us discuss
k Z 2+
. From the
definition in (4.29), it follows that ak 0. Hence it is clear from deadzone (4.4) that
| ek |= ak | ek | + | yk 1 ylk 1 | which means lim k ek = 0 due to Eqs. (4.27) and (4.28).

+
Z1 and Z 2+ are infinite sets. If k Z1+ then ak = 0. Following Case

lim k ek = 0 . Otherwise, ak 0, it follows from Case ii) that lim k ek = 0 .
Case iii).
Based on the discussions for the above three cases, we can conclude that
implies that
i) gives
lim k ak ek = 0
lim k ek = 0 . This completes the proof.
5. Conclusion
In this chapter, we have formulated and discussed the adaptive estimation and control
problems for a class of semi-parametric models with both parametric uncertainty and nonparametric uncertainty. For a typical semi-parametric system model, we have discussed new
ideas and principles in how to estimate the unknown parameters and non-parametric part
by making full use of a priori knowledge, and for a typical type of a priori knowledge on the
non-parametric part, we have proposed novel information-concentration estimator so as to
deal with both kinds of uncertainties in the system, and some implementation issues in
various cases have been discussed with applicable algorithm descriptions. Furthermore, we
have applied the ideas of adaptive estimation for semi-parametric model into two examples
of adaptive control problem for two typical semi-parametric control systems, and discussed
in details how to establish the closed-loop stability of the whole system with semiparametric adaptive estimator and controller. Our discussions have demonstrated that the
topic in this chapter is very challenging yet important due to its wide
background. Especially, for the closed-loop analysis problem of semi-parametric adaptive
control, the examples given in this chapter illustrate different methods to overcome the
difficulties.
In the first example of semi-parametric adaptive control, we have investigated a simple firstorder nonlinear system with both non-parametric uncertainties and parametric
uncertainties, which is largely motivated by the recent-year exploration of the capability and
limitations of the feedback mechanism. For this model, based on the principle of the
proposed IC estimator, we have constructed a unified adaptive controller which can be used
in both cases of b = 1 and b > 1. When the parametric part is of linear growth rate (b = 1), we
have proved the closed-loop stability under some assumptions and a simple algebraic
condition
ML 3
< + 2
M 2
, which reveals essential connections with the known magic
number
L=
3
+ 2
2
61
discovered in recent work [XG00] on the study of feedback
mechanism capability.
In the second example of semi-parametric adaptive control, we further assume that the
control gain is also unknown, yet the system is noise-free, and we have designed an
adaptive controller based on gradient-like estimation algorithm with time-varying deadzone
according to the a priori knowledge on the non-parametric part and the unknown controller
gain. In this example, although we cannot establish perfect results revealing the magic
number
3
+ 2
2
as in the first example, we can still establish good results of asymptotic
tracking performance under some mild conditions. This example has demonstrated yet
another method to deal with uncertainties in semi-parametric model.
Finally, we shall remark that the discussed topic in this chapter is still in its infant stage, and
many more chanlenging problems can be investigated in the future. These problems may
root in wide practical background where the system model is only partially known a priori,
that is to say, the major part of the system can be parameterized and the other part is
unknown and non-parameterized with only limited a priori knowledge. Solving such
problems can definitely improve our understanding to the whole feedback mechanism and
help us gain more insights on the capability of adaptive control, especially non-traditional
adaptive control methods which were not extensively addresed and studied in previous
study. Therefore, we expect more theoretical study in this new topic, i.e. semi-parametric
adaptive estimation and control.
6. Acknowledgement
The authors would like to thank Chenguang Yang for his helpful discussion and valuable
comments.
7. References
[ASL98] A. M. Annaswamy, F. P. Skantzes, and A.-P. Loh, Adaptive control of continuous
time systems with convex/concave parametrization, Automatica, vol. 34, no. 1, pp.
3349, 1998.
[AW89] K. Astrm and B. Wittenmark, Adaptive Control. Addison-Wesley Pupl. Comp., 1989.
[BDB95] R. M. Bethea, B. S. Duran, and T. L. Boullion, Statistical Methods for Engineers and
Scientists. CRC Press, 1995.
[Ber04] R. A. Berk, Regression Analysis: A Constructive Critique. SAGE, 2004.
[BJR08] G. E. P. Box, G. M. Jenkins, and G. C. Reinsel, Time Series Analysis: Forecasting and
Control, 4th ed. Wiley, Jun 2008.
[Bos95] J. D. Boskovic, Stable adaptive control of a class of first-order nonlinearly
parameterized plants, IEEE Transactions on Automatic Control, vol. 40, pp. 347350,
1995.
[CG91] H. F. Chen and L. Guo, Identification and Stochastic Adaptive Control. Boston, MA:
Birkhuser, 1991.
62
Adaptive Control
[DS98] N. R. Draper and H. Smith, Applied Regression Analysis, 3rd ed. Wiley-Interscience,
Apr 1998.
[FL99] G. Feng and R. Lozano, Adaptive Control Systems. Butterworth Heinemann, 1999.
[Fox97] J. Fox, Applied Regression Analysis, Linear Models, and Related Methods. SAGE, 1997.
[Gau09] C. F. Gauss, Theoria motus corporum coelestium in sectionibus conicis solem ambientium:
in sectionibus conicis solem ambientium (Theorie der Bewegung der Himmelskrper,
die die Sonne in Kegelschnitten umkreisen). Dover, New York, 1809, english
translation by C. H. Davis, reprinted 1963.
[GHZ99] S. S. Ge, C. C. Hang, and T. Zhang, A direct adaptive controller for dynamic
systems with a class of nonlinear parameterizations, Automatica, vol. 35, pp. 741
747, (1999.
[Gio85] A. A. Giordano, Least Square Estimation With Applications to Digital Signal Processing.
Wiley-Interscience, Mar 1985.
[GS84] G. Goodwin and K. Sin, Adaptive Filtering, Prediction and Control. Englewood Cliffs,
NJ: Prentice-Hall, 1984.
[Guo97] L. Guo, On critical stability of discrete-time adaptive nonlinear control, IEEE
Transactions on Automatic Control, vol. 42, no. 11, pp. 14881499, 1997.
[Hel63] Z. Hellwig, Linear regression and its application to economics. Pergamon Press, 1963.
[Kan94] I. Kanellakopoulos, A discrete-time adaptive nonlinear system, IEEE Transactions
on Automatic Control, vol. 39, Nov. 1994.
[LH74] C. Lawson and R. Hanson, Solving Least Squares Problems. Prentice-Hall, 1974.
[Lju98] L. Ljung, System IdentificationTheory for the User, 2nd ed. Upper Saddle River, NJ
07458: Prentice Hall, 1998.
[LQF03] A. P. Loh, C. Y. Qu, and K. F. Fong, Adaptive control of discrete time systems with
concave/convex parametrizations, IEEE Transactions on Automatic Control, vol. 48,
no. 6, 2003.
[LX06] C. Y. Li and L. L. Xie, On robust stability of discrete-time adaptive nonlinear
control, Systems and Control Letters, vol. 55, no. 6, pp. 452458, June 2006.
[Ma06] H. B. Ma, Capability and limitation of feedback mechanism in dealing with
uncertainties of some discrete-time control systems, Phd Thesis, Graduate School
of Chinese Academy of Sciences, Beijing, China, June 2006.
[Ma08a] , Further results on limitations to the capability of feedback, International
Journal of Control, vol. 81, no. 1, pp. 21 42, 2008. [Online]. Available:
http://www.informaworld.com/10.1080/00207170701218333
[Ma08b] , An impossibility theorem on a class of high-order discrete-time nonlinear
control systems, Systems & Control Letters, vol. 57, no. 6, pp. 497504, 2008.
[Online]. Available: http://dx.doi.org/10.1016/j.sysconle.2007. 11.008
[MG05] H. B. Ma and L. Guo, An impossibility theorem on second-order discrete-time
nonlinear control systems, in Proceedings of the 24th Chinese Control Conference.
South China University of Technology Press: Guangzhou, July. 2005, pp. 5761.
[MPV07] D. C. Montgomery, E. A. Peck, and G. G. Vining, Introduction to Linear Regression
Analysis, Solutions Manual, 4th ed., ser. Wiley Series in Probability and Statistics.
Wiley, John Sons, Incorporated, Feb 2007.
[MT95] R. Marino and P. Tomei, Nonlinear adaptive design: Geometric,adaptive, and robust.
London: Printice-Hall, 1995.
63
[SB89] S. Sastry and M. Bodson, Adaptive Control: Stability, Convergence, and Robustness.
Englewood, New Jersey: Prentice-Hall, 1989.
[Sok03] V. F. Sokolov, Adaptive suboptimal tracking for the first-order plant with lipschitz
uncertainty, IEEE Transactions on Automatic Control, vol. 48, no. 4, Apr 2003.
[Tsa05] R. S. Tsay, Analysis of Financial Time Series, 2nd ed., ser. Wiley Series in Probability
and Statistics. Wiley, John Sons, Incorporated, Aug 2005, no. 543.
[VV07] M. Verhaegen and V. Verdult, Filtering and System Identification: A Least Squares
Approach, 1st ed. Cambridge University Press, May 2007.
[web08] Statistics & data analysis for environmental science and management, 2008,
available
online
from
http://www.esm.ucsb.edu/academics/courses/206/Readings/re ader Ch8.pdf.
[Wei05] S. Weisberg, Applied Linear Regression, 3rd ed. Wiley-Interscience, Jan 2005.
[Wik08a] Wikipedia, Gauss-Markov theorem wikipedia, the free encyclopedia, 2008.
[Online].
Available:
http://en.wikipedia.org/w/index.php?title=Gauss%E2%80%93Markov
theorem&oldid=232973932
[Wik08b] , Generalized linear model wikipedia, the free encyclopedia, 2008.
[Online].
Available:
http://en.wikipedia.org/w/index.php?title=Generalized
linear model&oldid=237425065
[Wik08c] , Kernel regression wikipedia, the free encyclopedia, 2008. [Online].
Available:
http://en.wikipedia.org/w/index.php?title=Kernel
regression&oldid=234955359
[Wik08d] , Kriging wikipedia, the free encyclopedia, 2008. [Online]. Available:
http://en.wikipedia.org/w/index.php?title=Kriging&oldid=234876679
[Wik08e] , Least squares wikipedia, the free encyclopedia, 2008. [Online].
Available:
http://en.wikipedia.org/w/index.php?title=Least
squares&oldid=233230305
[Wik08f] , Linear least squares wikipedia, the free encyclopedia, 2008. [Online].
Available:http://en.wikipedia.org/w/index.php?title=Linearleastsquares&oldid=
23323034 42
[Wik08g] , Linear regression wikipedia, the free encyclopedia, 2008. [Online].
Available:
http://en.wikipedia.org/w/index.php?title=Linear
regression&oldid=237064808
[Wik08h] , Moore-Penrose pseudoinverse wikipedia, the free encyclopedia, 2008.
[Online].Available:http://en.wikipedia.org/w/index.php?title=MoorePenrosepseudoinver se&oldid=237863190
[Wik08i] , Nonparametric regression wikipedia, the free encyclopedia, 2008.
[Online].Available:http://en.wikipedia.org/w/index.php?title=Nonparametricreg
ression& oldid=235577025
[Wik08j] , Regression analysis wikipedia, the free encyclopedia, 2008. [Online].
Available:http://en.wikipedia.org/w/index.php?title=Regressionanalysis&oldid
=2322210 82
[XG99] L. L. Xie and L. Guo, Fundamental limitations of discrete-time adaptive nonlinear
control, IEEE Transactions on Automatic Control, vol. 44, no. 9, pp. 17771782, 1999.
[XG00] , How much uncertainty can be dealt with by feedback? IEEE Transactions on
Automatic Control, vol. 45, no. 12, pp. 22032217, 2000.
64
Adaptive Control
[XG01] F. Xue and L. Guo, Necessary and sufficient conditions for adaptive stabilizability
of jump linear systems, Communications in Information and Systems, vol. 1, no. 2, pp.
205224, Apr. 2001.
[YDG+08] C. G. Yang, S. L. Dai, S. S. Ge, Z. X. Jiao, and T. H. Lee, Adaptive asymptotic
tracking control of a class of strict-feedback discrete-time nonlinear systems, IEEE
Transactions on Automatic Control, 2008, to be submitted.
[ZG02] Y. X. Zhang and L. Guo, A limit to the capability of feedback, IEEE Transactions on
Automatic Control, vol. 47, no. 4, pp. 687692, 2002.
3
Adaptive output regulation of unknown linear
systems with unknown exosystems
Ikuro Mizumoto and Zenta Iwai
Department of Intelligent Mechanical Systems, Kumamoto University,

2-39-1 Kurokami, Kumamoto, 860-8555, Japan
1. Introduction
The problems of the output regulations and/or disturbance reductions have attracted a lot
of interest and have been actively researched in the consideration of the control problem for
systems which are required to have servomechanism and for vibration attenuation in
mechanical systems. It is well known that such problems are solvable using the Internal
Model Principle in cases where the system to be controlled and the exosystem which
generates the output reference signal and external disturbances are known. In the case
where the controlled system is unknown and/or the exosystem is unknown, adaptive
control strategies have played active roles in solving such problems for systems with
uncertainties. For known systems with unknown exosystems, solutions with adaptive
internal models have been provided in (Feg & Palaniswami, 1991), (Nikiforov, 1996) and
(Marino & Tomei, 2001). In (Marino & Tomei, 2001), an output regulation system with an
adaptive internal model is proposed for known non-minimum phase systems with
unknown exosystems. Adaptive regulation problems have also been presented for time
varying systems and nonlinear systems (Marino & Tomei, 2000; Ding, 2001; Serrani et al.,
2001). Most of these methods, however, assumed that either the controlled system or the
exosystem was known. Only few adaptive regulation methods for unknown systems with
unknown exosystems have been provided (Nikiforov, 1997a; Nikiforov, 1997b). The method
in (Nikiforov, 1997a) is an adaptive servo controller design based on the MRAC strategy, so
that it was essentially assumed that the order of the controlled system was known. The
method in (Nikiforov, 1997b) is one based on an adaptive backstepping strategy. In this
method, it was necessary to design an adaptive observer that had to estimate all of the
unknown system parameters depending on the order of the controlled system. Further, the
controller design based on the backstepping strategy essentially depends on the order of the
relative degree of the controlled system. As a result, the controller's structure was quite
complex in both methods for higher order systems with higher order relative degrees.
In this paper, the adaptive regulation problem for unknown controlled systems is dealt with
and an adaptive output feedback controller with an adaptive internal model is proposed for
single input/single output linear minimum phase unknown systems with unknown
exosystems. The proposed method is based on the adaptive output feedback control
Adaptive Control
66
utilizing the almost strictly positive real-ness (ASPR-ness) of the controlled system and the
controller is designed based on an expanded backstepping strategy with a parallel
feedforward compensator (PFC) (Mizumoto et al., 2005). It is shown that, under certain
assumptions, without a priori knowledge of the order of the controlled system and without
state variables, one can design an adaptive controller with a single step backstepping
strategy even when the system to be controlled has an unknown order and a higher order
relative degree. Using the proposed method, one can not attain perfect output regulation,
however, the obtained controller structure is relatively simple even if the system has a
higher order and a higher order relative degree.
2. Problem Statement
Consider the following single input/single output LTI system.
x& (t ) = Ax(t ) + bu(t ) + C d w(t )
y(t ) = cT x(t ) + d T w(t ),
(1)
where x = [x1 ,L , x n ] T R n is the state vector and u , y R are the input and the output,
respectively. Further w(t ) R m is an unknown vector disturbance.
We assume that the disturbances and the reference signal which the output y is required to
track are generated by the following unknown exosystem:
& (t ) = A d w(t )
w
T
y m (t ) = cm
w(t ),
(2)
where A d R mm is a stable matrix with all its eigenvalues on the imaginary axis. It is also
assumed that the characteristic polynomial of Ad is expressed by
det(I A d ) = m + m 1 m 1 + L + 1 + 0 .
(3)
The objective is to design an adaptive controller that has the output y(t) track the reference
signal ym(t) generated by an unknown exosystem given in (2) for unknown systems with
unknown disturbances generated by the unknown exosystem in (2) using only the output
signal under the following assumptions.
Assumption 1 The system (1) is minimum-phase.
Assumption 2 The system (1) has a relative degree of r.
Assumption 3 cT A r 1b > 0 , i.e. the high frequency gain of the system (1) is positive.
Assumption 4 The output y(t) and the reference signal ym(t) are available for measurement.
Adaptive output regulation of unknown linear systems with unknown exosystems
67
3. System Representation
From Assumption 2, since the system (1) has a relative degree of r, there exists a smooth
nonsingular variable transformation: z T , T
= x such that the system (1) can be
transformed into the form (Isidori, 1995):
0
z& (t ) = Azz(t ) + bzu(t ) + T (t ) + Dd w(t )
cz
0
& (t ) = Q(t ) + z1 (t ) + Fd w(t )
(4)
y = [1,0,L,0]z1 (t ) + dT w(t ),
where
0 Ir 1r 1
Az =
,
a 0 L a r 1
bz = [0,L, bz ], bz = cTAr 1b,
and cz R n r is an appropriate constant vector. From assumption 1, Q is a stable matrix
because & (t ) = Q (t ) denotes the zero dynamics of system (1).
3.1 Virtual controlled system

We shall introduce the following (r-1)th order stable virtual filter 1 f (s ) with a state space
representation:
z& f (t ) = Auf zf (t ) + buf u(t )

uf1 (t ) = cTuf zf (t ),
where z f = z f1 ,L , z fr 1
and
0 Ir 2r 2
Au f =
,
0 L r 2
bTu f = [0,L,1], cTu f = [1,0,L,0].
With the following variable transformation using the filtered signal z f i given in (5):
(5)
Adaptive Control
68
u(t )
u f1 (t )
1
f(s )
u(t )
f (s )
y(t )
Controlled
system
Virtual controlled system

Fig. 1. Virtual controlled system with a virtual filter
1 (t ) = z1 (t )
i (t ) = bz u fi 1 (t ) + z i (t ) +
i 1
c
j=1
j zi j
(6)
(t ),
where
c i = i ar i ,
c r = a 0 +
(1 i r 1)
r 1
j 1 c
j=1
1 = r 2
i = r i 1 +
i 1
r i + j 1c
j=1
the system (1) can be transformed into the following virtual system which has u f1 given
from a virtual input filter as the control input (Michino et al., 2004) (see Fig.1):
& 1 (t ) = z (t ) + c1T y (t ) + bz u f1 (t ) + cdT1 w(t )
& y (t ) = A y (t ) + c 1 (t ) + Cd w(t )
(7)
y(t ) = 1 (t ) + d T w(t ),
where y = T , T
, = [ 2 , 3 ,L , r ] T and c1T = [1,0 ,L ,0] , c = cT ,0 ,L ,0 ,1
. cd 1 and Cd
are a vector and a matrix with appropriate dimensions, respectively. Further, A is given by
the form of
A
A = uf
0
czT .
69
Since A u f and Q are stable matrices, A is a stable matrix.

3.2 Virtual error system
Now, consider a stable filter of the form:
z& c f (t ) = A c f z c f (t ) + cc f u f1 (t )
u f (t ) = T z c f (t ) + u f1 (t ) ,
(8)
where cc f = [0 ,L ,0 ,1] T and
0 Im 1m 1
Acf =
c 0 ,L , c m 1
T = 0 c 0 ,L , m 1 c m 1 .
c 0 , c 1 ,L , c m 1 are chosen such that A c f is stable.
Let's consider transforming the system (7) into a one with uf given in (8) as the input. Define
new variables X1 and X 2 as follows:
X 1 = (1m ) + m 1 (1m 1) + L + 1& 1 + 0 1
X 2 = (ym ) + m 1(ym 1) + L + 1& y + 0 y .
(9)
Since it follows from the Cayley-Hamilton theorem that

m 1
+ L + 1A m + 0 I = 0 ,
Am
m + m 1A m
(10)
we have from (2) and (7) that

X& 1 (t ) = z X 1 (t ) + c1T X 2 (t ) + bz uf (t )
X& (t ) = A X (t ) + c X (t ),
(11)
u f = u(f1m ) + m 1u(f1m 1) + L + 1u& f1 + 0 u f1
(12)
where
Further we have from (10) that

e( m ) + m 1e( m 1) + L + 1e& + 0 e = X 1 .
(13)
Adaptive Control
70
Therefore defining E = e, e& ,L , e(m 1)
, the following error system is obtained:
E& (t ) = A EE(t ) + X 1 (t )
X& 1 (t ) = z X 1 (t ) + c1T X 2 (t ) + bz uf (t )
X& (t ) = A X (t ) + c X (t )
2
e(t ) = [1,0 ,L ,0]E(t ) .
(14)
Obviously this error system with the input u f and the output e has a relative degree of m+1
and a stable zero dynamics (because A is stable).
Furthermore, there exists an appropriate variable transformation such that the error system
(14) can be represented by the following form (Isidori, 1995):
0
0
z& e (t ) = A z e z e (t ) + u f (t ) + T z e (t )
b
z
e
c z e
0
& z e (t ) = Q z e z e (t ) + z e 1 (t )
1
(15)
e(t ) = z e 1 (t ) ,
where z e = z e 1 ,L , z e m + 1
and z e R n 1 . Since the error system (14) has stable zero
dynamics, Q z e is a stable matrix.

Recall the stable filter given in (8). Since we have from (8) that
u(fm ) + c m 1 u(fm 1) + L + c 1 u& f + c 0 u f
= u(f1m ) + m 1u(f1m 1) + L + 1u& f1 + 0 u f1 = u f ,
(16)
the filter's output signal uf can also be obtained from

0
z& c f (t ) = A c f zc f (t ) + uf (t )
1
u f (t ) = [1,0 ,L ,0 ] zc f (t )
by defining zc f = u f , u& f ,L , u(fm 1)
.Using this virtual filter signal in the variable
transformation given in (6), the error system (15) can be transformed into the following form,
the same way as the virtual system (7) was derived, with uf as the input.
e& (t ) = e e(t ) + be u f (t ) + ceT e (t )

& e (t ) = Q e e (t ) + be(t ),
(17)
71
where
Virtual error system
ym
Virtual controlled system

u
1
f(s )
u f1
n d (s) u f
dd (s )
dd (s) u f1
n d (s)
f(s )
Controlled
system
Fig. 2. Virtual error system with an virtual internal model
A
Qe = cf
0
0
czTe .
Q z e
Since A c f and Q z e are stable matrices, Qe is a stable matrix. Thus the obtained virtual error
system (17) is ASPR from the input uf to the output e.
The overall configuration of the virtual error system is shown in Fig.2.
4. Adaptive Controller Design

Since the virtual error system (17) is ASPR, there exists an ideal feedback gain k such that
the control objective is achieved with the control input: u f (t ) = k e(t ) (Kaufman et al., 1998;
Iwai & Mizumoto, 1994). That is, from (8), if the filter signal u f1 can be obtained by
u f1 (t ) = k e(t ) T z c f (t ),
(18)
one can attain the goal. Unfortunately one can not design u f1 directly by (18), because u f1 is
a filter signal given in (8) and the controlled system is assumed to be unknown. In such
cases, the use of the backstepping strategy on the filter (5) can be considered as a
countermeasure. However, since the controller structure depends on the relative degree of
the system, i.e. the order of the filter (5), it will become very complex in cases where the
controlled system has higher order relative degrees. Here we adopt a novel design strategy
using a parallel feedforward compensator (PFC) that allows us to design the controller
through a backstepping of only one step (Mizumoto et al., 2005; Michino et al., 2004).
4.1 Augmented virtual filter
For the virtual input filter (5), consider the following stable and minimum-phase PFC with
an appropriate order nf :
Adaptive Control
72
y& f (t ) = a f1 y f (t ) + a Tf2 f (t ) + ba u(t )
(19)
& f (t ) = A f f (t ) + bf y f (t ),
uf1
1
f(s)
uf
nd (s)
dd (s)
e
Virtual error system
uaf
PFC
Fig. 3. Virtual error system with an augmented filter

where y f R is the output of the PFC. Since the PFC is minimum-phase Af is a stable
matrix.
The augmented filter obtained from the filter (5) by introducing the PFC (19) can then be
represented by
z& u f (t ) = A z f z u f (t ) + bz f u(t )
u a f (t ) = czTf z u f (t ) = u f1 (t ) + y f (t ),
where z u f = z Tf , y f , Tf
(20)
and
A u f
Azf = 0
0
0
a f1
bf
czTf = cu f ,1,0 ,L ,0
0
b u f
a Tf2 , bz f = ba ,
0
A f
Here we assume that the PFC (19) is designed so that the augmented filter is ASPR, i.e.
minimum-phase and a relative degree of one. In this case, there exists an appropriate
variable transformation such that the augmented filter can be transformed into the following
form (Isidori, 1995):
u& a f (t ) = a a 1 u a f (t ) + a aT2 a (t ) + ba u(t )
0
& a (t ) = A a a (t ) + u a f (t ),
1
where Aa is a stable matrix because the augmented filter is minimum-phase.
73
Using the augmented filter's output u a f , the virtual error system is rewritten as follows (see
Fig.3):
e& (t ) = e e(t ) + be u a f (t ) + T z c f (t ) y f (t ) + ceT e (t )

& e (t ) = Q e e (t ) + be(t ).
(21)
4.2 Controller design by single step backstepping

[Pre-step] We first design the virtual input 1 for the augmented filter output u a f in (21) as
follows:
1 (t ) = k (t )e(t ) (t )T z c f (t ) + 0 (t ) ,
(22)
where k(t) is an adaptive feedback gain and (t ) is an estimated value of , these are
adaptively adjusted by
k& (t ) = k e 2 (t ) k k (t ) , k > 0 , k > 0
&
(t ) = z c f (t )e(t ) (t ) , T = > 0 , > 0 .
(23)
Further, 0 (t ) is given as follows:
& (t ) = D(y ) a (t ) + b u(t )
0
a
f
f1 0
0 , if
D(y f ) =
1, if
yf yf
(24)
yf > yf
where y f is any positive constant.

Now consider the following positive definite function:
V0 =
1 2
1
1
e +
k 2 + T 1 + eT Pe e ,
2 be
2k
2
(25)
where
k = k (t ) k , = (t ) ,
k is an ideal feedback gain to be determined later and Pe is a positive definite matrix that
satisfies the following Lyapunov equation for any positive definite matrix Re.
Pe Q e + Q eT Pe = R e < 0.
Adaptive Control
74
Since Qe is a stable matrix, there exists such Pe.

The time derivative of V0 can be evaluated by
& k v e 2 ( [R ] )
V
0
0
min
e
1
e
+ 1e
{y f 0 }e k 2 k 2
k
[ ]
(26)
min 1 3 2 + R 0
with any positive constant 1 to 3 . Where 1 = u a f 1 and
v0 = e +
be
(c
+ 2 Pe b b e
4 b e2 1
[ ])
2 min 1
2 k
R0 = k 2 +
4 3
4 2 k
2
(27)
2
2
[Step 1] Consider the error system, 1 -system, between u a f and 1 . The 1 -system is
given from (21) by
& 1 = u&a f &1

= aa1 ua f + aaT2 a + bau &1 .
(28)
The time derivative of 1 is obtained as follows:

& 1 =
1
1 T
1
1 T
ee +
b e u f1 +
1 z c f +
c e e
e
e
e
e
1 & 1
1 &
+
k+
z& c +
k (t )
zcf f

(29)
+ D (y f ) a f1 0 + b a u ,
where 1T = be T . Taking (28) and (29) into consideration, the actual control input is
designed as follows:
u=
1
ba
c 1 1 + 0 u a f
+ a
+ ,
1
1 1 1
2
if
yf yf
2
1
2
c 1 1 + 0 u a f + a 1 + 1 1 1 2
b a y f
1
3
2
2
f y f + 2 f y f
0 , if
yf > yf
ba
bayf
(30)
75
where 0 to 3 and f are any positive constants, and 1 and 2 are given by
2
1 & 2 1
1 =
k +
k

&2
1
+
zcf
z& c f
+l
2
2 =

ee 1 1T zc f 1 be uf1 + 1 1 1 ,
e
e
e
e
where l is any positive constant and e , be , 1 , 1 are estimated values of e , be , 1 , 1 ,

respectively, and adaptively adjusted by the following parameter adjusting laws.
& e (t ) = 1 (t ) 1 e(t ) e (t )
e
&
1
be (t ) = b1 (t )
uf (t ) bbe (t )
e 1
&
1 (t ) = 1 zc f (t ) 1 1 (t ) 1 1 (t )
e
(31)
&

1 (t ) = 1 1 (t ) 2 1 1 1 (t )
e
where , b , 1 , , b , 1 , 1 are any positive constants and 1 = T1 > 0 .
4.3 Boundedness analysis
For the designed control system with control input (30), we have the following theorem
concerning the boundedness of all the signals in the control system.
Theorem 1 Under assumptions 1 to 3 on the controlled system (1), all the signals in the
resulting closed loop system with the controller (30) are uniformly bounded.
Proof: Consider the following positive and continuous function V1.
V0 +
V1 =
V0 +
1
1 2 1 T 1
e2
1 + 1 1 1 +
2
2
2
1
1
1
be2 +
12 + 2y f , if
+
2 b
2 1
2
yf yf
(32)
1 2 1 T 1
1
1 + 1 1 1 +
e2
2
2
2
+
1
1
1
be2 +
12 + y 2f , if
2
2 b
2 1
yf > y f ,
Adaptive Control
76
where
e = e (t ) e , be = be (t ) be
= (t ) , = (t ) ,
1
and y f is any positive constant.

From (26) and (32), the time derivative of V1 for y f y f can be evaluated by
& k v 1 e 2 ( [R ] )
V
1
0
min
e
1
0
e
4l
k 2 k 2 min 1 3 2
[ ] )
[ ] )
(
c1 12 1 min
1
1
2
1
(33)
2 e2 b 3 be2

2
1 4 1 (y f 0 (y f ))e + R 1

with any positive constants 0 to 4 . Where
R1 = R0 +
[ ])
2 min 11
3
+ 1
4 1
41
2 2
2 b2
2 e2
+ b e +
.
42 4 3 b 4 4
Here we have
(y + 0 ) + (y f 0 ) 2 + e 2
(y f 0 )e = 5 e f
5
2 5
45
(34)
& (t ) = 0 is held, there

with any positive constant 5 . Furthermore, for y f y f , since
0
exists a positive constant M such that y f (t ) 0 (t ) M .
Therefore the time derivative of V1 can be evaluated by
& V + R
V
1
a 1
1
for y f y f , where
(35)
77
5 , s a ,2
a = min 2 be k v0
4
l
min 3
[R ] 1 0
,
,2 k k 2 ,2
sa = min min e
mac 1
max [Pe ]
2 c 1 ,2
[ ] ) ,2
[ ]
1
1
1 min
max
1
1
[ ] )
[ ]
2 ,

2 b b 3 ,2 1 4

R1 = R1 +
2
M
+ 2y f .
45
For y f > y f , the time derivative of V1 is evaluated as

& k v 1 e 2 ( [R ] ) 2
V
1
0
min
e
1
0
e
4l
k 2 k 2 min 1 3 2 c1 12
k
1 min 11 1 12 2 e2
b 3 be2 1 4 12 + R 1
[ ] )
[ ]
(36)
a f1 y 2f + a f2 f y f f y 2f 2 f y 2f 3 02
+ 0 e y f e ,
and thus we have for y f > y f that
& V + R ,
V
1
2
b 1
(37)
where
1
1
1
b = min 2 be k v0
, s a ,2 f
4l a f1 4 3
R2 = R1 +
a f2
4 2
(38)
Adaptive Control
78
Finally, for an ideal feedback gain k which satisfies

k > v0 +
1
1
1
+ v1 , v1 = max 5 ,
,
4l
a f1 4 3
the time derivative of V1 can be evaluated by

& V + R ,
V
1
1
(39)
where = min[ a , b ] > 0 , R = max R 1 , R 2 . Consequently it follows that V1 is uniformly

bounded and thus the signals e(t ), 1 (t ), e (t ), y f (t ), f (t ) and adjusted parameters k (t ), (t ),
(t ), (t ), b (t ), (t ) are also uniformly bounded.
e
Next, we show that the filter signal z c f and the control input u are uniformly bounded.
Define new variable z 1 as follows:
z(m1 ) + c m 1 z(m1 1) + L + c 0 z 1 = 1
(40)
& 1 = z 1 + cT y + bz u f1 + cdT1 w
(41)
& y = A y + b 1 + Cd w ,
(42)
where 1 and y have been given in (7). Further define z 1 by

z& 1 = z z 1 + bz z c f + 1
(43)
(m1 ) + c m 1 (m1 1) + L + c 0 1 = cT y + cdT1 w ,
(44)
where z c f 1 = [1,0 ,L ,0]z c f and we set z(k1 ) (0 ) = z(k1 ) (0 ), k = 0 ,L , m .We have from (40) and (41)
that
+ 1)
& 1 z 1 = z(m
+ c m 1 z z(m1 )
1
+ c m 2 z c m 1 z(m1 1) + L
+ c 0 z c 1 z& 1 z z 1
= bz u f1 + cT y + cdT1 w.
Further, we have from (43), (44) and (8) that
(45)
79
z(m1 + 1) + c m 1 z z(m1 ) + c m 2 z c m 1 z(m1 1)
+ L + c 0 z c 1 z& 1 z z 1
=
bz u f1 + cT y
(46)
cdT1 w.
It follows from (45) and (46) that z(k1 ) = z(k1 ) , k = 0 ,L , m .
Define z = z 1 , z& 1 ,L , z(m1 1)
and z = z 1 , z& 1 ,L , z(m1 1)
. Since sm + c m 1 s m 1 + L + c 0
is a stable polynomial, we obtain from (40) that
z = z l 1 1 + l 2 ,
(47)
with appropriate positive constants l1, l2. From the boundedness of w(t ) and e(t), we have
1 (t ) is bounded and thus z is also bounded.
Furthermore defining = 1 , & 1 ,L , (m1 1)
, we have from (44) that
0
& (t ) = A c f (t ) + cT y (t ) + cdT1 w(t ) .
1
(48)
From (8) and (48), we obtain

bz z& c f (t ) + & (t )
0
0
= A c f bz z c f (t ) + (t ) + bz u f1 (t ) + cT y (t ) + cdT1 w(t ) .
1

1
(49)
Therefore bz z& c f (t ) + & (t ) can be evaluated from (48) and the fact that u f1 = 1 + 1 y f by
bz z& c f (t ) + & (t ) A c f bz z c f (t ) + (t )
+ bz 1 (t ) y f (t ) + bz 1 (t )
(50)
+ c y (t ) + cd 1 w(t ) .
Here, we have from (22) that
1 (t ) y f (t ) = k (t ) e(t )
T (t )
T (t )
bz z c f (t ) + (t ) +
(t ) + 0 (y f ) y f (t ).
bz
bz
(51)
Adaptive Control
80
Since it follows from (19) and (24) that

& (t ) = a (y (t ) (t )) + a T (t )
y& f (t )
0
0
f1
f
f2 f
(52)
for y f > y f and from the boundedness of f (t ) , there exists a positive constant such that
y f 0 (t ) M . Further, from the boundedness of w(t ) and e(t) i.e. 1 (t ) , we can confirm
that y (t ) and (t ) are also bounded from (7) and (48). Finally, taking the boundedness of
the signals e(t), 1 (t ), k (t ), (t ) and y (t ), (t ) into consideration, from (50) bz z& c f (t ) + & (t ) can
be evaluated by
bz z& c f (t ) + & (t ) l z 1 bz z c f (t ) + (t ) + l z 2
(53)
with appropriate positive constants l z 1 and l z 2 . Consequently, considering the system:

z& (t ) = z z + bz z c f + (t )
(54)
from (44) with bz z c f + (t ) as the input and z as the output, since this system is
minimum-phase and the inequality (53) is held, we have from the Output/Input Lp Stability
Lemma (Sastry & Bodson, 1989) that the input bz z c f + (t ) in the system (54) can be
evaluated by
bz z c f (t ) + (t ) l z 1 z (t ) + lz 2
(55)
with appropriate positive constants lz 1 and lz 2 . From the boundedness of z (t ) and (t ) ,
we can conclude that z c f (t ) is uniformly bounded and then the control input u(t) is also
uniformly bounded. Thus all the signals in the resulting closed loop system with the
controller (30) are uniformly bounded.
5. Simulation Results
The effectiveness of the proposed method is confirmed through numerical simulation for a
3rd order SISO system with a relative degree of 3, which is given by
0.5 0.5
1
0
1 0.1 0.1 0.1
z& = 1.5 2.5 0.5 z + 0 u + 0.1 0.1 1 0.1 w

2.5 0.5
1
0.1 0.1 0.1 1
1
y = z 1 + [0.1 0.1 1 0.1] w ,
where w is an unknown disturbance which has the following form:
(56)
sin (2t )
2 cos(2 t )
w=
0.5 sin (5t )
2.5 cos(5t )
81
(57)
Before designing a controller, we first introduce the following pre-filter:

b
s+a
(58)
in order to reduce the chattering phenomenon to be expected by switching the controller

given in (30). Therefore, the considered controlled system has a relative degree of 4.
Since the relative degree of the controlled system is 4, we consider a 3rd order input virtual
filter in (5). Further we consider a stable internal model filter (8) of the order of 4.
For the input virtual filter, in this simulation, we consider a first order PFC:
y& f = a f1 y f + ba u
in order to make an ASPR augmented filter.
The design parameters for the pre-filter (58), the input virtual filter (5) and the internal
model filter (8) are set as follows:
a = b = 1000
0 = 15 , 1 = 75, 2 = 125
c 0 = 20 , c 1 = 150 , c 2 = 500 , c 3 = 625
and the PFC parameters are set by

a f1 = 10 , ba = 0.01.
Further design parameters in the controller given in (23), (24), (30) and (31) are designed by
k = 500 , k = 0.01, y f = 10
l = 0.5, = 0.05 , 1 = a = b = 1 = 0.1
= 1 = 5000I 4 , a = b = = 100
c1 = 1000 , 0 = 1 = 2 = 0.01, 3 = f = 100.
Figure 4 shows the simulation results with the proposed controller. In this simulation, the
disturbance w is changed at 50 [sec]:
Adaptive Control
82
sin (2t )
2 sin (4 t )
2 cos(2 t )
w = 4 cos(4 t ) .
w=
0.5 sin (5t )
0.5 sin (20 t )
2.5 cos(5t )
2.5 cos(20 t )
Figure 5 is the tracking error and Fig.6 shows the adaptively adjusted parameters in the
controller.
output
Fig. 4. Simulation results with the proposed controller
Fig. 5. Tracking error with the proposed controller
input
feedback gain k(t)
Fig. 6. Adaptively adjusted parameters
83
Adaptive Control
84
A very good control result was obtained and we can see that a good control performance is
maintained even as the frequencies of the disturbances were changed at 50 [sec].
Figures 7 and 8 show the simulation results in which the adaptively adjusted parameters in
the controller were kept constant after 40 [sec]. After the disturbances were changed, the
control performance deteriorated.
output
input
Fig. 7. Simulation results without adaptation after 40 [sec].
85
Fig. 8. Tracking error without adaptation
6. Conclusions
In this paper, the adaptive regulation problem for unknown controlled systems with
unknown exosystems was considered. An adaptive output feedback controller with an
adaptive internal model was proposed for single input/single output linear minimum phase
systems. In the proposed method, a controller with an adaptive internal model was
designed through an expanded backstepping strategy of only one step with a parallel
feedforward compensator (PFC).
7. References
A, Isidori. (1995). Nonlinear Control Systems-3rd ed., Springer-Verlag, 3-540-19916-0, London
A, Serrani.; A, Isidori. & L, Marconi. (2001). Semiglobal Nonlinear Output Regulation With
Adaptive Internal Model. IEEE Trans. on Automatic Control, Vol.46, No.8, pp.
11781194, 0018-9286
G, Feg. & M, Palaniswami. (1991). Unified treatment of internal model principle based
adaptive control algorithms. Int. J. Control, Vol.54, No.4, pp. 883901, 0020-7179
H, Kaufman.; I, Bar-Kana. & K, Sobel. (1998). Direct Adaptive Control Algorithms-2nd ed.,
Springer-Verlag, 0-387-94884-8, New York
I, Mizumoto.; R, Michino.; M, Kumon. & Z, Iwai. (2005). One-Step Backstepping Design for
Adaptive Output Feedback Control of Uncertain Nonlinear systems, Proc. of 16th
IFAC World Congress, DVD, Prague, July
R, Marino. & P, Tomei. (2000). Robust Adaptive Regulation of Linear Time-Varying Systems.
IEEE Trans. on Automatic Control, Vol.45, No.7, pp. 13011311, 0018-9286
86
Adaptive Control
R, Marino. & P, Tomei. (2001). Output Regulation of Linear Systems with Adaptive Internal
Model, Proc. of the 40th IEEE CDC, pp. 745749, 0-7803-7061-9, USA, December,
Orlando, Florida
R, Michino.; I, Mizumoto.; M, Kumon. & Z, Iwai. (2004). One-Step Backstepping Design of
Adaptive Output Feedback Controller for Linear Systems, Proc. of ALCOSP 04, pp.
705-710, Yokohama, Japan, August
S, Sastry. & M, Bodson. (1989). Adaptive Control Stability, Convergence, and Robustness,
Prentice Hall, 0-13-004326-5
V, O, Nikiforov. (1996). Adaptive servocompensation of input disturbances, Proc. of the 13th
IFAC World Congress, Vol.K, pp. 175180, San-Francisco
V, O, Nikiforov. (1997a). Adaptive servomechanism controller with implicit reference model.
Int J. Control, Vol.68, No.2, pp. 277286, 0020-7179
V, O, Nikiforov. (1997b). Adaptive controller rejecting uncertain deterministic disturbances
in SISO systems, Proc. of European Control Conference, Brussels, Belgium
Z, Ding. (2001). Global Output Regulation of A Class of Nonlinear Systems with Unknown
Exosystems, Proc. of the 40th IEEE CDC, pp. 6570, 0-7803-7061-9, USA, December,
Orlando, Florida
Z, Iwai. & I, Mizumoto. (1994). Realization of Simple Adaptive Control by Using Parallel
Feedforward Compensator. Int. J. Control, Vol.59, No.6, pp. 15431565, 0020-7179
4
Output Feedback Direct Adaptive Control
for a Two-Link Flexible Robot Subject to
Parameter Changes
Selahattin Ozcelik and Elroy Miranda
Texas A&M University-Kingsville, Texas

USA
1. Introduction
Robots today have an ever growing niche. Many of todays robots are required to perform
tasks which demand high level of accuracy in end effector positioning. The links of the robot
connecting the joints are large, rigid, and heavy. These manipulators are designed with
links, which are sufficiently stiff for structural deflection to be negligible during normal
operation. Also, heavy links utilize much of the joint motors power moving the link and
holding them against gravity. Moreover the payloads have to be kept small compared to the
mass of the robot itself, since large payloads induce sagging and vibration in the links,
eventually bringing about uncertainty in the end effector position. In an attempt to solve
these problems lightweight and flexible robots have been developed. These lightweight
mechanical structures are expected to improve performance of the robot manipulators with
typically low payload to arm weight ratio. The ultimate goal of such robotic designs is to
accurate tip position control in spite of the flexibility in a reasonable amount of time. Unlike
industrial robots, these robot links will be utilized for specific purposes like in a space
shuttle arm. These flexible robots have an increased payload capacity, lesser energy
consumption, cheaper construction, faster movements, and longer reach. However, link
flexibility causes significant technical problems. The weight reduction leads the manipulator
to become more flexible and more difficult to control accurately. The manipulator being a
distributed parameter system, it is highly non-linear in nature. Control algorithms will be
required to compensate for both the vibrations and static deflections that result from the
flexibility. This provides a challenge to design control techniques that:
a) gives precise control of desired parameters of the system in desired time,
b) cope up with sudden changes in the bounded system parameters,
c) gives control on unmodeled dynamics in the form of perturbations, and
d) robust performance.
Conventional control system design is generally a trial and error process which is often not
capable of controlling a process, which varies significantly during operation. Thus, the quest
for robust and precise control led researchers to derive various control theories. Adaptive
control is one of these research fields that is emerging as timely and important class of
controller design. Area much argued about adaptive control is its simplicity and ease of
88
Adaptive Control
physical implementation on actual real-life systems. In this work, an attempt has been made
to show the simplicity, ease and effectiveness of implementation of direct model reference
adaptive control (DMRAC) on a multi input multi output (MIMO) flexible two-link system.
The plant comprises of a planar two-link flexible arm with rotary joints subject only to
bending deformations in the plane of motion. A payload is added at the tip of the outer link,
while hub inertias are included at actuated joints. The goal is to design a controller that can
control the distal end of the flexible links.
Probably the first work done pertaining to the control of flexible links was presented by
(Cannon & Schmitz, 1984). Considering a flexible link, which was only flexible in one
dimension (perpendicular to gravity), a Linear Quadratic Gaussian controller was designed
for the position control. Direct end point sensing was used and the goal was to execute a
robot motion as fast as possible without residual vibrations in the beam. Also, experiments
were carried out on end point control of a flexible one link robot. These experiments
demonstrated control strategies for position of one end to be sensed and precisely
positioned by applying torque at the other end. These experiments were performed to
uncover and solve problems related to the control of very flexible manipulators, where
sensors are collocated with the actuators.
(Geniele et al., 1995) worked on tip-position control of a single flexible link, which rotates on
a horizontal plane. The dynamic model was derived using assumed-modes method based
on the Euler-Bernoulli beam theory. The model is then linearized about an operating point.
The control strategy for this non-minimum phase linear time varying system consisted of
two parts. The first part had an inner stabilizing control loop that incorporates a
feedforward term to assign the systems transmission zeros at desired locations in the
complex plane, and a feedback term to move the systems poles to the desire positions in the
left half plane. In the second part, the other loop had a feedback servo loop that allowed
tracking of the desired trajectory. The controller was implemented on an experimental test
bed. The performance was then compared with that of a pole placement state feedback
controller.
(Park & Asada, 1992) worked on an integrated structure and control design of a two-link
non-rigid robot arm for the purpose of high speed positioning. A PD control system was
designed for the simple dynamic model minimizing the settling time. Optimal feedback
gains were obtained as functions of structural parameters involved in the dynamic model.
These parameters were then optimized using an optimization technique for an overall
optimal performance.
(Lee et al., 2001) worked on the adaptive robust control design for multi-link flexible robots.
Adaptive energy-based robust control was presented for both close loop stability and
automatic tuning of the gains for desired performance. A two-link finite element model was
simulated, in which each link was divided into four elements of same length. The controller
designed was independent of system parameters and hence possessed stability robustness
to parameter variations.
Variations in flexible links have also been researched. Control of a two-link flexible arm in
contact with a compliant surface was shown in (Scicliano & Villani, 2001). Here, for a given
tip position and surface stiffness, the joint and deflection variables are computed using
closed loop inverse kinematics algorithm. The computed variables are then used as the set
points for a simple joint PD control, thus achieving regulation of the tip position and contact
force via a joint-space controller.
Output Feedback Direct Adaptive Control for a Two-Link Flexible Robot Subject to
Parameter Changes
89
(Ider et al., 2002) proposed a new method for the end effector trajectory tracking control of
robots with flexible links. In order to cope with the non-minimum phase property of the
system, they proposed to place the closed-loop poles at desire locations using full state
feedback. A composite control law was designed to track the desired trajectory, while at the
same time the internal dynamics were stabilized. A two-link planar robot was simulated to
illustrate the performance of the proposed algorithm. Moreover the method is valid for all
types of manipulators with any degree of freedom.
(Green, A. & Sasiadek, J., 2004) presented control methods for endpoint tracking of a twolink robot. Initially, a manipulator with rigid links is modeled using inverse dynamics, a
linear quadratic regulator and fuzzy logic schemes actuated by a Jacobian transpose control
law computed using dominant cantilever and pinned-pinned assumed mode frequencies.
The inverse dynamics model is pursued further to study a manipulator with flexible links
where nonlinear rigid-link dynamics are coupled with dominant assumed modes for
cantilever and pinned-pinned beams. A time delay in the feedback control loop represents
elastic wave travel time along the links to generate non-minimum phase response.
An energy-based nonlinear control for a two-link flexible manipulator wasstudied in (Xu et
al., 2005). It was claimed that their method can provide more physical insights in nonlinear
control as well as provide a direct candidate for the Lyapunov function. Both simulation and
experimental results were provided to demonstrate the effectiveness of the controllers
A robust control method of a two-link flexible manipulator with neural networks based
quasi-static distortion compensation was proposed in (Li et al., 2005). The dynamics
equation of the flexible manipulator was divided into a slow subsystem and a fast
subsystem based on the assumed mode method and singular perturbation theory. A
decomposition based robust controller is proposed with respect to the slow subsystem, and
H control is applied to the fast subsystem. The proposed control method has been
implemented on a two-link flexible manipulator for precise end-tip tracking control.
In this work a direct adaptive controller is designed and the effectiveness of this adaptive
control algorithm is shown by considering the parametric variations in the form of additive
perturbations. This work emphasizes the robust stability and performance of adaptive
control, in the presence of parametric variations. This approach is an output feedback
method, which requires neither full state feedback nor adaptive observers. Other important
properties of this class of algorithms include:
a) Their applicability to non-minimum phase systems,
b) The fact that the plant (physical system) order may be much higher than the
order of the reference model, and
c) The applicability of this approach to MIMO systems.
Its ease of implementation and inherent robustness properties make this adaptive control
approach attractive.
2. Mathematical Modeling of the System

In this section mathematical model of the system is derived using Lagrange equations with
the assumed-modes method. The links are assumed to obey Euler-Bernoulli beam model
with proper boundary conditions. A payload has been added to the tip of the second link,
while hub inertias are included at the actuator joints.
Adaptive Control
90
2.1 Kinematic Modeling

A planar two-link flexible arm with rotary joints subject to only bending deformations in the
plane of motion is considered. The following coordinate frames, as seen in Fig. 1, are
established: the inertial frame ( X 0 , Y0 ), the rigid body moving frame associated to link i
( X i , Yi ), and the flexible body moving frame associated with link i ( X i , Yi ) (Brook, 1984).
Fig. 1. Planar Flexible Two-Link Arm

The rigid body motion is described by the joint angle, i , while yi ( xi ) denoted the
transversal deflection of link i at abscissa, 0 xi li , li being the link length. Let
pii ( xi ) = ( xi , yi ( xi ))T be the position of a point along the deflected link i with respect to frame
( X i , Yi ) and pi be the absolute position of the same point on frame ( X 0 , Y0 ). Also, rii+1 = pii (li )
indicates the position of the origin of frame ( X i +1 ,Yi +1 ) with respect to frame ( X i , Yi ), and ri
gives absolute positioning of the origin of frame ( X i , Yi ) with respect to frame ( X 0 , Y0 ). The
rotation matrix Ai for rigid body motion and the rotation matrix Ei for the flexible mode are,
respectively,
Ai =
cos i
sin
i
sin i
cos i
Ei =
1
'
y ie
y ' ie
where yie = ( yi / xi )|xi = li and for small deflections arctan( yie )
(1)
yie . Therefore, the previous
absolute position vectors can be expressed as,

i
p i = r1 + W i p i
E i = r i+1 = r 1 + W i
r i+1
(2)
where, Wi is the global transformation matrix from ( X 0 , Y0 ) to ( X i , Yi ), which obeys the

A and W
=I
recursive equation W = W E A = W
i
i 1
i 1
i 1
Parameter Changes
91
2.2. Lagrangian Modeling

The equations of motion for a planar n-link flexible arm are derived by using the Lagrange
equations. The total kinetic energy is given by the sum of the following contributions:
n
i 1
i 1
T = Thi + Tli +Tp
(3)
where the kinetic energy of the rigid body located at the hub i of mass mhi and the moment
of inertia J hi is
Thi =
1
1
mhi r&i + J hi i2
2
2
(4)
where & i is the (scalar) absolute angular velocity of frame ( X i , Yi ) given by

1
i 1
j =1
k =1
&i = &j + y& ke
(5)
Moreover, the absolute linear velocity of an arm is
p& = r&i + W&i pii + Wi i p& i
(6)
and r&ii+ 1 = p& ii (li ) . Since the links are assumed inextensible ( x& i = 0 ) , then p& ii ( xi ) = (0, y& i ( xi ))T .
The kinetic energy pertaining to link i of linear density i is
Tli =
1 li
i ( xi ) p& iT ( xi ) dxi
2 0
(7)
and the kinetic energy associated to a payload of mass mp and moment of inertia J p located
at the end of link n is
Tp =
1
1
)2
m p r&nT+1r&n +1 + J p (& n + y& ne
2
2
(8)
Now, in the absence of gravity (horizontal plane motion), the potential energy is given by
n
i =1
i =1
U = U i =
li
d 2 yi ( xi )
1
( EI ) i ( xi )
dxi
2
20
dxi
(9)
Where U i is the elastic energy stored in link i, and (EI )i being its flexural rigidity. No
Adaptive Control
92
discretization of structural link flexibility has been made so far, so the Lagrangian will be a
functional.
2.3. Assumed Mode Shapes
Links are modeled as Euler Bernoulli beams of uniform density i and constant flexural
rigidity (EI )i with the deformation yi ( xi , t ) satisfying the partial differential equation
( EI )i
4 yi ( xi , t )
2 yi ( xi , t )
+ i
= 0, i = 1,..., n.
4
xi
t 2
(10)
Boundary conditions are imposed at the base of and the end of each link to solve this
equation. The inertia of a light weight link is small compared to the hub inertia, and then
constrained mode shapes can be used. We assume each slewing link to be clamped at the
base
yi (0, t ) = 0, yi(0, t ) = 0, i = 1,..., n
(11)
For the remaining boundary conditions it is assumed that the link end is free of dynamic
constraints, due to the difficulty in accounting for time-varying or unknown masses and
inertias. However, we consider mass boundary conditions representing balance of moment
and shearing force, i.e.
2 yi ( xi ,t )
d 2 y ( x , t )
d2
( EI )i
= J Li 2 i i ( MD)i 2 ( yi ( xi ,t ) xi =li )
2
dt xi x =l
dt
xi
xi =li
i i
3 y ( x )
d2
d 2 y ( x , t )
( EI )i i 3 i ,t
= M Li 2 ( yi ( xi ,t ) xi =li ) ( MD)i 2 i i
x
dt
dt
xi xi =li
i
xi =li
i = 1,...., n
(12)
where, M Li and J Li are the actual mass and moment of inertia at the end of link i. ( MD)i
accounts for the contribution of masses of distal links, i.e. non-collocated at the end of link i.
A finite-dimensional model of link flexibility can be obtained by assumed modes technique.
Using this technique the link deflections can be expressed as
mi
yi ( xi , t ) = ij ( xi ) ij (t )
(13)
j =1
where ij (t ) are the time varying variables associated with the assumed spatial mode
shapes ij ( xi ) of link i. Therefore each term in the general solution of (10) is the product of a
time harmonic function of the form
ij (t ) = exp( jij t )
(14)
Parameter Changes
93
and of a space eigenfunction of the form
ij ( xi ) = C1,ij sin( ij xi ) + C2,ij cos( ij xi ) + C3,ij sinh( ij xi ) + C3,ij cosh( ij xi )

In (14)
(15)
ij is the jth natural angular frequency of the eigenvalue problem for link i, and in
(15) ij = ij2 i /( EI )i .
Application of the aforementioned boundary conditions allows the determination of the
constant coefficients in (15). The clamped link conditions at the link base yield
C3,ij = C1,ij , C4,ij = C2,ij
(16)
while, the mass conditions at the link end lead to homogeneous system of the form
C
F ( ij ) 1,ij
C2,ij
(17)
The so-called frequency equation is obtained by setting to zero the determinant of the (22)
matrix F ( ij ) that depends on explicitly on the values of M Li , J Li , and ( MD )i . The first mi
roots of this equation give the positive values of ij to be plugged in (15). Using this the
coefficients C1,ij and C2,ij are determined up to a scale factor that is chosen via a suitable
normalization. Further the resulting eigenfunctions
ij satisfy
a modified orthogonality
condition that includes the actual M Li , J Li , and ( MD )i . In an open kinematic chain

arrangement, M Li is the constant sum of all masses beyond link i, but J Li and ( MD )i
depend on the position of successive links. This will considerably increase the complexity of
model derivation and overload the computational burden of on-line execution. Thus, some
practical approximation leading to constant although nonzero boundary conditions at the
link end is done. Thus, a convenient position is set to ( MD )i = 0 and compute J Li for a fixed
arm configuration. In this case, it can be shown that det(F) = 0 results in the following
transcendental equation (De Luca & Scicliano, 1989)
(1 + cos ( l ) cosh ( l ) )
ij i
M Li ij
ij i
J Li ij3
( sin ( l ) cosh ( l ) cos ( l ) sinh ( l ))

ij i
ij i
ij i
ij i
( sin ( l ) cosh ( l ) + cos ( l ) sinh ( l ))

ij i
M Li J Li ij4
i 2
ij i
ij i
ij i
(1 cos ( l ) cosh ( l ) ) = 0
ij i
ij i
(18)
Adaptive Control
94
2.4. Closed-Form Equations of Motion

On the basis of the discretization introduced in the previous section, the Lagrangian L
becomes a function of set of N generalized coordinates qi(t) the dynamic model is obtained
satisfying the Lagrange-Euler equations
d L L
= fi ,
dt q&i qi
i = 1LN
(19)
where, fi are the generalized forces performing work on qi(t). Under the assumption of
constant mode shapes, it can be shown that spatial dependence present in the kinetic energy
term (7) can be resolved by the introduction of a number of constant parameters,
characterizing the mechanical properties of the (uniform density) links (De Luca, et. al. 1988,
Cetinkunt, et. al., 1986)
li
mi = i dxi = i li
0
di =
1
mi
li
1
2
i xi xi = li
li
1
J 0i = i xi2 dxi = mi li2
0
3
(20)
(21)
(22)
li
vij = i ij ( xi ) dxi
0
(23)
li
wij = i ij ( xi ) xi dxi
0
(24)
li
zijk = i ij ( xi ) ik ( xi ) dxi
0
(25)
li
kijk = ( EI )i ij ( xi ) ik ( xi ) dxi
0
(26)
where, mi is the mass of the link i, d is the distance of center of mass of link i from joint i
axis, J 0i is the inertia of link i about joint i axis, vij and ij are the deformation moments of
order zero and one of mode j of the link i. Also, kijk is the cross elasticity coefficient of
modes j and k of link i. The actual numerical values of the previous parameters are
calculated off-line. As a result of this procedure, the equations of motion for a planar n-link
arm can be written in a familiar closed form
B(q )q&& + h(q, q& ) + Kq = Qu
(27)
Parameter Changes
where q =
1 n 11 1, m1
n ,1 n , mn
95
is the N-vector of generalized coordinates ( N = n + mi ),

i
and u is the n-vector of joint actuator torques. B is the positive definite symmetric inertia
matrix, h is the vector of Corriolis and centrifugal forces, K is the stiffness matrix and Q is
the input weighting matrix that is of the form I nxn Onx ( N n )
due to the clamped link
assumptions. Joint viscous friction and link structural damping can be added as Dq& , where
D is a diagonal matrix. It is noted that orthonormalization of mode shapes implies

convenient simplification in the diagonal blocks of the inertia matrix relative to the
deflections of each link, due to the particular values attained by
zijk
in (25). Also the
stiffness matrix becomes diagonal ( K1 = L = K n = 0; K n +1 ,K , K N > 0 ) being kijk = 0 for
jk
in (27). The components of h can be evaluated through the Christoffel symbols given by
N N B
1 B jk
ij
hi =
2 q j
j =1 k =1 qk
q& j q&k
(28)
2.5. Explicit Dynamic Model of Two-Link Flexible Arm

Two assumed mode shapes are considered for each link ( m1 = m2 = 2 ). Thus, the vector of
Lagrangian coordinates reduces to q = (1 2 11 12 21 22 )T , i.e. N = 6. It can be shown

(Brook, 1984, De Luca et. al. 1988) that the contributions of kinetic energy due to deflection
variables are
{ factor of & } = z
2
i1
M
factor of 2&i1&i 2 = i1,e i1,e 1 Li
2 ( MD)i
{ factor of & } = z
2
i2
where,
(29)
i11
1
2
( MD)i i 2,e
+ zi12
J Li i2,e
(30)
(31)
i 22
ij ,e = ij ( xi ) |x =l and ij ,e = ij ( xi ) |x =l , i, j = 1, 2 . The above equations are

i
obtained expanding terms (7) and (8) by using (5) and (6). Accounting for separability (13)
then leads to expressions for the factors of the quadratic deflection rate terms, in which
parameters defined in (25) and the mass coefficients on the right hand side of (12) can be
identified. It is found for link-1:
M L1 = m2 + mh2 + m p
(32)
J L1 = J 02 + J h2 + J p + mP l22
(33)
( MD )1 = ( m2 d 2 + m p l2 ) cos2 ( v21 + m p 21,e ) 21 + ( v22 + m p 22,e )
22 sin2
(34)
Adaptive Control
96
Note that in the case of only two links,

ML2 = mp ,
J L1
is a constant. On the other hand for link-2:
J L2 = J p ,
( MD)2 = 0
(35)
A convenient normalization of mode shapes is accomplished by setting:
ziii = mi ,
i, j = 1,2
(36)
This also implies that the nonzero coefficients in the stiffness matrix K take on values wij2 mi .
It is stressed that, if the exact values for the boundary conditions in (12) were used the
natural orthogonality of the computed mode shapes would imply that { factor of 2&11&12 } is
zero for both links. For link-2 the use of (35) automatically ensures the correct
orthogonality of mode shapes. On the other hand, however for link-1, the off-diagonal term
( MD )1 varies with arm configuration. This implies that the mode shapes which are spatial
quantitieswould become implicit functions of time, thus conflicting with the original
separability assumption. It is seen that for different positions of second link, (MD)1 results in
variations of (34), so the actual mode shapes of the first link become themselves functions of
time-varying variables describing the deflection of the second link. A common
approximation in computing the elements of the inertia matrix for flexible structures is to
evaluate kinetic energy in correspondence to the undeformed configuration. In our case, it is
equivalent to neglecting the second term ( MD )1 in (34), which is an order of magnitude
smaller than the first term. Accordingly, ( MD )1 is constant for a fixed arm configuration.
Taking 2 = / 2 leads to ( MD )1 = 0 and thus the eigen-frequencies can be computed
through (19). This is equivalent to having zeroed only that portion of the { factor of 2&11&12 }
generated by constant diagonal terms, i.e.
M Li
'
11,
e
0
11,e
0 12,e
+ z112 = 0
J Li 12,
e
(37)
This will produce nonzero off-diagonal terms in the relative block of the inertia matrix. The
resulting model is cast in a computational advantageous form, where a set of constant
coefficients appear that depend on the mechanical properties of the arm. The inertia matrix
as well as other derivations can be found in (Miranda, 2004). Once having obtained the
expressions of the inertia matrix, the components of h can be evaluated using (28). Viscous
friction and passive structural damping are included in matrix D for improvement in arm
movement, and finally, the stiffness matrix K is of the form,
2 m w2 m w2 m , w2 m
K = diag 0, 0, w11
1, 12 1, 21 2 22 2
(38)
Then the equations of motion is given in its standard form as

B (q )&q& + h (q, q& ) + D q& + Kq = Qu
(39)
Parameter Changes
97
After tremendous of algebra and neglecting friction, (39) can be written as,
B11 &&1 + B12
B21 &&1 + B22
&&2 + B13
&&2 + B23
B31 &&1 + B32 &&2 + B33 &&11 + B34

B41 &&1 + B42 &&2 + B43 &&11 + B44
&&11 + B14
&& + B
11
24
&&12 + B15
&&12 + B25
&&21 + B16
&&21 + B26
&&12 + B35 &&21 + B36

&& + B && + B
12
45 21
46
&&22 + h1
&&22 + h 2
&&22 + h3 + K 3 11
&& + h + K
22
4
4 12
=
=
u p1
up2
= 0
= 0
(40)
B15 &&1 + B52 &&2 + B53 &&11 + B54 &&12 + B55 &&21 + B56 &&22 + h2 + K 5 21 = 0
B61 &&1 + B62 &&2 + B63 &&11 + B64 &&12 + B65 &&21 + B66 &&22 + h2 + K 6 22
= 0
where, up1 and up2 are input torques to joints 1 and 2, respectively. Plant outputs are
considered to be link tip displacements y1 and y2. As seen from above equations, the system
is highly nonlinear and of 12th order. For the flexible robot, the following physical
parameters were considered
1 = 2 = 0.2 kg / m
l1 = l2 = 0.5 m, d2 = 0.25 m
m1 = m2 = mp = 0.1 kg, mh2 = 1 kg
J01 = J02 = 0.0083 kgm 2
(41)
J h1 = J h2 = 0.1 kgm 2 , J p = 0.0005 kgm 2

( EI ) = ( EI ) = 1 Nm 2
1
2
The natural frequencies fij = wij/2 and the remaining parameters in the model coefficients
are computed as (Miranda, 2004):
f11 = 0.48 Hz, f12 = 1.80 Hz ,

f 21 = 2.18 Hz, f 22 = 15.91 Hz,
e = 0.657, 12,
e = 0.560,
11,e = 0.186, 12,e = 0.215, 11,
e = 2.641, 22,
e = 10.853,
21,e = 0.883, 22,e = 0.069, 21,
(42)
v11 = 0.007, v12 = 0.013, v21 = 0.033, v22 = 0.054,

w11 = 0.002, w12 = 0.004, w21 = 0.012, w22 = 0.016
In order to design the proposed adaptive controller, the plant needs to be linearized and the
transfer function matrix be obtained. After linearization, neglecting higher order terms, and
tremendous amount of algebra, it can be shown (Miranda, 2004) that the plant Gpo(s) =
yp(s)/up(s) with nominal parameters can be obtained as,
0.01641s10 + 7.061 10 5 s 8
y p1 ( s )
s12 + 3.68 10 3 s10
G p 0 (s)=

10
3 8
y p2 ( s ) 7.357s 9.636 10 s
12
3 10
s + 3.68 10 s
2.259s10 1.362 10 3 s 8
u s
s12 + 3.68 10 3 s10
p1 ( )
u p2 ( s )
1317 s10 + 0.674s 8
s12 + 3.68 10 3 s10
(43)
Adaptive Control
98
where, Gpo(s) is the nominal plant transfer function matrix. Now, performing minimal
realization, Gpo(s) can be reduced to
0.01641s 2 + 7.061 105
s 4 + 3.68 103 s 2
G po ( s ) =
7.357s 2 9.636 103
4
3 2
s + 3.68 10 s
2.259s 2 1.362 103
s 4 + 3.68 103 s 2
1317s 2 + 0.674
s 4 + 3.68 103 s 2
(44)
From (44), it is straight forward to obtain the actual plant in general form as
C111 s 2 + C011
y p ( s ) s 4 + B211 s 2
Gp ( s ) =
= 21 2
u p ( s ) C1 s + C021
4
21 2
s + B2 s
C112 s 2 + C012
s 4 + B212 s 2
C122 s 2 + C022
s 4 + B222 s 2
(45)
where, above coefficients of Gp(s) are functions of plant parameters and can vary with the
range as defined below:
ij
ij
ij
C p j C p j C p j
ij
ij
ij
Br j Br j Br j
i = 1, 2, j = 1, 2
i = 1, 2, j = 1, 2
(46)
The values of the nominal plant parameters are defined in the following table. The range
considered for each parameter is 30%.
Parameter
Nominal
Range
0.01641
0.011487 to 0.02133
C011
7.061*105
4.9427 *105 to 9.1793*105
B211 = B212 = B221 = B222
3.68*103
2.576*103 to 4.784*103
C112
2.259
1.5813 to 2.9367
C012
1.362*103
9.583*104 to 1.7797 *103
C121
7.357
5.1499 to 9.5641
C021
9.636*103
6.7452*101 to 12.5268*103
C122
1317
921.9 to 1712.1
C022
0.674
0.4718 to 0.8762
11
1
Table 1. Plant parameters, nominal values, and variation range.
Parameter Changes
99
For comparison reasons, uncompensated response of the nominal plant is given below
Fig. 2. Uncompensated response of the nominal plant
3. Controller Design
Consider now that the plant given by (45) is represented by the following state-space
equations:
x& p ( t ) = Ap x p ( t ) + B p u p (t )
y p (t ) = C p x p (t )
(47)
where xp(t) is the (n 1) state vector, up(t) is the (m1) control vector, yp(t) is the (q 1) plant
output vector, and Ap, Bp and Cp are matrices with appropriate dimensions. The range of the
plant parameters given by (46) is now given by
a ij a p (i , j ) a ij , i , j = 1, K , n
b ij b p (i , j ) b ij , i , j = 1, K , n
(48)
where ap(i, j) is the (i, j)th element of Ap and bp(i, j) is the (i, j)th element of Bp. Consider also
the following reference model, for which plant output is expected to follow the model
output without explicit knowledge of Ap and Bp.
x& m (t ) = A m x m (t ) + Bm u m (t )
y m (t ) = C m x m (t )
(49)
In light of this objective, consider now the following output feedback adaptive control law,
u p (t ) = K e (t )e y (t ) + K x (t ) xm (t ) + K u (t )u m (t )
(50)
Adaptive Control
100
where ey(t) = ym(t)yp(t) and Ke(t), Kx(t), and Ku(t) are adaptive gains defined below. The
control law consists of a feedback term from output error and a feedforward terms from
model states and inputs. The adaptive gains Ke(t), Kx(t), and Ku(t) are combination of
proportional and integral gains as given below,
K j (t ) = K pj (t ) + Kij (t )
j = e, x, u
(51)
and they are updated according to the following adaptation law (Kaufman, et. al. 1998,
Ozcelik & Kaufman, 1999)
K pj (t ) = ey (t )[ey (t ) + xm (t ) + um (t )]Tp
Kij (t ) = ey (t )[ey (t ) + xm (t ) + um (t )]Ti
j = e, x, u , Tp 0
j = e, x, u, Ti > 0
(52)
(53)
where Ti and Tp are constant proportional and integral weighting matrices, respectively. It is
seen from (53) that the term Kij(t) is a perfect integrator and may steadily increase whenever
perfect following (ey(t) = 0) is not possible. The gain may reach unnecessarily large values, or
may even diverge. Thus, a -term is introduced in order to avoid the divergence of integral
gains (Ionnou & Kokotovic, 1983). With the -term, Ki(t) is now from a first-order filtering
of ey(t)rT (t)Ti and therefore cannot diverge, unless ey(t) diverges. However, in this context,
the -term does more for the concept of adaptive control. The gains increase only if high
gains are needed and decrease if they are not needed any more. They are also allowed to
change at any rate without affecting stability, such that the designer can adjust this rate to fit
the specific needs of the particular plant. Thus, using -term we rewrite the equation (53) as
follows,
K ij (t ) = ey (t )[ey (t ) + xm (t ) + um (t )]Ti Kij (t )
j = e, x, u
(54)
For this adaptive control to work and for asymptotic tracking to be achieved, the plant is
required to be almost strictly positive real (ASPR) (Bar-Kana, 1994); that is, there exists a
gain matrix Ke, not needed for implementation, such that the closed-loop transfer function
G c (s) = [ I + G p (s)K e ]1 G p (s)
(55)
is strictly positive real (SPR). And that it can be shown that (Kaufman, et. al., 1998) a MIMO
system represented by a transfer function Gp(s) is ASPR if it:
a) is minimum phase (zeros of the transfer function are on the left-half plane),
b) has relative degree of m or zero (i.e., the difference in the degree of denominator
and numerator, (n-m=m) or (n-m=0)), and
c) has minimal realization with high frequency gain CpBp > 0 (positive definite).
Obviously, the plant given by (45) does not satisfy the so-called ASPR conditions and that
can not be applied. However, it has been shown in (Kaufman, et. al., 1998) and (Ozcelik,
Parameter Changes
101
2004) that there exist a feedforward compensator H(s) such that the augmented plant Ga(s) =
Gp(s) + H(s) is ASPR and the proposed adaptive algorithm can be implemented confidently.
3.1. Design of a Feedforward Compensator (FFC) for the Flexible Robot
From the above restrictions it is obvious that the plant given by (45) is not ASPR and that an
FFC has to be designed. Now consider the actual plant Gp(s) again,
C 111 s 2 + C 011
4
s + B 11 s 2
G p (s ) = 21 2 2 21
C1 s + C0
s 4 + B 21 s 2
2
C 112 s 2 + C 012
s 4 + B212 s 2
C 122 s 2 + C 022
s 4 + B222 s 2
(56)
Assuming that the nominal plant parameters are known, the parametric uncertainty of the
plant can be transformed into a frequency dependent additive perturbation of the plant by
representing of the actual plant Gp(s) as Gp(s) = Gp0(s) + a(s), with Gp0(s) being a nominal
plant and a(s) being a frequency dependent additive perturbation. Then, one can write
a (s) = G p (s) G p 0 (s)
(57)
From (57), the additive uncertainty transfer function can be obtained as
11 (s) 12 (s)
a (s ) =
21 (s) 22 (s)
(58)
where,
11 (s) =
12 (s) =
21 (s) =
(C 111 0.016411)s 6 + (3.68 * 103 C 111 + C 011 7 * 105 0.01641B211 )s 4 + (3.68 * 103 C 011 7 * 105 B211 )s 2
s 8 + (3.68 * 103 + B211 )s 6 + B211 3.68 * 103 s 4
(2.25 + C 112 )s 6 + (3.68 * 10 3 C 112 + C 012 + 1.36 * 10 3 + 2.25B212 )s 4 + (3.68 * 10 3 C 012 + 1.36 * 10 3 B212 )s 2
s 8 + (3.68 * 10 3 + B212 )s 6 + b1 3.68 * 10 3 s 4
(7.32 + C 121 )s 6 + (3.68 * 10 3 C 121 + C 021 + 9.63 * 10 3 + 7.35B221 )s 4 + (3.68 * 10 3 C 021 + 9.63 * 10 3 B221 )s 2
s 8 + (3.68 * 10 3 + B221 )s 6 + b1 3.68 * 10 3 s 4
22 (s ) =
(59)
(C 122 1317 )s 6 + (3.68 * 10 3 C 122 + C 022 0.674 1317 B221 )s 4 + (3.68 * 10 3 C 022 0.674B221 )s 2
s 8 + (3.68 * 10 3 + B221 )s 6 + B221 3.68 * 10 3 s 4
It is seen that the uncertainty is a function of plant parameters, which vary in a given range.
Thus, in the design of a feedforward compensator, the worst case uncertainty should be
taken into account. To this effect, the following optimization procedure is considered for
determining the worst case uncertainty at each frequency (suitable number of discrete
values). Define a vector whose elements are plant parameters, i.e.
Adaptive Control
102
v = C pij
C pij 1 L C 0ij
Brij
Brij 1 L B0ij
(60)
Then
maximize
14243 ij ( jw )
at each w
C ij C ij C ijp j
p j
p j
subject to :
ij
ij
ij
Br j Br j Br j
(61)
where ij is the ijth element of (jw). In other words, this optimization is performed for each
element of (jw). After having obtained the worst case (maximum) perturbation, we will
assume that the perturbation is not exactly known but its upper bound is known. In other
words, there exists a known rational function as an upper bound of the worst case
uncertainty. Now the upper bound is characterized by an element by element interpretation,
where the upper bound means that each entry of (jw) is replaced by its corresponding
bound. In other words, given the worst case uncertainty for each (jw), it is assumed that
there exists a known rational function wij(s) RH such that
w ij ( jw ) max ij ( jw ) w
(62)
Knowing that the plant parameters can vary within their lower and upper bounds, this
parametric uncertainty is formulated as an additive perturbation in the transfer function
matrix. It is important to note that the controller be designed with respect to worst case
uncertainty for each ij. This can be achieved by performing an optimization procedure
given by (61) for 200 frequencies. Here an element by element uncertainty bound model is
used for the characterization of upper bound of the uncertainty matrix. Then wij , which
satisfies (62) for each ij is given in matrix form as,
7 *102
2
W (s ) = 800s + 22s +3 0.05
425 * 10
150s 2 5.75s + 0.05
9 *104
80s + 4s + 0.05
9
2 * 10
2
25s + 3.75s + 0.125
2
(63)
The magnitude responses for each max(|ij|) and the corresponding (|wij|) are given in
Figures 3-6. Having obtained the nominal plant and formulated unmodeled dynamics, lets
have the following assumptions on the plant,
Assumption 1:
a) The nominal plant parameters are known.
b) The off-diagonal elements of Gpo(s) and a(s) are strictly proper.
c) a(s) RHmxm and satisfies (62)
Parameter Changes
Fig. 3. |11 (jw)| and |w11 (jw)|
Fig. 4. |12 (jw)| and |w12 (jw)|
Fig. 5. |21 (jw)| and |w21 (jw)|
Fig. 6. |22 (jw)| and |w22 (jw)|
103
Now, consider the augmented nominal plant with the parallel feedforward compensator
Gao (s ) = G po (s ) + H (s )
(64)
and the following lemma

Lemma 1:
Let the feedforward compensator H(s) be of the form,
h11
0
H (s ) =
M
h22
M
0
L
O
K
0
0
M
hmm
(65)
with each element hii(s) of a feedforward compensator being relative degree zero, then the augmented
nominal plant Gao(s) = Gpo(s)+H(s) will have positive definite high frequency gain and relative
McMillan degree zero (Ozcelik & Kaufman, 1999).
Adaptive Control
104
In other words, the new plant Gao(s) including H(s) becomes ASPR. Now that the ASPR
conditions are satisfied for the nominal plant case, we next need to guarantee that the ASPR
conditions are also satisfied in the presence of plant perturbations. To this effect, consider
the following theorem
Theorem 1: If H(s) is designed according to the following conditions, then the augmented plant
Ga(s) = Gp(s)+H(s) with the plant perturbations will be ASPR.
a) H(s) is stable with each hij (s) being relative degree zero
b) H1 (s) stabilizes the nominal closed loop system
c) % ( s ) RH and % ( s ) < 1
% ( s) is the uncertainty of the augmented plant and given in the following (Ozcelik &
where
Kaufman, 1999)
~
1
(s ) = (G p 0 (s ) + H (s )) W (s )
(66)
In light of the Theorem 1, we can readily determine the FFC as follows:

a) The order of each element hii(s) of a feedforward compensator is chosen to be
equal to the order of the corresponding diagonal element of the nominal plant
Gp0(s).
h1s 4 + h2 s 3 + h3 s 2 + h4 s + h5
4
3
2
H (s ) = s + 4 s + 6 s + 4 s + 1
4
3
2
h6 s + h7 s + h8 s + h9 s + h10
s 4 + 4 s 3 + 6s 2 + 4s + 1
0
(67)
Denominator of the compensator H(s) appears in the numerator of the closedloop transfer function and therefore, can be pre-determined such that its time
constant is fast enough that its dynamics is negligible.
b) Compensator parameters are determined from the following optimization
procedure:
%
minimize
14243 ( jw)
X
subject to : Re al [ roots(Z(s ))] < 0
(68)
where Z(s) is the characteristic polynomial of the nominal closed-loop system matrix and X
is a vector composed of the parameters of each Hij(s).
numerator of h11 = 3.7444e4 s 4 + 4.6805e 4 s 3 + 9.5966e 3s 2 + 1.1650e3 s + 4.0310e 4
numerator of h22 = 9.2748e5 s 4 + 5.7746e 4 s 3 + 2.3495e 4 s 2 + 2.9786e5 s + 5.6472e 4
(69)
Parameter Changes
105
With this FFC, all the conditions of Theorem 1 are satisfied. Thus, the augmented plant
satisfies the almost strictly positive real conditions over a wide range of plant parameter
variations. It is expected that the DMRAC augmented with this feedforward compensator
will be robust for the maximum deviations from the nominal plant. The block diagram of
the overall control system is depicted in Figure 7.
Fig. 7. DMRAC with Two-Link Arm
Adaptive Control
106
Vert Cat
Um
Zm=Ym+Rm
Rm
Step
Model compensator
Step1
Matrix
Concatenation4
em
Xm
Ym1
Ym2
Um
Ym
Reference model
error
Zm
ez
ez
Up1
Zp
Um
Up2
Xm
Adaptive Controller
Up1
Up2
Rp
Plant Compensator
Zp=Yp+Rp
3
Y_1
Vert Cat
Up1
Y_1
Yp
Up2
Y_2
Nonlinear Plant
4
Y_2
Matrix
Concatenation
Fig. 8. Simulink Block Diagram of the overall control system with nonlinear plant.
4. Simulation Results
The nonlinear equations were used in building the plant in Simulink/MATLAB (Figure 8)
and with the above described DMRAC algorithm, the following cases were simulated.
4.1 Case 1
All initial conditions were set to zero.

For both links the reference models were set to Gm1 (s ) = Gm 2 (s ) = 1/(30s + 1)
For both links tip displacements were set to=0.05m
Upper bound of plant parameters was used.
Parameter Changes
Fig. 9. Case 1: y1 and reference model (left), y2 and reference model (right)
Fig. 10. Case 1: 1 and 2 (left), errors ez1 and ez2 (right)
Fig. 11. Case 1: Control inputs, up1 and up2

4.2. Case 2
All initial conditions were set to zero.

Lower bound of plant parameters was used.
107
108
Adaptive Control

From the Figures 9-11 of Case 1 and the Figures 12-14 of Case 2, we can see that the tips of
both links in each case follow the reference input. In Case 2 we have used a faster reference
model to show the effectiveness of the DMRAC.
4.3. Case 3
1 (0) = 5o , 2 (0) = 0o
Parameter Changes
109
Nominal set of plant parameters were used.
From Figure 15 we see that at t=0, the tip position of link 2, y2 begins at zero, however due to
the initial condition for 1 tip position of first link, y1 is displaced with respect to the desired
reference model at t=0. From Figure 16 it is seen that 1 and 2 come to steady-state, while
errors approach to zero.
Adaptive Control
110
4.5. Case 4
1 (0) = 0o , 2 (0) = 0o
For both links tip displacements were set to= 0.01m
Nominal plant parameters were used
Fig. 19. Case 4: 1 and 2 (left), error ez (right)
Fig. 20. Case 4: Control inputs, up1 and up2.
Parameter Changes
111
5. Conclusions
Direct Model Reference Adaptive Control is utilized to control two-link flexible robot,
whose parameters vary. The feedforward compensator was designed for the system and it
was showed that the augmented plant satisfies the ASPR conditions over the range of
parameter variations. As seen from the results of the tip position response, the system
closely follows the reference model. During the simulations it was observed that DMRAC
was capable of controlling the plant despite the changes in the plant parameters. The ease of
its implementation and its robustness were demonstrated.
6. References
Cannon, R. H. & Schmitz, E. (1984). Initial Experiments on the End-Point Control of a
Flexible One-Link Robot. Int. J. Robotics Research, Vol 3, No. 3, pages 325-338.
Geniele, H, Patel, R.V. & Khorasani, K. (1995). Control of a Flexible Link Manipulator. IEEE
Int. Conf. Robotics and Automation, Vol. 1, pp 1217-1222, May 1995.
Ge, S. S., Lee, T.H. & Wang, Z. P. (2001). Adaptive Robust Controller Design for Multi-Link
Flexible Robots. Proceedings of the American Control Conference, Vol. 2, pp 947-952,
June, 2001.
Park, J. H. & Asada, H. (1992). Integrated Structure/Control Design of a Two-Link Nonrigid
Robot Arm for High Speed Positioning, IEEE Int. Conf. Robotics and Automation, pp
735-741, Nice, France, May 1992.
Green, A. & Sasiadek, J. (2004). Dynamics and Trajectory Tracking Control of a Two-Link
Robot Manipulator, Journal of Vibration and Control, Vol. 10, No. 10, 1415-1440.
Ider , K., S., Ozgoren, M., K. & Ay, O. (2002) Trajectory Tracking Control of Robots with
Flexible Links, Mechanism and Machine Theory, Vol 37, pp 1377-1394.
Siciliano, B. & Villani, L. (2001). An Inverse Kinematics Algorithm for Interaction Control of
a Flexible Arm with a Compliant Surface, Control Engineering Practice, Vol 9, pp 191198.
De Luca, A. & Scicliano, B. (1989). Trajectory Control of a Nonlinear One-Link Flexible Arm
International Journal of Control, Vol 50, pp 1699-1716.
Xu, B., Kenji, F. & Yoshikazun, H. (2005). An Energy-Based Nonlinear Control for a TwoLink Flexible Manipulator, Trans Japan Soc Aeronautical Space Sc., Vol 48, No 160, pp.
92-101.
De Luca, A, Lucibello, P. & Nicolo, F. (1988). Automatic Symbolic Modeling and Non Linear
Control of Robots with Flexible Links. Proc. IEE Work On Robot Control, pp. 62-70,
Oxford, UK, April 1988.
Li, Y., Liu, G, Hong, T. & Liu, K. (2005). Robust Control of a Two-Link Flexible Manipulator
with Quasi-Static Deflection Compensation Using Neural Networks, Journal of
Intelligent and Robotic Systems Vol 44 , No 3 pp. 263-276. Kluwer Academic
Publishers, MA, USA
Cetinkunt, S., Scicliano, B. & Book, J. W. (1984). Symbolic Modeling and Dynamic Analysis
of Flexible Manipulators Proc. IEEE Int. Conf. Syst., Man and Cyber.,pp. 798-803,
Georgia, USA.
Book, W., J. (1984). Recursive Lagrangian Dynamics of Flexible Manipulator Arms. Int.
Journal of Robotics Research, Vol. 3, No. 3, pp. 87-101.
112
Adaptive Control
Miranda, E. (2004). Direct Adaptive Control of a Flexible Manipulator, Thesis, Texas A&M
University, Kingsville, TX, USA.
Kaufman, H., Bar-Kana, I., & Sobel, K. (1998). Direct Adaptive Control Algorithms: Theory and
Applications, Springer Communication and Control Eng., 2nd Edition, SpringerVerlag.
Ozcelik, S. & Kaufman, H. (1999). Design of robust direct adaptive controllers for SISO: time
and frequency domain design conditions. Intl. Journal of Control, Vol.72, No. 6, pp.
517-530.
Ionnou, P. & Kokotovic, P. (1983). Adaptive Systems with reduced models, Springer-Verlag,
New York, USA.
Bar-Kana, I. (1994). Positive realness in multivariable stationary linear systems, Journal of Franklin
Institute ,No. 328, pp. 403-417.
Ozcelik, S. (2004). Robust direct adaptive control for MIMO Systems Using Q-Parameterization,
IFAC ALCOSP, pp 93-98, Yokohama, Japan.
5
Discrete Model Matching Adaptive Control for
Potentially Inversely Non-Stable ContinuousTime Plants by Using Multirate Sampling
S. Alonso-Quesada & M. De La Sen
Electricity and Electronics Department, Faculty of Science and Technology

Basque Country University, Spain
1. Introduction
Adaptive control theory has been widely applied for stabilizing linear time invariant plants
of unknown parameters (Goodwin & Sin, 1984). One of the more used methods for such a
purpose is based on the model reference adaptive control (MRAC) problem (Astrm &
Wittenmark, 1997). Such a method requires some assumptions relative to the plant to be
controlled in order to carry out the synthesis of a stable controller (Narendra &
Annaswamy, 1989). One of them is that the plant has to be inversely stable, what means that
its zeros have to be located within the stability domain. However, this information is not
always available to the designer when the system under control contains unknown
parameters. There are several alternatives to circumvent this drawback and carry out the
stable adaptive control design. Some of them consist on relaxing the control performance
from the model matching to that achievable from the closed-loop pole placement (AlonsoQuesada & De la Sen, 2004 and Arvanitis, 1999). In this way, the stabilization of the closedloop system can be ensured although its transient behaviour cannot be fixed to a predefined
one.
On one hand, the work (Alonso-Quesada & De la Sen, 2004) includes an estimates
modification in the estimation algorithm to ensure the controllability of the estimated plant
model without assuming any knowledge about the parameters of the plant to be controlled.
This controllability property is crucial to avoid pole-zero cancellations between the
estimated plant and the controller, which are both time-varying. In this context, a projection
of the estimated plant parameters into a region in the parameter space where the closedloop system is free of pole-zero cancellations for all time can be alternatively used provided
that the true plant is controllable and the knowledge of a region where the true plant
parameters belong to (Goodwin & Mayne, 1987).
On the other hand, the research (Arvanitis, 1999) proposes an adaptive pole-placement
control for linear systems using generalized sampled-data hold functions. Following such a
technique, gain controllers essentially need to be designed. Concretely, a periodic piecewise
constant gain controller is added in the feedback chain. In the non-adaptive case, such
constant gain values are those required so that the discretized closed-loop model under a
fundamental sampling period and a zero-order hold (ZOH) be stabilized. For such a
114
Adaptive Control
purpose, each sampling period is divided in a certain finite number of uniform subintervals
and the controller gain takes a different value within each of them in order to locate the
discretized poles at the stable desired locations. In other words, the controller consists of a
constant vector of gains. In this sense, the controller works with a sampling rate faster than
that used to discretize the plant to be controlled. In the adaptive case, an estimated model of
the discretized plant is on-line updated by means of an estimation algorithm. Such a model
is used to parameterize the controller gains vector which becomes time-varying and
converges asymptotically to a constant one.
Another alternative, which does not relax the MRAC objective, to overcome the
drawback of the unstable zeros in a continuous-time plant is the design of discrete-time
controllers which are synthesized from the discretization of the continuous-time plant by
means of a holder device combined with a multirate with fast input sampling rate (De la Sen
& Alonso-Quesada, 2007 and Liang & Ishitobi, 2004). The main motivation of this method is
that an inversely stable discretized model of the plant can be obtained with an appropriate
choice of the multirate gains. In this way, an adaptive controller can be designed to match a
discrete-time reference model since all the discretized plant zeros may be cancelled if suited.
In this context, a fractional-order hold (FROH) with a multirate input is used in this paper to
obtain an inversely stable discretized plant model from a possible non-inversely stable and unstable
time invariant continuous-time plant. Then, a control design for matching a discrete-time reference
model is developed for both non-adaptive and adaptive cases. Note that a FROH includes as
particular cases the ZOH and the FOH (first-order hold). In this way, the stabilization of the
continuous-time plant is guaranteed without any assumption about the stability of its zeros
and without requiring estimates modification in contrast with previous works on the
subject. In this sense, this paper is an extension of the work (De La Sen & Alonso-Quesada,
2007) where the same problem is addressed. The main contribution is related to the method used
to built the continuous-time plant input from the discrete-time controller output. In the present
paper, the FROH acts on the fundamental sampling period used to discretize the plant
(plant output sampling) while in the aforementioned paper the FROH acted on the sampling
period used to define the multirate device at the plant input. This later sampling period is
an integer fraction of the plant output one, i.e. an integer number of input samples takes
place within each output sampling period. Such an integer has to be suitably chosen for
disposing of the enough freedom degrees being necessary to place the discretized plant
zeros at desired locations, namely within the unity circle in order to guarantee the inverse
stability of the discretized plant model.
The assumptions about the plant to guarantee the closed-loop stability of the adaptive
control system are the following: (1) the stabilizability of the plant and (2) the knowledge of
the continuous-time plant order. The motivation for using a multirate sampling input
instead of the most conventional single rate one resides in the fact that the former, with the
appropriate multirate gains, provides an inversely stable discretized plant model without
requirements on either the stability of the continuous-time plant zeros or the size of the
sampling period. In this sense, a single rate input can only provide an inversely stable
discretized plant from an inversely stable continuous-time plant and, moreover, the
fundamental sampling period to discretized the plant has to be sufficiently small (Blachuta,
1999). Finally, the use of a FROH, instead of the most conventional ZOH, allows to
accommodate better some discrete adaptive techniques to the transient response of discretetime controlled continuous-time plants (Brcena et al., 2000 and Liang et al., 2003).
Discrete Model Matching Adaptive Control for Potentially Inversely Non-Stable Continuous-Time
Plants by Using Multirate Sampling
115
The paper is organized as follows. Section 2 formulates a discrete state-space

description under fast input sampling to then obtain an input-output discrete transfer
function for the running slow sampling rate, namely, that acting on the output signal. The
selection of the scalar gains that generate the fast sampled input so that the discrete plant
zeros are stable is focused on depending on the continuous-time plant parametrization.
Section 3 discusses the synthesis of a model-matching based controller with a possible
potential free design of all the zeros of the reference model. The case of known plant
parameters and the adaptive case for not fully known plant parameters are both considered.
Two alternatives are proposed to update on-line the time-varying multirate gains in the
adaptive case. The first one updates the multirate gains for all sampling instants in order to
maintain the zeros of the estimated discretized plant fixed within the stability domain. On
the contrary, the other one updates the multirate gains only when the change of gains is
crucial to guarantee the stability of the estimated discretized plant zeros. In this way, the
multirate gains are not updated for all sampling instants and then they became piecewise
constant. As a result, the zeros of the estimated discretized plant become time-varying
within the stability domain. Section 4 deals with the stability analysis of the adaptive control
system. Simulated examples which highlight the proposed design philosophy are provided
in Section 5. A comparison of the results obtained with the two different methods for
updating the multirate gains is presented. Finally, conclusions end the paper in Section 6.
2. Discretized Plant Representation

Consider a linear time-invariant, single-input single-output and strictly proper continuoustime plant described by the following state space equations:
x& (t) = Ax(t) + Bu(t) ; y(t) = Cx(t)
(1)
where u(t) and y(t) are, respectively, the input and output signals, x(t) n denotes the
state vector and A, B and C are constant matrices of suitable dimensions. A FROH and a
multirate sampling on the fast input sampling will be used in order to obtain an inversely
stable discretized plant model. The signal generated by such a device is given by,
u(k) u(k 1)
u(t) = j u(k) +
( t kT )
T
(2)
for t [ kT + ( j 1)T,kT + jT ) , j {1, 2, K , N} and FROH correcting gain [ 1,0 ) ( 0,1] ,
where T is the sampling period (slow sampling) which is divided in N equal subperiods of
length T = T N (fast sampling) to generate the multirate plant input, u(k) denotes the
value of a discrete-time controller output signal at the instant kT, for all non-negative
integer k, and j s are real constants. Note that the FROH device operates on the sequence
{u(k)}
defined at the slow sampling instants kT and then the input u(t) is generated over
each subperiod with the corresponding gain j , for j {1, 2, K , N} , via (2). By substituting
(2) into (1) and sampling the plant output y(t) over the sampling period T, the following
Adaptive Control
116
state space representation is obtained which corresponds to the discrete-time plant model
that relates the sequences {u(k)} and {y(k)} :
x(k + 1) = F x(k) + G 1 u(k) + G 2 u(k 1)
; y(k) = C x(k)
(3)
where F = N = (T) = eAT is the continuous-time state transition matrix computed for a
slow sampling period and,

N
N
G1 = l N l + n 1 ; G 2 = l N l n1
T
T
l=1
l=1
(4)
with,
T
= (T s) B ds n 1 ; = (T s) B s ds n 1
(5)
The transfer function of this discrete-time plant model is,

1 B(z)
H(z) = C(zI n N )1 G1 + G 2 =
z A(z)
(6)
where,
zI N
B(z) = C Adj(zI n N )C (z)g = Det n
C
A(z) = z Det(zI n ) = z
N
n +1
+ ai z
C (z)g n + 1 n i + 1
= bi z
0 i =1
(7)
n i + 1
i =1
with Adj() and Det() denoting, respectively, the adjoint matrix and the determinant of the
square matrix (.), I n denoting the n-th order identity matrix, and
g = [ 1 L N ] N + 1 ; C (z) L N 1 (z) L (z)

T
(z) nxN
(8)

(z) = z +
T
T
Note that the coefficients bi , for i {1, 2, K , n + 1} , of the polynomial B(z) in (7)
depend on the values j , for j {1, 2, , N} , which define the multirate device. This fact
lets to allocate the zeros of the discretized plant model at desired locations if a suitable
number of multirate gains is provided. In this sense, the multirate gains j , being the
components of the vector g , are calculated to guarantee that such zeros are maintained
within the stability domain, i.e., the unity circle. In particular, the coefficients bi can be
117
expressed as:
N
bi = bi , j j v = M g
(9)
j=1
where M = bi , j ( n + 1)N and v = [ b1
b 2 K b n + 1 ] . The coefficients bi , j depend on the

T
parameters of the continuous-time plant, the sampling period T and the correcting gain of
the FROH considered in the discretization process.
Assumptions 1.
(i) The plant is stabilizable, i.e. its transfer function does not possess unstable pole-zero
cancellations,
(ii) the plant order n is known, and
(iii) the correcting gain of the FROH device and the sampling period T are chosen such
that M is a full rank matrix.
***
Remark 1. The multirate gains vector g required to place the zeros of the discretized plant
transfer function (6) at desired locations may be calculated from (9) provided Assumptions 1
and that N n + 1 . In this sense, such locations are prefixed via a suitable choice of the
vector v composed by the coefficients of the desired polynomial for the transfer function
numerator. If N > n + 1 , different solutions can be obtained for g . Otherwise, i.e. if
N = n + 1 , there is a unique solution for the multirate gains vector from the linear algebraic
system (9) which places the discretized zeros at desired locations.
***
The discretized model (6) can be described by the following difference equation:
n
n +1
n +1 N
i =1
i =1
i =1 j=1
y(k) = a i y(k i) + bi u(k i) = a i y(k i) + bi , j j u(k i)

i =1
n
n+1 N
i =1
i =1 j=1
(10)
= a i y(k i) + bi , j u j (k i) = (k 1)
T
where,
= aT
T
T
T
b,1
b,2
K b,n
+1
a = [ a1 a2 K an ]
; (k 1) = Ty (k 1) uT (k 1) uT (k 2) K uT (k n 1)
; y (k 1) = [ y(k 1) y(k 2) K y(k n)]
b,i = bi ,1 bi ,2 K bi ,N
u j (k i) = j u(k i)
; u (k i) = [ u1 (k i) u2 (k i) K uN (k i)]
(11)
for all i {1, 2, K , n + 1} and all j {1, 2, K , N} . In the rest of the paper, the case
N = n + 1 will be considered for simplicity purposes.
Adaptive Control
118
3. Control Design
The control objective in the case of known plant parameters is that the discretized plant
B (z)
model matches a stable discrete-time reference model Hm (z) = m
whose zeros can be
A m (z)
freely chosen, where z is the Z-transform argument. Such an objective is achievable if the
discretization process uses the multirate sampling input with the appropriate multirate
gains, what guarantees the inverse stability of the discretized plant. Then, all the discretized
plant zeros may be cancelled by controller poles. In this way, the continuous-time plant
output tracks the reference model output at the sampling instants. The tracking-error
between such signals is zero at all sampling instants in the case of known plant parameters
while it is maintained bounded for all time while it converges asymptotically to zero as time
tends to infinity in the adaptive case considered when the plant parameters are fully or
partially unknown. A self-tuning regulator scheme is used to meet the control objective in
both non-adaptive and adaptive cases.
3.1 Known Plant
The proposed control law is obtained from the difference equation:
R(q) u(k) = T(q) c(k) S(q) y(k)
for all non-negative integer k, where
{c(k)}
(12)
is the input reference sequence and q is the
running sample rate advance operator being formally equivalent to the Z-argument used in
discrete transfer functions. The reconstruction of the continuous-time plant input u(t) is
made by using (2), with the control sequence {u(k)} obtained from (12), with the
appropriate multirate gains j , for j {1, 2, K , N} , to guarantee the stability of the
discretized plant zeros.

The discrete-time transfer function of the closed-loop system obtained from the
application of the control law (12) to the discretized plant (6) is given by:
Y(z)
B(z)T(z)
T(z)
=
=
C(z) A(z)R(z) + B(z)S(z) A(z) + S(z)
(13)
where the second equality is fulfilled if the control polynomial R(z) = B(z) . In this way, the
polynomial B(z) , which is stable, is cancelled. Then, the polynomials T(z) , R(z) and S(z)
of the controller (12) so that
Y(z) Bm (z)
are obtained from:
=
C(z) A m (z)
T(z) = B m (z)A s (z) ; R(z) = B(z) ; S(z) = A m (z)A s (z) A(z)
(14)
where A s (z) is a stable monic polynomial of zero-pole cancellations of the closed-loop

system. The following degree constraints are satisfied in the synthesis of the controller:
Deg [Am (z)] + Deg [As (z)] = Deg [A(z)] = n + 1
n i
Deg [ S(z)] = Deg [A(z)] 1 = n S(z) = si +1 z
i =0
Deg [ T(z)] = Deg [ B (z)] + Deg [A (z)] n

m
s
119
(15)
3.2 Unknown Plant

If the continuous-time plant parameters are unknown then the vector in (11) composed of
the discretized plant model parameters is also unknown. However, all the above control
design in the previous subsection remains valid if such a parameter vector is estimated by
an estimation algorithm. In this way, the controller parameterization can be obtained from
R(z,k) = B(z,k)
, with B(z,k)
denoting the estimated of B(z) at the current slow sampling
instant kT, and equations similar to (14) by replacing the discretized plant polynomial A(z)
(Alonso-Quesada & De la Sen, 2004). Note that

by its corresponding estimated one A(z,k)
T(z) in (14) has to be calculated once for all since Bm (z) and A s (z) are time-invariant while
S(z) is updated at each running sampling time since the polynomial A(z,k)
is time-
varying. The coefficients of the unknown polynomial B(z) depend, via (9), on the multirate
input gains j , for j {1, 2, K , N} , being applicable to calculate the input within the intersample slow period. However, the estimation algorithm provides an adaptation of each
parameter bi , j , namely b i , j (k) , for i, j {1, 2, K , N} and all non negative integer k. Then,
the j -gains have to be also updated in order to ensure the stability of the zeros of the
estimated discretized plant, i.e. the roots of B(z,k)

be stable. Then, the gains j become
time-varying, namely j (k) . The estimation algorithm for updating the parameters vector
(k) , which denotes the estimated of , and two different design alternatives for the
adaptation of the multirate gains are presented below. Also, the main boundedness and
convergence properties derived from the use of such algorithms are established.
3.2.1. Estimation algorithm
An a priori estimated parameters vector is obtained at each slow sampling instant by using
a recursive least-squares algorithm (Goodwin & Sin, 1984) defined by:
P(k) = P(k 1)
P(k 1) (k 1) T (k 1) P(k 1)
1 + T (k 1) P(k 1) (k 1)
0 (k) = 0 (k 1) +
P(k 1) (k 1) e 0 (k)
1 + T (k 1) P(k 1) (k 1)
(16)
T
T
for all integer k > 0 where e 0 (k) = 0 (k 1) (k 1) = % 0 (k 1)(k 1) denotes the a
priori estimation error and P(k) is the covariance matrix initialized as P(0) = P T (0) > 0 .
Adaptive Control
120
Such an algorithm provides an estimation 0 (k) of the parameters vector by using the
regressor (k 1) , defined in (11), built with the output and input measurements with the
multirate gains j (k 1) obtained at the previous slow sampling instant, i.e.
u j (k i) = j (k 1) u(k i) for all i {1, 2, K , n+1} . Then, an a posteriori estimates vector
is obtained in the following way:
Modification algorithm.
This algorithm consists of three steps:
0 (k) = b 0 (k) NN , for i, j {1, 2, K , N} , from the a priori
Step 1: Built the matrix M
i,j
0
estimates b ,i (k) , included in 0 (k) , of the corresponding b ,i defined in (11).
0 (k)
Step 2: M(k)
=M
0 then b ,i (k) = 0b ,i (k)

If Det M(k)
< 0
else while Det M(k)
M(k) = M(k) + I
end;
for i = 1 to N
(k)
b ,i (k) = M
i
end
end.
T
T
Step 3: (k) = 0a (k) Tb ,1 (k) bT,2 (k) K bT,N (k) ,
(k) denotes the i-th row

for some real positive constants << 1 and 0 << 1 , and where M
i
of M(k) .
***
Remark 2. Note that the estimate 0a (k) corresponding to the parameters of a is not
affected by the modification algorithm. Also, note that the while instruction part of the
second step is doing a finite number of times since there exists a finite integer number l
N
0
0
0
.
***
such that Det M(k)
= Det M (k) + lI N = ( l ) + f , b,1 (k), b,2 (k), K , b,N (k) 0
3.2.2. Updating of the time-varying multirate gains

Once the estimated parameters vector is obtained at each slow sampling instant the
multirate input gains have to be updated. Two alternative algorithms are considered to
carry out such an operation.
Algorithm 1.
A vector of multirate gains is updated at all slow sampling instants in order to maintain the
zeros of the estimated discretized plant fixed at desired locations within the stability domain
z < 1 . Such desired zeros are the roots of a predefined polynomial B(z) . For such a
121
purpose, the required vector g(k)

is obtained from the resolution of the following matrix
equation:
g(k)
=v
M(k)
at each slow sampling instant, where v = [ b1
(17)
b2 K bN ] is composed by the coefficients

T
of B(z) , M(k)
= b i , j (k) NN , with b i , j (k) denoting each of the a posteriori estimated
parameters corresponding to the components of the vectors b ,i defined in (11), and
g(k)
= [ 1 (k) 2 (k) K N (k)] . In this way, g(k)
is composed by the multirate gains
T
which make the numerator of the estimated discretized plant model be equal to the desired
polynomial B(z) . Note that the matrix equation (17) can be solved at all slow sampling
instants since the parameters modification added to the estimation algorithm ensures the
.
non-singularity of the matrix M(k)
Algorithm 2.
It consists of solving the equation (17) only when it is necessary to modify the previous
values of the multirate gains in order to guarantee the stability of the zeros of the estimated
discretized plant model. i.e., the multirate gains remain equal to those of the preceding slow
sampling instant if the zeros of the estimated discretized plant obtained with the current
estimated parameters vector, (k) , and the previous multirate gains, j (k 1) , are within
the discrete-time stability domain. Otherwise, the multirate gains are updated by the
resolution of the equation (17), which can be solved whenever it is necessary since the
is invertible at all slow sampling instant due to the modification included in

matrix M(k)
the estimation algorithm. In this way, the multirate gains are piecewise constant, the
estimated discretized plant zeros are time-varying and the computational burden associated
with the updating of the multirate gains is reduced with respect to that of Algorithm 1.
3.2.3. Properties of the estimated models
The parameter estimation algorithm, together with any of the considered adaptation
algorithms for the multirate gains, possesses the properties given in the following lemma,
whose proof is presented in Appendix A.
Lemma 1. Main properties of the estimation and multirate gains adaptation algorithms
(i) P(k) is uniformly bounded for all non-negative integer k, and it asymptotically
converges to a finite, at least semidefinite positive, limit as k .

(ii) 0 (k) and (k) are uniformly bounded and they asymptotically converge to a finite
limit as k .
of multirate gains is bounded and converges to a finite limit as k .

(iii) The vector g(k)
( e (k))
0
(iv)
1 + T (k 1) P(k 1) (k 1)
is uniformly bounded and it asymptotically converges to
Adaptive Control
122
zero as k .
(v) e 0 (k) asymptotically converges to zero as k .
(vi) Assuming that the external input c(k) is sufficiently rich such that (k 1) in (11) is
persistently exciting, 0 (k) tends to the true parameters vector as k . Then, (k)
T
tends to 0 (k) and e(k) = (k 1) (k 1) tends to zero as k .
***
Remark 3. The convergence of the estimated parameters to their true values in requires
that (k 1) is persistently exciting. In this context, (k 1) is persistently exciting if there
exists an integer l such that 1I m >
k0 + l
(k 1)
k =k0
(k 1) > 2 I m where 1 > 0 , 2 > 0 and
m = n + N 2 = n 2 + 3n + 1 is the number of components of the regressor (k 1) . Such a

condition may be ensured by chosing an external input sufficiently rich of order m , i.e. it
***
consists of at least m 2 frequencies in the frequency domain (Ioannou & Sun, 1996).
4. Stability Analysis
The plant discretized model can be written as follows,
n
n +1
i=1
i =1
y(k) = y(k)
+ e(k) = T (k 1)(k 1) + e(k) = a i (k 1)y(k i) + b i (k 1)u(k i) + e(k)
(18)
and the adaptive control law as,

u(k) =
n
1 n
( s 1 (k 1)a i (k 1) s i + 1 (k 1)) y(k i) s 1 (k 1)b i (k 1) + b i + 1 (k 1) u(k i)
b (k)
i =1
i =1
1
n +1
s (k)
s 1 (k 1)b n + 1 (k 1)u(k n 1) + bmi c(k i + 1) 1
e(k) + (k)
i =1
b 1 (k)
(19)
where (12) has been used with R(q) and S(q) substituted, respectively, by time-varying
polynomials R(z,k)
and S(z,k)
, which is the solution of the equation (14) for the
= B(z,k)
adaptive case, and,
(k) =
n
( s 1 (k) s 1 (k 1)) a i (k 1) ( s i + 1 (k) s i + 1 (k 1) ) y(k i)
b 1 (k) i = 1
1
( s 1 (k) s 1 (k 1) ) b i (k 1) + b i + 1 (k) b i + 1 (k 1) u(k i)
i=1
( s 1 (k) s 1 (k 1)) b n + 1 (k 1)u(k n 1)
By combining (18) and (19), the discrete-time closed-loop system can be written as:
(20)
123
(21)
x(k) = (k 1) x(k 1) + 1e(k) + 2 (k)
where (k) =
1 n+1
bmi c(k i + 1) s 1 (k) e(k) + (k) and,

b (k)
i
=
1
1
x(k-1)=[y(k-1)y(k- 2)Ly(k- n)u(k-1)u(k- 2)Lu (k- n -1)]

T
1 =[10L0] (2n+1)x1 ;2 = 00L0 1{ 0L0 (2n+1)x1
n+1
a1(k1)
1
0
M
0
(k 1)= f1(k1)
b1(k)
0
0
M
a2(k1)
an1(k1)
an (k1)
b (k1)
1
b (k1)
2
b (k1)
n
0
0
0
f (k1)
1
b (k)
1
0
0
f (k1)
n
b (k)
1
0
0
h 1(k1)
b (k)
1
1
0
h 2 (k1)
b (k)
1
0
1
f (k1)
n-1
b (k)
1
0
0
h n (k1)
b (k)
1
0
b (k1) (22)
n+1
0
0
0
h n+1(k1)
b (k)
1
0
with fi (k 1) = s 1 (k 1)a i (k 1) s i + 1 (k 1) , h i (k 1) = s i (k 1)b i (k 1) + b i + 1 (k 1) , for

i {1, 2, , n} , and h n + 1 (k 1) = s 1 (k 1)b n + 1 (k 1) .
N
Note that a i (k 1) and b i (k 1) = b i , j (k 1) j (k 1) are uniformly bounded from

j=1
Lemma 1 (properties ii and iii). Also, b 1 (k) 0 since the adaptation of the multirate gains
makes such a parameter fixed to a prefixed one which is suitably chosen and s i (k 1) is
uniformly bounded from the resolution of a equation being similar to that of (14) replacing
1) and S(z,k
1) ,
polynomials A(z) and S(z) by time-varying polynomials A(z,k
respectively.
The following theorem, whose proof is presented in Appendix B, establishes the main
stability result of the adaptive control system.
Theorem 1. Main stability result.
(i) The adaptive control law stabilizes the discrete-time plant model (6) in the sense that
{u(k)} and {y(k)} are bounded for all finite initial states and any uniformly bounded
reference input sequence {c(k)} subject to Assumptions 1,

(ii) {y(k)} converges to {y m (k)} as k tends to infinity, and
(iii) the continuous plant input and output signals, u(t) and y(t) , are bounded for all t. ***
Adaptive Control
124
5. Simulations Results
Some simulation results which illustrate the effectiveness of the proposed method are
shown in the current section. A continuous-time unstable plant of transfer function
s2
G(s) =
with an unstable zero, and whose internal representation is defined by
(s 1)(s + 3)
3 0
T
the matrices A =
, B = [ 1 1] and C = [ 1.25 0.25] , is considered. A suitable
0
1
multirate scheme with fast input sampling through a FROH device is used to place the zeros
of the discretized plant within the stability region and a discrete-time controller is
synthesized so that the discrete-time closed-loop system matches a reference model. The
results for the case of known plant parameters are presented in a first example and then two
more examples with the described adaptive control strategies are considered. The difference
among such adaptive control strategies relies on the way of updating the multirate gains for
ensuring the stability of the estimated discretized plant zeros.
5.1. Known Plant Parameters

The discretization of the continuous-time plant with a multirate, N = 3 , and a FROH device
with = 0.7 for a slow sampling time T = 0.3 is performed leading to the discrete transfer
function H(z) =
B(z) b1 (g)z 2 + b 2 (g)z + b 3 (g)

=
where b1 (g) = 0.0307 1 + 0.0693 2 + 0.13 3 ,
A(z) z(z 2 1.7564z + 0.5488)
b 2 (g) = (0.0788 1 + 0.1488 2 + 0.2631 3 ) and b 3 (g) = 0.0083 1 + 0.0343 2 + 0.0797 3 are the
coefficients of the transfer function numerator of the discretized model. Such coefficients
depend on the multirate gains i , for i {1, 2, 3} , included as components in the vector g .
The zeros of such a discretized plant can be fixed within the stability domain via a suitable
choice of the multirate gains. In this example such gains are 1 = 621.8706 , 2 = 848.4241
and 3 = 297.4867 so that B(z) = B(z) = z 2 + z + 0.25 and then both zeros are placed at
z0 = 0.5 . The control objective is the matching of the reference model defined by the
z 2 + z 0.272
. For such a purpose, the controller has to cancel the
(z + 0.2)3
discretized plant zeros, which are stable, and add those of the reference model to the
discrete-time closed-loop system. The values of the control parameters to meet such an
objective are s1 = 2.3564 , s 2 = 0.4288 and s 3 = 0.008 . A unitary step is considered as
transfer function Gm (z) =
external input signal. Figure 1 displays the time evolution of the closed-loop system output,
its values at the slow sampling instants and the sequence of the discrete-time reference
model output. Figure 2 shows the plant input signal. Note that perfect model matching is
achieved, at the slow sampling instants, without any constraints in the choice of the zeros of
the reference model G m (z) , in spite of the continuous-time plant possesses an unstable zero.
Furthermore, the continuous-time output and input signals are maintained bounded for all
time.
Fig. 1. Plant and reference model output signals
Fig. 2. Plant input signal
125
126
Adaptive Control
5.2. Unknown Plant Parameters

An adaptive version of the discrete-time controller designed in the previous example is
considered with the parameters estimation algorithm being initialized with
T
0 (0) = 10 2 [ 263.46 82.32 4.61 10.39 19.51 11.82 22.33 39.46 1.25 5.15 11.95]
and P(0) = 1000 I 11 . Furthermore, the values = 0 = 10 6 are chosen for the modification
algorithm included in such an estimation process. Two different methods are considered to
update the multirate gains. The first one consists of updating such gains at all the slow
sampling instants so that the discretized zeros are maintained constant within the stability
domain (Algorithm 1). The second one consists of changing the value of the multirate gains
only when at least one of the discretized zeros, which are time-varying, is going out of the
stability domain. Otherwise, the values for the multirate gains are maintained equal to those
of the previous slow sampling instant (Algorithm 2).
5.2.1. Algorithm 1: Discretized plant zeros are maintained constant
Figure 3 displays the time evolution of the closed-loop adaptive control system output, its
values at the slow sampling instants and the sequence of the discrete-time reference model
output under a unitary step as external input signal. Note that the discrete-time model
matching is reached after a transient time interval. Figures 4 and 5 show, respectively, the
plant output signal and the input signal generated from the multirate with the FROH
applied to the control sequence {u(k)} . It can be observed that both signals are bounded for
all time. Finally, Figures 6 and 7 display, respectively, the time evolution of the multirate
gains and the adaptive controller parameters. Note that the multirate gains and the adaptive
control parameters are time-varying until they converge to constant values.
Fig. 4. Plant output signal
127
128
Fig. 6. Multirate gains
Fig. 7. Adaptive control parameters
Adaptive Control
129
5.2.2. Algorithm 2: Discretized plant zeros are time-varying

The multirate gains are maintained constant to their values at the previous slow sampling
instant until at least one of the discretized plant zeros is going out of the stability domain. In
this sense, note that the discretized zeros vary when the values of the multirate gains are
maintained constant and eventually they can go out of the stability domain. When this
happens such gains are again calculated to place both discretized zeros at z0 = 0.5 . The
discrete-time model matching is reached after a transient time interval and the continuoustime plant output and input signals are bounded for all time as it can be observed from
Figures 8, 9 and 10 where the response to a unitary step is shown. The maximum values
reached by both continuous-time output and input signals are larger than those obtained
with the previous method (Algorithm 1) for updating the multirate gains. Figures 11 and 12
display, respectively, the evolution of the multirate gains and the controller parameters. The
adaptive control parameters are time-varying until they converge to constant values while
the multirate gains are piecewise constant and also they converge to constant values. Note
that this second method ensures a small number of changes in the values of the multirate
gains compared with the first method since such gains only vary when it is necessary to
maintain the zeros within the stability domain. This fact gives place to a less computational
effort to generate the control law than that required with the first method. However, the
behaviour of the continuous-time plant output and input signals is worse with the use of
this second alternative in this particular example. Finally, the evolution of the modules of
the discretized plant zeros and the coefficients of the time-varying numerator of such an
estimated model are, respectively, shown in Figures 13 and 14.
130
Fig. 9. Plant output signal
Adaptive Control
Fig. 11. Multirate gains
Fig. 12. Adaptive control parameters
Fig. 13. Modules of the estimated discretized plant zeros
131
132
Adaptive Control
Fig. 14. Coefficients of the estimated discretized plant numerator
6. Conclusion
This paper deals with the stabilization of an unstable and possibly non-inversely stable
continuous-time plant. The mechanism used to fulfill the stabilization objective consists of
two steps. The first one is the discretization of the continuous-time plant by using a FROH
device combined with a multirate input in order to obtain an inversely stable discretized
model of the plant. Then, a discrete-time controller is designed to match a discrete-time
reference model by such a discretized plant. There is not any restriction in the choice of the
reference model since the zeros of the discretized plant model are guaranteed to be stable by
the fast sampled input generated by the multirate sampling device.
An adaptive version of such a controller constitutes the main contribution of the
present manuscript. The model matching between the discretized plant and the discretetime reference model is asymptotically reached in the adaptive case of unknown plant. Also,
the boundedness of the continuous-time plant input and output signals are ensured, as it is
illustrated by means of some simulation examples. In this context, the behaviour of the
designed adaptive control system in the inter-samples period may be improved. In this
sense, an improvement in such a behaviour has been already reached with a multiestimation scheme where several discretization/estimation processes, each one with its
proper FROH and multirate device, are working in parallel providing different discretized
plant estimated models (Alonso-Quesada & De la Sen, 2007). Such a scheme is completed
with a supervisory system which activates one of the discretization/estimation processes.
Such a process optimizes a performance index related with the inter-sample behaviour. In
this sense, each of the discretization/estimation processes gives a measure of its quality by
means of such an index which may measure the size of the tracking-error and/or the size of
the plant input for the inter-sample period. The supervisor switches on-line from the current
process to a new one when the last is better than the former, i.e. the performance index of
the new process is smaller than that of the current one. Moreover, the supervisor has to
guarantee a minimum residence time between two consecutive switches in order to ensure
the stability of the adaptive control system.
133
7. Appendix A. Proof of Lemma 1

(i) P(k) is a monotonic non-increasing matrix sequence since P(k) P(k 1) 0 for all
integer
k>0
from (16). Moreover, if
P(k 1 + 1) P(k 1 ) = 0
P(k 1 ) = 0
for any integer
from (16) and then P(k) = 0
k1 > 0
then
for all integer k k 1 . Thus,
0 P(k) P(0) and P(k) asymptotically converges to a finite limit as k .

T
(ii) By considering the non-negative sequence V(k) = % 0 (k)P 1 (k)% 0 (k) and applying the
matrix inversion lemma (Goodwin & Sin, 1984) to (16) it follows that,
( e (k))
0
V(k) V(k 1) =
1 + T (k 1)P(k 1)(k 1)
(23)
where (16) and the definition of the estimation error have been used. Then, V(k) V(0)
and
{P(0)} % 0
% 0 (k) max
(0) <
min {P(0)}
where
max {P(0)} and
min {P(0)}
denote the
maximum and the minimum eigenvalues of P(0) , respectively. It implies that % 0 (k) ,
and then also 0 (k) , is uniformly bounded. Then, (k) is also bounded since the
modification algorithm guarantees the boundedness of M(k)

provided that 0 (k) is
bounded. On other hand, V(k) asymptotically converges to a finite limit as k from
its definition and the fact that such a sequence is non-negative and monotonic nonincreasing. Then, % 0 (k) , and also 0 (k) , converges to a finite limit as k since P(k)
also converges as it has been proved in (i). Then, M(k)

and (k) also converge to finite
limits as k .
(iii) The boundedness and convergence of the estimation model parameters vector together
with the non-singularity of the matrix M(k)

, guaranteed by the modification algorithm,
implies the boundedness and convergence of the vector g(k)

obtained by resolution of
equation (17).
(iv) It follows that
1+
i =1
( e (k))
0
( e (i))
0
(i 1)P(i 1)(i 1)
= V(0) V(k) V(0) < from (23), then
1 + T (k 1)P(k 1)(k 1)
is uniformly bounded and it converges to zero as k .
(v) It follows that lim {e 0 (k)} = 0 irrespective of the boundedness of (k 1) from the fact
k
e 0 (k) )
(
that lim
= 0 . On one hand, if (k 1) is bounded then
T
k 1 + (k 1)P(k 1)(k 1)
Adaptive Control
134
( e0 (k))
lim
{e0 (k)} = 0 . On the other
= 0 implies directly that lim
T
k
k 1 + (k 1)P(k 1)(k 1)
e 0 (k) )
(
hand, if (k 1) is unbounded then lim

= 0 implies that
T
k 1 + (k 1)P(k 1)(k 1)
% 0 (k 1) 2
2
0
% 0T
lim
% 0 (k 1) = 0 from
2 = 0 since e (k) = (k 1)(k 1) and then lim
k
k
P(k 1)
the fact that P(k) is uniformly bounded. Thus, lim {e 0 (k)} = 0 .

k
(vi) Provided that the external input sequence {c(k)} is sufficiently rich such that (k 1) is
persistently exciting, 0 (k) tends to the true parameters vector as k (Goodwin &
0 (k) from the modification algorithm and,

Sin, 1984). Then, M(k)
tends to M
***
consequently, (k) tends to 0 (k) and e(k) tends to zero as k .
8. Appendix B. Proof of Theorem 1

(i) (k) is bounded since the estimation model parameters a i (k) and b j (k) , and the
controller parameters s j (k) , for i {1, 2, K , n} and j {1, 2, K , n + 1} , are bounded
thanks to the estimated parameters vector (k) and the multirate gains vector g(k)
are
bounded for all integer k 0 . The eigenvalues of (k) are in z < 1 since they are the
roots of A (z) and A (z) , due to the designed control law, and the roots of B(z,k)
m
which are within the unit circle due to the suitable adaptation of the multirate gains.
Besides,
k
k = k 0 + 1
(k ) (k 1) 0 + 1 (k k 0 )
(24)
for all integers k and k 0 such that k > k 0 0 , and some sufficiently small positive real
constants 0 and 1 (Bilbao-Guillerna et al., 2005). Note that (24) is fulfilled with a slow
enough estimation rate via a suitable choice of P(0) in (16) so that 1 is sufficiently
small. Thus, the time-varying homogeneous system
x(k) = (k 1) x(k 1)
is
k 1
exponentially stable and its transition matrix (k,k ) = ( j) satisfies (k,k) 1k0 k
j=k
for all k k where 0 ( 0,1 ) and 1 is a non-dependent constant (Alonso-Quesada &

De la Sen, 2004). It follows from (21) that:
x(k) = (k,k 0 ) x(k 0 ) +
(k,k) ( e(k) + (k))

2
(25)
+ 3 e(k) + 4 (k) )
(26)
k = k 0
135
for all integer k k 0 0 . Then,

x(k) = 1k0 k 0 x(k 0 ) +
k = k 0
k k
0
for some positive real constants 2 , 3 and 4 , provided that the input reference
sequence
{c(k)}
is
lim b j (k) b j (k 1) = 0
k
bounded.
It
follows
that
for all i {1, 2, K , n} and
lim a i (k) a i (k 1) = 0
k
j {1, 2, K , n + 1}
and
from the
convergence property of the estimation algorithm. Then, lim s i (k) s i (k 1) = 0 as it

k
follows from the adaptive control resolution. Consequently, lim (k) = 0 . Besides,
k
lim e(k) = 0 from the estimation algorithm. Then, x(k) is bounded from (26), which
k
implies that sequences {u(k)} and {y(k)} are also bounded.

(ii) On one hand, the adaptive control law ensures that the estimated sequence
}
{y(k)
matches the reference model one {y m (k)} for all integer k 0 . On the other hand, the
estimation algorithm guarantees the asymptotic convergence of the estimation error
e(k) to zero. Then, the output sequence {y(k)} tends to {y m (k)} asymptotically as
k .
(iii) The adaptive control algorithm ensures that there is no finite escapes. Then, the
boundedness of the sequences {u(k)} and {y(k)} implies that the plant input and
output continuous-time signals u(t) and y(t) are bounded for all t.
9. Acknowledgment
The authors are very grateful to MCYT by its partial support through grants DPI 2003-0164
and DPI2006-00714.
10. References
Alonso-Quesada, S. & De la Sen, M. (2004). Robust adaptive control of discrete nominally
stabilizable plants. Appl. Math. Comput., Vol. 150, pp. 555-583.
Alonso-Quesada, S. & De la Sen, M. (2007). A discrete multi-estimation adaptive control
scheme for stabilizing non-inversely stable continuous-time plants using fractional
holds, Proceedings of 46th IEEE Conference on Decision and Control, pp. 1320-1325,
ISBN: 1-4244-1498-9, New Orleans, LA, USA, December 2007, Publisher:
Omnipress.
136
Adaptive Control
Arvanitis, K. G. (1999). An algorithm for adaptive pole placement control of linear systems
based on generalized sampled-data hold functions. J. Franklin Inst., Vol. 336, pp.
503-521.
Astrm, K. J. & Wittenmark, B. (1997). Computer Controlled Systems: Theory and Design,
Prentice-Hall Inc., ISBN: 0-13-736787-2, New Jersey.
Brcena, R., De la Sen, M. and Sagastabeitia, I. (2000). Improving two stability properties of
the zeros of sampled systems with fractional order hold, IEE Proceedings - Control
Theory and Applications, Vol. 147, No. 4, pp. 456-464.
Bilbao-Guillerna, A., De la Sen, M., Ibeas, A. and Alonso-Quesada, S. (2005). Robustly stable
multiestimation scheme for adaptive control and identification with model
reduction issues. Discrete Dynamics in Nature and Society, Vol. 2005, No. 1, pp. 31-67.
Blachuta, M. J. (1999). On approximate pulse transfer functions. IEEE Transactions on
Automatic Control, Vol. 44, No. 11, pp. 2062-2067.
De la Sen, M & Alonso-Quesada, S. (2007). Model matching via multirate sampling with fast
sampled input guaranteeing the stability of the plant zeros. Extensions to adaptive
control. IET Control Theory Appl., Vol. 1, No. 1, pp. 210-225.
Goodwin, G. C. & Sin, K. S. (1984). Adaptive Filtering, Prediction and Control, Prentice-Hall
Inc., ISBN: 0-13-004069-X, New Jersey.
Goodwin, G. C. & Mayne, D. Q. (1987). A parameter estimation perspective of continuoustime model reference adaptive control. Automatica, Vol. 23, No. 1, pp. 57-70.
Ioannou, P. A. & Sun, J. (1996). Robust Adaptive Control, Prentice-Hall Inc., ISBN: 0-13-4391004, New Jersey.
Liang, S., Ishitobi M. & Zhu, Q. (2003). Improvement of stability of zeros in discrete-time
multivariable systems using fractional-order hold, International Journal of Control,
Vol. 76, No. 17, pp. 1699-1711.
Liang, S. & Ishitobi, M. (2004). Properties of zeros of discretised system using multirate
input and hold, IEE Proc. Contr. Theory Appl., Vol. 151, No. 2, pp. 180-184.
Narendra, K. S. & Annaswamy, A. M. (1989). Stable Adaptive Systems, Prentice-Hall Inc.,
ISBN: 0-13-840034-2, New Jersey.
6
Ricardo Ribeiro & Kurios Queiroz
Federal University of Rio Grande do Norte

Brazil
1. Introduction
The purpose of this chapter is to redesign the standard adaptive control schemes by using
hybrid structure composed by Model Reference Adaptive Control (MRAC) or Adaptive Pole
Placement Control (APPC) strategies, associated to Variable Structure (VS) schemes for
achieving non-standard robust adaptive control strategies. The both control strategies is
now on named VS-MRAC and VS-APPC. We start with the theoretical base of standard
control strategies APPC and MRAC, discussing their structures, as how their parameters are
identified by adaptive observers and their robustness properties for guaranteeing their
stability. After that, we introduce the sliding mode control (variable structure) in each
control scheme for simplifying their design procedure. These design procedure are based on
stability analysis of each hybrid robust control scheme. With the definition of both hybrid
control strategies, it is analyzed their behavior when controlling system plants with
unmodeled disturbances and parameter variation. It is established how the adaptive laws
compensates these unmodeled dynamics. Furthermore, by using simple systems examples it
is realized a comparison study between the hybrid structures VS-APPC and VSMRAC and
the standard schemes APPC and MRAC. As the hybrid structures use switching laws due to
the sliding mode scheme, the effect of chattering is analyzed on the implementation and
consequently effects on the digital control hardware where sampling times are limiting
factor. For reducing these drawbacks it is also discussed possibilities which kind of
modifications can employ. Finally, some practical considerations are discussed on an
implementation on motor drive systems.
2. Variable Structure Model Reference Adaptive Controller (VS-MRAC)

The VS-MRAC was originally proposed in (Hsu et al., 1989) and extensively discussed in
(Hsu et al., 1994). The main features of this control scheme are the robustness of parameters
uncertainties and unmodeled disturbances, as well as good transitory response.
Consider the following first order plant
W (s ) =
bp
y
,
=
u
s + ap
(1)
138
Adaptive Control
where bp and a p are unknown or known with limited uncertainties. Admitting a reference
model given by
M (s ) =
ym
r
bm
s + am
(2)
in which km > 0 and am > 0 , the following output error variable can be defined as
e 0 = y ym .
(3)
The control objective is to force y(t ) to asymptotically track the reference output signal,
ym (t ) , by regulating e0 to be zero, while keeping all the closed-loop signals uniformly

bounded. The control law used for accomplished this is
u = 1y + 2r ,
(4)
which is the same as used in traditional model reference adaptive control. However, instead
of the integral adaptive laws for the controller parameters, switching laws are proposed in
order to improve the system transient performance and its robustness.
If bp and a p are known, the ideal controller parameters ( 1* and 2* ) can be founded using
the following condition
y
y
= m ,
r
r
(5)
which means that our control objective is achieved, i.e., the closed-loop system behaves like
the open-loop reference model. Consequently, the control law equation can be rewritten as
u = 1*y + 2*r .
(6)
Analyzing (1) and (2) in the time domain, we get
y = a p y + k p u ,
ym = am ym + km r .
(7)
(8)
Adding and subtracting terms related to the ideal control parameters in (4),
u = 1y + 2r 1*y 2*r + 1*y + 2*r ,
(9)
139
and then grouping some terms
u = (1 1* )y + (2 2* )r + 1*y + 2*r ,
(10)
u = 1y + 2r + 1*y + 2*r ,
(11)
we have
in which terms 1 and 2 are deviations of ideal controller parameters 1 and 2 .

Substituting the resulting equation (11) in (7),
y = a p y + bp (1y + 2r + 1*y + 2*r ) ,
(12)
we can rewrite this equation as
y = a p y + bp 1*y + bp 2*r + bp (1y + 2r ) ,
(13)
y = (a p bp 1* )y + bp 2*r + bp (1y + 2r ) .
(14)
which results in
From (6), the model input r can be defined as
r =
u 1*y
2*
(15)
Therefore, using (11) and (15) in (8), we get
ym = am y + bm r +
bm
(1y + 2r ) .
2*
(16)
Finally, comparing (14) and (16) due to the condition (5), we have the desired controller
parameters
1* =
a p am
bp
(17)
140
Adaptive Control
2* =
bm
bp
(18)
The above desired controller parameters assure that plant output converges to its reference
model, because bp and a p are known. This design criteria is named as The Matching
Conditions.
However, our interests are concerned with unknown plant parameters or with known plant
parameters with uncertainties, which require the use of adaptive laws for adjusting
controller parameters. Derivating the output error equation given in (3),
e0 = y ym
(19)
and using the condition (5), with equations (8), (16) and (19), we get
e0 = am y + bm r +
bm
(1y + 2r ) (am ym + bm r ) ,
2*
(20)
bm
(1y + 2r ) .
2*
(21)
which can be rearranged as
e0 = am (y ym ) +
Thus,
e0 = ame0 +
bm
(1y + 2r ) .
2*
(22)
Now, consider the Lyapunov function candidate given by
1
V (e0 ) = e02 > 0 ,
2
(23)
and its respective first time derivative
V (e0 ) = e0e0 .
By substituting (22) in (24), we obtain the following equation
(24)
b
V (e0 ) = ame0 + m (1y + 2r ) e0 ,
*
2
141
(25)
that can be rewritten as
b
V (e0 ) = ame02 + m ( 1 1* ) e0y + ( 2 2* )e0r .
2*
(26)
Using the switching laws,
1 = 1sgn(e0y ) ,
(27)
2 = 2sgn(e0r ) ,
(28)
b
V (e0 ) = ame02 m ( 1 | e0y | +1*e0y ) + ( 2 | e 0r | +2*e0r ) .
2*
(29)
we obtain,
If the conditions 1 >| 1* | and 2 >| 2* | are satisfied, the terms with indefinite signals
in (29) are dominated, and then
V (e0 ) ame02 < 0
(30)
which guarantees that e0 = 0 is a globally asymptotically stable (GAS) equilibrium point,

because (30) is a negative definite function.
3. Variable Structure Adaptive Pole Placement Control (VS-APPC)

As the VS-MRAC, the VS-APPC is the hybrid control structure obtained from the
association of Pole Placement Control (PPC) together with Variable Structure (VS).
Therefore, the theoretical development of this section starts from PPC control scheme and
then we introduce the VS concepts for achieving the proposed VS-APPC.
Considering the single input/single output (SISO) LTI plant
y = G (s )u ,
in which
(31)
142
Adaptive Control
bn 1s n 1 + ... + b1s + b0
Z (s )
,
=
R(s ) s n + an 1s n 1 + ... + a1s + a 0
G (s ) =
(32)
there are, as plant parameters, 2n elements, which are the coefficients of the numerator and
denominator of transfer function G (s ) . We can define the vector * as
T
* = bn 1 . . . b1 b0 an 1 . . . a1 a 0 .
(33)
From this, the following constraints must be observed:

S1. R(s ) is a monic polynomial whose degree n is known.
S2. Z (s ) , R(s ) are coprime and degree (Z ) < n .
Assumptions (S1) and (S2) allow Z (s ) , R(s ) to be non-Hurwitz in contrast to the MRC
(Model Reference Control) case, where Z (s ) is required to be Hurwitz.
We can also extend the PPC scheme for including the tracking objective, where output y is
required to follow a certain class of reference signals r , by using the internal model
principle (Ioannou & Sun, 1996). The uniformly bounded reference signal is assumed to
satisfy
Qm (s )r = 0 ,
(34)
where Qm (s ) , the internal model of r , is a known monic polynomial of degree q with nonrepeated roots on the j-axis and satisfies
S3. Qm (s ) , Z (s ) are coprime.
Considering the control law given by
Qm(s )L(s )u = P (s )y + M (s )r ,
(35)
where P (s ) , M (s ) , L(s ) are polynomials (with L(s ) monic) of degree q + n 1 ,
q + n 1 e n 1 , respectively, and Qm (s ) satisfies (34) and assumption (S3).

Applying (35) to the plant (31), we obtain the closed-loop plant equation
y =
whose characteristic equation is
Z (s )M (s )
r,
Qm(s )L(s )R(s ) + P (s )Z (s )
(36)
143
Qm(s )L(s )R(s ) + P (s )Z (s ) = 0 ,
(37)
and has order 2n + q 1 . The objective now is chosen P (s ) , L(s ) such that
Qm(s )L(s )R(s ) + P (s )Z (s ) = A* (s )
(38)
is satisfied for a given monic Hurwitz polynomial A* (s ) of degree 2n + q 1 . Because of

assumptions S2 e S3 which guarantee that Qm (s ) , R(s ) , Z (s ) are coprime, there is a
solution so that L(s ) and P (s ) satisfy (38) and this solution is unique (Ioannou & Sun,
1996).
Using (38), the closed-loop is described by
y =
Z (s )M (s )
A* (s )
r.
(39)
Similarly, from the plant (31) and the control law (35) and (38), we obtain
R(s )M (s )
u =
Because r is uniformly bounded and
A* (s )
r.
Z (s )M (s ) R(s )M (s )
A* (s )
A* (s )
(40)
are proper with stable poles,
y and u remain bounded whenever t for any polynomial M (s ) of degree

n + q 1 (Ioannou & Sun, 1996). Therefore, the pole placement objective is achieved by
the control law (35) without having any additional restrictions in M (s ) and Qm (s ) . When
r = 0 , (39) and (40) imply that y and u converge to zero exponentially fast. On the other
hand, when r 0 , the tracking error e = y r is given by
e=
Z (s )M (s ) A* (s )
*
A (s )
r =
Z (s )
*
A (s )
[M (s ) P (s )]r
L(s )R(s )
A* (s )
Qm(s )r .
(41)
In order to obtain zero tracking error, the equation above suggests the choice of
M (s ) = P (s ) to cancel its first term, while the second term can be canceled by using (34).
Therefore, the pole placement and tracking objective are achieved by using the control law
Qm(s )L(s )u = P (s )(y r ) ,
(42)
144
Adaptive Control
which is implemented as shown in Fig. 1 using n + q 1 integrators for the controller

realization. An alternative realization of (42) is obtained by rewriting it as
u =
LQm
P
u (y r ) ,
(43)
where is any monic Hurwitz polynomial of degree n + q 1 .
P(s)
r +
Qm (s)L(s)
G(s)
Fig. 1. Block diagram of pole placement control.

The PPC design supposes that the plant parameters are known, what not always is true or
possible. Therefore, integral adaptive laws can be proposed for estimating these parameters
and then used with PPC schemes. This new strategy is called Adaptive Pole Placement
Controller (APPC), where the certainty equivalence principle guarantees that the output plant
tracks the reference signal r , if the estimates converge to the desired values. In this section,
instead of these traditional adaptive laws, switching laws will be used for the the first order
plant case, according to (Silva et al., 2004).
Consider the plant,
y =
b
u,
s +a
(44)
and its respective time domain equation,
y = ay + bu ,
(45)
where the parameters a and b are unknown or known with uncertainties. Let be am a
positive constant, we may write (45) by adding and subtracting the term am y ,
y = am y + (am a )y + bu .
(46)
A model for the plant may be written as
,
y = am y + (am a)y + bu
(47)
145
where a and b are estimates for a and b , respectively (Ioannou & Sun, 1996).
We define the estimation error e0 as
e0 = y y ,
(48)
,
bu
e0 = ame0 + ay
(49)
a = a a ,
b = b b .
(50)
and with (46) and (47), we get
where
(51)
Choosing the following Lyapunov function candidate,
1
V (e0 ) = e02 > 0 ,
2
(52)
V (e0 ) = e0e0 ,
(53)
u.
0y be
V (e0 ) = ame02 + ae
0
(54)
we have
which can be rewritten using (49),
Expanding the above equation with (50) and (51),
V (e0 ) = ame02 + (a a )e0y (b b)e0u ,
(55)
and then using the switching laws,
we get,
a = a sgn(e0y ) ,
(56)
b = b sgn(e0u ) ,
(57)
146
Adaptive Control
V (e0 ) = ame02 (a e0y + ae0y ) (b e0u be0u ) .
(58)
Finally, if the conditions a > a and b > b are satisfied,
V (e0 ) ame02 < 0 ,
(59)
which guarantees that e0 = 0 is a globally asymptotic stable (GAS) equilibrium point.

Moreover, if we follow a similar procedure described in (Hsu & Costa, 1989), we can prove
that e0 = 0 reaches the sliding surface in a finite time t f ( e0 = 0 , t > t f ).
4. Application on a Current Control Loop of an Induction Machine

To evaluate the performance of both proposed hybrid adaptive schemes, we use an
induction machine voltage x current model as an experimental plant. The voltage equations
of the induction machine on arbitrary reference frame can be presented by the following
equations:

g
g
r rq
+ rd ,
r

g
l ls
rq
g
r rd
+ s
,
lm
r
l ls
g
= rs + s
vsd
r
g
di g
isd + ls sd
dt
l ls
g
g ls isq
s
lm
(60)
l ls
g
vsq
= rs + s
r
g
di g
isq + ls sq
dt
g
+ g ls isd
(61)
g
g
g
g
where vsd
, vsq
, isd
and isq
are dq axis stator voltages and currents in a generic reference
frame, respectively; rs , ls and lm are the stator resistance, stator inductance and mutual
inductance, respectively; g and r are the angular frequencies of the dq generic reference
2
/ lslr and r = lr / rr are the
frame and rotor reference frame, respectively; = 1 lm
leakage factor and rotor time constant, respectively.

The above model can be simplified by choosing the stator reference frame ( g = 0 ).
Therefore, equations (60) and (61) can be rewritten as
s
s
vsd
= rsr isd
+ ls
s
disd
dt
s
+ esd
,
(62)
147
s
s
vsq
= rsr isq
+ ls
s
disq
dt
s
+ esq
,
(63)
where s is the superscript related to the stator reference frame, rsr = rs + (ls ls ) / r ,
s
s
esd
and esq
are fcems of the dq machine phases given by
s (l ls )
s
s
= r rq
+ rd s
esd
,
r lm
(64)
and
s
esq
rq
s
= r rd
(l l )
s
s
,
lm
(65)
The current x voltage transfer function of the induction machine can be obtained from (62)
and (63) as
s
I sd
(s )
Vsds (s )
s
I sd
(s )
Vsds (s )
1 / rsr
s s + 1
(66)
s
s
(s ) and Vsqs ' (s ) = Vsqs (s ) Esq
(s ) . The fcems
where s = ls / rsr , Vsds (s ) = Vsds (s ) Esd
s
s
Esd
(s ) and Esq
(s ) are considered unmodeled disturbances to be compensated by the
control scheme.
Analyzing the current x voltage transfer functions of a standard machine, we can observe
that the time constant s has parameters which vary with the dynamic behavior of
machine. Moreover, this plant has also unmodeled disturbances. This justifies the use of this
control plant for evaluating the performance of proposed control schemes.
5. Control System
Fig. 2 presents the block diagram of a standard vector control strategy, in which the
proposed control schemes are employed for induction motor drive. Block RFO realizes the
s
and
vector rotor field oriented control strategy. It generates the stator reference currents isd
s
isq
, angular stator frequency o of stator reference currents from desired reference torque
Te , and reference rotor flux r , respectively. Blocks VS-ACS implement the proposed
robust adaptive current control schemes that could be the VS-MRAC strategy or the VS-
148
Adaptive Control
APPC strategy. Both current controllers are implemented on the stator reference frame.
Block dq s / 123 transforms the variables from dq s stationary reference frame into 123
stator reference frame.
Generically, the current-voltage transfer function given by equation (66) can be rewritten as
s
Wisdq
(s ) =
s
I sd
(s )
Vsds (s )
s
I sq
(s )
Vsqs (s )
bs
s + as
(67)
s
s
and esq
are considered
in which bs = 1 / ls and as = 1 / s . In this model, the fcems esd
unmodeled disturbances to be compensated by current controllers. The parameters as and
bs are known with uncertainties that can be introduced by machine saturation, temperature
changes or loading variation.
w*o
s
s
f*r
VS-ACS
RFO
IM
VS-ACS
123/dq
Fig. 2. Block diagram of the proposed IM motor drive system.
5.1 VS-MRAC Scheme
Consider that the linear first order plant of induction machine current-voltage transfer
s
given by (67) and a reference model characterized by transfer function
function Wisdq
s
M isdq
(s ) = km
N m (s )
Dm (s )
be
s + ae
(68)
which attends for the stability constraints that is the constant bs in (67) and be should have
positive sign, as mentioned before. The output error can be defined as
s
s
e0s sdq = isdq
imdq
,
(69)
149
s
s
s
where imdq
( imd
and imq
) are the outputs of the reference model. The tracking of the model
s
s
s
s
= imd
or isq
= imq
) is reached if the input of the control plant is defined
control signal ( isd
as
s
s
s
vsdq
= 1dq isdq
+ 2dq isdq
(70)
where 1d ( 1q ) and 2d ( 2q ) are the ideal controller parameters, that can be only
s
determined if Wisdq
(s ) is known. According to section 2, they can be determined as
1d = 1q =
as ae
bs
(71)
and
2d = 2q =
be
bs
(72)
s
Once Wisdq
(s ) is not known, the controllers parameters 1dq (t ) and 2dq (t ) are updated by
using switching laws as

s
idq = idq sgn(e0s sdq yisdq
)
(73)
s
s
s
is the reference currents isdq
or the output currents isdq
, and
where i = [1,2] and ysdq
idq > idq

are upper bounds which are assumed to be known, and the signal-function
sgn is defined as
1 if x > 0
.
sgn(x ) =
1 if x < 0
(74)
). It is
Introducing nominal values of controller parameters idq (nom ) (ideally idq (nom ) = idq
convenient to modify the control plant input given by (70) for the following
150
Adaptive Control
s
vsdq
with T = v1dq
s 1dq
v 2dq
v1dq
is
T
= T sdq + nom
v2dq
s
i
sdq
is
sdq ,
s
i
sdq
(75)
T
s 2dq , nom
= s 1dq (nom ) s 2dq (nom ) and
s
v1dq = v1dq + vsdq
s
v2dq = v2dq + isdq
(76)
in which
s
s 1dq = s 1dq sgn(e0s sdq isdq
) + s 1dq (nom )
s
s 2dq = s 2dq sgn(e0s sdq isdq
) + s 2dq (nom )
(77)
and
v1dq = v1dq sgn(e0s sdq v1dq )

v 2dq = v 2dq sgn(e0s sdq v2dq )
(78)
where s 1dq , s 2dq , v1dq and v 2dq are the controller parameters, s 1dq (nom ) and s 2dq (nom )
are the nominal parameters of the controller, and v1dq and v2dq are the system plant input
and output filtered signals, respectively. The constants s 1dq or s 2dq is chosen by
considering that
s 1dq > s1dq s 1dq (nom )

s 2dq > s2dq s 2dq (nom )
(79)
The input and output filters given by equation (76) are designed as proposed in (Narendra
& Annaswamy, 1989). The filter parameter is chosen such that N m (s ) is a factor of
det(sI ) . Conventionally, these filters are used when the system plant is the second
order or higher. However, it is used in the proposed controller to get two more parameters
for minimizing the tracking error e0s sdq .
151
Fig. 3. Block diagram of proposed VS-MRAC current controller.

The block diagram of the VS-MRAC control algorithm is presented in Fig. 3. The proposed
control scheme is composed by VS for calculating the controller parameters and a MRAC for
determining the system desired performance. The VS is implemented by the block Controller
Calculation, in which Equations (77) and (78) together are employed for determining s 1dq ,
s 2dq , v 1dq and v 2dq . These parameters are used by Controller blocks for generating the
s
control signals vsdq
. To reduce the chattering at the output of controllers, input filters,
represented by blocks Vid (s ) and Viq (s ) are employed. They use filter model represented
s
by Eqs. (76). These filtered voltages feed the IM which generates phase currents isdq
which
are also filtered by filter blocks Vod (s ) and Voq (s ) and then, compared with the reference
s
model output imdq
for generating the output error e0s sdq . The reference models are
implemented by two blocks which implements transfer functions (68). The output of these
s
blocks is interconnected by coupling terms o I mq
s
and o I md
, respectively. This
152
Adaptive Control
s
s
approach used to avoid the phase delay between the input ( I sdq
) and output ( I mdq
) of the
reference model.
5.1.1 Design of the Controller
To design the proposed VS-MRAC controller, initially is necessary to choose a suitable
s
reference model M isdq
(s ) . Based on the parameters of the induction machine used in
present study, given in Table 1, the reference model employed is

s
M isdq
(s ) =
550
,
s + 550
(80)
From this reference model, the nominal values can be determined by using equations (71)
and (72) which results in 1sd (nom ) = 1sq (nom ) = 3.7 and 2sd (nom ) = 2sq (nom ) = 55 .
Considering the restrictions given by (79), the parameters s 1dq and s 2dq , chosen for
achieving a control signal with minimum amplitude are s 1dq = 0.37 and s 2dq = 5.5 . It
is important to highlight that choice criteria determines how fast the system converges to
their references. Moreover, it also determines the level of the chattering verified at the
control system after its convergence. As mentioned before the use of input and output filters
are not required for control plant of fist order. They are used here for smoothing the control
signal. Their parameters was determined experimentally, which results in
= 1 , v1d = v1d = 2.0 and v 2d = v 2q = 0.1 . This solution is not unique and
different adjust can be employed on these filters setup which addresses to different overall
system performance.
5.2 VS-APPC Scheme
The first approach of VS-APPC in (Silva et al., 2004) does not deal with unmodeled
disturbances occurred at the system control loop like machine fems. To overcome this, a
modified VS-APPC is proposed here.
Let us consider the first order IM current-voltage transfer function given by equation (67).
The main objective is to estimate parameters as and bs to generate the inputs vsd and vsq
s
s
so that the machine phase currents isd
and isq
following their respective reference currents
s
s
isd
and isd
and, the closed loop poles are assigned to those of a Hurwitz polynomials
As (s ) given by
A (s ) = s 3 + 2s 2 + 1s + 0 ,
(81)
153
where coefficients 2 , 1 and 0 determine the closed-loop performance requirements.

To estimate the parameters as and bs , the respective switching laws are used
with the restrictions as > as
s
),
as = as sgn(e0s sdq isdq
(82)
s
),
bs = bs sgn(e0s sdq vsdq
(83)
and bs > bs
satisfied, as mentioned before. The pole
placements and the tracking objectives of proposed VS-APPC are achieved, if the following
control law is employed
s
s
s
Qm (s )L(s )Vsdq
(s ) = P (s )(I sdq
I sdq
)
(84)
which addresses to the implementation of the controller transfer function
C sd (s ) = C sq (s ) =
P (s )
.
Qm (s )L(s )
(85)
s
s
The polynomial Qm (s ) is choose to satisfy Qm (s )I sd
(s ) = Qm (s )I sq
(s ) = 0 . For the IM
current-voltage control plant (see equation (67)) and considering that the VS-APPC control
algorithms are implemented on the stator reference frame, which results in sinusoidal
reference currents, a suitable choice for the controller polynomials are Qm (s ) = s 2 + o2
(internal
model
of
sinusoidal
reference
signals
isd
and
),
isq
L(s ) = 1
and
P (s ) = p2s 2 + p1s + p0 , where o is the angular frequency of reference currents. This

choice results in a current controller with the following transfer functions
C sd (s ) = C sq (s ) =
p2s 2 + p1s + p0
s 2 + o2
(86)
where angular frequency o is generated by vector RFO control scheme and coefficients p2 ,
p1 and p0 are determined by solving the Diophantine equation for desired Hurwitz
polynomial As (see equation (81)) as follows
p2 =
2 as
b
s
(87)
154
Adaptive Control
p1 =
1 o2
b
(88)
0 o2as
b
(89)
p0 =
To avoid zero division on the equation (87)-(89), the switching law (83) is modified by
s
bs = bs sgn(e0s sdq vsdq
) + bs (nom )
in
which
bs (nom )
is
the
nominal
values
of
bs
and
(90)
the
stability
restriction
becomes bs > bs bs (nom ) .

s
s
and vsq
generated at the output of the proposed controller VS-APPC
The control signals vsd
can be derived from equation (86) which results in the following state-space model
s
x1ssdq = x 2ssdq + p1sdq
s
x2ssdq = o2x1ssdq + (p0 o2 p2 )sdq
s
s
vsdq
= x1ssdq + p2 sdq
(91)
(92)
(93)
s
s
s
(t ) = isdq
isdq
is the current error that is calculated from the measured
where sdq
quantities issued by data acquisition plug-in board as described next. Therefore, to generate
the output signal of the controllers it is necessary to solve the equations (91)-(93).
155
Fig. 4. Block diagram of proposed VS-APPC current controller.

The block diagram of the VS-APPC control algorithm for the machine current control loop is
presented in Fig. 4. The proposed adaptive control scheme is composed a SMC parameter
estimator and a machine current control loop subsystems. The SMC composed by blocks
system controller and plant model identifies the dynamic of the IM current-voltage model.
s
and isqs .
The output of this system generates the estimative of machine phase currents isd
The control loop subsystem composed by system controller and IM regulates the machine
s
s
s
s
and isq
and compensate the disturbances esd
and esq
. The comparison
phase currents isd
s
s
s
between the estimative currents ( isd
and isqs ) and the machine phase currents ( isd
and isq
)
determines the estimation errors e0s sd and e0s sq . These errors together with machine voltages
s
s
vsd
and vsq
, and VS-APPC algorithm set points as , bs and bs (nom ) are used for calculating
parameter estimative as and bs , from the use of equations (82) and (90). These estimates
update the plant model of the IM and are used by the controller calculation for together
with, the coefficients of the desired polynomial As and angular frequency o , determine
the parameters of the system controller p2 , p1 and p0 . The introduction of the IMP into
the controller modeling avoids the use of stator to synchronous reference frame
transformations. With this approach, the robustness for unmodeled disturbances is
achieved.
5.2.1 Design of the Controller
To design the proposed VS-APPC controller is necessary to choose a suitable polynomial
and to determine the controllers coefficients p2 , p1 , and p0 . A good choice criteria for
156
Adaptive Control
accomplishing the bound system conditions, is to define a polynomial which roots are
closed to the control plant time constants. The characteristics of IM used in this work are
listed in the Table 1. The current-voltage transfer functions for dq phases are given by
s
I sdq
(s )
s
Vsdq
(s )
10
s + 587
(94)
A possible choice for suitable polynomial As (s ) can be
As (s ) = (s + 587)3
(95)
According to Equations (82), (90) and (87)-(89), and based on the desired polynomial (95),
the estimative of the parameters of VS-APPC current controllers can be obtained as
1761 as
b
(96)
1033707 o2
b
(97)
202262003 o2as
b
(98)
p2 =
p1 =
p0 =
To define the coefficients of the switching laws it is necessary to take into account together
the stability restrictions as > as and bs > bs bs (nom ) . Based on the simulation and the
theoretical studies, it can be observed that the magnitude of the respective switching laws
( as and bs ) determine how fast the VS-APPC controllers converge to their respective
references. However, the choice of greater values, results in controllers outputs ( vsd and vsq )
with high amplitudes, which can address to the operation of system with nonlinear
behavior. Thus, a good design criteria is to choose the parameters closed to average values
of control plant coefficients as and bs . Using this design criteria for the IM employed in this
work, the following values are obtained bs (nom ) = 9 , bs = 2 and as = 600 . This solution is
not unique and different design adjusts can be tested for different induction machines. The
performance of these controllers is evaluated by simulation and experimental results as
presented next.
rs = 31.0
rr = 27.2
ls = 0.8042H
lr = 0.7992H
lm = 0.7534H
J = 0.0133kg.m 2
F = 0.0146kg.m
P =2
157
Table 1. IM nominal parameters
6. Experimental Results
The performance of the proposed VS-MRAC and VS-APPC adaptive controllers was
evaluated by experimental results. To realize these tests, an experimental platform
composed by a microcomputer equipped with a specific data acquisition card, a control
board, IM and a three-phase power converter was used. The data of the IM used in this
platform, are listed in Table 1. The command signals of three-phase power converter are
generated by a microcomputer with a sampling time of 100 s . The data acquisition card
employs Hall effect sensors and A/D converters, connected to low-pass filters with cutoff
frequency of fc = 2.5kHz . Figures 5(a) and 5(b) show the experimental results of VSMRAC control scheme. In these figures are present the graphs of the reference model phase
s
s
s
s
and imq
superimposed to the machine phase currents isd
and isq
. In this
currents imd
s
experiment, the reference model currents are settled initially in I mdq
= 0.8A and
fs = 30Hz . At the instant t = 0.15s , each reference model phase currents is changed by
s
I mdq
= 0.2A . In these results it can be observed that the machine phase currents follow the
model reference currents with a good transient response and a current ripple
s
0.05A .
of isdq
Figures 6-7 present the experimental results of VS-APPC control
s
superimposed
scheme. In the Fig. 6(a) are shown the graph of reference phase current isd
s
by its estimation phase current isd
. In this test, similar to the experiment realized to the VSs
= 0.8A and at instant
MRAC, the magnitude of the reference current is settled in I sdq
s
t = 0.15s , it is changed by I sdq
= 0.2A . These results show that the estimation scheme
employed in the VS-APPC estimates the machine phase current with small current ripple.
s
superimposed by its
Figure 6(b) shows the graphs of the reference phase current isd
s
. In this result, it can be verified that the machine
corresponded machine phase current isd
phase current converges to its reference current imposed by RFO vector control strategy.
Similar to the results presented before, Fig. 7(a) presents the experimental results of
s
superimposed by its estimation phase current isqs and Fig. 7(b)
reference phase current isq
s
shows the reference phase current isq
superimposed by its corresponded machine phase
s
. These results show that the VS-APPC also demonstrates a good performance. In
current isq
comparison to the VS-MRAC, the machine phase currents of the VS-APPC present small
current ripple.
158
Adaptive Control
(b)
(a)
Fig. 5. Experimental results of VS-MRAC phase currents
s
imd
s
(a) and imq
(b) superimposed
s
s
(a) and isd
(b), respectively.
to IM phase currents isd
(b)
(a)
Fig. 6. Experimental results of VS-APPC reference phase current
s
isd
superimposed to
s
s
estimation IM phase current isd
(a) and IM phase current isd
(b).
(b)
(a)
Fig. 7. Experimental results of VS-APPC reference phase current
s
(b).
estimation IM phase current isqs (a) and IM phase current isq
s
isq
superimposed to
159
7. References
Hsu, L. & Costa, R. R. (1989). Variable structure model reference adaptive control using only
input and output measurement: part I. International Journal of Control, Vol.49, No. 1,
pp. 399-416.
Hsu, L. (1990). Variable Structure model reference adaptive control using only Input and
output measurements - general case. IEEE Transactions on Automatic Control, Vol.35,
pp. 1238-1243.
Hsu, L.; Arajo, A. D. de & Costa, R. R. (1994). Analysis and design of i/o based variable
structure adaptive control. IEEE Transactions on Automatic Control, Vol. 39, No. 1,
pp. 4-21.
Ioannou, P. A. & Kokotovic, P. V. (1984). Instability analysis and improvement of robustness
of adaptive control. Automatica, Vol. 20, pp. 583-594.
Ioannou, P. A. & Tsakalis, K. S. (1986). Robust direct adaptive controller. IEEE Transactions
on Automatic Control, Vol. AC-31, pp. 1033-1043.
Ioannou, P. A. & Sun, J. (1996). Robust adaptive control. Prentice Hall, New Jersey, USA.
Kazmierkowski, M. P. & Malesani, L. (1998). Current control techniques for three-phase
voltage-source pwm converters: a survey. IEEE Transactions on Industry
Applications, Vol. 45, No 5, pp. 601-703.
Malesani, L.; Mattavelly, P. & Tomasin, P. (1997) . Improved constant-frequency hysteresis
current control of vsi inverters with simple feedforward bandwith prediction. IEEE
Transactions on Industry Applications, Vol. 33, No. 5, pp. 1194-1202.
Naik, S. M., Kumar, P. R. & Ydstie, B. E. (1992). Robust continuous-time adaptive control by
parameter projection. IEEE Transactions on Automatic Control, Vol. AC-37, pp. 182197.
Narendra , K. S. & Valavani, L. S. (1978). Stable adaptive controller design-direct control.
IEEE Transactions on Automatic Control, Vol. AC-23, pp. 570-583.
Narendra , K. S., Lin, Y. H. & Valavani, L. S. (1980). Stable adaptive controller design, part II:
proof of stability. IEEE Transactions on Automatic Control, Vol. AC-25, pp. 440-448.
Narendra, K. & Annaswamy, A. (1989). Stable adaptive control. Englewood Cliffs and NJ:
Prentice-Hall.
P. Vas (1998). Sensorless vector and direct torque control. Oxford University Press.
Plunkett, B. (1979). A current controlled pwm transistor inverter drive. In Proceedings of the
IEEE Industry Applications Society Annual Meeting, pp. 785-792.
Sastry, S. S. & Bodson, N. M. (1989). Adaptive control: stability, convergence and robustness.
Prentice-Hall, New Jersey.
Silva Jr, F. C; Arajo, A. D. de & Oliveira, J. B. (2004). A proposal for a variable structure
adaptive pole placement control. In Proceedings of the IEEE Workshop on Variable
Structure Systems, 2004.
Utkin, V. I. (1977). Variable structure system with sliding mode. IEEE Transactions on
Automatic Control, Vol.AC- 22, pp. 212-222.
Utkin, V. I. (1978). Sliding modes and their application in variable structure systems. Mir
Publishers, Moscow.
Utkin, V. I. (1992). Sliding Modes in Control and Optimization. Spring-Verlag, New York.
Wishart, M. T. & Harley, R. G. (1995). Identication and control of induction machine using
articial neural networks. IEEE Transactions on Industry Applications, Vol. 31, No. 3,
pp. 612-619.
160
Adaptive Control
Zhang, C. J. & Dunnigan, M. W. (2003). Robust adaptive stator current control for an
induction machine. In Proceedings of the IEEE Conference Control Applications, pp.
779-784.
7
Adaptive Control for Systems with Randomly
Missing Measurements in a Network
Environment
1Department
Yang Shi1 and Huazhen Fang1
of Mechanical Engineering, University of Saskatchewan

Canada
1. Introduction
Networked control systems (NCSs) are a type of distributed control systems, where the
information of control system components (reference input, plant output, control input, etc.)
is exchanged via communication networks. Due to the introduction of networks, NCSs have
many attractive advantages, such as reduced system wiring, low weight and space, ease of
system diagnosis and maintenance, and increased system agility, which motivated the
research in NCSs. The study of NCSs has been an active research area in the past several
years, see some recent survey articles (Chow & Tipsuwan, 2001; Hespanha & Naghshtabrizi,
2007; Yang, 2006) and the references therein. On the other hand, the introduction of
networks also presents some challenges such as the limited feedback information caused by
packet transmission delays and packet loss; both of them are due to the sharing and
competition of the transmission medium, and bring difficulties for analysis and design for
NCSs. The information transmission delay arises from by the limited capacity of the
communication network used in a control system, whereas the packet loss is caused by the
unavoidable data losses or transmission errors. Both the information transmission delay and
packet loss may result in randomly missing output measurements at the controller node, as
shown in Fig. 1. So far different approaches have been used to characterize the limited
feedback information. For example, the information transmission delay and packet losses
have been modeled as Markov chains (Zhang et al., 2006). The binary Bernoulli distribution
is used to model the packet losses in (Sinopoli et al., 2004; Wang et al., 2005 a & 2005 b).
The main challenge of NCS design is the limited feedback information (information
transmission delays and packet losses), which can degrade the performance of systems or
even cause instability. Various methodologies have been proposed for modeling, stability
analysis, and controller design for NCSs in the presence of limited feedback information. A
novel feedback stabilization solution of multiple coupled control systems with limited
communication is proposed by bringing together communication and control theoretical
issues in (Hristu & Morgansen, 1999). Further the control and communication codesign
methodology is applied in (Hristu-Varsakelis, 2006; Zhang & Hristu-Varsakelis, 2006) a
method of stabilizing linear NCSs with medium access constraints and transmission delays
by designing a delay-compensated feedback controller and an accompanying medium
162
Adaptive Control
access policy is presented. In (Zhang et al., 2001), the relationship of sampling time and
maximum allowable transfer interval to keep the systems stable is analyzed by using a
stability region plot; the stability analysis of NCSs is addressed by using a hybrid system
stability analysis technique. In (Walsh et al., 2002), a new NCS protocol, try-once-discard
(TOD), which employs dynamic scheduling method, is proposed and the analytic proof of
global exponential stability is provided based on Lyapunovs second method. In (AzimiSadjadi, 2003), the conditions under which NCSs subject to dropped packets are mean
square stable are provided. Output feedback controller that can stabilize the plant in the
presence of delay, sampling, and dropout effects in the measurement and actuation
channels is developed in (Naghshtabrizi & Hespanha, 2005). In (Yu et al., 2004), the authors
model the NCSs with packet dropout and delays as ordinary linear systems with input
delays and further design state feedback controllers using Lyapunov-Razumikhin function
method for the continuous-time case, and Lyapunov-Krasovskii based method for the
discrete-time case, respectively. In (Yue et al., 2004), the time delays and packet dropout are
simultaneously considered for state feedback controller design based on a delay-dependent
approach; the maximum allowable value of the network-induced delays can be determined
by solving a set of linear matrix inequalities (LMIs). Most recently, Gao, et al., for the first
time, incorporate simultaneously three types of communication limitation, e.g.,
measurement quantization, signal transmission delay, and data packet dropout into the
NCS design for robust H state estimation (Gao & Chen, 2007), and passivity based
controller design (Gao et al., 2007), respectively. Further, a new delay system approach that
consists of multiple successive delay components in the state, is proposed and applied to
network-based control in (Gao et al., 2008).
However, the results obtained for NCSs are still limited: Most of the aforementioned results
assume that the plant is given and model parameters are available, while few papers
address the analysis and synthesis problems for NCSs whose plant parameters are
unknown. In fact, while controlling a real plant, the designer rarely knows its parameters
accurately (Narendra & Annaswamy, 1989). To the best of our knowledge, adaptive control
for systems with unknown parameters and randomly missing outputs in a network
environment has not been fully investigated, which is the focus of this paper.
Fig. 1. An NCS with randomly missing outputs.

It is worth noting that systems with regular missing outputs a special case of those with
randomly missing outputs can also be viewed as multirate systems which have uniform
Adaptive Control for Systems with Randomly Missing Measurements in a Network Environment
163
but various input/output sampling rates (Chen & Francis, 1995). Such systems may have
regular-output-missing feature. In (Ding & Chen, 2004a), Ding, et al. use an auxiliary model
and a modified recursive least squares (RLS) algorithm to realize simultaneous parameter
and output estimation of dual-rate systems. Further, a least squares based self-tuning
control scheme is studied for dual-rate linear systems (Ding & Chen, 2004b) and nonlinear
systems (Ding et al., 2006), respectively. However, network-induced limited feedback
information unavoidably results in randomly missing output measurements. To generalize
and extend the adaptive control approach for multirate systems (Ding & Chen, 2004b; Ding
et al., 2006) to NCSs with randomly missing output measurements and unknown model
parameters is another motivation of this work.
In this paper, we first model the availability of output as a Bernoulli process. Then we
design an output estimator to online estimate the missing output measurements, and further
propose a novel Kalman filter based method for parameter estimation with randomly
output missing. Based on the estimated output or the available output, and the estimated
model parameters, an adaptive control is proposed to make the output track the desired
signal. Convergence of the proposed output estimation and adaptive control algorithms is
analyzed.
The rest of this paper is organized as follows. The problem of adaptive control for NCSs
with unknown model parameters and randomly missing outputs is formulated in Section 2.
In Section 3, the proposed algorithms for output estimation, model parameter estimation,
and adaptive control are presented. In Section 4, the convergence properties of the proposed
algorithms are analyzed. Section 5 gives several illustrative examples to demonstrate the
effectiveness of the proposed algorithms. Finally, concluding remarks are given in Section 6.
Notations: The notations used throughout the paper are fairly standard. E denotes the
expectation. The superscript T stands for matrix transposition; max/min ( X ) represents the
Maximum/minimum eigenvalue of X ; |X|= det( X ) is the determinant of a square matrix
X; X
= tr ( XX
T
+
+
) stands forthe trace of XX . If 0 R and k0 Z , | f ( k )| 0 g( k )
for k k0 , then f ( k ) = O ( g( k )) ; if f ( k ) / g( k ) 0 for k , then f ( k ) = o ( g( k )) .
2. Problem Formulation
The problem of interest in this work is to design an adaptive control scheme for networked
systems with unknown model parameters and randomly missing outputs. In Fig. 2, the
output measurements y k could be unavailable at the controller node at some time instants
because of the network-induced limited feedback information, e.g., transmission delay
and/or packet loss. The data transmission protocols like TCP guarantee the delivery of data
packets in this way: When one or more packets are lost the transmitter retransmits the lost
packets. However, since a retransmitted packet usually has a long delay that is not desirable
for control systems, the retransmitted packets are outdated by the time they arrive at the
controller (Azimi-Sadjadi, 2003; Hristu-Varsakelis & Levine, 2005). Therefore, in this paper,
it is assumed that the output measurements that are delayed in transmission are regarded as
missed ones.
The availability of y k can be viewed as a random variable k . k is assumed to have Bernoulli
distribution:
Adaptive Control
164
E ( k s ) = E k E s for k s ,
(1)
k , if k = 1,
Prob( k ) =
1
k , else if k = 0,
where 0 < k 1 .
Consider a single-input-single-output (SISO) process (Fig. 2):
A z x k = Bzuk , y k = x k + vk
(2)
where uk is the system input, y k the output and vk the disturbing white noise with variance
rv . Az and Bz are two backshift polynomials defined as
Az = 1 + a1 z 1 + a22 + L + ana z na ,
Bz = b0 + b1 z 1 + b2 z 2 + L + bnb z nb .
The polynomial orders na and nb are assumed to be given. Eqn. (2) can be written
equivalently as the following linear regression model:
(3)
y k = 0Tk + vk ,
where
T
0 k = xk 1 x k 2 L x k n uk uk 1 L uk n ,
a
= a1 a2 L an b0 b1 L bn .
a
Vector 0 k represents systems excitation and response information necessary for parameter
estimation, while vector contains model parameters to be estimated.
Fig. 2. Output-error (OE) model structure.

For a system with the output-error (OE) model placed in a networked environment subject
to randomly missing outputs, the objectives of this paper are:
1. Design an output estimator to online estimate the missing output measurements.
2. Develop a recursive Kalman filter based identification algorithm to estimate unknown
model parameters.
165
3. Propose an adaptive tracking controller to make the system output track a given
desired signal.
4. Analyze the convergence properties of the proposed algorithms.
3. Parameter Estimation, Output Estimation, and Adaptive Control Design

There are two main challenges of the adaptive control design for a networked system as
depicted in Fig. 1: (1) randomly missing output measurements; (2) unknown system model
parameters. Therefore, in this section, we first propose algorithms for missing output
estimation and unknown model parameter estimation, and then design the adaptive control
scheme.
3.1 Parameter estimation and missing output estimation
Consider the model in (3). It is shown by (Cao & Schwartz, 2003) and (Guo, 1990) that the
corresponding Kalman filter can be conveniently used for parameter estimation. In
combination with an auxiliary model, the Kalman filter based parameter estimation
algorithm for an OE model is given by
k = k 1 + K a , k ( y k aT, kk 1 ),
Ka,k =
Pa , k 1 a , k
rv + aT, k Pa , k 1 a , k
Pa , k = Pa , k 1
(4)
(5)
(6)
Pa , k 1 a , k aT, k Pa , k 1
,
rv + aT, k Pa , k 10 k
x a , k = aT, kk ,
(7)
T
a , k = x a , k 1 x a , k 2 L x a , k n uk uk 1 L uk n ,
a
(8)
where k represents the estimated parameter vector at time instant k .

It is worth to note that the above algorithm as shown in (4)-(8) cannot be directly applied to
the parameter estimation of systems with randomly missing outputs in a network
environment, as y k in (4) may not be available. This motivates us to develop a new
algorithm that can simultaneously online estimate the unavailable missing output and
estimate system parameters under the network environment. The proposed algorithm
consists of two steps.
Step 1: Output estimation
Albertos, et al. propose a simple algorithm that uses the input-output model, replacing the
unknown past values by estimates when necessary (Albertos et al., 2006). Inspired by this
work, we design the following output estimator:
Adaptive Control
166
zk = k y k + (1 k )y k ,
(9)
with
y k = kTk 1 .
In (9), k is a Bernoulli random variable used to characterize the availability of y k at time
instant k at the controller node, as defined in (1). With the time-stamp technique, the
controller node can detect the availability of the output measurements, and thus, the values
of k (either 1 or 0) are known. The knowledge of their corresponding probability k is not
used in the designed estimator. The structure of the designed output estimator is intuitive
and simple yet very effective, which will be seen soon from the simulation examples.
Step 2: Model parameter estimation
Replacing y k in the algorithm (4)-(8) by zk , defining a new k , and considering the random
variable k , we readily obtain the following algorithm:
k = k 1 + K k ( zk kTk 1 ),
Kk =
(10)
Pk 1 k
,
rv + kT Pk 1 k
Pk = Pk 1 k
(11)
(12)
Pk 1 k kT Pk 1
,
rv + kT Pk 1 k
xb , k = kTk ,
(13)
T
k = xb , k 1 xb , k 2 L xb , k n uk uk 1 L uk n .
a
(14)
Remark 3.1. Consider two extreme cases. If the availability sequence { 1 ,L , k } constantly assumes
1, then no output measurement is lost, and the algorithm above will reduce to the algorithm (4)-(6).
On the other hand, if the availability sequence k constantly takes 0, then all output measurements
are lost, and the parameter estimates just keep the initial values.
3.2 Adaptive control design
Consider the tracking problem. Let y r , k be a desired output signal, and define the output
tracking error
k := y k y r , k .
167
If the control law uk is appropriately designed such that y r , k = 0Tk , then the average
tracking error zk approaches zero finally. Replacing by k 1 and 0 k by k yields
yr , k
na
nb
i =1
i =0
kTk 1 = i , k 1x k i + n
+ i + 1, k 1 k i
= a 1, k 1 xb , k 1 L a na , k 1 xb , k na + b0 , k 1uk + L + bnb , k 1uk nb .
Therefore, the control law can be designed as

uk =
na
nb
1
y r , k + a i , k 1 x k i bi , k 1uk i .
b
i =1
i =1
0 ,k 1
(15)
The proposed adaptive control scheme consists of the missing output estimator [Equation
(9)], model parameter estimator [Equations (10-14)], and the adaptive control law [Equation
(15)]. The overall control diagram is shown in Fig. 3.
4. Convergence Analysis
This section focuses on the analysis of some convergence properties. Some preliminaries are
first summarized to facilitate the following convergence analysis of parameter estimation in
(10)-(12) and of output estimation in (9). Inspired by the work in (Chen & Guo, 1991; Ding &
Chen, 2004a; Ding et al., 2006), the convergence analysis is carried out under the stochastic
framework.
Fig. 3. Adaptive control diagram.

4.1 Preliminaries
To facilitate the convergence analysis, directly applying the matrix inversion formula (Horn
Adaptive Control
168
& Johnson, 1991)

( A + BCD)1 = A1 A1 B(C 1 + DA1 B)1 DA1 ,
the proposed parameter estimation algorithm in Section 3.1 [(10)-(12] can be equivalently
rewritten as:
k = k 1 + rv1 Pk k ( zk kTk 1 ),
(16)
Pk1 = Pk11 + rv1 k k kT .
(17)
Suppose that Pk is initialized by p0 I , where p0 is a positive real value large enough, and
define rk = tr(Pk1 ) . The relation between rk and |Pk1 | can be established in the following
lemma.
Lemma 4.1. The following relation holds:
ln E|Pk1 |= O ( ln Erk ) .
(18)
Proof: Using the formulae

n
i =1
i =1
tr( X ) = i ( X ) and|X|= i ( X ),
where n is the dimension of X , we have
E|Pk1 | (Erk )n .
This completes the proof.
The next lemma shows the convergence of two infinite series that will be useful later.
Lemma 4.2. The following inequalities hold:
t
r E (
i =1
1
i v
Pi ) ln E|Pk1 |+n0 ln p0 a.s.,
T
i i
irv1
i =1
E (iT Pii )
( ln E|P |)
1
< a.s.,
(19)
(20)
where c > 1 .
Proof: The proof can be done along the similar way as Lemma 2 in (Ding & Chen, 2004b)
and is omitted here.
The following is the well-known martingale convergence theorem that lays the foundation
for the convergence analysis of the proposed algorithms.
169
Theorem 4.1. (Goodwin & Sin, 1984) Let {X k } be a sequence of nonnegative random variables
adapted to an increasing -algebras { F k } . If
E ( X k + 1 |F k ) (1 + k )X k k + k , a.s.,
where k 0 , k 0 , EX0 < , i = 0 i
<
and
i=0
i < almost surely (a.s.), then X k
converges a.s. to a finite random variable and

N
lim i < , a.s.
i =0
4.2 Convergence analysis

To carry out the convergence analysis of the proposed algorithms, it is essential to
appropriately construct a martingale process satisfying the conditions of Theorem 4.1. Main
results on the convergence properties of the proposed algorithm are summarized in the
following Theorem.
Theorem 4.2. For the system considered in (3), assume that
(A1) { vk , F k } is a martingale difference sequence satisfying
(A2)
E ( vk |F k 1 ) = 0, a.s.,
(21)
E ( vk2 |F k 1 ) = rv < , a.s.;
(22)
1 1
is strictly positive real;
Az 2
(A3) Bz is stable; i.e., zeros o f Bz are inside the closed unit disk.
Suppose the desired output signal is bounded:| yr , k |< . Applying the missing output estimator
[Equation (9)], model parameter estimator [Equations (10-14)], and the adaptive control law
[Equation (15)], then the output tracking error has the property of minimum variance, i.e.,
1 k
( y r , i y i + vi )2 = 0, a.s.;
k k
i =1
(1) lim
1 k
i E {( zi yr , i )2 |F i 1 } = rv < , a.s.
k k
i =1
(2) lim
Proof: As pointed out in (Goodwin & Sin, 1984; Chen & Guo, 1991), from (A2) it follows that
1 k 2
c k
ui O(1) + O 1 y i2 , a.s.
k i =1
k i =1
(23)
Adaptive Control
170
Here, c1 is a positive constant. Define the following vectors:

ek = zk Tk 1 ,
k = y k xb , k ,
k = k k ,
k = y r , k y k + vk ,
k = k k .
From (2), (3), (16) and (16), it follows that
k = k ( xk xb , k + vk ),
(24)
k = (1 + rv1 kT Pk 1 k )1 ek ,
(25)
ek = k + k vk .
(26)
Also define the parameter estimation error vector and a Lyapunov-like function as
%k = k ,
Vk = %kT Pk1%k .
From (9), (16) and (25), we obtain
%k =%k 1 + rv1 Pk k ek = %k 1 + rv1 Pk 1 k k .

With (17) and (27), Vk can be further evaluated as
Vk = Vk 1 + rv1 k ( kT%k )2 + 2 rv1 kT%k k rv2 kT Pk k (1 rv1 kT Pk k )ek2 .
Let us define
u% k = kT%k ,
1
y% k = kT%k + ( k vk ).
2
(27)
171
Then we have
Vk
Vk 1 2 rv1 k u% k y% k + 2 rv1 k kT%k v k rv2 kT Pk k (1 rv1 kT Pk k )e k2

= Vk 1 2 rv1 k u% k y% k + 2 rv1 k kT%k 1 v k + 2 rv2 kT Pk k ( e k k v k )T v k + k v k2
rv2 kT Pk k (1 rv1 kT Pk k ) k2 + 2 rv2 kT Pk k (1 rv1 kT Pk k ) k v k
rv2 kT Pk k (1 rv1 kT Pk k ) v k2
Vk 1 2 rv1 k u% k y% k + 2 rv1 k kT%k 1 v k + 2 rv2 kT Pk k ( e k k v k )T v k + k v k2
rv2 kT Pk k (1 rv1 kT Pk k ) k2 + 2 rv2 kT Pk k (1 rv1 kT Pk k ) k v k .
=
(28)
Note that kT%k 1 , ek k vk , kT Pk k and k are uncorrelated with vk and F k 1 -measurable. Thus
taking the conditional expectation of both sides of (28) with respect to F k 1 gives
E (Vk |F k 1 ) Vk 1 2 rv1 k E ( u% k y% k ) + 2 rv1 k E ( kT Pk k )
rv2 k E kT Pk k (1 rv1 kT Pk k ) k2 .
(29)
Consider that
A z ( k v k )
Az ( y k xb ,k )
Bz u k A z x b , k
kT%k = u% k .
Therefore, we have
1 1
u% k .
y% k =
Az 2
1 1
is positive real, which indicates
In (A2), it is assumed that
Az 2
k
Sk := 2rv1 k u% k y% k 0, a.s.
i =1
(30)
Adding Sk to both sides of (29) yields

E (Vk + Sk | F k 1 )
Vk 1 + Sk 1 + 2 rv1 k E ( kT Pk k )
rv2 k E kT Pk k (1 rv1 kT Pk k ) k2 .
Define a new sequence:
(31)
Adaptive Control
172
Vk + Sk
Wk =
( ln E|P |)
1
k
(32)
, c > 1.
Since ln E|Pk1 | is nondecreasing and kT Pk k = o(1) , there exists a k0 such that if k k0 we

have
Vk 1 + Sk 1
E ( Wk |F k 1 )
( ln E|P |)
1
k
2rv1 k E ( kT Pk k )
( ln E|P |)
1
k
rv2 k E kT Pk k (1 rv1 kT Pk k ) k2
( ln E|P |)
2 r E ( P )
W +
( ln E|P |)
r E ( 1 r P )
.
( ln E|P |)
c
1
k
1
v
T
k
k 1
2
v
1
v
T
k
(33)
1
k
2
k
1
k
From (12) we have

E ( 1 rv1 kT Pk k ) > 0.
Also note that by Lemma 4.2 the summation of the third term in (33) from 0 to is finite.
Therefore, Theorem 4.1 is applicable, and it gives
rv2 k E ( 1 rv1 kT Pk k ) k2
k =1
( ln E|P |)
1
k
< a.s.
(34)
Further, Lemma 4.1 indicates
rv2 k E ( 1 rv1 kT Pk k ) k2
k =1
( ln Erk )
< a.s.
(35)
As 1 rv1E ( kT Pk k ) is positive and nondecreasing, it holds that 1 = O 1 rv1E ( kT Pk k ) .

Hence,
i =1
i2
( ln Eri )
< a.s.
Since lim k ln Erk = , then from the Kronecker lemma [15] it follows that
(36)
173
lim k = 0, a.s.,
k
where
k :=
1
(ln Erk )c
i =1
2
i
With
1
rk =
k
n
+ r iiTi
p0 i = 1 v
and (23), we obtain

1 k 2
i
k i =1
k
O (ln Erk )c
k
k
O [ E( rk )]
k
k
k
k n
O + na i E 2 ( zi ) + nb ui2
k p0
i =1
i =1
=
=
=
(37)
1 k
kO y i2
k
i =1
By (22) we have
1 k 2
1 k
yi = O(1) + O i2 .
k i =1
k i =1
Substituting (37) into (38) gives
1 k 2
yi = O(1), a.s.,
k i =1
which implies together with (37) that
1 k 2
i = 0, a.s.,
k k
i =1
lim
or equivalently
(38)
Adaptive Control
174
lim
k
1 k
( yr ,i yi + vi )2 = 0, a.s.
k i =1
(39)
Since
E {( y r , k y k + vk )2 |F k 1 } = E[( yr , k y k )2 + 2 yr , k vk 2 y k vk + vk2 |F k 1 ]
= E[( y r , k y k )2 |F k 1 ] + 0 2 rv + rv
= E[( y r , k y k )2 |F k 1 ] rv , a.s.,
and k zk = k y k , we have
1 k
1 k
i E {( zi yr , i )2 |F i 1 } = lim i E {( yi yr , i )2 |F i 1 } = rv , a.s.
k k
k k
i =1
i =1
lim
This completes the proof.
5. Illustrative Examples
In this section, we give three examples to illustrate the adaptive control design scheme
proposed in the previous sections.
The OE model shown in Fig. 2 in the simulation is chosen as
yk =
b0 + b1 z 1 + b2 z 2
uk + v k ,
1 + a1 z 1 + a2 z 2
which is assumed to be placed in a network environment (Fig. 1) with randomly missing

output measurements and unknown model parameters. { vk } is a Gaussian white noise
sequence with zero mean and variance rv = 0.052 . The parameter vector = [ a1 a2 b0 b1 b2 ]T is
to be estimated. Here, true values of are
= [ 0.3 0.6 0.5 0.2 0.34]T .

For simulation purposes, we assume that: (1) is unknown and initialized by ones; (2) the
output measurement { y k } is subject to randomly missing when transmitted to the controller
node; (3) the availability of the output measurements ( y k ) at the controller node is
characterized by the probability k ; (4) The desired output signal to be tracked is a square
wave alternating between -1 and 1 with a period of 1000. Mathematically, it is given by
y r ,( 500 i + j ) = ( 1)i + 1 , i = 0,1, 2,L , j = 1, 2,...500.
In the following simulation studies, we carry out experiments for three different scenarios
175
regarding the availability of the output measurements at the controller node and the
parameter variation, and examine the control performance, respectively. According to the
proposed adaptive control scheme shown in Fig. 3, we apply the algorithms of the missing
output estimator, model parameter estimator, and the adaptive control law to the
networked control system.
Example 1: k = 0.85 . In the first example, 85% of all the measurements are available at the
controller node after network transmission from the sensor to the controller. The output
response is shown in Fig. 4, from which it is observed that the output tracking performance
is satisfactory. In order to take a closer observation on the model parameter estimation and
output estimation, we define the relative parameter estimation error as
It is shown in Fig. 5 (solid blue curve) that par % is becoming smaller with k increasing.
Comparison between the estimated outputs and true outputs during the time range
501 t 550 is illustrated in Fig. 6: The dashed lines are corresponding to the time instants
when data missing occurs, and the small circles on the top of the dashed lines represent the
estimated outputs at these time instants. From Fig. 6 it can be found that the missing output
estimation also exhibits good performance.
Fig. 4. Example 1: Output response when k = 0.85 .
Fig. 5. Comparison of relative Parameter estimation errors for Example 1 and Example 2:
176
Adaptive Control
Blue solid line for Example 1; red dotted line for Example 2.
Example 2: k = 0.65 . In the second example, a worse case subject to more severe randomly
missing outputs is examined: Only 65% of all the measurements are available at the
controller node. The output response is shown in Fig. 7. Even though the available output
measurements are more scarce than those in Example 1, it is still observed that the output is
tracking the desired signal with satisfactory performance. The relative parameter estimation
error, par % , is shown in Fig. 5 (dashed red curve). Clearly, it is decreasing when k is
increasing. The estimated outputs and the true outputs are illustrated in Fig. 8, from which
we can see good output estimation performance.
For the comparison purpose, the relative parameter estimation errors of these two examples
are shown in Figure 5. We can see that the parameter estimation performance when
k = 0.85 is better than that when k = 0.65 .It is no doubt that the estimation performance
largely depends on data completeness that is characterized by k .
Fig. 6. Example 1: Comparison between estimated and true outputs when k = 0.85 (The
dashed line represents output missing).
Fig. 7. Example 2: Output response when k = 0.65 .
177
Example 3: Output tracking performance subject to parameter variation. In practice, the model
parameters may vary during the course of operation due to the change of load, external
disturbance, noise, and so on. Hence, it is also paramount to explore the robustness of the
designed controller against he influence of parameter variation. In this example, we assume
that at k = 2500 , model parameters are all increased by 50%. The output response is shown
in Fig. 9. It can be seen that: At k = 2500 , the output response has a big overshoot because of
the parameter variation; however, the adaptive control scheme quickly forces the system
output to track the desired signal again.
Observing Figs. 4, 7, and 9 in three examples, we notice that the tracking error and
oscillation still exist. This is mainly due to (1) the missing output measurements, and, (2) the
relatively high noise-signal ratio (around 25%). On the other hand, it is desirable to develop
new control schemes to further improve the control performance for networked systems
subject to limited feedback information, which is worth to do extensive research.
Fig. 8. Example 2: Comparison between estimated and true outputs when k = 0.65 (The
dashed line represents output missing).
Fig. 9. Example 3: Output response subject to parameter variation: At time instant k = 2500,
all parameters are increased by 50%.
178
Adaptive Control
6. Conclusion
This paper has investigated the problem of adaptive control for systems with SISO OE
models placed in a network environment subject to unknown model parameters and
randomly missing output measurements. The missing output estimator, Kalman filter based
model parameter estimator, and adaptive controller have been designed to achieve output
tracking. Convergence performance of the proposed algorithms is analyzed under the
stochastic framework. Simulation examples verify the proposed methods. It is worth
mentioning that the proposed scheme is developed for SISO systems in this work, and the
extension to multi-input-multi-output (MIMO) systems is a subject worth further
researching
7. References
Albertos, P.; Sanchis, R. & Sala A. (1999). Output prediction under scarce data operation:
control applications. Automatica, Vol. 35, No. 10, page numbers (16711681), ISSN
0005-1098.
Azimi-Sadjadi, B. (2003). Stability of networked control systems in the presence of packet
losses. Proceedings of 42nd IEEE Conf. Decision & Control, pp. 676681, Hyatt Regency
Maui, December 2003.
Branicky, M. S.; Liberatore, V. & Phillips, S. M. (2003). Networked control system cosimulation for co-design, Proceedings of Amer. Control Conf., pp. 33413346, Denver,
June 2003.
Cao, L. & Schwartz, H. M. (2003). Exponential convergence of the Kalman filter based
parameter estimation algorithm. Int. J. Adapt. Contr. & Signal Process., Vol. 17, No.
10, page numbers (763784), ISSN 1099-1115.
Chen, H. F. & Guo, L. (1991). Identification and Stochastic Adaptive Control, Birkhauser, ISBN
0-8176-3597-1, Boston.
Chen J. M. & Chen B. S. (2000). System parameter identification with input/output noisy
data and missing measurements. IEEE Trans. Signal Processing, Vol. 48, No. (6), page
numbers (15481558), ISSN 1053-587X.
Chen, T. & Francis, B. (1995). Optimal Sampled-data Control Systems, Springer, ISBN 3-54019949-7, London.
Chow, M. Y. & Tipsuwan, Y. (2001). Network based control systems: a tutorial, Proceedings of
IECON01: 27th Ann. Conf. IEEE Industrial Electronics Soc., pp. 15931602, Denver,
Nov 2001.
Ding, F. & Chen, T. (2004a). Combined parameter and output estimation of dual-rate
systems using an auxiliary model. Automatica,Vol. 40, No. 10, page numbers (1739
1748), ISSN 0005-1098.
Ding, F. & Chen, T. (2004b). Identification of dual-rate systems based on finite impulse
response models. Int. J. Adapt. Contr. & Signal Process., Vol. 18. No. 7, page numbers
(589598), ISSN 1099-1115.
Ding, F.; Chen, T. & Iwai, Z. (2006). Adaptive digital control of Hammerstein nonlinear
systems with limited output sampling. SIAM J. Control Optim., Vol. 45, No. 6, page
numbers (22572276), ISSN 1095-7138.
179
Gao, H. & Chen, T. (2007). H estimation for uncertain systems with limited communication
capcaity. IEEE Trans. Automat. Contr., Vol. 52, No. 11, page numbers (20702084),
ISSN 0018-9286.
Gao, H.; Chen, T. & Chai, TY. (2007). Passivity and passification for networked control
systems. SIAM J. Control Optim., Vol. 46, No. 4, page numbers (12991322), ISSN
1095-7138.
Gao, H.; Chen, T. & Lam, J. (2008). A new delay system approach to network-based control.
Automatica, Vol. 44, No. 1, page numbers (3952), ISSN 0005-1098.
Goodwin, G. C. & Sin, K. S. (1984) Adaptive Filtering, Prediction and Control, Prentice Hall,
ISBN-13: 978-0130040695, Englewood Cliffs.
Guo, L. (1990). Estimating time-varying parameters by the Kalman filter based algorithm:
stability and convergence. IEEE Trans. Automat. Contr., Vol. 35, No. 2, page
numbers (141147), ISSN 0018-9286.
Hespanha, J. P.; Naghshtabrizi, P. & Xu, P. (2007). A survery of recent results in networked
control systems. Proc. of the IEEE, Vol. 95, No. 1, page numbers (138162), ISSN
0018- 9219.
Horn, R. A. & Johnson, C. R. (1991). Topics in Matrix Analysis, Cambridge University Press,
ISBN-13: 9780521467131, Cambridge.
Hristu, D. & Morgansen, K. (1999). Limited communication control. Systems & Control
Letters, Vol. 37, No. 4, page numbers (193205), ISSN 0167-6911.
Hristu-Varsakelis, D. & Levine, W. S. (2005). Handbook of Networked and Embedded Control
Systems, Birkhauser, ISBN: 978-0-8176-3239-7.
Hristu-Varsakelis, D. (2006). Stabilization of networked control systems with access
constraints and delays. Proceedings of 45th IEEE Conf. Decision & Control, pp. 1123
1128, San Diego, December 2006.
Isaksson, A. J. (1993). Identification of ARX-models subject to missing data. IEEE Trans.
Automat. Contr, Vol. 38, No. 5, page numbers (813819), ISSN 0018-9286.
Kao, C. Y. & Lincoln, B. (2004). Simple stability criteria for systems with time-varying
delays, Automatica, Vol. 40, No. 8, page numbers (14291434), ISSN 0005-1098.
Lian, F. L.; Moyne, J. & Tilbury, D. (2002). Network design consideration for distributed
control systems. IEEE Trans. Contr. Syst. Technol, Vol. 10, No. 2, page numbers (297
307), ISSN 1558-0865.
Naghshtabrizi, P. & Hespanha, J. P. (2005). Designing an observer-based controller for a
network control system. Proceedings of 44th IEEE Conf. Decision & Control, and
European Control Conf., pp. 848853, Seville December 2005.
Narendra, K. S. & Annaswamy, A. M. (1989). Stable Adaptive Systems, Dover Publications,
ISBN 9780486442266.
Nilsson, J. (1998). Real-time Control Systems with Delays, Ph.D. Thesis, Lund Institute of
Technology.
Shi, Y.; Fang, H. & Yan, M. (2008). Kalman filter-based adaptive control for networked
systems with unknown parameters and randomly missing outputs. Int. J. Robust
Nonlinear Contol. In press, DOI: 10.1002/rnc.1390.
Sinopoli, B.; Schenato, L.; Franceschetti, M.; Poolla, K.; Jordan, M. I. & Sastry, S. S. (2004).
Kalman filtering with intermittent observations. IEEE Trans. Autom. Contr. Vol. 49,
No. 9, page numbers (14531464), ISSN 1558-0865.
180
Adaptive Control
Wallin, R.; Isaksson, A. J. & Noreus, O. (2001). Extension to Output prediction under scarce
data operation: control applications. Automatica, Vol. 37, No. 12, page numbers
(20692071), ISSN 0005-1098.
Walsh, G. C. & Ye, H. (2001). Scheduling of networked control systems. IEEE Control Syst.
Mag. Vol. 21, No. 1, page numbers (57-65), ISSN 0272-1708.
Walsh, G. C.; Ye, H. & Bushnell, L. G. (2002). Stability analysis of networked control
systems. IEEE Trans. Contr. Syst. Technol., Vol. 10, No. 3, page numbers (438446),
ISSN 1558-0865.
Wang, Z. D.; Yang, F.; Ho, D. W. C. & Liu, X. (2005a). Robust finite-horizon filtering for
stochastic systems with missing measurements. IEEE Signal Process. Lett., Vol. 12,
No. 6, page numbers (437440), ISSN: 1558-2361.
Wang, Z. D.; Yang, F.; Ho, D. W. C. & Liu, X. (2005b). Variance-constrained control for
uncertain stochastic systems with missing measurement. IEEE Trans. Syst. Man
Cybern. A, Syst. Humans, Vol. 35, No. 5, page numbers (746753), ISSN: 1083-4427.
Wu, J. & Chen, T. (2007). Design of networked control systems with packet dropouts. IEEE
Trans. Automat. Contr., Vol. 52, No. 7, page numbers (13141319), ISSN 1558-0865.
Yang, T. C. (2006). Networked control system: a brief survey. IEE Proc. Contr. Theory Appl.,
Vol. 153, No. 4, page numbers (403412), ISSN 1350-2379.
Yu, M.; Wang, L.; Chu, T. & Hao, F. (2004). An LMI approach to networked control systems
with data packet dropout and transmission delays, Proceedings of 45th IEEE Conf.
Decision & Control, pp. 35453550, San Diego, December 2005.
Yue, D.; Han, Q. L. & Peng, C. (2004). State feedback controller design of networked control
systems. IEEE Trans. Circuits Syst. II, Vol. 51, No. 11, page numbers (640644), ISSN
1549-7747.
Zhang, L. & Hristu-Varsakelis, D. (2006). Communication and control co-design for
networked control systems. Automatica, Vol. 42, No. 6, page numbers (953958),
ISSN 0005-1098.
Zhang, L.; Shi, Y.; Chen, T. & Huang, B. (2005). A new method for stabilization of networked
control systems with random delays. IEEE Trans. Automat. Contr., Vol. 50, No. 8,
page numbers (11771181), ISSN 1558-0865.
Zhang, W.; Branicky, M. S. & Phillips S. M. (2001). Stability of networked control systems.
IEEE Control Syst. Mag., Vol. 21, No. 1, page numbers (8499), ISSN 0272-1708.
8
Sun Wei, Zhang Lujin, Zou Jinhai and Miao Siyi
College of Electrical and Information Engineering, Hunan University

Changsha City, Hunan Province, P. R. China
1. Introduction
Neural network has good nonlinear function approximation ability. It can be widely used to
identify the model of controlled plant. In this chapter, the theories of modeling uncertain
plant by using two kinds of neural networks: feed-forward neural network and recurrent
neural network are introduced. And two adaptive control strategies for robotic tracking
control are developed. One is recurrent fuzzy neural network based adaptive control
(RFNNBAC), and another is neural network based adaptive robust control (NNBARC). In
RFNNBAC, a kind of recurrent fuzzy neural network (RFNN) is constructed by using
recurrent neural network to realize fuzzy inference, In which, temporal relations are
embedded in the network by adding feedback connections on the first layer of the network.
Two RFNNs are used to identify and control plant respectively. Base on the Lyapunov
stability approach, the convergence of the proposed RFNN is analyzed. In NNBARC, A
robust controller and a neural network are combined into an adaptive robust robotic
tracking control scheme. Neural network is used to approximate the modeling uncertainties
in robotic system. Then the disadvantageous effects on tracking performance, due to the
approximating error of the neural network and non-measurable external disturbances in
robotic system, are attenuated to a prescribed level by robust controller. The robust
controller and the adaptation law of neural network are designed based on Hamilton-JacobiIssacs (HJI) inequality theorem. The weights of NN are easily tuned on-line by a simple
adaptation law, with no need of a tedious and lengthy off-line training phase.
This chapter is organized in the following manner. In the first section a robust robotic
tracking controller based on neural network is designed and its effectiveness is proved by
applying it to control the trajectories of a two-link robot. Secondly, a recurrent fuzzy neural
network based adaptive control is proposed and simulation experiments are made by
applying it on robotic tracing control problem to confirm its effectiveness. Finally, some
conclusions are drawn.
2. A robust robotic tracking controller based on neural network

In the past decades, there has been much research on the applications of nonlinear control
theory to control robots, and many useful properties of robot dynamics such as the skewsymmetry property were discovered. There are basically two strategies to control such
uncertain nonlinear systems: the robust control strategy and the adaptive control strategy. A
Adaptive Control
182
convenient point of robust control strategy is that it can attenuate disadvantageous effects of
various uncertainties (e.g., structured parametric uncertainties and unstructured
disturbances) to a required level, provided that the upper bound of uncertainties is well
known (Abdallah et al. 1991). However, since this strategy use max-min method to design
the controller, it can not yield good transient performance. On the other hand, regressor
matrixes are always used in the design of adaptive control systems for robot manipulators
(Ortega & Spong 1989). In this situation, the unknown nonlinear dynamics of robotic
systems are always assumed to be linearly parametrisable. However, there are some
potential difficulties associated with this classical adaptive control design. For example, the
unknown parameters may be quickly varying, the linear parametrisable property may not
hold, computation of the regressor matrix is a time-consuming task, and implementation
also requires a precise knowledge of the structure of the entire robot dynamic model (Saad
et al. 1994; Sanner & Slotine 1998; Spooner & Passino 1996).
It has been shown that multi-layer neural networks can approximate any continuous
function as accurately as possible. Based on this universal approximation property, many
important adaptive neural-network-based control schemes have been developed to solve
highly nonlinear control problem (Sanner & Slotine 1998; Spooner & Passino 1996; Narenra
& Parthasarathy 1990; Polycarpou 1996). But most of these schemes use grads-descent
method to train the weights, which can not ensure the stability of whole closed-loop system.
In the recent years, researchers began to develop the neural-network-based controller with
closed-loop stability based on the Lyapunov method. A controller based on forward
propagation network was developed in (Carelli et al. 1995), but it didnt consider the effects
of uncertainties. An adaptive neural network control strategy with guaranteed stability was
proposed in (Behera et al. 1996) on the assumption that the approximation error of the
neural network is known and bounded.
In the first part of this chapter, we will propose a neural-network-based robust robotic
tracking controller according to HJI inequation theorem presented by Shen in (Shen 1996). A
neural network equipped with a robust learning algorithm is introduced firstly to learn the
modeling uncertainties in robotic system. Then the disadvantageous effects on tracking
performance caused by neural network approximating error and non-measurable external
disturbances in robotic system will be attenuated to a prescribed level by the designing a
robust controller.
This section is organized as follows. In subsection 2.1, HJI inequation theorem is introduced.
In subsection 2.2 the dynamics of robot system and its properties are described. The neural
network based robust control strategy is proposed in subsection 2.3, where the structure of
robust controller and the robust learning algorithm of neural network are derived.
Simulations for a two-link robot are presented in subsection 2.4.
2.1 HJI inequation theorem

A system with non-measurable disturbance d can be formulated as:
x& = f(x) + g(x)d
(1)
For evaluating the disturbance restraint performance of system (1), an evaluation signal
z = h(x) is introduced to represent the signals need to be concerned, such as error. And a
183
performance index signal can be defined as:

J = sup
d 2 0
(2)
Obviously, smaller J means better disturbance restraint performance. The robust design
problem of system (1) can be solved by designing a controller to make J less than a
prescribed level.
HJI(Hamilton-Jacobi-Isaacs)InequationTheorem: Given an positive constant > 0 , if there
exists an derivable function, V(x)0, which satisfies the following HJI inequation:
& = V x& = V f( x ) + V g(x)d 1 2 d

V
x
2
x
x
}, d
(3)
then the performance index signal of system (1) is less than , that is to say, J .
2.2 Problem statement

The kinetics equation of a robotic manipulator with uncertainties can be expressed as:
&& + V(q, q& )q& + G(q) + T(q, q& ) + d R = T
M(q)q
(4)
&& R n is the joint position, velocity, and acceleration vectors; M(q) R nn

where q, q& , q
denotes the moment of inertia; V(q, q& )q& are the Coriolis and centripetal forces; G(q) includes
the gravitational forces; T is the applied torque; T(q, q& ) represents the modelling
uncertainties in robotic system, and d R is external non-measurable disturbance.
It is well known that the robot dynamics has the following properties.
Property 1 Boundedness of the Inertia matrix: The inertia matrix M(q) is symmetric and
positive definite, and satisfies the following inequalities:
0 < m M(q) M
(5)
where m and M are known positive constants.

Property 2Skew symmetry: The inertia and centripetal-Coriolis matrices have the
following property:
& (q) 2V(q, q& )} = 0 , R n
T {M
(6)
Property 1 is very important in generating a positive definite function to prove the stability
of the closed-loop system. Property 2 will help in simplifying the controller.
The aim of this paper is to design a neural-network-based robust controller (NNBRC) for the
robot system under uncertainties, such that closed-loop system is guaranteed to be stable
Adaptive Control
184
and the joint position q(t) can track the desired trajectory q d (t) rapidly and accurately.
2.3 Design of NNBRC

A NNBRC is proposed in this section. In the proposed strategy, a neural network (NN) is
firstly used for identifing modelling uncertainties T(q, q& ) , then, a robust learning algorithm
and a robust controller are designed based on HJI equation theorem to counteract the
disadvantageous effects caused by approximation error of the NN and external disturbance
dR .
2.3.1 Construction of the neural network

(1)
(1)
A three-layer NN is shown in Fig.1.Using u i , o i
to denote the input and output of the ith
node in the lth layer separately, the signal propagation and the operation functions of the
nodes in each layer are introduced as follows.
x2
1
1
1
xm
x1
w11
y1
y2
yn
wnk
Fig. 1. Structure of three-layer NN

Layer 1 Input Layer:
(1)
oi
(1)
= ui
= x i , i = 1,2, L , m
(7)
Layer 2 Hidden Layer:

(2)
(1)
u j = o i , j = 1,2, L , k
i =1
(2)
(2)
o j = j = 1 [1 + exp( u j )] , j = 1,2, L , k
(8)
(9)
185
Layer 3 Output Layer:

(3)
(3)
(2)
y h = o h = u h = w hj o j , h = 1,2, L , n
j=1
Let
w 11
w
21
W=
M
w n1
w 12
w 22
M
w n2
(10)
L w 1k
1
y1

y
L w 2k
2
2
, = , Y=
M
M
M
M

L w nk
k
y n
then the outputs of the three-layer NN can be written as:

Y = W
(11)
In this paper, the three-layer NN described above will be used to identify the modeling
uncertainties T(q, q& ) in robotic system. Using T to denote the network approximation
error, then the modeling uncertainties can be denoted by:
T (q, q& ) = WT T + T T(q, q& )
(12)
where WT is the weight matrix, T is the activation function vector.

Substitute (12) into (4), then the dynamics of the robot manipulator with a NN identifier can
be formulated as:
&& + V(q, q& )q& + G(q) + WT T + T + d R = T
M(q)q
(13)
Regarding T as another external disturbance of robotic system, and using R = T + d R ,

then (13) can be rewritten as:
&& + V(q, q& )q& + G(q) + WT T + R = T
M(q)q
(14)
For attenuating disadvantageous effects caused by R to a prescribed level, a robust

learning algorithm of NN and a robust controller can be designed based on HJI equation as
below 2.3.2.
2.3.2 Robust controller and NN learning algorithm

At first, we introduce a control signal u, which satisfies:
&& d + V(q, q& )q& d + G(q) + u = T
M(q)q
(15)
&& d R n is desired joint position, velocity, and acceleration vectors separately.

where q d , q& d , q
Adaptive Control
186
Thus, the closed-loop robot control system can be constructed by substituting (15) into (14).
Let e = q q d , the closed-loop system can be formulated as:
M(q)&e& + V(q, q& )e& + WT T + R = u
(16)
By regarding R as external disturbance and introducing the evaluation signal z R = pe ,

where p is a positive constant, we can define the index signal as:
J R = sup
R 2 0
zR
(17)
& such that

The idea of NNBRC is to design controller u and the NN learning algorithm W
T
J R is less than a prescribed level, .
Define two state variables as:

x1 = e
x 2 = e& + e
(18)
where is an prescribed positive constant. Thus, system (16) can be rewritten as:
x& 1 = x 2 x 1
&
Mx 2 = Vx 2 + WT T R + u
(19)
where = Me& + Ve , WT is a n k matrix that can be described as:

w T 11
w
T 21
WT =
M
w T n1
w T 12
w T 22
M
w T n2
L w T 1k
L w T 2k
= [w T1 w T2 L w Tk ]
M
M
L w T nk
Theorem 1: Considering system (19), if the learning algorithm of NN is:

& = W
W
T
T
(20)
The controller u is designed as:

1
u = x 1 + WT T 2 +
2Y 2
x 2
and the parameter p in the evaluation signal , z R = pe = px 1 , satisfies:
(21)
187
1 2
p = 1
2
(22)
where 1 , 2 and are all prescribed positive constant, then the disturbance restraint index
signal of system (19), J R , is less than .

Proof: Considering system (19), we define the following derivable function:
L=
1
1 T
1
x 1 x 1 + x T2 M x 2 + WT
2
2
2
(23)
Thus,
m
1
& x + wT w
L& = x 1T x& 1 + x T2 Mx& 2 + x T2 M
2
Ti & Ti
2
i =1
= x 1T (x 2 x 1 ) + x T2 ( WT T R + u )
+
m
1 T &
& Ti
x 2 (M - 2V )x 2 + w TTi w
2
i =1
According to Property 2 of the robot dynamics, the above equation can be rewritten as:
m
& Ti
L& = x 1T (x 2 - x 1 ) + x T2 ( - w T T - R + u) + w TTi w
= x 1T x 1
x T2 (x 1
i =1
m
& Ti
+ - w T T - R + u) + w TTi w
i =1
Substituting (20) into above equation, then

L& = x 1T x 1 + x T2 (x 1 + - WT T - R + u) - WT
Regarding R as external disturbance, let
1
H = L& 2 R
2
= x 1
WT
1
zR
2
+ x T2 ( x 1 + W T T + u ) x T2 R
2
1 2
R
2
= - p 2 x1
2
1 2
p x1
2
+ x T2 (x1 + W T T + u )
x T2 R WT
1 2
R
2
Adaptive Control
188
1
= 2 x 2

= 2x T2 R + 2 R
+ 2x T2 R + 2 R
1
= x 2 + R

=
Q 2x T2 R 2 R
x 2 + R
x2
x T2 R
x2
x2
x2
1 2
R
2
H - p 2 x 1
2
= - p 2 x 1
2
x T2
x2
(x 1 + W T T
+ u ) WT
1
2 2
x2
1
x WT
+ x T2 x 1 + W T T + u +
2 2
Substituting (21), (22) into above inequation, then

H 1 x 1
2 x2
1
L& 2 R
2
zR
WT
2
According to HJI equation theorem, we can conclude that the disturbance restraint
performance index signal of system (19), J R , is less than . The structure of the proposed
neural network based robust control strategy is illustrated in Fig. 2.
Fig. 2. Structure of the NN-based robust tracking control system
189
2.4 Simulation example

In this section, the proposed control strategy will be applied to control the trajectory of a
two-link robot (see Fig. 3) for proving its effectiveness.
m2
l2
m1
t2
t1
l1
1
Fig. 3. Two-link robot

In Fig.3, m1and m2 are masses of arm1 and arm2 respectively; l1 and l2 are lengths of arm1
and arm2; t1 and t2 are torques on arm1 and arm2; 1 and 2 are positions of arm1 and
arm2. The dynamics model of two-link robot is same as (4).
Let
q = [ 1
&& = && 1
q
2 ]T , q& = & 1
&&
2
, T = [t 1
(24)
t 2 ]T
(25)
& 2
c i cos i , s i sin i , c ij cos( i + j )
(26)
then M, V,G in (4) can de described as:
m l 2 + m 2 (l 12 + l 22 + 2l 1 l 2 c 2 ) m 2 l 22 + m 2 l 1 l 2 c 2
M(q) = 1 1
m 2 l 22 + m 2 l 1 l 2 c 2
m 2 l 22
2m l l s q&
2 1 2 2 2
V(q, q& ) =
m 2 l 1 l 2 s 2 q& 1
m 1 l 1 l 2 s 2 q& 2
m 2 l 2 gc 12 + (m 1 + m 2 )l 1 gc 1
G(q) =
m 2 l 2 gc 12
(27)
(28)
(29)
Adaptive Control
190
In this paper, the parameters of the two-link robot are m 1 = 10 kg, m 2 = 2 kg, l 1 = 1.1 m,
and l 2 = 0.8 m. The Initial states are q(0) = [0.5 0.5] rad, q& (0) = [0 0]T rad/s, and
&&(0) = [0 0]T rad/s2. The desired trajectories can be described as:
q
q d ( t ) = [sin(2t) cos(2t)]T rad
(30)
q& d ( t ) = [2cos(2t) - 2sin(2t)]T rad/s
&& d ( t ) = - 4 2 sin(2t) - 4 2 cos(2t)

q
rad/s2
(31)
(32)
The model error due to friction is assumed as:
0.5sign (e& 1 )[0.1 + exp( e& 1 )]

T =
Nm
sign (e& 2 )[0.2 + exp( e& 2 )]
(33)
The external disturbance, d R = [d 1 d 2 ]T is a random signal which amplitude is less than

10Nm. In simulations, the NNBRC can be designed based on (21), in which = 50 ,
1 = 0.1 , 2 = 0.1 , = 0.05 , p = 9 . The NN learning algorithm is designed according to (20),
where = 0.1 .
Fig.4 and Fig.5 present the simulation experiment results, in which, proposed control
strategy is compared to traditional robust control (TRC) strategy. From these results, we can
conclude that the NN-based robust tracking control strategy proposed in this paper can
counteract disadvantageous effects caused by uncertainties in robotic system efficiently, and
can achieve better transient performance than traditional robust control.
Fig. 4. Robot trajectories
191
Fig. 5. Robot tracking errors
3. A Recurrent Fuzzy Neural Network Based Adaptive Control

Recently, much research has been done on using neural networks (NN) to identify and
control dynamic systems (Park et al. 1996; Narendra & Parthasarathy 1990; Brdys &
Kulawski 1999). NN can be classified as feed forward neural networks and recurrent neural
networks. Feed forward neural networks can approximate a continuous function to an
arbitrary degree of accuracy. However, feed forward neural network is a static mapping; it
can not represent a dynamic mapping. Although this problem can be solved by using
tapped delays, feed forward neural network requires a large number of neurons to represent
dynamical responses in the time domain. Moreover, since the weight updates of feed
forward neural network is irrelative to the internal information of neural network, the
function approximation is sensitive to the training data. On the other hand, recurrent neural
networks (Ku & Lee 1995; Ma & Ji 1998; Sundareshan & Condarcure 1998; Liang & Wang
2000) are able to represent dynamic mapping very well and store the internal information
for updating weights later. Recurrent neural network has an internal feedback loop; it
captures the dynamical response of a system without external feedback through delays.
Recurrent neural network is a dynamic mapping and demonstrates good performance in the
presence of uncertainties, such as parameter variations, external disturbance, unmodeled
and nonlinear dynamics. However, the drawbacks of recurrent neural network, which are
same as neural network, are that the function of the network is difficult to interpret and few
efficient constructive methods can be found for choosing network structure and determining
the parameters of neurons.
As is widely known, both fuzzy logic systems and neural network systems are aimed at
exploiting human-like knowledge processing capability. In recent years, researchers started
to recognize that fuzzy control has some similarities to neural network (Jang & Sun 1993;
Hunt et al. 1996; Buckley et al. 1993; Reyneri 1999). Fuzzy neural network (FNN), which uses
NN to realize fuzzy inference, combines the capability of fuzzy reasoning in handling
uncertain information and the capability of neural networks in learning from processes. It is
Adaptive Control
192
possible to train NN using the experience of human operators expressed in term of linguistic
rules, and interpret the knowledge that NN acquired from training data in linguistic form.
And it is very easy to choose the structure of NN and determine the parameters of neurons
from linguistic rules. However, a major drawback of the FNN is that its application domain
is limited to static problems due to its feed forward network structure.
Recurrent fuzzy neural network (RFNN) is a modified version of FNN, which use recurrent
network for realizing fuzzy inference and can be constructed from a set of fuzzy rules. It
inherits all characteristics of FNN such as fuzzy inference, universal approximation and
convergence properties. Moreover, with its own internal feedback connections, RFNN can
temporarily store dynamic information and cope with temporal problems efficiently. For
this ability to temporarily store information, the structure of RFNN is much simpler than
FNN. Fewer nodes are required in RFNN for system identification.
In this section, a recurrent fuzzy neural network structure is proposed, in which, the
temporal relations are embedded by adding feedback connections on the first layer of FNN.
Back propagation algorithm is used to train the proposed RFNN. To guarantee the
convergence of the RFNN, the Lyapunov stability approach is applied to select appropriate
learning rates. For control problem, an adaptive control scheme is proposed, in which, two
proposed RFNN are used to identify and control plant respectively. Finally, simulation
experiments are made by applying proposed adaptive control scheme on robotic tracking
control problem to confirm its effectiveness.
This section is organized as follows. In subsection 3.2, RFNN is constructed. The
construction of RFNNBAC is presented in subsection 3.3. Learning algorithms of RFNN are
derived in subsection 3.4. Stability of RFNN is analyzed in subsection 3.5. In subsection 3.6
proposed RFNNBAC is applied on robotic tracking control and simulation results are given.
Finally, some conclusions are drawn in subsection 3.7.
3.1 Construction of RFNN

The structure of the proposed RFNN is shown in Fig. 6, which comprises n input variables,
m term nodes for each input variable, l rule nodes, and p output nodes. This RFNN thus
consists of four layers and n + ( n m ) + l + p nodes.
Using u ki , O ki to denote the input and output of the ith node in the kth layer separately, the
signal propagation and the operation functions of the nodes in each layer are introduced as
follows.
Layer 1 (Input Layer): This layer accepts input variables. Its nodes transmit input values to
the next layer. Feedback connections are added in this layer to embed temporal relations in
the network. For every node in this layer, the input and output are represented as:
u 1i (k ) = x 1i (k ) + w 1i O 1i (k - 1), O 1i (k ) = u 1i (k ), i = 1,2, L , n
(34)
where k is the number of iterations; w 1i is the recurrent weights.

Layer 2 (Membership Layer): Nodes in this layer represent the terms of respective linguistic
variables. Each node performs a Gaussian membership function
u ij2 =
193
(O 1i - a ij ) 2
(b ij ) 2
, O ij2 = exp(u ij2 )
(35)
where i = 1,2, L , n , j = 1,2, L , m ; a ij and b ij are the mean and the standard deviation of the
Gaussian membership function; the subscript ij indicates the jth term of the ith input
variable.
Fig. 6. Structure of four-layer RFNN

Layer 3(Rule Layer): This layer forms the fuzzy rule base and realizes the fuzzy inference.
Each node is corresponding to a fuzzy rule. Links before each node represent the
preconditions of the corresponding rule, and the node output represents the firing
strength of corresponding rule.
If the qth fuzzy rule can be described as:
Adaptive Control
194
q
qth rule: if x 1 is A 1 , x 2 is A 2 , , x n is A nq then y 1 is B 1 , y 2 is B 2 , , y p is B pq ,

q
where A i is the term of the ith input in the qth rule; B j is the term of the jth output in the
qth rule.
Then, the qth node of layer 3 performs the AND operation in qth rule. It multiplies the input
signals and output the product.
2
Using O iq
to denote the membership of x i to A i , where q i {1,2, L , m} , then the input
i
q
and output of qth node can be described as:

2
u q3 = O iq
, O q3 = u q3 , i = 1,2, L , n; q = 1,2, L , l
i
i
(36)
Layer 4(Output Layer): Nodes in this layer performs the defuzzification operation. the input
and output of sth node can be calculated by:
4
u s4 = w sq
O q3 , O s4 =
q
u s4
(37)
Oq
q
4
where s = 1,2, L , p , q = 1,2, L , l , w sq
is the center of B j , which represents the output
action strength of the sth output associated with the qth rule.
From the above description, it is clear that the proposed RFNN is a fuzzy logic system with
memory elements in first layer. The RFNN features dynamic mapping with feedback and
more tuning parameters than the FNN. In the above formulas, if the weights in the feedback
unit w 1i are all equal to zero, then the RFNN reduces to an FNN. Since a fuzzy system has
clear physical meaning, it is very easy to choose the number of nodes in each layer of RFNN
and determine the initial value of weights. Note that the parameters w 1i of the feedback
units are not set from human knowledge. According to the requirements of the system, they
will be given proper values representing the memorized information. Usually the initial
values of them are set to zero.
3.2 Structure of RFNNBAC

In this section, the structure of RFNNBAC will be developed below, in which, two proposed
RFNN are used to identify and control plant respectively.
3.2.1 Identification based on RFNN

Resume that a system to be identified can be modeled by an equation of the following form:
y(k ) = f y(k 1), L y k n y , u(k ), L , u(k n u )
(38)
195
where u is the input of the system, n y is the delay of the output, and n u is the delay of the
input.
Feed forward neural network can be applied to identify above system by using y(k-1),
,y(k-n-1), u(k), , u(k-m) as inputs and approximating the function f.
For RFNN, the overall representation of inputs x and the output y can be formulated as
y(k) = g(O 11 (k), L , O 1n (k))
(39)
Where
O 1i (k ) = x i (k ) + w 1i (k )O 1i (k 1)
= x i (k ) + w 1i (k ) x i (k 1) + w 1i (k 1)O 1i (k 2 )
M
= x i (k ) + w 1i (k )x i (k 1) + w 1i (k )w 1i (k 1)x i (k 2 ) + L
+ w 1i (k )w 1i (k 1)L w 1i (1)x i (0 )
Using the current input u(k) and the most recent output y(k-1) of the system as the inputs of
RFNN, (39) can be modified as:
y(k ) = f (y(k 1), L , y(0 ), u(k ), L , u(0 ))
(40)
By training the RFNN according to the error e(k) between the actual system output and the
RFNN output, the RFNN will estimate the output trajectories of the nonlinear system (38).
The training model is shown in Fig.7.
Fig. 7. Identification of dynamic system using RFNN
Adaptive Control
196
From above description, For Using RFNN to identify nonlinear system, only y(k-1) and u(k)
need to be fed into the network .This simplifies the network structure, i. e., reduces the
number of neurons
3.2.2 RFNNBAC
The block diagram of RFNNBAC is shown in Fig. 8. In this scheme, two RFNNs are used as
controller (RFNNC) and identifier (RFNNI) separately. The plant is identified by RFNNI,
which provides the information about the plant to RFNNC. The inputs of RFNNC are e(k)
and e& (k) . e(k) is the error between the desired output r(t) and the actual system output
y(k). The output of RFNNC is the control signal u(k), which drives the plant such that e(k) is
minimized. In the proposed system, both RFNNC and RFNNI have same structure.
Fig. 8. Control system based on RFNNs
3.3 Learning Algorithm of RFNN

For parameter learning, we will develop a recursive learning algorithm based on the back
propagation method
3.3.1 Learning algorithm for identifier

For training the RFNNI in Fig.8, the cost function is defined as follows:
J I (k ) =
p
1 p
2
2
(e I s (k )) = (y s (k ) y I s (k ))
2 s =1
s =1
(41)
where y s (k) is the sth output of the plant, y I s (k ) = O s4 is the sth output of RFNNI, and
e I s (k ) is the error between y s (k) and y I s (k ) for each discrete time k.
By using the back propagation (BP) algorithm, the weights of the RFNNI is adjusted such
197
that the cost function defined in (41) is minimized. The BP algorithm may be written briefly
as:
WI (k + 1) = WI (k) + WI (k)
J I (k)
= WI (k) + I
WI (k)
(42)
where I represents the learning rate and WI represents the tuning weights, in this case,
which are w I4sq , a I iq i , b I iqi , and w 1I i . Subscript I represents RFNNI.
According to the RFNNI structure (34)~(37), cost function (41) and BP algorithm (42), the
update rules of RFNNI weights are
w I4sq (k + 1) = w I4sq (k ) Iw4
J I (k )
(43)
w I4sq (k )
a I iq i (k + 1) = a I iq i (k ) Ia
J I (k )
a I iq i (k )
(44)
b I iq i (k + 1) = b I iq i (k ) Ib
J I (k )
b I iq i (k )
(45)
w 1I i (k + 1) = w 1I i (k ) Iw1
J I (k )
(46)
w 1I i (k )
Where
J I (k )
w I4sq (k )
= e I s (k )
O I3q
3
OI q
q
2 O 1I i a I iq i
w I4sq O I4s
J I (k )
3
O
= e I s (k )
Iq
3
a I iq i (k )
b I iq i 2
s
OI q
q
w I4sq O I4s
2 O I1i a I iq i
J I (k )
3
= e I s (k )
O
I
q
3
b I iq i (k )
b I iq 3
s
OI q
q
J I (k )
w I1i
(k )
= e I s (k )
q s
w I4sq O I4s
3
OI q
q
O I3q
2 O I1i a I iq i
(b I iqi )
)2
) O 1 (k 1)
Ii
Adaptive Control
198
3.3.2 Learning algorithm for controller

For training RFNNC in Fig. 8, the cost function is defined as
J C (k ) =
p
1 p
2
2
(e s (k )) = (rs (k ) y s (k ))
2 s =1
s =1
(47)
where rs (k ) is the sth desired output, y s (k ) is the sth actual system output and e s (k ) is
the error between rs (k ) and y s (k ) .
Then, the gradient of J C is
y s
J C
J
= C
WC s y s WC
y s (k ) u o (k )
,
u o (k ) WC
s
u (k )
= e s (k )yu so (k ) o
WC
s
= e s (k )
(48)
where u o is the oth control signal, which is also the oth output of RFNNC, and
yu so (k ) = y s (k ) u o (k ) denotes the system sensitivity. Thus the parameters of the RFNNC
can be adjusted by
WC (k + 1) = WC (k) + WC (k)
= WC (k) + C (
J C (k)
)
WC (k)
(49)
Note that the convergence of the RFNNC cannot be guaranteed until yu so (k ) is known.
Obviously, the RFNNI can provide yu so (k ) to RFNNC. Resume that the oth control signal
is also the oth input of RFNNI, then yu so (k ) can be calculated by
3
2
4
1
O Iq
O Ioq
y s (k)
O Is
o O Io
=
3
2
1
u o
u o (k) q O Iq O Ioq
O Io
o
4
w I4sq O Is
O Iq
q
3
O Iq
- 2(O 1Io - a Ioq o )
(50)
(b Ioq o ) 2
3.4 Stability analysis of the RFNN
Choosing an appropriate learning rate is very important for the stability of RFNN. If the
value of the learning rate is small, convergence of the RFNN can be guaranteed, however,
199
the convergence speed may be very slow. On the other hand, choosing a large value for the
learning rate can fasten the convergence speed, but the system may become unstable.
3.4.1 Stability analysis for identifier

For choosing the appropriate learning rate for RFNNI, discrete Lyapunov function is
defined as
L I (k ) = J I (k ) =
1
2
(e I s (k ))
2 s
(51)
Thus the change of the Lyapunov function due to the training process is
L I (k ) = L I (k + 1) L I (k )
1
1
= (e I s (k + 1))2 (e I s (k ))2
2 s
2 s
1
= (e I s (k + 1))2 (e I s (k ))2
2 s
1
[(e I s (k + 1) + e I s (k )) (e I s (k + 1) e I s (k ))]
2 s
(52)
1
[(2e I s (k ) + e I s (k )) e I s (k )]
2 s
1
= (e I s (k ))2 + 2e I s (k )e I s (k )
2 s
1
1
= (e I s (k ))2 + [2e I s (k )e I s (k )]
2 s
2 s
=
The error difference due to the learning can be represented by

e I s (k ) = e I s (k + 1) e I s (k )
e I s (k )
WI (k )
WI (k )
Where
J I (k )
J I (k ) e I s (k )
= I
WI (k )
s e I s (k ) WI (k )
e (k )
= I e I s (k ) I s
WI (k )
s
WI (k ) = I
So (52) can be modified as
(53)
Adaptive Control
200
L (k ) =
J I (k )
e (k )
J I (k )
1 e I s (k )
1
+ 2e I s (k ) I s
I
I
WI (k )
WI (k )
WI (k )
2 s WI (k )
2 s
2
J I (k ) e I s (k )
J I (k )
e (k )
1
I

I
e I s (k ) I s
WI (k ) s
WI (k )
2 WI (k ) s WI (k )
(54)
2
e (k )
J (k )
J I (k )
1
I
Is
I I
W (k )
2 WI (k ) s WI (k )
I
2
2
e (k )
1 J (k )
2
I I s
= I I
2 WI (k ) s WI (k )
To guarantee the convergence of RFNNI, the change of Lyapunov function L I (k ) should

be negative. So learning rate must satisfy the following condition:
2
e (k )
.
0 < I (k ) < 2 I s
s WI (k )
(55)
For the learning rate of each weight in RFNNI, the condition (22) can be modified as
2

(
)
e
k
Is
0 < Iw4 (k ) < 2 max

q s w 4 (k )
I sq

0 < Ia (k ) < 2
0 < Ib
(k ) <
0 < Iw1
2

e I s (k )
max

q,i s a
(k )
I
iq
i

2

e I s (k )
max

q,i s b
(k )
I
iq
i

2

e I s (k )
(k ) < 2 max 1 .
i s w (k )
Ii

3.4.2 Stability analysis for controller

Similar to (51), the Lyapunov function for RFNNC can be defined as
(56)
(57)
(58)
(59)
201
L C (k ) = J C (k ) =
1
2
(e s (k ))
2 s
(60)
So, similar to (56)-(59), the learning rates for training RFNNC should be chosen according to
the following rules:
0 < Cw4
2

e s (k )
(k ) < 2 max 4
q s w
(k )
C sq
(61)
2

e s (k )
max

q,i s a
(k )
C
iq
i

(62)
0 < Cb (k ) < 2
2

e s (k )
max

q,i s b
(
k)

C iq i
(63)
2

e s (k )
0 < Cw1 (k ) < 2 max
i s w 1 (k )
Ci
(64)
0 < Ca
(k ) <
3.5 Simulation Experiments

Dynamics of robotic manipulators are highly nonlinear and may contain uncertain elements
such as friction and load. Many efforts have been made in developing control schemes to
achieve the precise tracking control of robot manipulators. Among available options, neural
networks and fuzzy systems (Er & Chin 2000; Llama et al. 2000; Wang & Lin 2000; Huang &
Lian 1997) are used more and more frequently in recent years. In the simulation experiments
of this chapter, the proposed RFNNBAC is applied to control the trajectory of the two-link
robotic manipulator described in chapter 2.4 to prove its effectiveness.
In the simulation, the parameters of manipulator are m 1 =4 kg, m 2 =2 kg, l 1 =1 m, l 2 =0.5
m, g =9.8 N/kg. Initial conditions are given as (0 ) =0 rad, (0 ) =1 rad, & (0 ) =0,
1
and & 2 (0 ) =0 rad/s. The desired trajectory is given by 1 (t ) = sin (2t ) and 2 (t ) = cos(2t ) .
The friction and disturbance terms in (4) are assumed to be
5cos(5t)
dR =
Nm, T(q, q& ) = 0.5sign(q& ) Nm.
5cos(5t)
Adaptive Control
202
Simulation results are shown in Fig.9 ~Fig.14. Fig.9 and Fig.10 illustrate the trajectories of
two joints; the two outputs of identifier (RFNNI) are shown in Fig.11 and Fig.12 separately;
the cost function for RFNNC is shown in Fig.13; and Fig.14 shows the cost function for
RFNNI.
From simulation results, it is obvious that the proposed RFNN can identify and control the
robot manipulator very well.
Fig. 9. Trajectory of joint1
Fig. 11. Identifier (RFNNI) output1
Fig. 13. Cost function for RFNNC
Fig. 10. Trajectory of joint2
Fig. 12. Identifier (RFNNC) output2
Fig. 14. Cost function for RFNNI
203
4. Conclusion
In this paper, the adaptive control based on neural network is studied. Firstly, a neural
network based adaptive robust tracking control design is proposed for robotic systems
under the existence of uncertainties. In this proposed control strategy, the NN is used to
identify the modeling uncertainties, and then the disadvantageous effects caused by neural
network approximating error and external disturbances in robotic system are counteracted
by robust controller. Especially the proposed control strategy is designed based on HJI
inequation theorem to overcome the approximation error of the neural network bounded
issue. Simulation results show that proposed control strategy is effective and has better
performance than traditional robust control strategy. Secondly, an RFNN for realizing fuzzy
inference using the dynamic fuzzy rules is proposed. The proposed RFNN consists of four
layers and the feedback connections are added in first layer. The proposed RFNN can be
used for the identification and control of dynamic system. For identification, RFNN only
needs the current inputs and most recent outputs of system as its inputs. For control, two
RFNNs are used to constitute an adaptive control system, one is used as identifier (RFNNI)
and another is used as controller (RFNNC). Also to prove the proposed RFNN and control
strategy robust, it is used to control the robot manipulator and simulation results verified
their effectiveness.
5. References
Abdallah, C., Dawson, D., Dorato, P. & Jamshidi, M. (1991). Survey of the robust of rigid
robots, IEEE Control Systems Magazine, Vol. 11, No. 2, pp. 24-30.
Ortega, R. & Spong, M. W. (1989). Adaptive motion control of rigid robots: a tutorial,
Automatica, Vol. 25, No. 3, pp. 877-888.
Saad, M., Dessaint, L. A., Bigras, P. & Haddad, K. (1994). Adaptive versus neural adaptive
control: application to robotics, International Journal of Adaptive Control and Signal
Processing, Vol. 8, No. 2, pp. 223-236.
Sanner, R. M. & Slotine, J. J. E. (1992). Gaussian networks for direct adaptive control, IEEE
Transactions on. Neural Network, Vol. 3, No. 4, pp. 837-863.
Spooner, J. T. & Passino, K. M. (1996). Stable adaptive control using fuzzy systems and
neural networks, IEEE Transactions on Fuzzy system, Vol. 4, No. 2, pp. 339-359.
Narenra, K. S. & Parthasarathy, K. (1990). Identification and control of dynamical systems
using neural networks, IEEE Transactions on Neural networks, Vol. 1, No. 1, pp. 4-27.
Polycarpou, M. M. (1996). Stable adaptive neural control scheme for nonlinear systems, IEEE
Transactions on Automatic Control, Vol. 41, No. 2, pp. 447-451.
Carelli, R., Camacho, E. F. & Patino, D. (1995). A neural network based feedforward
adaptive controller for robot, IEEE Transactions on Systems, Mman and Cybernetics,
Part B: Cybernetics, Vol. 25, No. 6, pp. 1281-1288.
Behera, L., Chaudhury, S. & Gopal, M. (1996). Neuro-adaptive hybrid controller for robotmanipulator tracking control, IEE Proceedings Control Theory Applications, Vol.143,
No.1, pp.2710-275.
Shen, T. L. (1996). H control theory and its applications, ISBN 7302022151, Tsinghua Press,
Beijin, China.
204
Adaptive Control
Park, Y. M., Choi, M. S. & Lee, K. Y. (1996). An optimal tracking neuro-controller for
nonlinear dynamic systems, IEEE Transactions on Neural Networks, Vol. 7, No. 5, pp.
1099-1110.
Narendra, K. S. & Parthasarathy, K. (1990). Identification and control of dynamical systems
using neural networks, IEEE Transactions on Neural Networks, Vol. 1, No. 1, pp. 4-27.
Brdys, M. A. & Kulawski, G. J. (1999). Dynamic neural controllers for induction motor, IEEE
Transactions on Neural Networks, Vol. 10, No. 2, pp. 340-355.
Ku, C. C. & Lee, K. Y. (1995). Diagonal recurrent neural networks for dynamic systems
control, IEEE Transactions on Neural Networks, Vol. 6, No. 1, pp. 144-156.
Ma, S. & Ji, C. (1998). Fast training of recurrent neural networks based on the EM algorithm,
IEEE Transactions on Neural Networks, Vol. 9, No. 1, pp. 11-26.
Sundareshan, M. K. & Condarcure, T. A. (1998). Recurrent neural-network training by a
learning automation approach for trajectory learning and control system design,
IEEE Transactions on Neural Networks, Vol. 9, No. 3, pp. 354-368.
Liang, X. B. & Wang, J. (2000). A recurrent neural network for nonlinear optimization with a
continuously differentiable objective function and bound constraints, IEEE
Jang, J. S. R. & Sun, C. T. (1993). Functional equivalence between radial basis function
networks and fuzzy inference systems, IEEE Transactions on Neural Networks, Vol. 4,
No. 1, pp. 156-159.
Hunt, K. J., Hass, R. & Munay-Smith, R. (1996). Extending the functional equivalence of
radial basis function networks and fuzzy inference systems, IEEE Transactions on
Neural Networks, Vol. 7, No. 3, pp. 776-781.
Buckley, J. J., Hayashi, Y. & Czogala, E. (1993). On the equivalence of neural nets and fuzzy
expert systems, Fuzzy Sets and Systems, Vol. 53, No. 2, pp. 129-134.
Reyneri, L. M. (1999). Unification of neural and wavelet networks and fuzzy systems, IEEE
Er, M. J. & Chin, S. H. (2000). Hybrid adaptive fuzzy controller of robot manipulators with
bounds estimation, IEEE Transactions on Industrial Electronics, Vol. 47, No. 5, pp.
1151-1160.
Llama, M. A., Kelly, R. & Santibanez, V. (2000). Stable computed-torque control of robot
manipulator via fuzzy self-tuning, IEEE Transactions on Systems, Man and
Cybernetics, Part B: Cybernetics, Vol. 30, No. 1, pp. 143-150.
Wang, S. D. & Lin, C. K. (2000). Adaptive tuning of the fuzzy controller for robots, Fuzzy Sets
Systems, Vol. 110, No. 3, pp. 351-363.
Huang, S. J. & Lian, R. J. (1997). A hybrid fuzzy logic and neural network algorithm for
robot motion control, IEEE Transactions on Industrial Electronics, Vol. 44, No. 3, pp.
408-417.
9
Adaptive control of the electrical drives with the
elastic coupling using Kalman filter
Krzysztof Szabat and Teresa Orlowska-Kowalska
Wroclaw University of Technology
Poland
1. Introduction
The control problem of the two-mass system originally derives from rolling-mill drives
(Sugiura & Hori, 1996), (Ji & Sul, 1995), (Szabat & Orlowska-Kowalska, 2007). Large inertias
of the motor, rolls and long shaft create an elastic system. The motor speed is different from
the load side and the shaft undergoes large torsional torque. A similar problem exists in the
field of conveyer drives (Hace et al., 2005). Also the performance of the machines used in
textile industry is reduced by the non-ideal characteristics of the shaft (Beineke et al., 1997),
(Wertz et al., 1999). An analogous problem appears in the paper machine sections
(Valenzuela et al., 2005) and in modern servo-drives (Vukosovic & Stojic, 1998), (OSullivan
et al., 2007), (Shen & Tsai, 2006). Moreover, torsional vibrations decrease the performance of
the robot arms (Ferretti et al., 2004), (Huang & Chen, 2004). This problem is especially
important in the field of space robot manipulators. Due to the cost of transport, the total
weight of the machine must be drastically reduced. This reduces the stiffness of the
mechanical connections which in turn influences the performance of the manipulator in a
negative way (Katsura & Ohnishi, 2005), (Ferretti et al., 2005). The elasticity of the shaft
worsens the performance of the position control of deep-space antenna drives (Gawronski et
al., 1995). Vibrations affect the dynamic characteristics of computer hard disc drives (Ohno
& Hara, 2006) and (Horwitz et al., 2007).
Torsional vibrations can appear in a drive system due to the following reasons:
- changeability of the reference speed;
- changeability of the load torque;
- fluctuation of the electromagnetic torque;
- limitation of the electromagnetic torque;
- mechanical misalignment between the electrical motor and load machine;
- variations of load inertia
- unbalance of the mechanical masses;
- system nonlinearities, such as friction torque and backlash.
The simplest method to eliminate the oscillation problem (occurring while the reference
speed changes) is a slow change of the reference velocity. Nevertheless, it causes the
decrease of the drive system dynamics and does not protect it against oscillations appearing
when the disturbance torque changes. The conventional control structure based on the PI
206
Adaptive control
speed controller, tuned by the classical symmetric criterion, with a single feedback from the
motor speed is not effective in damping the speed oscillations. One of the simplest ways to
improve the torsional vibrations ability of the classical structure is presented in (Zhang &
Furusho, 2000). It is based on the suitable selection of the system closed-loop poles. However,
this method improves the drive performance only for a limited range of the system
parameters.
When the resonant frequency of the system excides hundreds of Hertz, the application of the
digital filters is an industrial standard. The Notch-filter is usually mentioned as a tool
ensuring the damping of the oscillations (Vukosovic & Stojic, 1998), (Ellis & Lorenz, 2000).
Rarely a low-pass filter or Bi-filter is used. The digital filters can damp the torsional vibration,
yet the dynamics of the system may be affected.
To improve performances of the classical control structure with the PI controller, the
additional feedback loop from one selected state variable can be used. The additional
feedback allows setting the desired value of the damping coefficient, but the free value of the
resonant frequency cannot be achieved simultaneously (Szabat & Orowska-Kowalska, 2007).
According to the literature, the application of the additional feedback from the shaft torque is
very common (Szabat & Orowska-Kowalska, 2007). The design methodology of that system
can be divided into two groups. In the first framework the shaft torque is treated as the
disturbance. The simplest approach relies on feeding back the estimated shaft torque to the
control structure, with the gain less than one. The more advanced methodology, called
Resonance Ratio Control (RRC) is presented in (Hori et al., 1999). The system is said to have
good damping ability when the ratio of the resonant to antiresonant frequency has a
relatively big value (about 2). The second framework consists in the application of the modal
theory. Parameters of the control structure are calculated by comparison of the characteristic
equation of the whole system to the desired polynomial. To obtain a free design of the
control structure parameters, i.e. the resonant frequency and the damping coefficient, the
application of two feedbacks from different groups is necessary. The design methodology of
this type of the systems is presented in (Szabat & Orowska-Kowalska, 2007).
The control structures presented so far are based on the classical cascade compensation
schemes. Since the early 1960s a completely different approach to the analysis of the system
dynamics has been developed the state space methodology (Michels et al., 2006). The
application of the state-space controller allows to place the system poles in an arbitrary
position so theoretically it is possible to obtain any dynamic response of the system. The
suitable location of the closed-loop system poles becomes one of the basic problems of the
state space controller application. In (Ji & Sul, 1995) the selection of the system poles is
realized through LQ approach. The authors emphasize the difficulty of the matrices selection
in the case of the system parameter variation. The influence of the closed-loop location on the
dynamic characteristics of the two-mass system is analyzed in (Qiao et al., 2002), (Suh et al.,
2001). In (Suh et al., 2001) it is stated that the location of the system poles in the real axes
improve the performance of the drive system and makes it more robust against the parameter
changing.
In the case of the system with changeable parameters more advanced control concepts have
been developed. In (Gu et al., 2005), (Itoh et al., 2004) the applications of the robust control
theory based on the H and -synthesis frameworks are presented. The implementation of
the genetic algorithm to setting of the control structure parameter is shown in (Itoh et al.,
2004). The author reports good performance of the system despite the variation of the inertia
207
of the load machine. The next approach consists in the application of the sliding-mode
controller. For example, in paper (Erbatur et al., 1999) this method is applied to controlling
the SCARA robot. A design of the control structure is based on the Lyapunov function. The
similar approach is used in (Hace et al., 2005) where the conveyer drive is modelled as the
two-mass system. The authors clam that the design structure is robust to the parameter
changes of the drive and external disturbances. Other application examples of the slidingmode control can be found in (Erenturk, 2008). The next two frameworks of control
approach relies on the use of the adaptive control structure. In the first framework the
controller parameters are adjusted on-line on the basis of the actual measurements. For
instance in (Wang & Frayman, 2004) a dynamically generated fuzzy-neural network is used to
damp torsional vibrations of the rolling-mill drive. In (Orlowska-Kowalska & Szabat, 2008b)
two neuro-fuzzy structures working in the MRAS structure are compare. The experimental
results show the robustness of the proposed concept against plant parameter variations. In
the other framework changeable parameters of the plant are identified and then the
controller is retuned in accordance with the currently identified parameters. The Kalman
filter is applied in order to identify the changeable value of the inertia of the load machine
(Orlowska-Kowalska & Szabat, 2008a). This value is used to correct the parameters of the PI
controller and two additional feedbacks. A similar approach is presented in (Hirovonen et
al., 2006). In the paper (Cychowski et al., 2008) the model predictive controller is applied o
ensure the optimal control of the system states taking the system constrains into
consideration. In order to reduce the computational complexity the explicit version of the
controller is suggested to real-time implementation.
This paper is divided into seven sections. After an introduction, the mathematical model of
the two-mass drive system and utilised control structure are described. In section IV, the
mathematical model of the NEKF is presented. The simulation results of the non-adaptive
and adaptive NEKF are demonstrated in sections V. The proposed adaptation mechanism is
described and the analysed algorithms are compared. After a short description of the
laboratory set-up, the experimental results are presented in section VI. Conclusions are
presented at the end of the paper.
2. The mathematical model of the two-mass system and the control structure
In technical papers there exist many mathematical models, which can be used for the
analysis of the plant with elastic couplings. In many cases the drive system can be modelled
as a two-mass system, where the first mass represents the moment of inertia of the drive and
the second mass refers to the moment of inertia of the load side. The mechanical coupling is
treated as an inertia free. The internal damping of the shaft is sometimes also taken into
consideration. Such a system is described by the following state equation (Szabat &
Orlowska-Kowalska, 2007) (with non-linear friction neglected):
D
1 (t ) J 1
d
D
2 (t ) =
dt
J
M s (t ) 2
K
c
D
J1
D
J2
Kc
1
1
0
J 1 (t )
1
J
1
1
1
2 (t ) + 0 [M e ] + [M L ]
J2
J2

0 M s (t ) 0
0
(1)
Adaptive control
208
where: 1- motor speed, 2- load speed, Me motor torque, Ms shaft (torsional) torque, ML
load torque, J1 inertia of the motor, J2 inertia of the load machine, Kc stiffness coefficient,
D internal damping of the shaft.
The described model is valid for the system in which the moment of inertia of the shaft is
much smaller than the moment of the inertia of the motor and the load side. In other cases a
more extended model should be used, such as the Rayleigh model of the elastic coupling or
even a model with distributed parameters. The suitable choice of the mathematical model is
a compromise between the accuracy and calculation complexity. As can be concluded from
the literature, nearly in all cases the simplest shaft-inertia-free model has been used.
To simplify the comparison of the dynamical performances of the drive systems of different
power, the mathematical model (1) is expressed in per unit system, using the following
notation of new state variables:
1 =
M
M
M
1
2 = 2 me = e m s = s m L = L
N
M
M
M
N
N
N
N
(2)
where: N nominal speed of the motor, MN nominal torque of the motor, 1, 2 motor
and load speeds, me, ms, mL electromagnetic, shaft and load torques in per unit system.
The mechanical time constant of the motor T1 and the load machine T2 are thus given as:
T1 =
N J1
J
T2 = N 2
MN
MN
(3)
The stiffness time constant Tc and internal damping of the shaft d can be calculated as
follows:
Tc =
D
MN
d= N
MN
Kc N
(4)
Taking into account the equations (3)-(5) the state equation of the two-mass system in perunit value is represented as:
d
dt
1 (t )
(t )
2
m s (t )
T1
d
=
T2
1
T c
d
T1
d
T2
1
Tc
1
1
T 1 (t ) T 1
1
1
2 (t ) + 0
T2
m s (t ) 0
0
1m e
T 2 m L
0
(5)
Usually, due to its small value the internal damping of the shaft d is neglected in the
analysis of the two-mass drive system.
3. Adaptive control structure

A typical electrical drive system is composed of a power converter-fed motor coupled to a
209
mechanical system, a microprocessor-based controllers, current, rotor speed and/or position

sensors used as feedback signals. Typically, cascade speed control structure containing two
major control loops is used, as presented in Fig 1.
Fig. 1. The classical cascade control structure of the two-mass system

The inner control loop performs a motor torque regulation and consists of the power
converter, electromagnetic part of the motor, current sensor and respective current or torque
controller. As this control loop is designed to provide sufficiently fast torque control, it can be
approximated by an equivalent first order term with small time constant. If the control is
ensured, the driven machine could be an AC or DC motor, with no difference in the outer
speed control loop. The outer loop consists of the mechanical part of the motor, speed sensor,
speed controller, and is cascaded to the inner loop. It provides speed control according to the
reference value (Szabat & Orlowska-Kowalska, 2007).
Such a classical structure in not effective enough in the case of the two-mass system. To
improve the dynamical characteristics of the drive, the modification of the cascade structure
is necessary. In this paper the structure with the state controller which allows the free
location of the closed-loop poles is considered. So it requires the additional information of
the shaft torque and the load speed. The parameters of the control structures are set using
pole-placement methods, with the methodology presented in (Szabat & Orlowska-Kowalska,
2007), according to the following equations:
(6)
k I = T1T2Tc 04
k1 = 4T1 r 0
1
1
k 2 = T1Tc 2 02 + 4 r2 02
T
T
T
2 c
1Tc
k 3 = k1 02T2Tc 1
(7)
(8)
(9)
where: r - required damping coefficient, 0- required resonant frequency of the system.

In the industrial applications, the direct measurement of the shaft torque ms and the load
speed 2 is very difficult. For that reason, in this paper the Nonlinear Extended Kalman
Filter (NEKF) is used to provide the information about non-measurable mechanical state
variables. Additionally, the time constant T2 of the load side is also estimated and used to
on-line retuning the control structure parameters, according to Eq. (6)-(9). The estimated
Adaptive control
210
value of T2e is also used to change the element q55 of the covariance matrix Q in the way
presented in the next section (Eq. (21)). The considered control structure is presented in Fig.
2. The proposed adaptive control structure ensure the desired characteristic of the drives
despite the changes of the time constant of the load machine.
Fig. 2. The block diagram of the state-feedbacks adaptive control structure
3. Mathematical model of the nonlinear extended Kalman filter (NEKF)

In the presence of the time-varying load machine inertia T2, there is a need to extend the
two-mass system state vector (1) with the additional element 1/T2 and non-measurable load
torque mL :
T
1
(t ) .
x R (t ) = 1 (t ) 2 (t ) m s (t ) m L (t )
T2
(10)
The extended, nonlinear state and output equations can be written in the following form:
1
d
x R (t ) = A R (t )x R (t ) + B R u(t ) + w(t ) = f R (x R (t ), u(t )) + w(t )
dt
T2
y R (t ) = CR x R (t ) + v(t )
where matrices of the system are defined as follows (in [p.u.]):
(11a)
(11b)
0
1

A R
(t ) = 1
T2

Tc
0
0
0
1
Tc
0
0
1
T1
1
(t )
T2
1
(t )
T2
1
0
T
0
1

0
C R = 0
0
BR = 0

0
0
0
0
0
0

211
(12)
and w(t), v(t) - represent process and measurement errors (Gaussian white noise), according
to the Kalman Filter (KF) theory.
The matrix AR depends on the changeable parameter T2. It means that in every calculation
step this matrix must be updated due to the estimated value of T2. The input and the output
vectors of the drive system (and NEKF) are electromagnetic torque and motor speed
respectively:
u = m e y = 1
(13)
After the discretization of Eq. (11) with Tp sampling step, the state estimation using NEKF
algorithm is calculated:
x R (k + 1 / k + 1) = x R (k + 1 / k ) + K (k + 1)[y R (k + 1) CR (k + 1)x R (k + 1 / k )]
(14)
where the gain matrix K is obtained by the suitable numerical procedure.

In the first step the estimation of the filter covariance matrix is calculated:
P(k + 1 / k ) = FR (k )P(k )FRT (k ) + Q(k )
(15)
where:
FR (k ) =
f R x R ( k / k ), u ( k ), k
x ( k / k )
(16)
x R = x R k / k
and Q is a state noise covariance matrix. FR is the state matrix of the nonlinear dynamical
system (11) after its linearization in the actual operating point, which must be updated in
every calculation step:
F R (k ) = 1
Tp
Tc
0
0
1
1
Tp
T
c
0
0
1
Tp
T1
1
(k )T p
T2
1
(k )T p
T2
T p (m s (k ) m L (k ))
(17)
Adaptive control
212
The filter gain matrix K of the NEKF and the update of the covariance matrix of the state
estimation error P are calculated using the following equations:
K (k + 1) = P(k + 1 / k )C TR (k + 1) C R (k + 1)P(k + 1 / k )C TR (k + 1) + R (k )
P(k + 1 / k + 1) = [I K(k + 1)CR (k + 1)]P(k + 1/ k )
(18)
(19)
where: R the output noise covariance matrix.

The quality of the state estimation depends on the suitable choice of the covariance matrices
Q and R. However, according to the technical literature, the analytical guidelines which
ensure proper setting of these matrices do not exist. Usually the trial and error procedure is
used. However, this process is time-consuming and does not ensure the optimal
performances of NEKF. In this paper elements of covariance matrices have been set using
the genetic algorithm (Szabat & Orlowska-Kowalska, 2008), with the following cost function:
j
j
j
j
F = ms mse 2 2e mL mLe T2 T2e

1
1
1
1
(20)
where: ms, 2, mL, T2 real variables and parameter of the two-mass system; mse, 2e , mLe, T2e
estimated variables and parameter, j total number of samples. The cost function defined
in this way ensures the optimal setting of covariance matrices Q and R for changeable time
constant of the load machine.
4. Simulation results
4.1 Open-loop system
In simulation tests the estimation quality of all system state variables is investigated. The
shaft torque and the load speed are taken for the closed-loop structure with the direct
feedback from system state variables (Fig.1). The electromagnetic torque and the motor
speed, used as the input and output vectors of NEKF, are disturbed with white noises. In
Fig. 3. the transients of the electromagnetic torque and motor speed are presented.
a)
b)
Fig. 3. Transients of the electromagnetic torque (a) and the motor speed (b)
213
The drive system works in the reverse condition with the electromagnetic torque limit set to
3 [p.u.] in the considered case is tested. The state estimator working outside the control
structure is tested. The transients of all the real and estimated variables and theirs
estimation errors are demonstrated In Fig 4.
The NEKF starts work with a misidentified value of the time constant of the load machine
(initial value of the T2 is set to101.5ms Fig 4.g). Then at the time t1=2s the time constant of
the load machine T2 and the load torque mL begin to change (Fig. 4c,g). Those two variables
vary in a smooth sinusoidal way. The NEKF estimates all the system states simultaneously.
As can be seen from Fig. 4, the transients of all estimates contain high-frequency noises. The
steady state level of the estimation error is about 0.02 (Fig. 4e) for the load speed and about
0.10 (Fig. 4e) for the shaft torque. The biggest errors exist in the transients of the load torque
and of time constant of the load machine (Fig. 4h). The initial estimation error of T2, cause by
the misidentified value of the time constant of the load machine is eliminated after 500ms.
The typical disruptions can be seen in the estimated transient. They appear when the
direction of the motor speed is rapidly changed. The characteristic feature of the NEKF is
the fact that the estimation of the time constant of the load machine is only possible when
the load speed is changing. Therefore, the biggest estimation errors occurs when the time
constant of the load side is varied and the load speed is constant (Fig. 4g,h). The next NEKF
feature is that the estimate of the T2 contains bigger frequency noises in the case when the
real value of the T2 is larger. Because the load torque and time constant of the load machine
have been varied in a smooth way good estimation accuracy has been achieved in the
simultaneous estimation of all the states.
a)
b)
c)
d)
e)
f)
g)
h)
Fig. 4. Transients of the real and estimated state variables and their estimation errors: load
speed (a,d) shaft torque (b,e), load torque (c,f) and time constant of the load side (g,h)
214
Adaptive control
Then the case of the rapid changing of the load torque and time constant of the load
machine is considered. The input (electromagnetic torque) and output vector (motor speed)
of the NEKF are presented In Fig. 5. As the previously the drive is working under reverse
condition and the limit level of the electromagnetic torque is also set to 3 [p.u.]. The
electromagnetic torque and the motor speed are disrupted by white noises, which emulate
the measurements noises. The real and estimated variables and their estimation errors for
rapid changes of the load torque and the load side inertia are presented in Fig. 6.
Similarly as in the previous case, the drive system starts working with a misidentified time
constant of the load machine T2=101.5ms (Fig. 6g). Then at the time t=1s and 3s the time
constant of the load machine and the load torque change rapidly (Fig. 6c,g). Next, at the time
t=5, 6 and 8s only the load torque and at the time t= 4, 6.5 and 8.5s only the time constant of
the load machine vary quickly. The following work cycle allows to examine the quality of
the variables estimation under different conditions. The average level of the estimation error
is about 0.014 (Fig. 6e) for the load speed and about 0.06 for the shaft torque (Fig. 6f).
However, the simultaneous alternation of the load torque and time constant of the load
machine bring about the rise of the big, quickly damped estimation errors of the load speed
(Fig. 6b) and shaft torque (Fig. 6d). A single change of the above-mentioned variables cause
the increase of the estimation errors, but for a smaller extent than in the pervious case. The
last two estimated variables, i.e. the load torque and the time constant of the load machine
depend on each other significantly. The rapid change of one variable brings about a
significant increase of the estimation error of the other variable (Fig. 6f,h).
a)
b)
Fig. 5. Transients of the electromagnetic torque (a) and the motor speed (b)
Similarly as in the previous case, the drive system starts working with a misidentified time
From the transients presented in Fig. 4 and Fig. 6 the following remarks can be formulated:
-the estimation of the time constant of the load machine is possible only when the motor
speed is changing;
-the estimates of the load torque and the time constant of the load machine are correlated:
the change of the load torque causes the rise of the error of the load machine time constant
and vice versa. This is especially clearly visible in the transient presented in Fig. 6;
-the noise level of the of the estimated load machine time constant of the strictly depends on
the actual value of the real time constant and the value of the covariance matrix element q55;
when of the value of the T2 is smaller, the element q55 should have a bigger value and vice
versa.
215
The dynamic characteristics of the non-adaptive NEKF strictly depends on the proper
setting of the covariance matrix values. In the case of the changeable time constant of the
load machine the element q55 is a compromise between the slow covariance for a small value
of T2 and a large noise level when value of T2 is big. The modification of the estimating
procedure is related to this feature. Because the noise level in the estimated variable
depends on the real value of the T2, the NEKF with the changeable element q55 of the
correlation matrix Q is proposed. The element q55 adopts to the estimating value of the time
constant of the load machine according to the following formula:
T
q 55 = q 55 N 2 N
T2 e
(21)
where: q55N- the value of q55 selected for the nominal parameters of the drive (using genetic
algorithm), T2N nominal time constant of the load machine, T2e estimated time constant
of the load machine, n power factor.
a)
d)
g)
b)
c)
e)
f)
h)
Fig. 6. Transients of the real and estimated state variables and their estimation errors: load
speed (a,d)shaft torque (b,e), load torque (c,f) and time constant of the load side (g,h)
Adaptive control
216
Then the adaptive NEKF is tested under the same conditions as previously but with the
adaptation formula (21). Because the biggest difference is visible in the time constant of the
load machine only the transients of those variables are presented below. In Fig 7 the
transients for smooth (case 1- a) and rapid (case 2- b) changes of the load torque and time
constant of the load machine for power factor n=3 are presented.
The difference between the non-adaptive and adaptive NEKF algorithm is clearly visible
when the Fig. 4, 6 and 7 are compared. The estimate of T2 has a smaller estimation error and
noise level than for the non-adaptive NEKF. The rapid changing of the load torque does not
influence the estimate of T2 so significantly as in the previous non-adaptive NEKF case.
Also the estimate of the load torque has better accuracy in the adaptive NEKF case.
Similarly, the fast variation of the time constant of the load machine causes a smaller error in
the estimate of load torque in the adaptive NEKF.
a)
b)
Fig. 7. Transients of the real and estimated time constant of the load side for the adaptive
NEKF with power factor n=3, case-1 (a), case-2 (b)
In order to compare the performance of the non-adaptive and adaptive NEKFs, the
estimation errors of all estimated have been calculated using of the following equation:
N
vv
i =1
(22)
where: N total number of samples, real variable,

e estimating variable.
The estimation errors of all state variables for non-adaptive (n=0) and adaptive NEKF (n =3) are
presented in the Table 1.
ms
T2
mL
Case 1
n=0
0.0092
0.0456
0.0180
0.0942
Case 1
n=3
0.0086
0.0442
0.0159
0.0907
Case 2
n=0
0.0140
0.0605
0.0301
0.1073
Case 2
n=3
0.0123
0.0570
0.0224
0.0975
217
Table 1. The estimation errors of the state variables for the case 1 and case 2 for the adaptive
and non-adaptive NEKF
The application of the adaptation mechanism decreases the estimation error in all estimated
variables. This feature is especially evident when the time constant of the load machine and
the load torque change rapidly (case -2). For instance, the application of the adaptation
mechanism ensures the reduction of estimation error of the T2e by approximately 25%.
3.2 Closed-loop system
First, the effectiveness of the proposed control structure has been investigated in the
simulation study. The non-measurable state variables, e.g. shaft torque, load speed and load
torque, are delivered to the control structure by the NEKF.
a)
b)
c)
d)
e)
f)
g)
h)
i)
Adaptive control
218
j)
k)
l)
Fig. 8. Transients of the electromagnetic torque (a), motor speed (d), real and estimated state
variables and their estimation errors: load speed (b,e), shaft torque (c,f), load torque (g,j),
time constant of the load machine (h,k) adaptive control structure parameters (i,l) in the
control structure with simultaneous estimation of the time constant of the load machine and
load torque
The estimated time constant of the load machine is used in the adaptation law in order to
retune the control structure coefficients in accordance with (6)-(9). The adaptation formula
(21) is used to improve the NEKF performance. However, in order to ensure the stable work
of the control structure the coefficients of the covariance matrices are decreased in
comparison to the previous section. The desired values of the resonant frequency of the
system and the damping coefficient are 0=45s-1 and r=0.7 respectively. The transients of
the system states as well as the control structure coefficient are presented In Fig 8.
a)
b)
c)
d)
e)
f)
g)
h)
i)
j)
k)
l)
219
Fig. 9. Transients of the electromagnetic torque (a), motor speed (d), real and estimated state
variables and their estimation errors: load speed (b,e), shaft torque (c,f), load torque (g,j),
time constant of the load machine (h,k) adaptive control structure parameters (i,l) in the
control structure with modified estimation algorithm
The system starts work with a misidentified value of the time constant of the load machine
T2e =101ms (Fig. 8h) which results oscillations in the estimated load torque transient. Despite
this no visible oscillations appear in the transients of the load speed. After 2s, the estimate of
the time constant of the load machine reaches its real value. The rapid changing of the load
torque causes the oscillations in the estimate of T2e which are noticeable visible at the time
t=9s. Still, a such big estimation error can not be accepted in the high performance drive
system
In order to improve the control structure performance the following modifications of the
standard NEKF algorithm improving the quality of the estimation have been implemented.
Firstly, the estimation of the time constant T2 is active only when the motor speed is
changing. Secondly, during this time the estimation of the load torque mL is blocked. In the
NEKF algorithm the last estimated value of the mL is used. Also, when motor speed is not
changing, the estimate of T2 is stopped and the estimate of the mL becomes active. During
this time, the last estimated value of the time constant T2 is utilized in the algorithm. This
modification allows to increase the values of the covariance matrices of the NEKF.
All system states are reconstructed well and their estimation errors are very small and do
not influence the system dynamics negatively (Fig. 9). The time constant of the load machine
is estimated accurately with a small steady-stay error. The moments when the estimate of
mLe is stopped are visible in the load torque transient (Fig. 9g). Thus, the adaptive system
with adaptive NEKF work properly.
Adaptive control
220
5. Experimental results
All theoretical considerations have been confirmed experimentally in the laboratory set-up
composed of a 0.5kW DC-motor driven by a static converter. The motor is coupled to a load
machine by an elastic shaft (a steel shaft of 5mm diameter and 600mm length). The speed
and position of the driven and loading motors have been measured by incremental encoders
(36000 pulses per rotation). The mechanical system has a natural frequency of
approximately 9.5Hz. The nominal parameters of the system are T1=203ms, T2=203ms, Tc
=2.6ms. The picture of the experimental set-up is presented in Fig. 10.
a)
b)
Fig. 10. The mechanical part of the laboratory set-up (a) and the general view of the
laboratory set-up (b)
a)
b)
c)
d)
e)
f)
221
Fig. 11. Real transients of the: motor and load speeds (a), real and estimated load speeds and
its estimation error (b), electromagnetic and estimated shaft and load torque (c), estimated
time constant of the load side (d), control structure parameters (e,f) for the reference value
of the speed r=0.5
First the performance of the drive system has been tested for the nominal value of the time
constant of the load machine T2=0.203s. The electromagnetic torque limit has been set to 2.
a)
b)
Adaptive control
222
c)
d)
e)
f)
Fig. 12. Real transients of the: motor and load speeds (a), real and estimated load speeds and
its estimation error (b), electromagnetic and estimated shaft and load torque (c), estimated
time constant of the load side (d), control structure parameters (e,f) for the reference value
of the speed r=1
223
The system works with the reference value of the speed set to 0.5. According to the
adaptation procedure described in the previous section during start-up the estimate of the
mLe is blocked and the estimate of the T2e is activated which is observable in Fig. 11c,d. When
the control error decreases below 0.05, the estimate T2e is blocked and the mLe. At the time
t1=0.4s the nominal load torque is applied to the system. This affects the system speed in a
negative way and some disruption is visible in its transients. The load torque is switched off
at the time t2=0.8s and the non-zero value of the estimate of the mLe comes from the friction
torques. At the time t3=1s the system begins to reverse. When the value of the system speed
is negative, no external torque is applied to the system. The drive reverses again at the time
t4=2s and then the work cycle is repeated. Clearly, the adaptive control structure with the
NEKF works properly. The load speed as well as the time constant of the load machine are
estimated with small errors. The transients of the control structure parameters are presented
in Fig. 11 e,f. They vary (except k1) with the estimated value of the T2e.
Next the control structure with the electromagnetic torque limit set to 3 has been examined.
The work cycle is identical as previously. But the reference speed is set to the nominal value.
The transients of the system are presented in Fig. 12.
Similarly as before, the initial value of the time constant of the load machine is set to
T2e=0.1015s. After the start-up it reaches its real value almost without an error. During the
next reversal the estimate of the T2 oscillates around the real value. However, it should be
pointed out that the estimation error does not exceed a few percent of the real value. The
estimate of the T2 is reconstructed very well. Small errors appear in its transient during the
time when the load torque is switched on and off and during the reversal. The adaptive
control structure with the state controller works in a stable way.
6. Conclusion
In order to damp the torsional vibrations, which could destroy the mechanical coupling
between the driven and loading machine, the control structure with state controller is
applied. The control structure coefficients depend on the time constant of the load side
machine. In the case of the system with changeable load side inertia, there is a need to
estimate this parameter and adapt the control structure gains in accordance with the actual
estimated value. The application of the adaptive control structure ensures the required
transient of the load speed despite the changeable load side inertia. In order to use the
adaptive control structure, there is a need to choose a state estimator, which has to estimate
the non-measurable system state variables and changeable parameters of the system. In this
paper, the non-adaptive and adaptive nonlinear extended Kalman filter (NEKF) is tested.
Parameters of the covariance matrices Q and R are selected using the genetic algorithm with
special cost function. The application of the global optimization technique allows to reach
the global solution according to the defined cost function. However, the application of the
genetic algorithm is possible only as an off-line process due to a long calculation time. To
ensure the optimal values of the covariance matrix Q, despite the load side parameter
changes, the adaptation mechanism is developed. The suitable on-line change of the
covariance matrix element q55 is proposed, according to the estimated value of the load side
time constant. It is proved by simulation and experimental tests that the proposed control
structure is effective for damping the torsional oscillation of two-mass drive system, also in
224
Adaptive control
the case of wide range changes of load side inertia.
7. References
Beineke, S., Schtte, F. & Grotstollen H. (1997). Comparison of Methods for State Estimation
and On-Line Identification in Speed and Position Control Loops, Proc. of the Intern.
Conf. European Power Electronics EPE97, pp. 3.364-3.369, Norway.
Cychowski M. T., Delaney K. and Szabat K. (2008), Explicit Model Predictive Control of the
Drive System with Elastic Coupling, Proc. of 13th International Conference on Power
Electronics and Motion Control EPE-PEMC 2008, on. cd, Poland
Erbatur, K., Kaynak, O. & Sabanovic A. (1999). A Study on Robustness Property of Sliding
Mode Controllers: A Novel Design and Experimental Investigations, IEEE
Transaction on Industrial Electronics, Vol. 46, No. 5 , pp. 1012-1018.
Erenturk, K. (2008). Nonlinear two-mass system control with sliding-mode and optimised
proportional and integral derivative controller combined with a grey estimator,
Control Theory & Applications, IET, Vol. 2, No. 7, pp. 635 642.
Ellis, G. & Lorenz, R.D. (2000), Resonant load control methods for industrial servo drives.
Proc. of the IEEE Industry Application Society Annual Meeting, pp. 1438-1445.
Ferretti, G., Magnoni, G. A. & Rocco, P. (2004). Impedance Control for Elastic Joint Industrial
Manipulators, IEEE Trans. on Robotics and Automation, Vol. 20, pp. 488-498.
Ferretti, G., Magnoni, G. A., Rocco, P., Vigano, L. & Rusconi, A. (2005). On the Use of Torque
Sensors in a Space Robotics Application, : Proc. on the IEEE/RSJ International
Conference on Intelligent Robots and Systems IROS 2005, pp. 1947- 1952, Canada.
Gawronski, W., Racho, C. S. & Mellstrom, J. A. (1995). Application of the LQG and
Feedforward Controllers to the Deep Space Network Antennas, IEEE Trans. on
Control System Technology, Vol. 3, No. 4, pp. 417-421.
Gu D. W., Petkov P. H., Konstantinov M. M. (2005). Robust Control Design with Matlab,
Springer.
Hace. A., Jezernik, K. & Sabanovic, A. (2005). Improved Design of VSS Controller for a
Linear Belt-Driven Servomechanism, IEEE/ASME Trans. on Mechatronic, Vol. 10, No.
4, pp. 385-390.
Hirovonen, M., Pyrhonen, O. & Handroos H. (2006). Adaptive nonlinear velocity controller
for a flexible mechanism of a linear motor, Mechatronic, Elsevier, Vol. 16, No. 5, pp.
279-290.
Hori, Y., Sawada, H. & Chun, Y. (1999). Slow resonance ratio control for vibration suppression
and disturbance rejection in torsional system, IEEE Trans. on Industrial Electronics, Vol.
46, No. 1, pp. 162-168.
Horwitz, R., Li, Y., Oldham, K., Kon, S. & Huang, X. (2007), Dual-stage servo systems and
vibration compensation in computer hard disk drives, Control Engineering Practice,
Vol. 15, pp. 291-305.
Huang, A.,C. & Chen, Y., C. (2004). Adaptive Sliding Control for Single-Ling Flexible-Joint
Robot With Mismatched Uncertainties, IEEE Trans. on Control System Technology,
Vol. 12, pp. 770-775.
Itoh D., Iwasaki M., Matsui N. (2004). Optimal Design of Robust Vibration Suppression
Controller Using Genetic Algorithms, IEEE Transaction on Industrial Electronics, Vol.
51, No. 5, pp. 947-953.
225
Ji, J. K. & Sul, S. K. (1995). Kalman Filter and LQ Based Speed Controller for Torsional
Vibration Suppression in a 2-Mass Motor Drive System, IEEE Trans. on Industrial
Electronics, Vol. 42, No. 6, pp. 564-571.
Katsura, S. & Ohnishi, K. (2005). Force Servoing by Flexible Manipulator Based on
Resonance Ratio Control, Proc. of the IEEE International Symposium on Industrial
Electronics ISIE 2005, pp. 1343-1348, Croatia.
Michels, K., Klawonn, F., Kruse, R. & Nrnberger, A. (2006). Fuzzy Control Fundamentals,
Stability and Design of Fuzzy Controllers, Springer.
Ohno, K. & Hara, T. (2006). Adaptive Resonant Mode Compensation for hard Dick Drives, ,
IEEE Trans. on Industrial Electronics, Vol. 53, No. 2, pp. 624-629.
Orlowska-Kowalska, T. & Szabat, K. (2008). Damping of Torsional Vibrations in Two-Mass
System Using Adaptive Sliding Neuro-Fuzzy Approach, IEEE Transactions on
Industrial Informatics, Vol. 4, No. 1, pp. 47-57.
OSullivan, T., Bingham, C. C. & Schofield, N. (2007), Enhanced Servo-Control Performance
of Dual-Mass System, IEEE Trans. on Ind. Electronics, Vol. 54, No. 3, pp. 1387-1398.
Qiao, R., Zhu, Q. M., Li, S. Y. & Winfield, A. (2002). Torsional Vibration Suppression of a 2Mass Main Drive System of Rolling Mill with KF Enhanced Pole Placement, Proc. of
the 4th World Congress on Intelligent Control and Automation, pp. 206-210, China.
Shen, B. H. & Tsai, M. C. (2006) Robust dynamic stiffness design of linear servomotor drives,
Control Engineering Practice, Vol. 14, pp. 1325-1336.
Sugiura, K. & Hori, Y. (1996). Vibration Suppression in 2- and 3-Mass System Based on the
Feedback of Imperfect Derivative of the Estimated Torsional Torque, IEEE Trans. on
Industrial Electronics, Vol. 43, No. 2, pp. 56-64.
Suh, G., Hyun, D. S., Park, J. I., Lee, K. D. & Lee, S. G. (2001), Design of a Pole Placement
Controller for Reducing Oscillation and Settling Time in a Two-Inertia System,
Proc. of 24th Annual Conference of the IEEE Industrial Electronics Society IECON01,pp.
1439-1444, USA.
Szabat, K. & Orowska-Kowalska, T. (2007). Vibration Suppression in Two-Mass Drive
System using PI Speed Controller and Additional Feedbacks Comparative Study,
IEEE Trans. on Industrial Electronics, Vol. 54, No. 2, pp.1193-1206.
Szabat, K. & Orlowska-Kowalska, T. (2008). Performance Improvement of Industrial Drives
With Mechanical Elasticity Using Nonlinear Adaptive Kalman Filter, IEEE
Transactions on Industrial Electronics, Vol. 55, No. 3, pp. 1075-1084.
Valenzuela, M. A., Bentley, J. M. & Lorenz, R. D. (2005). Evaluation of Torsional Oscillations
in Paper Machine Sections, IEEE Trans. on Industrial Applications, Vol. 41, No. 2, pp.
493-501.
Vukosovic, S., N. & Stojic, M. R., (1998). Suppression of Torsional Oscillations in a HighPerformance Speed Servo Drive, IEEE Trans. on Industrial Electronic, Vol. 45, No. 1,
pp. 108-117.
Wertz H., Beineke S., Frhleke N., Bolognani S., Unterkofler K., Zigliotto M. & Zordan M.
(1999) Computer Aided Commissioning of Speed and Position Control for
Electrical Drives with Identification of Mechanical Load, Proc. of the Thirty-Fourth
IAS Annual Meeting Industry Applications Conference, pp. 4.1372-4.2379, USA.
Wang L., Frayman Y. (2002). A Dynamically Generated Fuzzy Neural Network and its
Application to Torsional Vibration Control of Tandem Cold Rolling Mill Spindles,
Engineering Applications of Artificial Intelligence, Vol.15, No. 6, pp. 541-550.
226
Adaptive control
Zhang, G. & Furusho, J. (2000). Speed Control of Two-Inertia System by PI/PID Control,
IEEE Trans. on Industrial Electronics, Vol. 47, No. 3, pp. 603-609.
10
Adaptive Control of Dynamic Systems with
Sandwiched Hysteresis Based on Neural
Estimator
Yonghong Tan1, Ruili Dong1,2 & Xinlong Zhao3
1. Shanghai Normal University 2. Shanghai Jiaotong University &

3. Zhejiang Sci-Tech University
China
1. Introduction
The so-called Sandwich system with hysteresis is a class of systems in which a hysteretic
subsystem is sandwiched between two smooth dynamic blocks. In engineering, many
practical processes can be considered as the sandwich systems with hysteresis. In the
following, two typical examples will be presented.
1.1 Ultra-precision moving positioning stage
A typical ultra-precision moving positioning stage is often used in ultra-precision
manufacturing system for its nanometer displacement and fast linear moving speed.
Usually, such platform consists of electric amplifiers, piezoelectric actuators and loads. As
hysteresis is inherent in piezoelectric actuator, the amplifier and load can be considered as
smooth dynamic subsystems. Therefore, this platform can be considered as a typical
sandwich system with hysteresis. Fig.1 shows the architecture of such system.
Fig. 1. Architecture of ultra-precision moving stage with piezoelectric actuator

1.2 Mechanical Transmission System
Mechanical transmission system often exists in machine tools or many other mechanical
systems. A typical mechanical transmission system is shown in Fig.2. In this system, the
servomotor is used to drive a gearbox connected with a mechanical work platform through
a screw. In this system, u is the servomotor angle, x is the angle of the gearbox, and y is the
displacement of the work platform. The servomotor and the work platform can be
considered as smooth dynamic subsystems. However, the gearbox and screw in this system
is a typical hysteresis nonlinearity due to the tear and wear of the gear teeth. Obviously, this
mechanical system can be described by the sandwich system with hysteresis.
228
Adaptive Control
Fig. 2. Mechanical transmission system

Although, sandwich systems with hysteresis often exist in engineering practice, there are
only several research reports found on the control of them. Taware & Tao (1999) presented
an analysis on the control of such systems with backlash-type hysteresis. Tao & Ma (2001)
proposed an optimal control for the systems with sandwiched backlash. In their methods, an
optimal control scheme is employed for backlash compensation. Then, the nonlinear
feedback control law is used for the control of nonlinear dynamics. Zhao & Tan (2006)
proposed a neural adaptive control for sandwich systems with hysteresis. The neural
network based hysteresis compensator is developed to compensate for the effect of the
hysteresis. Furthermore, Zhao et. al. (2007) presented an adaptive control strategy for
sandwich systems with dynamic hysteresis based on Duhem hysteretic operator. Corradini
et. al. (2007) proposed a variable structure control of nonlinear uncertain sandwich systems
with hysteretic block. Therefore, the control of sandwich systems with hysteresis has
become one of the interesting topics in control engineering domain.
It is known that the existence of hysteresis in actuators often leads to oscillation and
undesirable inaccuracy. Therefore, the main purpose of design a control scheme for
sandwich system with hysteresis is to eliminate the side effect of hysteresis inherent in the
system and force the system to track the reference trajectory. Note that hysteresis is a nondifferentiable nonlinear system with multi-valued mapping. Moreover the structure of the
sandwich system is rather complex. Hence, it is not easy to construct a compensator for the
hysteresis in such system. Therefore it is a real challenge to develop a control strategy for the
dynamic systems with sandwiched hysteresis.
In this chapter, a mathematical description of the sandwich systems with hysteresis will be
described in section 2. Then, in section 3, the control architecture for the sandwich systems
with hysteresis will be illustrated. In this architecture, a neural network based inverse model
is constructed to cancel the effect of the first dynamic block of sandwich system. Then, the
sandwich system can be transformed to a nonlinear system preceded by hysteresis which
can be described by a Hammerstein model with hysteresis. In Section 4, a neural network
based estimator will be developed in terms of a proposed expanded input space with
hysteretic operator. The developed neural hysteretic estimator can be used to compensate
for the system residual caused by the effect of hysteresis. Section 5 will present an adaptive
control strategy based on pseudo inverse control technique for the obtained Hammerstein
system with hysteresis. One of advantages of the controller is that it does not need to
construct the hysteresis inverse to cancel hysteretic effect. The neural control strategy and
the corresponding adaptive law based on the Lyapunov stability theory will be developed.
Adaptive Control of Dynamic Systems with Sandwiched Hysteresis Based on Neural Estimator
229
Furthermore, Comparison of the simulation results between the proposed method and the
PID control strategy will be illustrated in Section 6. Section 7 will present the remarks and
conclusions of this Chapter.
2. Mathematical Description of Sandwich Systems with Hysteresis

The structure of the sanwich system with hysteresis is shown in Fig.3. Suppose the
nonlinear single-input-single-output (SISO) system with sandwiched hysteresis is described
by
Li : fi [v ( n ) , v ( n 1) ,L , v (1) , v, r ( m ) , r ( m 1) ,L , r (1) , r ] = 0
(1)
(n)
(m)
where r is the input, v is the output, v is the n-th order derivative of v , r is m-th order
derivative of r , m and n ( m n ) are the orders of the input and output respectively.
H : u = H (v )
(2)
where H presents the hysteresis nonlinearity.
x&1 = x2
x& = x
3
2
Lo : L
x& = x
n
n 1
x& n = f o ( x ) + go ( x )u
(3)
and
y = x1
(4)
where x = [ x1 , x2 ,L , xn ] is the system state vector, u is the input, y is the output, v is the
control input and u is the actuator output. It is assumed that f o ( x ) and go ( x ) are sufficiently
T
smooth but unknown functions and satisfy
fo
u
0 and
g o
u
0 . Moreover, assume that f o is
invertible. Notation H [.] denotes that the hysteresis nonlinearity is not dependent on an
instantaneous value v(t ) but the trajectory, v(t ) C 0 [0, t ] . Assume that all the control and
input variables, i.e. v ( n ) , v ( n 1) ,L , v (1) , v, r ( m ) , r ( m 1) ,L , r (1) , r are known.
Fig. 3. The structure of sandwich system with hysteresis
Adaptive Control
230
3. Control Architecture for Sandwich System with Hysteresis

From Fig. 3, it is known that the architecture of the sandwich system with hysteresis is
rather complex. It would be convenient for us to design a control strategy for such system if
we could find a method to simplify the structure of the system. In this section, a control
architecture for the sandwich system will be discussed. In this architecture, a neural
networks (NN) based inverse system Li 1 will be constructed. By connecting the NN based
inverse with the system Li can form an approximate pseudo-linear unit compensator which
leads to Li 1 Li 1 . Then the sandwich system can be transform to a pseudo-linear unit
system connected with a nonlinear system preceded with hysteresis nonlinearity which is
shown in Fig.4. The obtained the system can be considered as a Hammerstein System with
hysteresis.
^ -1
Li
Li
Lo
Fig. 4. The sandwich system with a pseudo-linear unit compensation

With the above-mentioned NN based inverse, the effect of Li would be cancelled. So we can
design the controller for the system Lo preceded by a hysteresis nonlinearity. Usually, the
model uncertainty of the neural network based compensator exists. That implies the NN
based compensator cannot completely compensate for the effect of Li . Therefore, a model
residual should be added to system Lo . That is Li 1 Li = 1 + , where is a bounded modeling
error. Hence, the obtained system preceded by a hystersis can be described as follows:
H : u = H (v ) ,
x&1 = x2
x& = x
3
2
Lo : L
,
x& = x
n
n 1
&
xn = f o ( x ) + go ( x )u +
(5)
(6)
and
y = x1 .
(7)
The control objective is to design a control law v (t ) to force y (t ) , the plant output, to track a
smooth prescribed trajectory yd (t ) with an acceptable accuracy. The desired state vector is
defined as xd (t ) = [ yd , y& d ,L yd( n 1) ]T where yd( n 1) is the (n 1)th order derivative. Moreover, the
tracking error vector is defined as e = x xd . It is assumed that the desired states are
231
bounded, i.e. xd X d . Moreover, denotes bounded disturbance caused by NN based

inverse, where N , and N > 0 .
Define the filtered tracking error as
= [1 , 2 L n 1 ,1]e = [ T ,1]e
(8)
T
where = [1 , 2 L n 1 ] is a parameter vector to be designed. Suppose
s n 1 + n 1 s n 2 + L + 1
is Hurwitz. Differentiating (8) and using (6), it results in
& = x&n ydn + [0, T ]e = fo ( x) + go ( x)u ydn + [0, T ]e +
(9)
As u is the output of hysteresis which is usually unknown, an invertible function f ( x, v) is

introduced to approximate f o ( x ) + go ( x )u . Adding and subtracting f ( x, v) to and from the
right hand side of (9), it yields
& = + f o ( x ) + go ( x )u f ( x, v ) yd( n ) + [0, T ]e +

= + F ( x, u ) f ( x, v ) yd( n ) + [0, T ]e +
= + f% ( x, v, u ) y ( n ) + [0, T ]e +
(10)
where = f ( x, v ) is the so called pseudo-control (Calis & Hovakimyan, 2001) and

(Hovakimyan & Nandi ,2002), F ( x, u) = fo ( x) + go ( x)u and f% ( x, v, u) = F ( x, u) f ( x, v) is the
system residual. As f ( x, v) is invertible with respect to v and satisfies (Calis & Hovakimyan,
2001):
F u
f
,
= sgn
u v
v
(11)
f 1 F u
>
>0.
v 2 u v
(12)
1. sgn
and
2.
In order to design the corresponding control strategy, the approximation of the nonlinear
residual f% ( x, v, u ) is required. Neural networks would be one of the recommended
alternatives to model this residual. However, f% ( x, v, u ) involves the characteristic of
Adaptive Control
232
hysteresis, the traditional nonlinear identification methods such as neural modeling

technique usually cannot be directly applied to the modeling of it since the hysteresis is a
non-linearity with multi-valued mapping (Adly & Abd-El-Hafiz, 1998). In Section 4, we will
present a method to construct the neural estimator for f% ( x, v, u ) to compensate for the effect
of hysteresis. Moreover, a corresponding adaptive control method based on the control
archieture stated-above will be illustrated in Section 5.
4. Neural Estimator for System Residual

In order to approximate the system residual, neural network can be considered as an
alternative. However, the system residual contains the characteristic of hysteresis which is a
system with multi-valued mapping. In this section, a hysteretic operator is proposed to
construct an expanded input space so as to transform the multi-valued mapping of
hysteresis into a one-to-one mapping (Zhao & Tan, 2008). Thus, the neural networks can be
used for modeling of hysteresis based on the expanded input space with the hysteretic
operator. The proposed hysteretic operator is defined as:
h( x) = (1 e
| x x p |
)( x x p ) + h( x p ) ,
(13)
where x is the current input, h( x ) is the current output, x p is the dominant extremum
adjacent to the current input x . h( x p ) is the output of the operator when the input is x p .
Lemma 1: Let x(t ) C ( R + ) , where R + = {t | t 0} and C ( R + ) are the sets of continuous
+
functions on R . If there exist two time instants t1 , t2 and t1 t2 , such that x(t1 ) = x(t2 ) ,
x(t1 ) and x (t2 ) are not the extrema, then h x ( t1 ) h x ( t2 ) .
Proof: For x(t ) decreases or increases monotonically, (13) becomes

( x xp )
]( x x p ) + h( x p ),
hin ( x ) = [1 e
h( x ) =
x xp
)( x x p ) + h( x p ),
hde ( x) = (1 e
h 'in ( x) = e
( x xp )
( x x p ) + [1 e
= 1 [1 ( x x p )] e
> 11 e
x xp
x xp
( x xp )
x& (t ) > 0
x& (t ) < 0
(14)
]
(15)
>0
Therefore, hin ( x) is monotonic. Similarly one can obtain that hde ( x) is monotonic. It is noted
that hin ( x) is obtained from hin 0 ( x) = (1 e x ) x ( x 0) . That means its origin moves
from (0, 0) to ( x p , h( x p )) . Similarly hde ( x) is obtained from hde 0 ( x) = (1 e x ) x ( x 0) . It
represents that its origin moves from (0, 0) to ( x p , h( x p )) . As hin 0 ( x) = hde 0 ( x) , it implies
233
that hin ( x) and hde ( x) are antisymmetric. Therefore it can be concluded that hin ( x) and hde ( x)
intersect only at extrumum point ( x p , h( x p )) . That is, if x(t1 ) and x (t2 ) are not the extrema,
x(t1 ) = x(t2 ) , then h x ( t1 ) h x ( t2 ) .

Remark: If both h( x ) and H [] are fed with the same input v(t ) , the curve of h[v (t )] exhibits
similarity to that of H [v (t )] such as ascending, turning and descending. Moreover,
since x(t1 ) = x (t2 ) , x(t1 ) and x(t2 ) are not the extrema, h x ( t1 ) h x ( t2 ) , the pair
(v(t ), h[v (t )]) will uniquely correspond to one of the output values of hysteresis H [v(t )] .
Lemma 2: If there exist two time instants t1 , t 2 and t1 t2 , such that h[ x(t1 )] h[ x(t2 )] 0 , then
x (t1 ) x (t 2 ) 0 .
Proof:
hin x ( t1 ) hin x ( t2 )
x ( t1 ) x ( t2 )
= k , k (0, + ) ,
(16)
and
x ( t1 ) x ( t2 ) =
hin x ( t1 ) hin x ( t2 )
k
(17)
It is clear that if hin x ( t1 ) hin x ( t2 ) 0 , then x(t1 ) x(t2 ) 0 . Similarly, it is obtained that
if hde x ( t1 ) hde x ( t2 ) 0 , then x(t1 ) x(t2 ) 0 .Thus, it leads to the following theorem, i.e.:
Theorem 1: For any hysteresis, there exists a continuous one-to-one mapping : R 2 R , such
that H [v(t )] = (v(t ), h[v(t )]) , where {v (t ), h[v(t )]} is an expanded input space with
hysteresis operator.
Proof: The proof can be divided into two cases, i.e.
Case 1: If v(t ) is not the extrema. Based on Lemma1, if there exist two time instants t1 , t 2
and t1 t2 , then (v ( t1 ) , h v ( t1 ) ) (v ( t2 ) , h v ( t2 ) ) . Therefore, the pair (v (t ), h[v(t )]) uniquely
corresponds to an output value of H [v(t )] .
Case 2: If v(t ) is the extrema, then (v ( t1 ) , h v ( t1 ) ) = (v ( t2 ) , h v ( t2 ) ) . According to the principle
of the classical Preisach modeling, i.e. H [v(t1 )] = H [v(t2 )] , then the pair
corresponds to an output value of H [v (t )] .
uniquely
Combining the above-mentioned two cases, there exists a mapping : R 2 R such that
H [v (t )] = (v (t ), h[v (t )]) .
In theorem 1, the obtained mapping (.) is a continuous function. According to Lemma 2,
from v(t1 ) v(t2 ) 0 , it leads to H [v(t1 )] H [v(t2 )] 0 . Also, from h v ( t1 ) h v ( t2 ) 0 , it
Adaptive Control
234
yields v(t1 ) v(t2 ) 0 . Then, it results in H [v(t1 )] H [v(t2 )] 0 . Therefore, it is derived that
is a continuous function. Moreover, Theorem1 indicates that the multi-valued mapping of
hysteresis can be transformed to a one-to-one mapping. It can be proved that the obtained
mapping is a continuous mapping, i.e.
h
v
Let T = [t0 , ) R , V = {v | T
R} be the input sets. Given ti T
R} . Also let F = {h | T
it is obvious that v(ti ) < + and h[v(ti )] < + . So that (v (ti ), h[v(ti )]) R 2 . Thus, it is obtained that
= {(v(ti ), h[v(ti )]) | v(ti ) V , h[v(ti )] F } is a compact set .

Hence, it provides a premise to apply neural networks to modeling of the behavior of
hysteresis. Based on the proposed expanded input space with hysteretic operator, a neural
network is used to approximate the system residual, i.e. f% ( x, v, u ) :
f% ( x.v, u ) = W T (V T xnn ) + ( xnn )
(18)
where () is activation function, V is the first-to-second layer interconnection weights,
W is the second-to-third layer interconnection weights, xnn = ( x T , , u ) , is the NN

functional reconstruction error, ( xnn ) N , and N > 0 .
The above-mentioned neural network based on the expanded input space with hysteretic
operator can be used to construct the corresponding neural estimator for the system residual
f% ( x, v, u ) . Thus, it can be used for the compensation for the effect of the hysteresis inherent
in the sandwich system.
5. Adaptive Control Strategy

In section 3, we introduce an architecture of the control strategy for the sandwich system
with hysteresis. In the control structure, a neural inverse model is used to compensate for
the effect of Li in the architecture of the sandwich system with hysteresis. After the
compensation, the sandwich system with hysteresis is approximately tranformed into a
Hammerstein system with hysteresis. In this section, an adaptive control strategy is
developed for the obtained Hammerstein system with hysteresis.
Assumption 1: If the weight matrices, i.e. V and W of the neural estimator are respectively
bounded by V p > 0 and W p > 0 , i.e. W
W p and V V p ,where F represents Frobenius
norm. Then, the corresponding pseudo-control can be chosen as
= yd( n ) K [0, T ]e vad + vr
(19)
where vr is the term for robust design, K is a design parameter, vad is the output of neural
T
T
network, i.e. vad = W (V xnn ) where W and V are the estimated values of W and V .
235
From (10) and (19), notice that f% ( x, v, u ) depends on vad through . However, vad has to
be designed to cancel
the effect of f% ( x, v, u ) . This should assume that the mapping
ad a f% is a contraction over the entirely interested input domain. It has been proven by
Hovakimyan and Nandi (2002) that the assumption is held when (11) and (12) are satisfied.
Using (18) and (19), (10) can be written as
& = K W T (V T xnn ) + W T (V T xnn ) + vr + + .
(20)
Difine
V% = V V and W% = W W .
The Taylor series expansion of (Vxnn ) for a given
(21)
xnn can be written as
% + o(Vx
% )2
) + '(Vx
)Vx
(Vxnn ) = (Vx
nn
nn
nn
nn
(23)
2
T
where '( z ) = d ( z ) / dz |z = z and o( z% ) is the term of order two. Denoting = (V xnn ) ,
= (V T xnn ) , and ' = '(V T xnn ) , with the procedure as Appendix, we have
& = K + W% T ( 'V T xnn ) + W T 'V% T xnn + vr + + + w
(24)
where
w = W T ( ) + W T 'V T xnn W T 'V T xnn .
(25)
An upper bound for w can be presented as:
w W 1 + W 'V T xnn + V
xnnW T '
(26)
or
w ww (W ,V , xnn )
T
T
where w = 1 + 'V xnn + xnnW '
and w = max( W 1 , W , V
(27)
) .
Theorem 2: Let the desired trajectory be bounded. Consider the system represented by (5), (6)
and (7), if the control law and adaptive law are given by
Adaptive Control
236
v = f 1 ( x, )
(28)
= yd( n ) K [0, T ]e vad + vr
(29)
&
W = F [( 'V T xnn ) kW ]
(30)
&
V = R[ xnnW T ' kV ]
(31)
&
= [ (w + 1) k ]
(32)
and

0
(w + 1) ,
vr =
0,
0
(33)
where F = F T > 0 , R = R T > 0 , > 0 , = max[ w ,( N + N )] , and % = ; then the

signals e , W , V , and in the closed-loop system are ultimately bounded.
Proof: Consider the following Lyapunov function candidate, i.e.
1
1
1
1
L = 2 + tr (W% T F 1W% ) + tr (V% T R 1V% ) + % T 1%
2
2
2
2
The derivative of
(34)
will be
&
L& = & + tr (W% T F 1W&% ) + tr (V% T R 1V%& ) + % T 1%
(35)
Substituting (20) into (35), it yields
&
L& = K 2 + vr + ( w + + ) + %T 1% + trW% T [ F 1W&% + ( 'V T xnn ) ]
.
+ trV% T ( R 1V&% + x W T ')
(36)
nn
&
&
&
&
Substituting W% = W and V% = V into (30) and (31) , (36) can be rewritten as
&
L& = K 2 + vr + ( w + + ) + % T 1% + k [tr (W% T W ) + tr (V% TV )] .
(37)
237
Considering (27) and = max[ w ,( N + N )] ,we obtain
% 1& + k [tr (W% T W ) + tr (V% TV )] .

L& K 2 + vr + (w + 1)
(38)
Substituting (32) and (33) into (38), it results in
L& K 2 + k [tr (W% T W ) + tr (V% TV ) + % T ]
(39)
Defining
W%
%
Z =0
0
W
0
,
0 Z = 0
%
0
0
V%
0
0
V
0
0
W
and
Z =0
0
0
V
0
0
0 ,
(39)
can be rewritten as
L& K 2 + k tr ( Z% T Z ) .
T
As tr ( Z% Z ) Z%
Z%
2
F
(40)
it leads to
L& K 2 + k ( Z%
2
Z% ) .
(41)
That is
L& [ K + k ( Z%
Thus, L& is negative as long as either >
k Z
4K
)2
k Z
4
2
F
].
(42)
2
F
or Z%
> Z
. This demonstrate that ,
W% , V% , and % are ultimately bounded. According to Assumption 1 and the definition of

and , we can obtain that the variables e , W , V and in the closed-loop system are
ultimately bounded.
6. Simulation
In order to illustrate that the proposed approach is applicable to nonlinear system with
sandwiched hysteresis, we consider the following nonlear system:
Adaptive Control
238
Li , v& = 0.2(sin v cos v )
v
+ (0.4 sin v cos v 2 + 0.8) r , v (0) = 0
1 + v2
H , The hysteresis is generated by the sum of N = 50 backlash operators, i.e. ,

N
u = H [v(t )] = ui ,
and
i =1
v&(t )
u&i = v&(t )
di
2
di
v&(t ) < 0, ui (t ) = v(t ) +
2
otherwise
v&(t ) > 0, ui (t ) = v(t )
where ui and di are respectively the output and the dead-band width of i - th backlash
operator where i = 1, 2,L N ( N > 0 is a positive integer).The values of the dead-band
widths are evenly distributed within [0.02,1] . All the initial outputs of the operators are set
to zero. Fig. 5 shows the response of the hysteresis contained in the system.
x&1 = x2
Lo :
2
x& 2 = (1 x1 ) x2 x1 + u
and
y = x1 .
The design procedure of the controller for the snadwich system with hysteresis will be
shown in the following.
1)
Construction of nerual network based Li inverse. An artificial neural network unit

inverse , i.e. L i is constructed to cancel the effect of the first dynamic block, i.e. Li .
1
The system is excited by the input rl (t ) = sin 2t + cos t . Then, 500 input/output
pairs of data {rl , (vl , v&l )} are obtained. Using these data as learning samples, a
1
neural network based inverse L i is constructed. The architecture of neural
network based inverse model consists of 2 input nodes, 10 hidden neurons and 1
output neuron. The sigmoid function and linear function are respectively used as
activation function in the hidden layer and the output layer. The conjugate
gradient algorithm with Powell-Beale restarted method (Powell, 1977) is used to
239
train the neural network. The compensation result of the NN based Li 1 is shown in
Fig. 6. It is known that there are some larger error happened in the beginning. Then
it is gradually reduced as the control proceeded.
2
1.5
u(t)
0.5
-0.5
-1
-1.5
-2
-0.2
-0.15
-0.1
-0.05
0.05
0.1
0.15
0.2
0.25
v(t)
Fig. 5. The hysteresis in the system

4
compensation error
-1
-2
-3
-4
10
12
14
16
18
20
time
Fig. 6. The compensation error of NN based Li 1

2) Neural approximator of system residual: The neural network used to approximate f% ( x, v, u )
consists of 4 input nodes, 35 hidden neurons and 1 output neuron. The input of the NN
is xnn = ( x T , , u ) . The activation function is ( x ) =
1
.
1 + ex
3) The selection of the controller parameters: The other parameters of the controller are
respectively chosen as 1 = 2 , K = 11 , k = 0.001 , = 0.1 , f ( x, v ) = v , F = 8 I , and R = 5 I ,
where I is the unit matrix.
4) PID control for comparison: In order to compare the control performance of the proposed
control strategy with the PID controller , we choose
t
v(t ) = 22e1 + e1dt 13e2

0
Adaptive Control
240
where e1 = y yd , e2 = y& y& d . Moreover,
the desired output of the system
is yd (t ) = 0.1 [sin 2t cos t ] .
0.6
0.4
0.2
y yd
-0.2
-0.4
-0.6
system output y
desired output yd
-0.8
10
12
14
16
18
20
time
Fig. 7. The control response of the proposed method

From Fig.7, it is known that the control performance of the proposed controller has
achieved good control response. Also, Fig.8 illustrates the control performance of the PID
controller. It can be seen that the PID control strategy has led to larger control error when
the reference signal achieves its local extreme. However, the proposed control strategy
obtained better control performance. It can obviously derive more accurate control result.
7. Conclusions
An adaptive control strategy for nonlinear dynamic systems with sandwich hysteresis is
presented. In the proposed control scheme, a neural network unit inverse is constructed to
compensate for the effect of the first smooth dynamic subsystem. Thus, the sandwich system
with hysteresis can be transformed to a Hammerstein type nonlinear dynamic system
preceded by hysteresis. Considering the modified structure of the sandwich system, an
adaptive controller based on the pseduo-control technique is developed. In our method, a
neural network is used to approximate the system residual based on the proposed expanded
input space with hysteretic operator. The advantage of this method can avoid constructing
the hysteresis inverse. Then, the adaptive control law is derived in terms of the Lyapunov
stability theorem. It has been proved that the ultimate boundedness of the closed-loop
control error is guaranteed. Simulation results have illustrated that the proposed scheme has
obtained good control performance.
0.6
0.4
0.2
y yd
-0.2
-0.4
-0.6
system output y
desired output yd
-0.8
10
12
14
time
Fig. 8. The control response of the PID control method
16
18
20
241
8. Appendix
From (20), the approximation error can be written as:
W T W T
T
T
T
T
= W W + W W
T
T
= W% + W ( )
(A1)
Substituting (23) into (A1), it yields
W T W T
T
T
T
2
T
T
T
2
= W% ( + 'V% xnn + o(V% xnn ) ) + W ( 'V% xnn + o(V% xnn ) )
= W% T + W% T 'V% T x + W T 'V% T x + W T o(V% T x ) 2
nn
nn
nn
(A2)
T
T
T
T
T
T
T
T
T
2
= W% + W% 'V xnn W% 'V xnn + W 'V% xnn + W o(V% xnn )
T
T
T
T
T
T
T
T
2
= W% ( 'V x ) + W% 'V x + W 'V% x + W o(V% x ) .
nn
nn
nn
nn
Defining w = W% T 'V T xnn + W T o(V% T xnn ) 2 , (A2) becomes
W T W T
T
T
T
T
= W% ( 'V xnn ) + W 'V% xnn + w .
So that
w = W T W T W% T ( 'V T xnn ) W T 'V% T xnn

= W T W T + W% T 'V T x W T 'V% T x
nn
nn
T
T
T
T
T
= W ( ) + W 'V xnn W 'V xnn W 'V% xnn
T
T
T
T
T
= W ( ) + W 'V xnn W 'V xnn
T
9. Acknowledgement
This research is partially supported by the Innovation Program of Shanghai Municipal
Education Commission (Grant No.:09ZZ141), the National Science Foundation of China
(NSFC Grant No.: 60572055) and the Advanced Research Grant of Shanghai Normal
University (Grant No: DYL200809).
10. References
Taware, A. & Tao, G. (1999). Analysis and control of sandwich systems, Proceeding of the 38th
conference on decision and control, pp.1156-1161, Phoenix, Arizona, USA, December
1999
242
Adaptive Control
Tao, G. & Ma, X.(2001).Optimal and nonlinear decoupling control of system with
sandwiched backlash, Automatica, Vol.37, No.1, 165-176.
Hovakimyan, N.& Nandi, F.(2002). Adaptive output feedback control of uncertain nonlinear
systems using single-hidden-layer neural networks, IEEE Transactions on Neural
Networks, Vol.13, No.6, 1420-1431
Calis, A.& Hovakimyan, N. (2001). Adaptive output feedback control of nonlinear systems
using neural networks, Automatica, Vol. 37, 1201-1211.
Powell, M.(1977). Restart procedures for the conjugate gradient method, Mathematical
Programming, Vol. 12, 241-254.
Zhao, X.; Tan, Y. & Zhao, T.(2008). Adaptive control of nonlinear system with sandwiched
hysteresis using Duhem operator, Control and Decision, Vol. 22, No. 10, 1134-1138
Corradini, M., Manni, A & Parlangeli, G. (2007). Variable structure control of nonlinear
uncertain sandwich systems with nonsmooth nonlinearities,Proceedings of the 46th
IEEE Conference on Decision and Control, pp. 2023-2038
Zhao, X. & Tan, Y.(2006). Neural adaptive control of dynamic sandwich systems with
hysteresis, Proceedings of the 2006 IEEE International Symposium on Intelligent Control,
pp. 82-87
Adly, A.A.& Abd-El-Hafiz, S.K. (1998). Using neural networks in the identification of
Preisach-type hysteresis models. IEEE Trans. on Magnetics, Vol. 34, No.3, 629-635
Zhao, X. & Tan, Y. (2008). Modeling hysteresis and its inverse model using neural networks
based on expanded input space method,IEEE Transactions on Control Systems
Technology, Vol. 16, No. 3, pp. 484-490
11
High-Speed Adaptive Control Technique Based
on Steepest Descent Method for Adaptive
Chromatic Dispersion Compensation in Optical
Communications
Ken Tanizawa and Akira Hirose
Department of Electronic Engineering, The University of Tokyo

Japan
1. Introduction
The traffic volume of the data transmission is increasing each year with the explosive
growth of the Internet. The networking technologies supporting the data transmission are
optical fiber transmission technologies. In the physical layer, the networks are classified into
three networks, the long-haul network that connects city to city, the metropolitan area
network that connects the central station in the city to the neighboring base station, and the
access network that connects the base station to the home. In order to adapt to the increase
of the data transmission, we need to achieve high-speed transmission and increase the
capacity of transmission in each network.
In the access network, many kinds of passive optical networks (PON) are studied to offer a
high-speed access to the Internet at low cost. In the metropolitan area network, we
contemplate the update of the network structure from the conventional peer-to-peer
transmission to the ring or mesh structure for the high-capacity and highly reliable networks.
In the long-haul network, the study on multilevel modulation such as the differential
quadrature phase shift keying (DQPSK) is a recent popular topic for the high-capacity
transmission because the multilevel modulation utilizing the phase information offers highspeed transmission without increasing the symbol rate. Other modulation and multiplexing
technologies are also studied for the high-capacity networks. The orthogonal frequency
division multiplexing (OFDM) is one of the wavelength division multiplexing methods and
achieves high spectral efficiency by the use of orthogonal carrier frequencies. The optical
code division multiple access (OCDMA) is a multiplexing technique in the code domain.
These techniques are developed in the wireless communication and modified for the optical
transmission technologies in these days.
In the long-haul and the metropolitan area networks whose transmission distance is over 10
km in 40 Gb/s, chromatic dispersion (CD) is one of the main factors which limits the
transmission speed and the advances of the network structure. The CD is a physical
phenomenon that the group velocity of light in the fiber depends on its wavelength
(Agrawal, 2002). The CD causes the degradation of the transmission quality as the optical
Adaptive Control
244
signals having a spectral range are distorted by the difference of the transmission speed in
the wavelength domain. The effect of dispersion increases at a rate proportional to the
square of the bit-rate.
In the high-speed optical transmission over 40 Gb/s, we have to compensate for the CD
variation caused by the change of strain and temperature adaptively in addition to the
conventional static CD compensation because the dispersion tolerance is very small in such
a high-speed transmission. Also, in metropolitan area networks employing reconfigurable
networking technology such as the mesh or ring network, the transmission route changes
adaptively depending on the state of traffic and the network failure. As the CD value
depends on the length of the transmission fiber, we have to compensate for the relatively
large CD variation caused by the change of the transmission distance.
With the aforementioned background, many researches and demonstrations have been
conducted in the field of the adaptive CD compensation since around 2000 (Ooi et al., 2002;
Yagi et al., 2004). The adaptive compensations are classified into two major groups, the
optical compensations and the electrical compensations. In the electrical compensation, we
utilize the waveform equalizer such as the decision feedback equalizer (DFE), the feed
forward equalizer (FFE) or the maximum likelihood sequence equalizer (MLSE) after
detection (Katz et al., 2006). These equalizers are effective for the adaptive CD compensation
because they act as a waveform reshaping. The compensation based on DEF and FFE has
advantages that the equalization circuit is compact and implemented at low cost. However,
the compensation range is limited because the phase information of the received signal is
lost by the direct detection. The MLSE scheme is very effective in 10 Gb/s transmission.
However it is difficult to upgrade high bit-rate over 40 Gb/s because the scheme requires
high-speed A/D converter in implementation.
In the optical domain, the adaptive CD compensation is achieved by the iterative feedback
control of a tunable CD compensator with a real-time CD monitoring method as shown in
Fig. 1. Many types of tunable CD compensators are researched and developed recently. The
tunable CD compensator is implemented by the devices generating arbitral CD value. Also,
many kinds of CD monitoring methods are studied and demonstrated for the feedback
control of tunable CD compensators. While the compensation devices and the dispersion
monitoring methods are studied with keen interest, the adaptive control algorithm, how to
control the tunable CD compensator efficiently, has not been fully studied yet in the optical
domain CD compensation. When the tunable CD compensator is controlled iteratively for
the adaptive CD compensation, the control algorithm affects the speed of the compensation
to a great degree as well as the response time of the compensation devices and the
monitorings. Although the simple hill-climbing method and the Newton method are
employed as a control algorithm in many researches and demonstrations, these algorithms
are not always the best control algorithm for the adaptive CD compensation.
Tunable CD
compensator
Real-time CD
monitoring
Feedback control (search control algorithm)

Fig. 1. Adaptive CD compensation in the receiver.
High-Speed Adaptive Control Technique Based on Steepest Descent Method for Adaptive
Chromatic Dispersion Compensation in Optical Communications
245
In this chapter, we report the adaptive CD compensation employing adaptive control

technique in optical fiber communications. We propose a high-speed and low cost adaptive
control algorithm based on the steepest descent method (SDM) for feedback control of the
tunable CD compensator. The steepest descent approach has an ability to decrease the
iteration number for the convergence. We conducted transmission simulations for the
evaluation of the proposed adaptive control technique, and the simulation results show that
the proposed technique achieves high-speed compensation of the CD variation caused by
the change of the transmission distance in 40 Gb/s transmission.
The organization of this chapter is as follows. In Section 2, we explain the fundamentals of
CD and adaptive CD compensation in optical fiber communications for the background
knowledge of this research. Then we propose the adaptive control technique based on the
SDM for adaptive CD compensation in Section 3. In Section 4, we show the demonstrations
and performance analysis of the proposed technique in 40 Gb/s transmission by simulations.
Finally, we summarize and conclude this paper in Section 5.
2. Chromatic Dispersion in Optical Fiber Communications

2.1 Fundamental of chromatic dispersion
The group velocity of the light depends on its wavelength when the light is propagating in
mediums. This phenomenon is called CD or group velocity dispersion (GVD). In optical
communications utilizing the optical fiber as a transmission medium, the optical pulse is
affected by the CD as the propagation time depends on the constituent wavelength of the
optical pulse as shown in Fig. 2. The CD has two contributions, material dispersion and
waveguide dispersion in a single mode fiber (SMF). The material dispersion is attributed to
the characteristics of silica that the refractive index changes with the optical wavelength. The
waveguide dispersion is caused by the structure of optical fiber, the core radius and the
index difference.
Considering optical propagation in the fiber, the propagation constant is a function of the
angular frequency and expanded by Taylor expansion as follows.
( ) = 0 + 1 ( 0 ) +
1
1
2 ( 0 ) 2 + 3 ( 0 )3 + L
2
6
(1)
Here, 0 is a center angular frequency, and 0, 1, 2, and 3 are Taylors coefficients. The
time required for the propagation of unit length is obtained by differentiating partially the
propagation constant as follows.
( ) = 1 + 2 ( 0 ) +
1
3 ( 0 ) 2 + L
2
(2)
It is confirmed from (2) that the required time is angular frequency dependent; the
propagation time of optical pulse depends on the wavelength in optical communications.
The coefficients 2 and 3 are first-order and second-order constants indicating the degree of
the angular frequency dependence, respectively. Assuming that the second-order CD is
negligible, the CD parameter is defined as
Adaptive Control
246
input
1
t
OpticalFiber
Fig. 2. Schematic diagram of chromatic dispersion.
D=
d
2c
= 2 2
d
(3)
where c is the speed of light. The unit of the CD parameter is ps/nm/km.

In SMF, the CD parameter is zero at around 1300 nm and about 20 ps/nm/km at the typical
wavelength used for optical communications, around 1550 nm. We have many
characteristics of optical fibers such as dispersion shifted fiber (DSF) whose CD parameter is
zero at around 1550 nm for the reduction of CD effect in optical fiber communications, and
dispersion compensating fibers (DCF) whose CD parameter is minus value for the purpose
of static CD compensation.
In optical fiber communications, the optical pulse is affected by the CD as it has relatively
wide spectral range corresponding to the bit-rate. Assuming that the optical pulse is a
Gaussian waveform for the simplicity, the waveform in time-domain is expressed as
T2
U (0, T ) = exp
2
2T0
(4)
where T0 is a full width at half maximum (FWHM) of the pulse. When the pulse is
transmitted for arbitral distance z, the waveform is affected by the CD and distorted as
U ( z, T ) =
T0
(T02 j 2 z )1 / 2
T2
exp
2
2(T0 j 2 z )
(5)
247
1
2z = 0
2z = 10000
2z = 20000
Optical Intesity [a.u.]
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
Time [a.u.]
Fig. 3. Optical pulses affected by chromatic dispersion.
x 10
O pticalpulse
Fig. 4. Interference of neighboring pulses in optical communication.

where we neglect the second-order CD for simplicity as the first-order CD is dominant.
Figure 3 shows the waveforms of optical pulse when we change the product of 2 and z
under the condition that T0=100 ps. The larger the product of 2 and z is, the wider the
FWHM of the transmitted waveform is; the effect of CD is larger in the case that the
transmission distance is longer and the CD parameter is larger. If the FWHM of the optical
pulse gets wider, the possibility of the inter symbol interference (ISI) is higher as shown in
Fig. 4. As the ISI causes code error in optical communications, the transmission distance is
limited by the CD. Also, the maximum transmission distance is reduced according to the bit
rate of the transmission B because the FWHM of the optical pulse T0 is decreased when the
bit rate increases. We can also understand it from the fact that the spectral width is wide in
short optical pulse. The effect of CD on the bit rate B can be estimated and the CD tolerance
DT, the limitation of CD that the quality of the transmission is assured, is expressed as
Adaptive Control
248
DT <
1
Bz
(6)
where is the range of wavelength in the optical pulse. The CD tolerance is inversely
proportional to the bit rate and the transmission distance and the wavelength range of the
input pulse.
2.2 Adaptive chromatic dispersion compensation
As mentioned in Section 1, the adaptive CD compensation is an essential technology for
high-speed optical fiber communications as the CD tolerance is very small in the systems
whose transmission speed is over 40 Gb/s. Many researches have been conducted for the
adaptive CD compensation in optical communications. The principle of the CD
compensation is very simple as shown in Fig. 5. We can achieve the CD compensation by
placing a transmission medium which has the inverse CD value of the transmission fiber in
the transmission line. The adaptive CD compensation is achieved by changing the
compensating CD value adaptively according to the CD in the transmission fiber. The
conventional setup of the adaptive CD compensation is shown in Fig. 1; the tunable CD
compensator is feedback controlled with the real-time CD monitoring. In this section,
tunable CD compensators and CD monitoring techniques are briefly introduced for the
background information of the adaptive control algorithm to be proposed.
We have many types of tunable CD compensators for the adaptive compensation. They are
basically implemented by the dispersive medium with the function of tunability, for
example, chirped fiber Bragg grating (CFBG) with heater elements (Matsumoto et al., 2001;
Eggleton et al., 2000), micro-electro mechanical system (MEMS) (Sano et al., 2003), ring
resonator (Takahashi et al., 2006), and so on. We adopt a virtually imaged phased array
(VIPA) compensator in the following research. The VIPA compensator is a tunable CD
compensator, which is consisted of the combination of a VIPA plate and a three dimensional
adjustable surface mirror (Shirasaki, 1997; OOi et al., 2002). The VIPA plate operates as a
grating, and the specific spectral components of light is reflected by the mirror to induce CD.
group delay
group delay
wavelength
group delay of
opticalfiber
group delay
wavelength
group delay of
com pensator
wavelength
group delay after com pensation
Fig. 5. Principle of chromatic dispersion compensation.
Response
time
Clock power level

monitoring
method
Clock phase
detection method
Eye-diagram
Cost
Relationship
between
transmission
quality
and
monitoring signal
Monitoring
range
Good
Good
Fair
Good
Good
Fair
Good
Poor
Good
Fair
Excellent
Good
BER
Fair
Poor
Excellent
Good
Table 1. Performances of feedback signals in adaptive CD compensation
249
Pilot
signal
Not
required
Not
required
Not
required
Required
We can generate arbitral CD value as the change of the geometry of the three dimensional
mirror. In the VIPA compensator, wide compensation range, 1800 ps/nm in 10 and 40
Gb/s, is achieved by the appropriate design of the three dimensional mirror.
Also, many kinds of CD monitoring methods are studied and demonstrated for the feedback
control of the tunable CD compensators. The typical monitoring signals are bit error rate
(BER), eye-diagram, clock power level (Sano et al., 2001), and phase difference of clock
signals (Qian et al., 2002). We show the performance comparison of the feedback signals for
adaptive control of the tunable CD compensator in Table 1. The requirement of pilot signal
is the disadvantage for the BER as the monitoring signal. If we consider each characteristic
of the feedback signal, the extracted-clock power level or the eye-diagram is better for the
feedback signal in adaptive CD compensation. We adopt the eye-opening value obtained
from the eye-diagram as the feedback signal in the adaptive control method to be proposed.
3. High-Speed Adaptive Control Method Based on Steepest Descent Method

In this section, we propose a method of high-speed adaptive control of tunable CD
compensator for adaptive CD compensation. We apply the steepest descent method to the
adaptive control algorithm in order to reduce the compensation time. The approximation of
partial derivative for the steepest descent approach is proposed and applied to the control of
the VIPA compensator.
3.1 Steepest descent-based control algorithm for adaptive chromatic dispersion
compensation
The adaptive control system must be low cost, high-speed, and applicable to wide
dispersion ranges for the adaptive CD compensation in optical communications. Most
control systems require high-cost measuring instruments for the CD monitoring. We
therefore propose the feedback control method that does not require high-cost CD
monitoring. In our proposal, the feedback signal is a received waveform in the time domain.
The tunable CD compensator is controlled repeatedly to reshape the waveform. The
measurement of the waveform is relatively easy and uninfluential in the transmission
Adaptive Control
250
conditions such as pilot-signal requirements. Conventional feedback control is based on the

hill-climbing method, which requires a lot of time for optimization. We have therefore
applied the steepest descent method to the feedback control for high-speed compensation.
Figure 6 shows an optical dynamic routing network with the adaptive CD compensation.
Transmitted signals are passed through a route that is chosen arbitrarily among optical
paths, being affected by the CD. In the receiver part, the degraded signals are fed into the
tunable CD compensator and the dispersion is compensated. The adaptive dispersion
compensation is achieved by the combination of a tunable CD compensator and a controller.
The compensated signals are received by a photodiode and demodulated.
Transmitter
Laser
All-Optical Routing Network
Receiver
Tunable dispersion
EDFA
Photo diode
compensator
Modulator
OXC
OXC
OXC
OXC
OXC
OXC
OXC
OXC
OXC
control signal
Controller
fref
fout :Received signal

fref :Memorized reference signal
fout
Amplifier
Demodulator
(received signal without dispersion)

OXC: Optical cross connect
Fig. 6. Schematic diagram of all-optical dynamic routing network with the adaptive
dispersion compensation technique.
Received signal: fout
Pout
Controller
Memorized reference signal: fref
Pout
Pref
Pref
Calculate error value

Cauculate partial derivative of error value
Update control parameters by steepest
descent method
Fig. 7. Procedure of proposed steepest-descent-based control.
251
The tunable compensator is controlled by our proposed adaptive control method based on
the steepest descent method. The proposed procedure of the controller is shown in Fig. 7,
where Pout and Pref are the eye-opening values (normalized as Pref = 1) of the received and
reference signals, fout and fref respectively. In this method, we measure and register the
reference signal, fref which is a received signal unaffected by the CD. The reference signal is
determined from the characteristics of the transmitter-receiver set. Therefore, we can copy
the reference signal to other receivers after it has been measured once.
The first step is a calculation of an error value: Er. The error value is defined as the
difference between the eye-opening values, Pout and Pref.
Er =
1
( Pref Pout ) 2
2
(7)
The next step is a calculation of partial derivatives of Er in terms of the control parameters,
xi (i=1,2,, n), for the update based on the steepest descent method.
Pout
Er
= Pout Pref
xi
xi
(8)
We need to measure small changes in Pout when xi changes slightly in order to get the
accurate partial derivatives of Pout with respect to xi. However, this is unrealistic as it takes a
lot of time for the measurement and our goal is to achieve quick CD compensation.
Therefore, we approximate the partial derivatives of Pout with respect to xi. The
approximation is to be mentioned at the next subsection.
In the final step, the control parameters are update as
xi xi
Er
xi
(9)
where is an appropriate constant concerning the speed and accuracy of the convergence.
We repeat this procedure until the transmission quality becomes optimal. The required
number of update iterations is fewer than that of the normal feedback control based on the
hill-climbing method due to the steepest descent approach. In practical all-optical dynamic
routing networks, the procedure is repeated all the time as the transmission route changes at
frequent intervals.
3.2 Approximation of partial derivatives for steepest descent approach
To approximate the partial derivatives of Pout with respect to xi, we need to know the change
in one-bit waveforms of the received signal, wout(t), caused by the change in xi. When we
assume that the waveform of the transmitted signal is a Gaussian-like pulse (the peak level
is unity) just like in the approximation in return-to-zero transmissions and that the
transmission is affected only by the CD, the waveform, wout(t) is calculated analytically in
terms of the CD values of the transmission fiber, Dfiber ps/nm and TDC, DTDC ps/nm, as
Adaptive Control
252
v peak 2 t 2
wout (t ) = v peak exp
2
T
FWHM
TFWHM
v 2peak =
TFWHM
(10)
2
2
D fiber +
DTDC
+
2c
2c
(11)
where TFWHM is the FWHM of the transmitted signal, is the center wavelength , t is time,
and c is the speed of light. The partial derivative of wout(t) with respect to xi is calculated
from (10) and (11).
2
2
2
2
2
v peak
wout (t )
2 2t v peak
exp t v peak
=
1
1
peak
2
2
2
T
T
xi
TFWHM
FWHM
FWHM
2 D
TDC
2c xi
(12)
Equation (10) shows that the value vpeak is the peak level of wout(t). We can measure it in a
practical system. Therefore, (12) shows that we can obtain the approximated partial
derivative of wout(t) with respect to xi because TFWHM and are known parameters. We obtain
the partial derivative of the peak value in wout(t) by substituting 0 for t.
2
v peak
v peak
wout (t )
2 DTDC
2
=
1
=
peak
2
xi
2c xi
xi t =0
TFWHM
(13)
The value of vpeak corresponds to the eye-opening value in nonreturn-to-zero (NRZ)

transmission approximately. Therefore, the partial derivative of Pout with respect to xi is
approximated as follows.
2
Pout
P
2 DTDC
2
= out 2 1 Pout
xi
2c xi
TFWHM
(14)
3.3 Detailed control algorithm for VIPA compensator

In the simulations described in the next section, we employ a VIPA compensator as the
tunable CD compensator. The VIPA compensator has a single control parameter, i.e. CD S
ps/nm. The detailed control procedure of the VIPA compensator is as follows. In general,
we can apply the proposed method to any kind of tunable CD compensators.
(i) Initialize the parameter of the VIPA compensator: S ps/nm
S = 0 ps/nm
(ii) Calculate the error value: Er
(15)
253
The error value, Er, is derived from (7).

If Pout is zero, we go to (iii), otherwise to (iv).
(iii) Update S by the hill-climbing method
S S S
(16)
where S ps/nm is an appropriate small constant. We then go on to (v).

(iv) Update S by the steepest descent method
We calculate the partial derivative of Er from (8) and (14).
The partial derivative of Pout with respect to S is approximated as
2
P
Pout
2
2
= out 2 1 Pout
2c
S
TFWHM
(17)
The parameter S is updated as
S S
Er
S
(18)
where is an appropriate constant. We then go to (v).

(v) Judge the error value: Er
We calculate Er again by using (7). If Er increases or becomes small enough, the
procedure stops. Otherwise, we go back to (ii) and repeat the same process.
However, in practical all-optical dynamic routing networks, the compensation
process is repeated all the time as the dispersion value changes frequently.
4. Transmission Simulations at 40 Gb/s

4.1 Simulation results in NRZ-OOK transmission at 40 Gb/s
Numerical transmission simulations using OptiSystem were conducted to verify the
application of the proposed technique to 40 Gb/s optical fiber transmission system. In the
proposed control method, we have to set the constants for search, and S, appropriately.
They were adjusted for the 40 Gb/s transmission and set at 3105 and 30, respectively. The
output power of a distributed feedback laser diode (DFB-LD) at the transmitter was 0 dBm.
We supposed that the modulation format were NRZ-OOK. The central wavelength of the
transmitted signal was 1550 nm. The transmission path was a non-zero dispersion shifted
fiber (NZ-DSF). We assumed that CD, polarization mode dispersion (PMD), self-phase
modulation (SPM), and other nonlinearity affect the transmitted signal. The power loss was
amplified to 0 dBm by an erbium-doped fiber amplifier (EDFA) after both of the fiber
transmission and the dispersion compensation by the VIPA compensator. The EDFA, the
Adaptive Control
Bit Error Rate
254
10-0
10-1
10-2
10-3
10-4
10-5
10-6
10-7
10-8
10-9
10-10
10-11
10-12
10-13
10-14
10-15
10-16
Update iteration number
Bit Error Rate
(a)
10-1
10-2
10-3
10-4
10-5
10-6
10-7
10-8
10-9
10-10
10-11
10-12
10-13
10-14
10-15
10-16
Update iteration number

(b)
Fig. 8. BERs at every update of the compensator (a) 020km, (b) 2025km.
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
255
(a)
(b)
Fig. 9. Eye-diagrams at every update of the compensator (a) 020km, (b) 2025km.
receiver, and other optical components were assumed to have moderate levels of noise.
Transmission simulation results are shown in Fig. 8 and Fig. 9. In these simulations, the
initial value of S was set at 0. Fig. 8 shows the BERs at every update of the VIPA
compensator when the transmission distance changes (a) from 0 to 20 km (dispersion: 0
Adaptive Control
256
100 ps/nm) and (b) from 20 to 25 km (dispersion: 100 125 ps/nm). The compensation
improved the BER (a) from 1.0100 to 1.010-15 and (b) from 7.4110-3 to 1.010-15,
respectively. As shown in Fig. 9, the eye-diagrams measured by a sampling oscilloscope
were found reshaped. These results show that the CD compensation with the proposed
control method improve the transmission quality.
The update iteration number to achieve a sufficiently low BER (<10-9) were four and two,
respectively in these two cases. A single calculation of the next dispersion value requires less
than 10 ms. The response time of the VIPA compensator is 2 ms for every 1 ps/nm
compensation. Therefore, the time required for 400 ps/nm CD compensation is about 1 s,
which is determined practically by the response time of the VIPA compensator. This
technique achieves a high-speed adaptive control of a tunable CD compensator in 40 Gb/s
transmission since the update iteration number is small and the calculation time with the
proposed approximation is short enough. The proposed technique is more effective if the
response time of the tunable CD compensator is faster as the required iteration number is
decreased by our proposed adaptive control technique based on the steepest descent
method.
4.2 Compensation range and required iteration number
Figure 10 shows the compensation range of the proposed method at 40 Gb/s. We measured
BERs before and after compensation when the CD value changed from 0 ps/nm to arbitrary
value. The compensation range in which the BER after compensation is less than 10-9 is
about from 450 to 450 ps/nm, corresponding to a NZ-DSF path-length change of about 90
km. This range is wide enough for compensating the change of CD caused by dynamic
routing. In this wide compensation range, the iteration number required for error free
transmission (BER<10-9) is less than 15. The fast adaptive CD compensation is also achieved
by the proposed adaptive control technique as the required iteration number is small.
Bit Error Rate
Before compensation
After compensation
100
10-1
10-2
10-3
10-4
10-5
10-6
10-7
10-8
10-9
10-10
10-11
10-12
10-13
10-14
10-15
10-16
-600
-400
-200
200
Dispersion [ps/nm]
Fig. 10. Compensation range of the proposed method at 40 Gb/s.
400
600
257
5. Conclusion
In this chapter, we have proposed high-speed adaptive CD compensation with the adaptive
control method based on the steepest descent method and reported the performances
evaluated by numerical simulations.
The simulation results show that the proposed control method based on the steepest descent
method controls the tunable CD compensator quickly and effectively for a wide dispersion
range in 40 Gb/s transmission. The range is up to 450 ps/nm, and the required
compensation time is about 1 s for the CD variation within 400 ps/nm at 40 Gb/s. These
achievements are valuable for the future optical networks employing dynamic routing
technique.
6. References
Agrawal, G. (2002). Fiber-Optic Communication Systems, Wiley-Interscience, ISBN 0-47121571-6, USA
Eggleton, B. J.; Ahuja, A.; Westbrook, P. S.; Rogers, J. A.; Kuo, P.; Nielsen, T. N. & Mikkelsen,
B. (2000). Integrated Tunable Fiber Gratings for Dispersion Management in HighBit Rate Systems. IEEE/OSA Journal of Lightwave Technology, Vol. 18, No. 10, (Oct.
2000) pp. 1418-1432
Katz, G.; Sadot, D. & Tabrikian, J. (2006). Electrical Dispersion Compensation Equalizers in
Optical Direct- and Coherent-Detection Systems. IEEE Transanctions on
Communications, Vol. 54, No. 11, (Nov. 2006) pp. 2045-2050
Matsumoto, S.; Ohira, T.; Takabayashi, M.; Yoshiara, K. & Sugihara, T. (2001). Tunable
Dispersion Equalizer with a Divided Thin-Film Heater for 40-Gb/s RZ
Transmission. IEEE Photonics Technology Letters, Vol. 13, (Aug. 2001) pp. 827-829
Ooi, H.; Nakamura, K.; Akiyama, Y.; Takahara, T.; Terahara, T.; Kawahata, Y.; Isono, H. &
Ishikawa, G. (2002). 40-Gb/s WDM Transmission With Virtually Imaged Phased
Array (VIPA) Variable Dispersion Compensators. IEEE/OSA Journal of Lightwave
Technology, Vol. 20, (Dec. 2002) pp. 2196-2203
Sano, T.; Iwashima, T.; Katayama, M.; Kanie, T.; Harumoto, M.; Shigehara, M.; Suganuma,
H.; & Nishimura, M. (2003). Novel Multichannel Tunable Chromatic Dispersion
Compensator Based on MEMS and Diffraction Grating. IEEE Photonics Technology
Letters, Vol. 15, No. 8, (Aug. 2003) pp. 1109-1111
Shirasaki, M. (1997). Chromatic-Dispersion Compensator Using Virtually Imaged Phased
Array. IEEE Photonics Technology Letters, Vol. 9, No. 12, (Dec. 1997) pp. 1598-1600
Tanizawa, K. & Hirose, A. (2007). Adaptive Control of Tunable Dispersion Compensator
That Minimizes Time-Domain Waveform Error by Steepest Descent Method. IEEE
Photonics Technology Letters, Vol. 18, No. 13, (Jul. 2006) pp. 1466-1468
Tanizawa, K. & Hirose, A. (2007). Performance Analysis of Steepest-Descent-Based
Feedback Control of Tunable Dispersion Compensator for Adaptive Dispersion
Compensation in All-Optical Dynamic Routing Networks. IEEE/OSA Journal of
Lightwave Technology, Vol. 24, No. 4, (Apr. 2007) pp. 1086-1094
Yagi, M.; Satomi, S.; Tanaka, S.; Ryu, S.; & Asano, S. (2005). Field Trial of Automatic
Chromatic Dispersion Compensation for 40-Gb/s-Based Wavelength Path
258
Adaptive Control
Protection. IEEE Photonics Technology Letters, Vol. 17, No. 1, (Jan. 2005) pp. 229-231
12
Adaptive Control of Piezoelectric Actuators
with Unknown Hysteresis
Wen-Fang Xie, Jun Fu, Han Yao and, C.-Y. Su
Department of Mechanical & Industrial Engineering
Concordia University
Canada
1. Introduction
Hysteresis phenomenon occurs in all smart material-based sensors and actuators, such as
shape memory alloys, piezoceramics and magnetostrictive actuators (Bank & Smith, 2000;
Tan & Baras, 2004). In order to study this phenomenon, different models were proposed
(Brokate & Sprekels, 1996; Visintin, 1994). Normally, hysteresis models are classified into
two categories, physics-based model such as Jiles-Atherton model (Jiles & Atherton, 1986)
and phenomenology-based model such as Preisach operator (Brokate & Sprekels, 1996;
Visintin, 1994) and Duhem model (Visintin, 1994). From control systems point of view,
hysteresis is generally non-differentiable, nonlinear, and unknown. As a result, systems
with hysteresis usually exhibit undesirable inaccuracies or oscillation and even instability.
Mitigating the effect of hysteresis becomes necessary and important, thus it has received
increasing attention in recent years (Tao & Kokotovic, 1995, Su, et al, 2000, Su, et al, 2005).
Many of these studies are related to modeling of hysteresis and their control issues.
With the development of artificial intelligent (AI), AI is being applied to dealing with
nonlinearities in systems (Ge & Wang, 2002). Only a few studies have been carried out by
using NN to tackle hysteresis modeling and compensation (Makaveev, et al, 2002;
Kuczmann & Ivanyi, 2002; Beuschel et al, 1998; Zhao & Tan, 2006). In the paper (Makaveev,
et al, 2002), a NN model is used to describe the hysteresis behavior in different frequencies
with the knowledge of some properties of magnetic materials, such as loss separation
property to allow the separate treatment of quasi-static and dynamic hysteretic effects.
Beuschel et al used (Beuschel et al, 1998) a modified Luenberger observer and NN are used
to identify a general model of hysteresis. These researches demonstrate that NN can work as
an unknown function approximator to describe the characteristics of hysteresis. Recently,
two papers (Zhao & Tan, 2006; Lin et al 2006) applied the approximation property of NN to
coping with the identification of Preisach-type hysteresis in piezoelectric actuator, and the
hysteresis estimation problem for piezo-positioning mechanism based on hysteresis friction
force function, respectively. It should be noted that the aforementioned results share a
common assumption that the output of hysteresis is measurable.
In practical systems, smart actuators are integrated into the systems, which makes the
measurement of output of hysteresis hard. Hence it is a challenge to design an observer for
Adaptive Control
260
the unavailable output of hysteresis. Due to the unavailability of the output of hysteresis,
the major obstacle of pre-inversion compensator for hysteresis is the lack of effective
observer design methods for piezoelectric actuators. Especially, the traditional Luenbergertype nonlinear observer design (Krener & Isidori, 1983) or the high-gain observer
(Krener & Kang, 2003) cannot be applied directly, since the hysteresis is highly nonlinear.
The sliding-mode observer was developed to estimate the internal friction states of LuGre
model for the servo actuators with friction (Xie, 2007). This observer needs a low-pass filter
to remove the high-frequency components in the estimated state variable, which is not
applicable in this paper. Yang and Lin (Yang & Lin, 2004) proposed homogeneous
observers design for a class of n-dimensional inherently nonlinear systems whose Jacobian
linearization is neither controllable nor observable.
Inspired by NNs universal approximation property, and the aforementioned facts in
observer design, we propose an observer-based adaptive control of piezoelectric actuators
with unknown hysteresis in this paper. The main contribution of this paper is the following:
First, it applies the NN to on-line approximate complicated piecewise continuous unknown
nonlinear functions in the explicit solution to Duhem model. Second, an observer is
designed to estimate the output of hysteresis of piezoelectric actuator based on the system
input and output. Third, the stability of the controlled piezoelectric actuator with the
observer is guaranteed by using Lyapunov extension (Kuczmann & Ivanyi, 2002).
The organization of the paper is as follows. In Section II, a Duhem model of hysteresis and
the problem statement are given. The main results on NN-based compensator for hysteresis
are presented in Section III. Section IV provides an example to show the feasibility of the
proposed method. Conclusions are given in Section V.
2. Preliminaries
2.1 Duhem model of hysteresis
Many different mathematic models are built to describe the hysteresis behavior, such as
Preisach model, Prandtl-Ishlinkii model and Duhem model (Coleman & Hodgdon, 1987;
Macki et al, 1993). Considering its capability of providing a finite-dimensional differential
model of hysteresis, we adopt classical Duhem model to develop the adaptive controller for
the piezoelectric actuator.
The Duhem model is a rate independent operator, with input signal v , v& and output
signal . The Duhem model describes hysteresis H (t ) by the following mathematical
model (Coleman & Hodgdon, 1987; Macki et al, 1993).
d
dv
dv
=
[ f (v ) ] +
g (v )
dt
dt
dt
(1)
where is a positive number, f (v) and g (v) are prescribed real-valued functions on
(,+) .
It can also be represented as (Coleman & Hodgdon, 1987; Macki et al, 1993):
[ f ( v ) ] + g ( v ), v& > 0
d
=
dv
[ f ( v ) ] + g ( v ), v& < 0
(2)
Adaptive Control of Piezoelectric Actuators with Unknown Hysteresis
261
where is the same positive number in (1), g (v) is the slope of the model, and f (v) is the
average value of the difference between upward side and downward side.
Property 1 (Coleman & Hodgdon, 1987; Macki et al, 1993): f (v) is a piecewise smooth,
monotone increasing, odd function with a derivative f (v) , which is not identically zero. For
large value of input v(t ) , there exists a finite limit f () ;
f (v) = f (v) , lim f (v) <
(3)
Property 2 (Coleman & Hodgdon, 1987; Macki et al, 1993): g (v) is a piecewise continuous,
even function with
g (v) = g (v) , lim g (v) <
(4)
It has been shown that Duhem model can describe a large class of hysteresis in various
smart materials, such as ferromagnetically soft material, or piezoelectric actuator by
appropriately choosing f (v) and g (v) (Coleman & Hodgdon, 1987; Macki et al, 1993). One
widely used pair of functions of f (v) and g (v) are
a1 v s + a2 (v vs )
f (v ) =
a1 v
a v + a ( v + v )
1
2
s
s
for
for
for
v >v s
(5),
v v s
v < v s
g (v) = a3
(6)
where v s > 0 , a1> 0 , a 2 > 0 , 1 >a 3 > 0 , a1 and a2 satisfy a1 , a2 [amin
amax ] , amin and
amax are known constants. Substituting the f (v) and g (v) into (2), we have
v&[a1 vs +a2(vvs) ]+a3 v&
v&[a1 v]+a3 v&
& =
v&[ a1 v]+a3 v&
v&[a1 vs a2(v+vs ) +]+a3 v&
v > vs , v& > 0

0 < v vs , v& > 0
vs v < 0, v& < 0
v < vs , v& < 0
(7)
The above equation can be solved for
a2 v
a v
= 1
a1 v
a2 v
f 21
v > v s , v& > 0

f 22 0 < v v s , v& > 0
f 23 v s v < 0, v& < 0
f 24
v < v s , v& < 0
(8)
Adaptive Control
262
with
f = (a v ) e (vv0 ) ev v (a a ) e d (a a ) v
s
2 0
0
1
2
v0 3 2
21
v
(vv0 )
ev (a3 a1) e d
f22 = (a1 v0 0 ) e
v0
v
f23 = (a1 v0 0 ) e (vv0 ) ev (a3 a1) e d
v0
(vv0 )
v

f
=
(
a
e
24
2
0
0
v0 (a3 a2 ) e d + (a1 a2 ) vs
v>vs,v&>0
0<vvs,v&>0
vs v<0,v&<0
v<vs,v&<0
(9)
In order to describe the piezoelectric actuator, we choose the same functions f (v) and g (v)
as those in (Banning, et al, 2001), which is a special case of the foregoing choice of f (v)
and g (v) , i.e. a1 = a , a2 = 0 and a3 = b when v vs .
a vs
f (v ) = a v
a v
s
g (v) = b
0
where v s > 0 ,
satisfies a [amin
for
v >v s
for
for
v v s
(10)
v < v s
for
v >v s
for
for
v v s
(11)
v < v s
b > 0 and a > b a / 2 . Suppose

amax ] , amin and amax are known constants.
a>0 ,
the
parameter a
Substituting (5) and (6) into (1), we have
v&[a vs ]
v&[a v ] +b v&
&
=
v&[ a v] +b v&
v&[a vs +]
v > vs , v& > 0

0 < v vs , v& > 0
vs v < 0, v& < 0
v < vs , v& < 0
(12)
Equation (7) can be solved for
f 21
a v f
22
=
a
v
f
23
f 24
v > vs , v& > 0

0 < v vs , v& > 0
vs v < 0, v& < 0
v < vs , v& < 0
(13)
263
where
f21 = 0e(vv0)1 avs

v > vs , v& > 0
v
f22 =(a v0 0) e(vv0) ev (b a) e d 0 < v v , v& > 0
s
v0
v
<
0, v& < 0
f23 =(a v0 0) e(vv0) ev (b a) e d s
v0
v < vs , v& < 0

f = e(vv0)1 +av
s
24 0
(14)
Equation (8) can be also expressed as:
=a 1 v(f21 2 3 + f22 1 3 + f23 1 4 + f24 2 4)
(15)
where i (i = 1,2L 4) are indicator functions defined as:
1 v vs
0 v vs
1 v& 0
0 v& 0
, 2 =
, 3 =
, 4 =
.
0 v >vs
1 v > vs
0 v& <0
1 v& < 0
1 =
Following the definition of the indicator functions, we get
1 2 =0, 1 + 2 = 1, 3 4 =0, 1 + 2 = 1, k2 =k,k =1,2,3,4

By defining &1 = & 2 = 0 , we have
& =a 1 v& ( f&21 2 3 + f&22 1 3 + f&23 1 4 + f&24 2 4)

Let
F2 = f&21 2 3 + f&22 1 3 + f&23 1 4 + f&24 2 4

and K a = a1 . We can also write the derivative of as
& = K a v& F2
(16)
2.2 Augmented Multilayer Perceptron (MLP) Neural Network

The MLP NN has been explored to approximate any function with arbitrary degree of
accuracy (Hornik et al, 1989). However, it needs a large number of NN nodes and training
iterations to approximate non-smooth functions (i.e. piecewise continuous), such as friction,
hysteresis, backlash and other hard nonlinearities. For these piecewise continuous functions,
the MLP needs to be augmented to work as a function approximator. Results for
approximation of piecewise continuous functions or functions with jumps are given in the
Adaptive Control
264
paper (Selmic & Lewis, 2000). We use the augmented NN to approximate the piecewise
continuous function in hysteresis model.
n
n
Let S be a compact set of R and define C ( S ) be the space such that the
map
f (x) : S R n
n
is
piecewise
continuous.
The
NN
can
approximate
function f ( x) C ( S ) , x R , which has a jump at x = c and is continuous from the right

as
f ( x) = W T (V T x) + W f T [V f T ( x c)] + ( x )
where
(x)
is a functional restructure error vector, W T , W f
(17)
and
, V
are
nominal constant weight matrices. () and () are activation functions for hidden
neurons.
For the hysteresis model (16), the piecewise continuous function F2 will be approximated by
the augmented NN. In this paper, it is assumed that there exists weight matrix W such that
( x) N with constant N > 0 , for all x R n , and the Frobenius norm of each matrix is
bounded by a known constant W WN with WN > 0 .
3. NN-based compensator and controller design

Given the augmented MLP NN and hysteresis model, a NN-based pre-inversion
compensator for the hysteresis is designed to cancel out the effect of hysteresis. In this
section, a novel approach is developed to compensate the hysteretic nonlinearity and to
guarantee the stability of integrated piezoelectric actuator control system.
3.1 Problem statement
Consider a piezoelectric actuator subject to a hysteresis nonlinearities described by Duhem
model. It can be identified as a second-order linear model preceded by hysteretic
nonlinearity as follows:
m &y&(t ) + b y& (t ) + k y(t ) = k c pr (t )
pr (t ) = H [v(t )]
(18)
where v(t) is the input to piezoelectric actuator, y (t ) denotes the position of piezoelectric
actuator, m , b , k denote the mass, damping and stiffness coefficients, respectively,
H () represents the Duhem model (1).
In order to eliminate the effect of hysteresis on the piezoelectric actuator system, a NNbased hysteresis compensator is designed to make the output from hysteresis model pr
approach the designed control signal pd . After the hysteresis is compensated by the NN, an
265
adaptive control for piezoelectric actuator is to be designed to ensure the stability of the
overall system and the boundedness of output tracking error of the piezoelectric actuator
with unknown hysteresis.
We consider the tracking problem, in which y (t ) is to asymptotically track a reference
signal yd (t ) having the properties that yd (t ) and its derivatives up to second derivative are
bounded, and &y&d (t ) is piecewise continuous, for all t 0 . The tracking error of the
piezoelectric actuator is defined as
e p (t ) = yd (t ) y (t ) .
(19)
A filtered error is defined as
r p ( t ) = e& p ( t ) + p e p ( t )
(20)
where p > 0 is a designed parameter.

Differentiating rp (t ) and combining it with the system dynamics Eq. (18), one may obtain:
m
b
m
r&p =
rp pr +
( &y&d + p e& p )
k c
k c
k c
1
k
b
+
( y& d + ( p ) e p ) + yd .
b
c
k c
(21)
The tracking error dynamics can be written as
m
b
r&p =
r p pr + Yd T p
k c
k c
Y d = &y&d + (
where
and
m
=
k c
p min pi p max
b
k c
k
) e& p
b
1
c
R3
y& d +
is
ep
yd
unknown
(22)
is
regression
parameter
vector
vector
with
i = 1, 2, 3 where p min and p max are some known real numbers.
3.2 NN-based Compensator for Hysteresis

In presence of the unknown hysteresis nonlinearity, the desired control signal pd for the
piezoelectric actuator is different from the real control signal pr . Define the error as
~p = pd pr
(23)
Adaptive Control
266
Differentiating (23), yields
~& p = & pd & pr
(24)
~& p = & pd K a v& + F2
(25)
thus, we have
Here we utilize a second first-layer-fixed MLP to approximate the nonlinear function F2 .

F2 = W 2 T (V 2 T h ) + W f 21T 21 [V f 21T ( h c 21 )]
(26)
+ W f 22 T 22 [V f 22 T ( h c 22 )]
+ W f 23 T 23 [V f 23 T ( h c 23 )] + 2 ( h )
where h = pd
p0
v& 1 T , p 0 is the initial value of the control signal, V2 , V f 21 ,
T
V f 22T , and V f 23T are input-layer weight matrices, W2 , W f 21T , W f 22T , and W f 23T are
output-layer weight matrices, 0, vs , and vs are jump points on the output layer, and
() , 21 () , 21 () , and 23 () are the activation functions, and 1 (h) is the functional

restructure
matrices W2
error
T
T
, W f 21
in
which
T
, W f 22 and
inversion
W f 23T
error
is
included.
Output-layer
weight
are trained so that the output of NN approximates
to the nonlinear function F2 .

Let
(h,vs ) =[(V2T h) 21(Vf 21T h) 22(Vf 22T hvs ) 23(Vf 23T h+vs )]T
and W1T = [W 2 T
W f 21T
W f 22 T
W f 23 T ] . The nonlinear function F2 is expressed as:
F2 = W1T ( h, vs ) + 1 ( h)
(27)
It is assumed that the Frobenius norm of weight matrix W1 is bounded by a known constant
W1
W1N with W1N > 0 and 1 (h) 1N with constant 1N > 0 , for all x R n .
The estimated nonlinear function F2 is constructed by using the neural network with the
weight matrix W1 :
F2 = W1T (h, vs ) .
Hence the restructure error between the nonlinear functions F2 and F2 is derived as:
267
~
~
F2 = F2 F2 = W1T (h, vs ) + 1 (h) .
~
where W 1 T = W 1 T W1 T .
Remark 1 When the input changes its sign derivative (Beuschel et al, 1998), the augmented
MLP can approximate the piecewise continuous functions. In the process, the jump
functions leads to vertical segments in the feed-forward pre-inversion compensation,
where the functional restructure error can be confronted by the adaptive controller in
Section III.C (Selmic & Lewis, 2000).
A hysteresis pre-inversion compensator is designed:
v& = {kb ~p +&pd + W1T (h, vs ) + rp}

where =
a min
a
is an estimated constant, which satisfies 0 < 1 with the known
boundary of a [amin
and W1T = [W 2 T
(28)
W f 21 T
amax ] , kb is a positive constant, a is the estimated values of a ,

W f 22 T
W f 23 T ] is the estimated output-layer weight matrix
W1T .
Define error matrix as:
~
W 1 T = W 1 T W 1 T .
Inserting (26), (28) into (25), we obtain
~&p = kb Ka ~p + (1 Ka ) &pd + (1 Ka )
~ ~
W1T (h, vs ) Ka rp +W1T (h, vs ) + 1(h)
(29)
We choose weight matrix update rule as
&
W1 = (h, vs ) ~p + k p1 ~p W1
(30)
where is a positive adaptation gain diagonal matrix, and k p1 is a positive constant.

Design the update rule of parameter in pre-inversion compensator v& as
& = Proj( , ~p [& pd + W1T (h, vs ) + rp ]

where is positive constant, Proj(.) is a projection operator, which is defined as follows:
& = Pr oj ( ,
~ pr [& pd + ( h , v s )] =
(31)
Adaptive Control
268
(h, v )] < 0
if = 1 and ~pr [& pd +
0
s
if min < < 1

a
max
(h, v )]
(h, v )] 0
or = 1 and ~pr [& pd +
~pr [& pd +
s
s
(h, v )] 0
or = min
and ~pr [& pd +
s
a
max
min
(h, v )] > 0
if
and ~pr [& pd +

=
0
s
a
max
(32)
The adaptive NN-based pre-inversion compensator v& is developed to drive the adaptive
control signal pd to approach the output of hysteresis model pr so that the hysteretic
effect is counteracted.
3.3 Controller Design Using Estimated Hysteresis Output
It is noticed that the output of hysteresis is not normally measurable for the plant subject to
unknown hysteresis. However, considering the whole system as a dynamic model preceded
by Duhem model, we could design an observer to estimate the output of hysteresis based on
the input and output of the plant.
The velocity of the actuator y& (t ) is assumed measurable. Define the error between the
outputs of actuator and observer as
e1 = y y
(33)
The observed output of hysteresis is denoted as pr and the error between the output of
hysteresis pr and the observed pr is defined as e2 = pr pr . Then the observer is
designed as:
y& = y& + L1e1
(34)
& pr = K a v& F2 + L2e1 K pr pr
(35)
The error dynamics of the observer is obtained based on the actuator model and hysteresis
model.
e&1 = L1e1 = L1e1

~
~
e&2 = K a v& F2 L2e1 + K pr pr
~
where the parameter error is defined as K a = K a K a .
(36)
269
By using the observed hysteresis output pr , we may define the signal error between the
adaptive control signal pd and the estimated hysteresis output as:
~pe = pd pr
(37)
The derivative of the signal error is:
~& pe = & pd K a v& + F2 L2e1 + K pr pr .
(38)
A hysteresis pre-inversion compensator is designed:
v& = { k b ~ pe + & pd + F 2 + r p } .
(39)
T
By substituting the neural network output F2 = W1 ( h, vs ) and pre-inversion compensator
output into the derivative of the signal error, one obtains:
~&pe = (1 Ka)&pd K a kb~pe +(1 K a)W1T(h,vs ) K a rp L2e1 + Kprpr
(40)
The weight matrix update rule is chosen as:
&
W1 = (h, vs ) ~pe + k p1 ~pe W1
(41)
And the update rule of parameter in pre-inversion compensator v& is designed with the
same projection operator as (32):
& = Proj(, ~pe [&pd +W1T (h,vs ) + rp ]) .
(42)
The update rule of parameter K a in the observer (35) is designed with the same projection
operator as (32):
&
K a = Proj(K a , ~pe [&pd +W1T (h, vs ) + rp ] + v& ~pe) .
(43)
Hence we design the adaptive controller and update rule of control parameter as:
pd = k pd rp + Yd T p
&
p = P roj (p , Yd rp )
p
(44)
(45)
Adaptive Control
270
where the projection operator is
{P roj ( p ,
p
Yd rp )}i =
0
if pi = p max and (Yd rp ) i < 0
if p min < pi < p max
(Yd rp ) i or pi = p max and (Yd rp ) i 0
or pi = p min and (Yd rp ) i 0
0
if pi = p min and (Yd rp ) i > 0
With the adaptive robust controller, pre-inversion hysteresis compensator and hysteresis
observer, the overall control system of integrated piezoelectric actuator is shown in Fig. 3.
The stability and convergence of the above integrated control system are summarized in
Theorem 1.
Theorem 1 For a piezoelectric actuator system (18) with unknown hysteresis (1) and a
desired trajectory yd (t ) , the adaptive robust controller (44), NN based compensator (39) and
hysteresis observer (34) and (35) are designed to make the output of actuator to track the
desired trajectory yd (t ) . The parameters of the adaptive robust controller and the NN based
compensator are tuned by the updating rules (41)-(43) and (45). Then, the tracking error
e p (t ) between the output of actuator and the desired trajectory yd (t ) converge to a small
neighborhood around zero by appropriately choosing suitable control gains k pd , kb and
observer gains L1 , L2 and K pr .
Proof: Define a Lyapunov function
V2 =
+
~
~
1 m
1
1
1
rp 2 + ~pe 2 + tr (W1T 1W1 ) +

(1 K a ) 2
2 k c
2
2
2 K a
1
1
1
1
( p p )T ( p p ) + e12 + e22
( K a K a ) 2 +
2
2
2
2
The derivative of Lyapunov function is obtained:

~ 1
1
m
&
&
rp r&p +~pe ~& pe tr(W1T 1W1) (1 Ka )& (Ka K a )K a
V&2 =
k c
1
&
( p p )T p + e1 e&1 + e2e&2
Introducing control strategies (39), (44) and the update rules (41)-(43), (45) into above
equation, one obtains
~
b
V&2 = (
+ k pd ) rp2 kb K a ~pe2 + 1(h)~pe k p1 ~pe tr(W1TW1)
k c
~
e r L e ~ + K ~ L e2 (L e + F )e + K e
2 p
2 1 pe
pr pr pe
11
2 1
pr pr 2
By using pr = pr e2 , F 1N and inequality: ab
1 2 1 2
a + b , one has:
2
2

~
b
+ k pd ) rp2 k b K a ~pe 2 + 1 (h)~pe k p1 ~pe tr (W1T W1 )
V&2 = (
k c
1
1
1
1 2 1 ~2 1 2 2 1 2 ~2 1 2
+ e22 + rp2 + L22 e12 + ~pe
+ pe + K pr pr + K pr pe + e2
2
2
2
2
2
2
2
2
1
1
L1e12 + ( L2 e1 + 1N ) 2 + e22 + K pr pr e2 K pr e22
2
2
271
(46)
By using the inequality 1 ( a + b ) 2 a 2 + b 2 , we can derive the following inequality:

2
b
+ k pd
V&2 (
k c
~
k ~ W (W
p1
pe
1
1
) rp 2 (k b K a K 2pr 1) ~pe 2 + ~pe 1N
2
2
~
3 2 2
2
2
2 2
L
)
e
)
(
1N
1
1
2
1 ( K pr 2)e2 + 1N + K pr pr
2
From the Property 1 of Chapter 2 in the recent book (Ikhouane & Rodellar, 2007), we know
2pr is bounded (say, 2pr M 2 where M is a constant), and then define a constant
= 12N + K 2pr M 2 > 12N + K 2pr 2pr such that

b
1
1 2
V&2 (
+ k pd ) rp 2 (kb K a K pr
1) ~pe 2 + ~pe 1N
k c
2
2
~
~
3
k p1 ~pe W1 (W1N W1 ) ( L1 L2 2 )e12 ( K pr 2)e22 +
2
(47)
We select the control parameters k pd , kb and observer parameters L1 , L2 and K pr

satisfying the following inequalities:
b
1
+ k pd > 0
k c
2
,
K pr > 2
,
kb amax K a
L1 >
1 2
K pr 1 > 0
2
,
3 2
L2
2
.
Let km = kb amax K a
~ pe >
1 2
K pr 1 . If we have
2
k p1 W1 N 2 4 + 1 N
km
~
W1 > W1N / 2 + W12N / 4 1N / k p1
(48)
we can easily conclude that the closed-loop system is semi-globally bounded (Su &
Stepanenko, 1998).
Adaptive Control
272
Hence, the following inequality holds
p1
W 1N
4 + 1N
km
< br
where br > 0 represents the radius of a ball inside the compact set C r of the tracking error
~pe (t ) .
Thus, any trajectory ~pe (t ) starting in compact set Cr = r r br converges within Cr

and is bounded.
Then the filtered error of system rp (t ) and the tracking error of the
hysteresis ~pe (t ) converge to a small neighborhood around zero. According to the standard
Lyapunov theorem extension (Kuczmann & Ivanyi, 2002), this demonstrates the UUB
~
(uniformly ultimately bounded) of rp (t ) , ~pe (t ) , W1 , e1 and e2 .
Remark 2 It is worth noting that our method is different from (Zhao & Tan, 2006; Lin et al
2006) in terms of applying neural network to approximate hysteresis. The paper (Zhao &
Tan, 2006) transformed multi-valued mapping of hysteresis into one-to-one mapping,
whereas we sought the explicit solution to the Duhem model so that augmented MLP neural
networks can be used to approximate the complicated piecewise continuous unknown
nonlinear functions. Viewed from a wavelet radial structure perspective, the WNN in the
paper (Lin et al 2006) can be considered as radial basis function network. In our scheme, the
unknown part of the solution was approximated by an augmented MLP neural network.
4. Simulation studies
In this section, the effectiveness of the NN-based adaptive controller is demonstrated on a
piezoelectric actuator described by (18) with unknown hysteresis. The coefficients of the
dynamic system and hysteresis model are m =0.016kg, b =1.2Ns/m, k =4500N/ m, c =0.9
m /V, a =6, b =0.5, vs =6 m /s, = 0.1 , k pd = 50 .
The input reference signal is chosen as the desired trajectory: yd = 3 sin(0.2 t ) .
The
control objective is to make the output signal y follow the given desired trajectory y d . From
Fig. 1, one may notice that relatively large tracking error is observed in the output response
due to the uncompensated hysteresis.
The Neural Network has 10 hidden neurons for the first part of neural network and 5
hidden neurons for the rest parts of neural network with three jumping points (0, vs , vs ).
The gains for updating output weight matrix are all set as = diag {10}25 X 25 . The activation
function () is a sigmoid basis function and activation function () has the

k
1 e x
x 0 , otherwise zero. The parameters for the observer are

1 + e x
set as: K a = 20 , kb = 100 , = 0.1, = 0.1 , K pr = 10 , L1 = 100 , L2 = 1 and initial

definition () =
273
conditions are y (0) = 0 , (0) = 0 . The system responses are shown in Fig.2, from which it
is observed that the tracking performance is much better than that of adaptive controlled
piezoelectric actuator without hysteretic compensator.
The input and output maps of NN-based pre-inversion hysteresis compensator and
hysteresis are given in Fig. 3, respectively. The desired control signal and real control signal
map (Fig. 3c) shows that the curve is approximate to a line which means the relationship
between two signals is approximately linear with some deviations.
In order to show the effectiveness of the designed observer, we compare the observed
hysteresis output pr and the real hysteresis output pr in Fig. 4. The simulation results
show that the observed hysteresis output signal can track the real hysteresis output.
Furthermore, the output of adaptive hysteresis pre-inversion compensator v (t ) is shown in
Fig.5. The signal is shown relatively small and bounded.
(a)
(b)
Fig. 1 Performance of NN controller without hysteretic compensator (a) The actual control
signal (dashed line) with reference (solid) signal; (b) Error y1 yd
4
Reference
Actual
2
0
-2
-4
0
10
15
20
Time (s)
(a)
(b)
Fig. 2. Performance of NN controller with hysteresis, its compensator and observer (a) The
actual control signal (dashed line) with reference (solid) signal; (b) Error y1 yd
Adaptive Control
274
Hysteresis
3
2
1
0
-1
-2
-3
-1
-0.5
0
v (t)
0.5
(a)
1
Pre-inversion Hysteresis Compensator
v
(t)
0.5
-0.5
-1
-3
-2
-1
(b)
3
Desired and Estimated Control Signal
2
1
0
-1
-2
-3
-3
-2
-1
(c)
Fig. 3. (a) Hysteresiss input and output map pr vs. v ; (b) Pre-inversion compensators
input and output map v vs. pd ; (c) Desired control signal and Observed control signal
curve pr vs. pd .
60
Actual Ouput
Observed Output
40
20
0
-20
0
10
20
30
Time (s)
40
50
Fig. 4. Observed Hysteresis Ouput pr and Real Hysteresis Output pr

4
3
2
1
0
-1
-2
0
10
20
30
Tim e (s)
40
Fig. 5. Adaptive Hysteresis Pre-inversion Compensator v (t )
50
275
5. Conclusion
In this paper, an observer-based controller for piezoelectric actuator with unknown
hysteresis is proposed. An augmented feed-forward MLP is used to approximate a
complicated piecewise continuous unknown nonlinear function in the explicit solution to
the differential equation of Duhem model. The adaptive compensation algorithm and the
weight matrix update rules for NN are derived to cancel out the effect of hysteresis. An
observer is designed to estimate the value of hysteresis output based on the input and
output of the plant. With the designed pre-inversion compensator and observer, the stability
of the integrated adaptive system and the boundedness of tracking error are proved. Future
work includes the compensator design for the rate-dependent hysteresis.
6. References
Banks, H. T. & Smith, R. C. (2000). Hysteresis modeling in smart material systems, J. Appl.
Mech. Eng, Vol. 5, pp. 31-45.
Tan, X. & Baras, J. S. (2004). Modelling and control of hysteresis in magnetostrictive
actuators, Automatica, Vol. 40, No. 9, pp. 1469-1480.
Brokate, M. & Sprekels, J. (1996). Hysteresis and Phase Transitions, New York: Springer-Verlag.
Visintin, A. (1994). Differential Models of Hysteresis, New York: Springer-Verlag.
Jiles, D. C. & Atherton, D. L. (1986). Theory of ferromagnetic hysteresis, J. Magnet. Magn.
Mater, Vol. 61, pp. 48-60.
Tao, G. & Kokotovic, P. V. (1995). Adaptive control of Plants with Unknown Hysteresis,
IEEE Transactions on Automatic Control, Vol. 40, No. 2, pp. 200-212.
Su, C. Y.; Stepanenko, Y.; Svoboda, J. & Leung, T. P. (2000). Robust Adaptive Control of a
Class of Nonlinear Systems with Unknown Backlash-Like Hysteresis, IEEE
Su, C. Y.; Wang, Q.; Chen, K. & Rakheja, S. (2005). Adaptive Variable Structure Control of a
Class of Nonlinear Systems with Unknown Prandtl-Ishlinskii Hysteresis, IEEE
Tan, X. & Baras, J. S. (2005). Adaptive Identification and Control of Hysteresis in Smart
Materials, IEEE Transactions on Automatic Control,Vol.50, No.(6), pp. 827-839.
Ge, S. S. & Wang, C. (2002). Direct adaptive NN control of a class of nonlinear systems, IEEE
Makaveev, D.; Dupre, L, & Melkebeek, J. (2002). Neural-Network-Based Approach to
Dynamic Hysteresis for Circular and Elliptical Magnetization in Electrical Steel
Sheet, IEEE Transactions on Magnetics, Vol. 38, No.5, pp. 3189-3191.
Kuczmann, M. &Ivanyi, A. (2002). A New Neural-Network-Based Scalar Hysteresis Model,
IEEE Transactions on Magnetics, Vol. 38, No.2, pp. 857-860.
Beuschel, M.; Hangl, F, & Schroder, D. (1998). A General Approach for Hysteresis Modeling
and Identification Using Neural Networks, Proceeding of the 1998 IEEE world
Congress on Computational Intelligence & The 1998 IEEE International Joint Conference,
Vol.3, pp. 2425-2428.
Zhao, X. & Tan, Y. (2006). Neural network based identification of Preisach-type hysteresis in
piezoelectric actuator using hysteretic operator, Sensors and Actuators A: Physical,
Vol.126, pp.306-311.
276
Adaptive Control
Lin, F.; Shieh, H. & Huang P. (2006). Adaptive Wavelet Neural Network Control with
Hysteresis Estimation for Piezo-Positioning Mechanism, IEEE Transactions on
Neural Network, Vol. 17, No. 2, pp. 432-444.
Yang, B. & Lin, W. (2004). Homogeneous Observers, Iterative Design, and Global
Stabilization of High-Order Nonlinear Systems by Smooth Output Feedback, IEEE
Krener, A. J. & Isidori, A. (1983). Linearization by output injection and nonlinear observer,
Systems & Control Letters, Vol.3, pp. 47-52.
Krener, A. J. & Kang, W. (2003). Locally convergent nonlinear observers, SIAM J. Control
Optim., Vol. 42, No. 1, pp. 157-177.
Xie, W. F. (2007). Sliding-Mode-Observer-Based Adaptive Control for Servo Actuator with
Friction, IEEE Transactions on Industrial Electronics, Vol. 54, No. 3, pp. 1517-1528.
Kuczmann, M. & Ivanyi, A. (2002). A New Neural-Network-Based Scalar Hysteresis Model,
IEEE Transactions on Magnetics, Vol.38, No. 2, pp. 857-860.
Coleman, B. D. & Hodgdon, M. L. (1987). On a class of constitutive relations for
ferromagnetic hysteresis, Arch. Rational Mech. Anal., pp: 375-396.
Macki, J. W.; Nistri, P. & Zecca, P. (1993). Mathematical models for hysteresis, SIAM Review,
Vol.35, No. 1, pp. 94-123.
Banning, R.; Koning, W. L.; Adriaens, J. J. & Koops, R. K. (2001). State-space analysis and
identification for a class of hysteretic systems, Automatica, Vol. 37, No. 4, pp. 18831892.
Hornik, K.; Stinchcombe, M. & White, H. (19989). Multilayer feedforward networks are
universal approximators, Neural Networks, Vol.2, pp. 359-366.
Selmic, R. R. & Lewis, F. L. (2000). Deadzone Compensation in Motion Control Systems
Using Dynamical Neural Networks, IEEE Transactions on Automatic Control, Vol.45,
No. 4, pp. 602-613.
Ikhouane, F. & Rodellar, J. (2007). Systems with Hysteresis-Analysis, Identification and Control
using the Bouc-Wen Model, England: John Wiley.
Su, C. Y. & Stepanenko, Y. (1998). Redesign of hybrid adaptive/robust motion control of
rigid-link electrically-driven robot manipulators, IEEE Transactions on Robotics and
Automation, Vol.14, No. 4, pp. 651-655.
13
On the Adaptive Tracking Control of 3-D
Overhead Crane Systems
Yang, Jung Hua
National Pingtung University of Science and Technology

Pingtung, Taiwan
1. Introduction
For low cost, easy assembly and less maintenance, overhead crane systems have been
widely used for material transportation in many industrial applications. Due to the
requirements of high positioning accuracy, small swing angle, short transportation time,
and high safety, both motion and stabilization control for an overhead crane system
becomes an interesting issue in the field of control technology development. Since the
overhead crane system is underactuated with respect to the sway motion, it is very difficult
to operate an overhead traveling crane automatically in a desired manner. In general,
human drivers, often assisted by automatic anti-sway system, are always involved in the
operation of overhead crane systems, and the resulting performance, in terms of swiftness
and safety, heavily depends on their experience and capability. For this reason, a growing
interest is arising about the design of automatic control systems for overhead cranes.
However, severely nonlinear dynamic properties as well as lack of actual control input for
the sway motion might bring about undesired significant sway oscillations, especially at
take-off and arrival phases. In addition, these undesirable phenomena would also make the
conventional control strategies fail to achieve the goal. Hence, the overhead crane systems
belong to the category of incomplete control system, which only allow a limited number of
inputs to control more outputs. In such a case, the uncontrollable oscillations might cause
severe stability and safety problems, and would strongly constrain the operation efficiency
as well as the application domain. Furthermore, an overhead crane system may experience
a range of parameter variations under different loading condition. Therefore, a robust and
delicate controller, which is able to diminish these unfavorable sway and uncertainties,
needs to be developed not only to enhance both efficiency and safety, but to make the
system more applicable to other engineering scopes.
The overhead crane system is non-minimum phase (or has unstable zeros in linear case) if a
nonlinear state feedback can hold the system output identically zero while the internal
dynamics become unstable. Output tracking control of non-minimum phase systems is a
highly challenging problem encountered in many practical engineering applications such as
aircraft control [1], marine vehicle control [2], flexible link manipulator control [3], inverted
pendulum system control [4]. The non-minimum phase property has long been recognized
to be a major obstacle in many control problems. It is well known that unstable zeros cannot
278
Adaptive Control
be moved with state feedback while the poles can be arbitrarily placed (if completely
controllable). In most standard adaptive control as well as in nonlinear adaptive control, all
algorithms require that the plant to be minimum phase. This chapter presents a new
procedure for designing output tracking controller for non-minimum phase systems (The
overhead crane systems).
Several researchers have dealt with the modeling and control problems of overhead crane
system. In [5], a simple proportional derivative (PD) controller is designed to asymptotically
regulate the overhead crane system to the desired position with natural damping of sway
oscillation. In [6], the authors propose an output feedback proportional derivative controller
that stabilizes a nonlinear crane system. In [7], the authors proposed an indirect adaptive
scheme, based on dynamic feedback linearization techniques, which was applied to
overhead crane systems with two control input. In [8], Li et al attacked the under-actuated
problem by blending four local controllers into one overall control strategy; moreover,
experimental results delineating the performance of the controller were also provided. In [9],
a nonlinear controller is proposed for the trolley crane systems using Lyapunov functions
and a modified version of sliding-surface control is then utilized to achieve the objective of
cart position control. However, the sway angle dynamics has not been considered for
stability analysis. In [10], the authors proposed a saturation control law based on a
guaranteed cost control method for a linearized version of 2-DOF crane system dynamics.
In [11], the authors designed a nonlinear controller for regulating the swinging energy of
the payload. In [12], a fuzzy logic control system with sliding mode Control concept is
developed for an overhead crane system. Y. Fang et al. [13] develop a nonlinear coupling
control law to stabilize a 3-DOF overhead crane system by using LaSalle invariance theorem.
However, the system parameters must be known in advance. Ishide et al. [14] train a fuzzy
neural network control architecture for an overhead traveling crane by using
back-propagation method. However, the trolley speed is still large even when the
destination is arrived, which would result in significant residual sway motion, low safety,
and poor positioning accuracy. In the paper [15], a nonlinear tracking controller for the load
position and velocity is designed with two loops: an outer loop for position tracking, and an
inner loop for stabilizing the internally oscillatory dynamics using a singular perturbation
design. But the result is available only when the sway angle dynamics is much faster than
the cart motion dynamics. In the paper [16], a simple control scheme, based on second-order
sliding modes, guarantees a fast precise load transfer and swing suppression during the
load movement, despite of model uncertainties. In the paper [17], it proposes a stabilizing
nonlinear control law for a crane system having constrained trolley stroke and pendulum
length using the Lyapunovs second method and performs some numerical experiments to
examine the validity of the control law. In the paper [18], the variable structure control
scheme is used to regulate the trolley position and the hoisting motion towards their
desired values. However the input torques exhibit a lot of chattering. This chattering is not
desirable as it might shorten the lifetime of the motors used to drive the crane. In the paper
[19], a new fuzzy controller for anti-swing and position control of an overhead traveling
crane is proposed based on the Single Input Rule Modules (SIRMs). Computer simulation
results show that, by using the fuzzy controller, the crane can be smoothly driven to the
On the Adaptive Tracking Control of 3-D Overhead Crane Systems
279
destination in a short time with low swing angle and almost no overshoot. D. Liu et al. [20]
present a practical solution to analyze and control the overhead crane. A sliding mode fuzzy
control algorithm is designed for both X-direction and Y-direction transports of the
overhead crane. Incorporating the robustness characteristics of SMC and FLC, the proposed
control law can guarantee a swing-free transportation. J.A. Mendez et al. [21] deal with the
design and implementation of a self-tuning controller for an overhead crane. The proposed
neurocontroller is a self-tuning system consisting of a conventional controller combined
with a NN to calculate the coefficients of the controller on-line. The aim of the proposed
scheme is to reduce the training-time of the controller in order to make the real-time
application of this algorithm possible. Ho-Hoon Lee et al. [22] proposes a new approach for
the anti-swing control of overhead cranes, where a model-based control scheme is designed
based on a V-shaped Lyapunov function. The proposed control is free from the
conventional constraints of small load mass, small load swing, slow hoisting speed, and
small hoisting distance, but only guarantees asymptotic stability with all internal signals
bounded. This paper also proposes a practical trajectory generation method for a near
minimum-time control, which is independent of hoisting speed and distance. In this paper
[23], robustness of the proposed intelligent gantry crane system is evaluated. The evaluation
result showed that the intelligent gantry crane system not only has produced good
performances compared with the automatic crane system controlled by classical PID
controllers but also is more robust to parameter variation than the automatic crane system
controlled by classical PID controllers. In this paper [24], the I-PD and PD controllers
designed by using the CRA method for the trolley position and load swing angle of
overhead crane system have been proposed. The advantage of CRA method for designing
the control system so that the system performances are satisfied not only in the transient
responses but also in the steady-state responses, have also been confirmed by the simulation
results.
Although most of the control schemes mentioned above have claimed an adaptive
stabilizing tracking/regulation for the crane motion, the stability of the sway angle
dynamics is hardly taken into account. Hence, in this chapter, a nonlinear control scheme
which incorporates both the cart motion dynamics and sway angle dynamics is devised to
ensure the overall closed-loop system stability. Stability proof of the overall system is
guaranteed via Lyapunov analysis. To demonstrate the effectiveness of the proposed
control schemes, the overhead crane system is set up and satisfactory experimental results
are also given.
2. Dynamic Model of Overhead Crane
The aim of this section is to drive the dynamic model of the overhead crane system. The
model is dived using Lagrangian method. The schematic plotted in Figure 1 represents a
three degree of freedom overhead crane system. To facilitate the control development, the
following assumptions with regard to the dynamic model used to describe the motion of
overhead crane system will be made. The dynamic model for a three degree of freedom
(3-DOF) overhead crane system (see Figure 1) is assumed to have the following postulates.
A1: The payload and the gantry are connected by a mass-less, rigid link.
A2: The angular position and velocity of the payload and the rectilinear position and
280
Adaptive Control
velocity of the gantry are measurable.

A3: The payload mass is concentrated at a point and the value of this mass is exactly
known; moreover, the gantry mass and the length of the connecting rod are exactly known.
A4: The hinged joint that connects the payload link to the gantry is frictionless.
Fig. 1. 3-D Overhead Crane System

The 3-D crane system will be derived based on Lagrange-Euler approach. Consider the
3-dimensional overhead crane system as shown in Figure 1. The cart can move horizontally
in x-y plane, in which the moving distance of the cart along the X-rail is denoted as x(t) and
the distance on the Y-rail measured from the initial point of the construction frame is
denoted as y(t). The length of the lift line is denoted as l. Define the angle between the lift
line and its projection on the y-z plane as (t ) and the angle between the projection line
and the negative axis as (t ) . Then the kinetic energy and potential energy of the system
can be found in Equation (1.1) and (1.2), respectively and be expressed as the following
equations.
K=
1
1
1
m1 x& 2 + (m1 + m2 ) y& 2 + mc ( x& c2 + y& c2 + z& c2 )
2
2
2
V = mgl cos cos
(1)
(2)
where x c , y c are the related positions of the load described in the Cartesian coordinate,
which can be mathematically written as
x c = x + l sin
(3)
281
y c = y + l cos sin
(4)
z c = l cos cos
(5)
The following equations express the velocities by taking the time derivative of above
equations
x& c = x& + l& cos
(6)
y& c = y& l& sin sin + l& cos cos
(7)
z& c = l& sin cos l& cos sin
(8)
By using the Lagrange-Euler formulation,
d L
dt q& i
L

= i , i = 1,2,3,4.
qi
(9)
where L = K V , q i is the element of vector q = [ x y ] and i is the

corresponding external input to the system, we have the following mathematical
representation which formulates the system motion
T
M (q )q&& + C (q, q& ) + G (q ) =

44
(10)
41
where M ( q ) R
is inertia matrix of the crane system, C ( q, q& ) R
is the
41
nonlinear terms coming from the coupling of linear and rotational motion, G ( q ) R
T
is the terms due to gravity, and = [u x u y 0 0] is the input vector.
As mentioned previously, the dynamic equation of motion described the overhead crane
system also have the same properties as follows
n
P1:The inertia matrix M (q ) is symmetric and positive definite for all q R .
4
P2:There exists a matrix B ( q, q& ) such that C ( q, q& ) = B ( q, q& ) q& , and x R
T
x ( M& 2 B) x = 0 , i.e., M& 2 B is skew-symmetric. B(q, q& x )q& y = B(q, q& y )q& x .
P3:The parameters of the system can be linearly extracted as
M (q)q&& + C (q, q& ) + G (q) = W f (q, q& , q&&) f
(11)
&&) is the regressor matrix and f is a vector containing the system

where W f ( q, q& , q
parameters.
Dynamic Model of Overhead Crane
In this section, an adaptive control scheme will be developed for the position tracking of an
overhead crane system.
282
Adaptive Control
2.1 Model formulation

For design convenience, a general coordinate is defined as follows
q T = [q Tp
where
q Tp = [ x
qT ]
y ] , qT = [
and using the relations in P2, the dynamic equation of an overhead crane (10) is partitioned
in the following form
M pp
M T
p
M p q&&p B pp
+
M q&& Bp
B p q& p G p (q ) u p
+
=
B q& G (q ) 0
(12)
where M pp , M p , M , B pp , B p , Bp , B are 22 matrices partitioned from

the inertia matrix M (q ) and the matrix B ( q, q& ) , respectively, G p , G are 21
T
vectors, and u p = [u x u y ] . Before investigating the controller design, let the error
signals be defined as
e = q q d = [e Tp
eT ]T
(13)
and the stable hypersurface plane is defined as
e& + K p e p s p
s = e& + Ke = p
=
e& + K e s
(14)
where
e p = q p q pd = [ x x d
y y d ]T [e x
e = q qd = [ d
d ]T [e
k
Kp = 1
0
0
k 3
,
=
K
0
k 2
e y ]T ,
e ]T ,
0
k 4
and x d , y d , d and d are defined trajectories of x , y , and respectively,

and K p , K are some arbitrary positive definite matrices.
Then, after a lot of mathematical arrangements, the dynamics of the newly defined signal
vectors s p , s can be derived as
283
M PP
M T
P
M P s& p B PP
+
M s& BP
B P s p p + u P
=
B s
(15)
where
p = M pp (q&&pd + k p e& p ) + M p (q&& + k e& ) + Bpp (q& pd + k p ep ) + Bp (q& +k pe )
(16)
= Mp (q&&pd + k pe& p ) + M (q&& + k e& ) + Bp (q& pd + k pep ) + B (q& + k e )
(17)
Remark 1: The desired trajectories x d , y d , d and d should be carefully chosen so as

to satisfy the internal dynamics, as shown in the lower part of equation (15), when the
control objective is achieved, i.e.,
&
x&
&&
&x&
T
M P (qd ) d + M (qd ) &&d + BP (q, q&) d + B (q, q&) &d + G (q) = 0
d
y& d
d
&y&d
(18)
Without loss of generality, we always choose an exponentially-convergent trajectories with

final constant values for x d , y d and zero for d , d .
2.2 Adaptive Controller Design
In this subsection, an adaptive nonlinear control will be presented to solve the tracking
control problem.
q p , q& p , q , q&
q p , q& p , q , q&
Fig. 2. An Adaptive Self-tuning Controller Block Diagram

As indicated by property P3 in section 1.2, the dynamic equations of an overhead crane
have the well-known linear-in-parameter property. Thus, we define
284
Adaptive Control
11 = M pp (q&&pd + k p e& p ) + M p (q&&d k e& ) + Bpp (q& pd + k p e p ) + Bp k pe p
(19)
2 2 = M p (q&&pd + k p e& p ) + M (k e& ) + Bp (q& pd + k p e p ) + B k e
(20)
where 1 , 2 are regressor matrices with appropriate dimensions, and 1 , 2 are their
corresponding vectors of unknown constant parameters, respectively. As a majority of the
adaptive controller, the following signal is defined
2( Z x a x (t ) + bx (t )),
Z& x = 2bx (t ), Z x (t ) = 0,
, Z (t ) = 0,
x x
where
Z x (t ) > 0
bx (t ) > 0
bx (t ) 0
(21)
is some small positive constant and
a x (t ) =
bx (t ) =
sp
sp
sp
T
T
( s 22 s K v s )
(22)
( s 22 s K v s )
(23)
Remark 2: Note that (21) is simply to define a differential equation of which its variable
Z x (t ) remains positive. Let another signal k(t) be defined to be its positive root, i.e.,
k = Z x , It can be shown that
2
1 k sp +
k&(t ) =
(
)( sT 22 sT K v s )
k (t ) s 2 +
p
k0
(24)
In the sequel, we will first assume that there exists a measure zero set of time sequences
t i i =1 such that Z (t i ) = 0 or k (t i ) = 0 , i = 1,2,3,... , and then, verify the existence

assumption valid.
Now let the adaptive control law be designed as
{}
u P = 11 v K vp s p
(25)
(k 1) s
p
T
T
( s 2 2 s K v s )
v =
s 2 +
285
(26)
where
11 = M pp (q&&pd + k p e& p ) + M p (k e& ) + B pp (q& pd + k p e p ) + B p k p e p
(27)
22 = M p (q&&pd + k p e& p ) + M (k e& ) + Bp (q& pd + k p e p ) + B k e
(28)
and 1 , 2 are the estimates of

obtained as
M PP M P s& p BPP
M T M s& + B
P
P
1 , 2 respectively,
BP s p K vp
+
B s 0
then the error dynamics can be
~
0 s p 11 v
=
K v s 22 + K v s
(29)
or more compactly as
~
11 v
M (q)s& + h(q, q& )s + Ks =
22 + K v s
(30)
where
~
1 1 1
~ =
2 2 2
(31)
Moreover, let the adaptation laws be chosen as
&
1 = k a 1 s p
&
2 = k b 2 s
(32)
where k a , k b are some positive definite gain matrices. In what follows we will show that
the error dynamics (30) along with the adaptive laws (32) constitutes an asymptotically
stable closed-loop dynamic system. This is exactly stated in the following theorem.
286
Adaptive Control
Theorem : Consider the 3-D overhead crane system as mathematically described in (10) or (12) with
all the system parameters unknown. Then, by applying control laws (25)-(28) and adaptive laws (32),
the objective for the tracking control problem can be achieved, i.e., all signals inside the closed-loop
system (29) are bounded and e x , e , e y , e 0 asymptotically in the sense of Lyapunov.
Proof: Define the Lyapunov function candidate as
1
1~
~ 1~
~ 1
V (t ) = s T M (q)s + 1T k a11 + 2T kb12 + Z x
2
2
2
2
1
1~
~ 1~
~ 1
= s T M (q) s + 1T k a11 + 2T k b1 2 + k 2
2
2
2
2
It is obvious that, due to the quadratic form of system states as well as the definition of
Z x (t ) , V(t) is always positive-definite and indeed a Lyapunov function candidate. By
taking the time derivative of V we get
1
~& T 1 ~ ~& T 1 ~
V& (t ) = s T M (q ) s& + s T M& (q) s + 1 k a 1 + 2 k b 2 + kk&
2
~
11 v 1 T
~
~
T
T
= s ( B(q, q& ) s K vp s +
) + s M& (q ) s + s p 11 + s 2 2
2 2 + K v s 2
T
+(
k sp
sp
2
2
+
+
)( s 22 s K v s )
T
= sKs s p 11 (
T
+(
k sp
sp
2
2
+
+
(k 1) s p
2
sp +
~
T
T
T
T
)( s 22 s K v s ) + s 22 + s p 11
~
T
T
T
T
)( s 22 s K v s ) + s 2 2 + s K v s
= s T Ks
(33)
287
~ ~
It is clear that V& (t ) < 0 as long as K > 0 , which then implies s, k , 1 , 2 L Now,
assume that k (t ) = 0 instantaneously at t i . Because the solution Z x (t ) of the equation
(21) is well defined and is continuous for all t 0 , k(t) is continuous at t i , i.e.,
k (t i ) = k (t i +) . Since V is a continuous function of k , it is clear that V (t) remains
to be continuous at t i , i.e. , V (t i ) = V (t i + ) . Form then hypothesis, V& (ti ) < 0 and
+
V& (t i ) < 0, we hence can conclude that V is nonincreasing in t including t i , which then
readily implies that s, k L . Therefore, e, v and L directly from equation
&
(13) and definitions of v and . It then follows from

t i i =1 (30) that s L . On the other
hand,
if
the
set
of
time
instants
is
measure
zero,
then
2
0 V&dt = V (0) V () < or equivalently that 0 s dt < so that s L2 .
Form the error dynamics, we can further conclusion that s& L . Then by Barbalats
lemma we readily obtain that s 0 as t asymptotically as t and therefore,
e, e& 0 as t Note that in the above proof we have used the property
( M& (q) 2 B(q, q& )) is skew- symmetric. Finally, to complete the proof in theory, we
need to show that the above hypothesis that the set of time instants t i i =1 is indeed
measure zero. However, it is quite straightforward to conclude the result from (21) by
simply using the fact that all signals are bounded. This completes our proof.
{}
{}
Remark 3: From the robustness point of view, it would be better if additional feedback term
k q s is included in the control law (24). With such an inclusion, the sway stabilization
result subject to external disturbance can also be maintained as the cart arrived at its
destination. This can be easily checked from the stability proof given in the theorem.
Proof: Let the Lyapunov function candidate be chosen as
V (t ) =
1 T
1 ~ T 1 ~
1 ~ T 1 ~
1
s M ( q ) s + 1 k a 1 + 2 k b 2 + Z x
2
2
2
2
1 ~T
1 T
1 ~ T
1
1 ~
1 ~
s M (q )s + 1 k a 1 + 2 k b 2 +
k
2
2
2
2
and take the time derivative of V to get
1
~& T 1 ~ ~& T 1 ~
V& (t ) = s T M (q) s& + s T M& (q) s + 1 k a 1 + 2 k b 2 + kk&
2
~
11 v k q s 1 T &
~
~
T
T
= s (B(q, q&)s K vp s +
) + s M (q)s + s p 11 + s 22
22 + K v s 2
T
288
Adaptive Control
+(
k sp +
2
sp +
T
T
)(s 22 s Kv s )
= s T Ks k q s s p
V& (t ) = s T K v s k q s s p
1
2
2
2
2
V& (t ) min ( K v )( s p + s ) + k q ( s p + s )
2
1
2
2
= (min ( K v ) k q )( s p + s )
2
Thus, the same conclusion can be made as preciously if
min ( K v ) >
1
kq
2
3. Computer Simulation
In this subsection, several simulations are performed and the results also confirm the
validity of our proposed controller. The desired positions for X and Y axes are 1 m. Figure 3
shows the time response of X-direction. Figure 5 show the time responses of Y-direction. It
can be seen that the cart can simultaneously achieve the desired positions in both X and Y
axes in approximately 6 seconds with the sway angles almost converging to zero at the same
time. Figure 4 and Figure 6 show the response of the sway angle with the control scheme.
Figure 7 and Figure 8 show the velocity response of both X-direction and Y-direction. Figure
9 and Figure 10 show the control input magnitude. In Figure 11~14, the parameter estimates
are seen to converge to some constants when error tends to zero asymptotically and the time
response of the tuning function k(t) is plotted in Figure 15.
The control gains are chosen to be
1.5 0
2.35 0
kp =
, k =
,
1
0 1
0
1.5 0
1.35 0
k vp =
, k v =
0 1.8
0 1.2
The corresponding adaptive gains are set to be
Fig. 3. Gantry Tracking Response
Fig. 4. Sway Angle Response
x(t )
(t )
k a = kb = 1
with Adaptive Algorithm
y (t )
289
290
Adaptive Control
(t )
Fig. 7. Gantry Velocity Response
x& (t )
Fig. 8. Gantry Velocity Response
y& (t )
Fig. 9. Force Input
Fig. 10. Force Input
ux
uy
Fig. 11. Estimated Parameters 1 x (t )
291
292
Adaptive Control
Fig. 12. Estimated Parameters
1 y (t )
2 (t )
Fig. 15. Response Trajectory of
293
k (t )
4. Experimental Verification
In this section, to validate the practical application of the proposed algorithms, a three
degree-of-freedom overhead crane apparatus, is built up as shown in Figure 16. Several
experiments are also performed and indicated in the subsequent section for demonstration
of the effectiveness of the proposed controller.
Fig. 16. Experimental setup for the overhead crane system

The control algorithm is implemented on a xPC Target for use with real time Workshop
manufactured by The Math Works, Inc., and the xPC target is inserted in a Pentium4
294
Adaptive Control
2.4GHz PC running under the Windows operating system. The sensing system includes the
two photo encoders and two linear position sensors. The cart motion X-direction and
Y-direction motion measured by linear potentiometer. Two potentiometers are connected to
the travel direction and the traverse direction. An AC servo motor with 0.95 N-m maximum
torque and 3.8N-m maximum torque output is used to drive the cart motion X direction and
Y direction. The servomotors are set in torque control mode so as to output the desired
torques.
In the experimental study, the proposed control algorithms have been tested and compared
with the conventional PD controller. From the experimental results, it is found that our
proposed algorithms indeed outperform the conventional control scheme in all aspects. A
schematic description of the experimental system is draw in Figure 17.
Fig. 17. A Schematic Overview of the Experimental Setup
4.1 Experiments for Conventional PD control as a comparative study

In the experiments, a simple PD control scheme with only position and velocity feedback is
first tested for the crane control. Figure 18 and Figure 20 show the control responses. From
Figure 19 and Figure 21 it is observed that the sway oscillation can not be rapidly damped
by using only conventional PD control, although the tracking objective is ultimately
achieved.
x(t )
with Conventional PD Control
(t )
(t )
295
296
Adaptive Control
y (t )
4.2 Experiments for the Proposed Adaptive Control Method with Set-point Regulation
In the subsection, the developed adaptive controller is applied. The following controlled
gains are chosen for experiments.
2
kp =
0
1.5
k vp =
0
0
3 0
, k =
,
1
0 1
0
1.35 0
, k v =
3
2
0
The corresponding adaptive gains are set to be 1 i.e., k a = k b = 1 . Figure 22~31 depict the
experimental results for the crane system with the adaptive control law. Figure 22 and
Figure 24 demonstrate the tracking performance in X and Y directions. It is experimentally
demonstrated that the sway angle can be actively damped out by using our proposed
adaptive schemes, as shown in Figure 23 and Figure 25 with maximum swing angle about
0.05 rad and 0.06 rad, respectively. Figure 26 and Figure 27 show the input torques from
each AC servo motors, whereas Figure 28~30 plot the associated adaptive gain turning
trajectories. The trajectory of coupling gain k(t) is also in Figure 31 with initial value 0.05.
The initial values of other state variable are all zero. Apparently the tracking and damping
performances by applying the adaptive control algorithm are much better than the ones
resulting from the PD control.
Fig. 22. Gantry Tracking Response with Adaptive Algorithm X(t)
Fig. 23. Sway Angle Response with Adaptive Algorithm
(t)
Fig. 24. Gantry Tracking Response with Adaptive Algorithm Y(t)
297
298
Fig. 25. Sway Angle Response with Adaptive Algorithm
Fig. 26. Force Input Ux
Fig. 27. Force Input Uy
Adaptive Control
(t)
1x (t)
2 (t) and 2 (t)
1 y (t)
299
300
Fig. 31. Trajectory of
Adaptive Control
k (t)
4.3 Experiments for the Proposed Adaptive Control with Square Wave Tracking
To prove the prevalence of our controllers, experiments on the tracking of square wave, as
shown in Figure 6, is also conducted. The gains are kept the same as in the previous
experiments. Figure 6(a) and Figure 6(c) demonstrate the tracking performance in X and Y
directions, respectively while Figure 6(b) and Figure 6(d) show the suppression results of
sway angles. It is found that good performance can still be preserved is spite of the sudden
change of desired position.
Fig. 32. Desired Trajectory
Fig. 33. Tracking Response
x(t )
Fig. 35. Tracking Response
(t )
y (t )
301
302
Adaptive Control
(t )
Fig. 37. Trajectory of k(t)
5. Conclusion
In this chapter, a nonlinear adaptive control law has been presented for the motion control
of overhead crane. By utilizing a Lyapunov-based stability analysis, we can achieve
asymptotic tracking of the crane position and stabilization of payload sway angle for an
overhead crane system which is subject to both underactuation and parametric
uncertainties. Comparative simulation studies have been performed to validate the
proposed control algorithm. To practically validate the proposed adaptive schemes, an
overhead crane system is built up and experiments are also conducted. Both simulations
and experiments show better performance in comparison with the conventional PD control.
6. References
Philippe Martin, Santosh Devasia, and Brad Paden, 1996, A
Different Look at Output
303
Tracking: Control of a VTOL Aircraft. Automatica, Vol. 32, No. 1, pp. 101-107.
Yannick Morel, and Alexander Leonessa, 2003, Adaptive Nonlinear Tracking Control of an
Underactuated Nonminimum Phase Model of a Marine Vehicle Using Ultimate
Boundedness. Proc. of the 42nd IEEE Conference on Decision and Control, Maui,
Hawaii USA, pp. 3097-3102.
Xuezhen Wang, and Degang Chen, 2006, Output Tracking Control of a One-Link Flexible
Manipulator via Causal Inversion. IEEE Transactions on Control Systems Technology,
Vol. 14, No. 1, pp. 141-148.
Qiguo Yan, 2003, Output Tracking of Underactuated Rotary Inverted Pendulum by
Nonlinear Controller. Proc. of the 42nd IEEE Conference on Decision and Control,
Maui, Hawaii USA, pp. 2395-2400.
Y. Fang, E. Zergeroglu, W. E. Dixon, and D.M. Dawson, 2001, Nonlinear Coupling
Control Laws for an Overhead Crane System. Proc. of the 2001 IEEE International
Conference on Control Applications, pp. 639-644.
B. Kiss, J. Levine, and Ph. Mullhaupt, 2000, A Simple Output Feedback PD controller for
Nonlinear Cranes. Proc. of the 39th IEEE Conference on Decision and Control, Sydney,
Australia, pp. 5097-5101.
F. Boustany, and B. dAndrea-Novel, 1992, Adaptive Control of an Overhead Crane Using
Dynamic Feedback Linearization and Estimation Design. Proc. of the 1992 IEEE
International conference on Robotics and Automation, pp. 1963-1968.
W. Li, and Q. Tang, 1993, Control Design for a Highly Nonlinear System. Proc. of the
ASME Annual Winter Meeting, Vol. 50, Symposium on Mechatronics, New Orleans, LA,
pp. 21-26.
Barmeshwar Vikramaditya, and Rajesh Rajamani, 2000, Nonlinear Control of a trolley
crane system. Proc. of the American Control Conference Chicago, Illinois, pp.
1032-1036.
Kazunobu Yoshida, and Hisashi Kawabe, 1992, A Design of Saturating Control with a
Guaranteed Cost and Its Application to the Crane Control System. IEEE
Chung Choo Chung, and John Hauser, 1995, Nonlinear Control of a Swinging Pendulum.
Automatica, Vol. 31, No. 6, pp. 851-862.
Dian-Tong Liu, Jian-Qiang Yi, and Dong-Bin Zhao, 2003, Fuzzy tuning sliding mode
control of transporting for an overhead crane. Proc. the second International
Conference on Machine Learning and Cybernetics, Xian, China, pp. 2541-2546.
Y. Fang, W. E. Dixon, D. M. Dawson, and E. Zergeroglu, 2001, Nonlinear coupling control
laws for 3-DOF overhead crane system. Proc. the 40th IEEE Conference on Decision
and Control, Orlando, Florida, USA, pp. 3766-3771.
T. Ishide, H. Uchida, and S. Miyakawa, 1991, Application of a fuzzy neural network in the
automation of crane system. Proc. of the 9th Fuzzy System Symposium, pp. 29-33.
J. Yu, F. L. Lewis, and T. Huang, 1995, Nonlinear feedback control of a gantry crane. Proc.
of the American Control Conference, Seattle, Washington, USA, pp. 4310-4315.
Giorgio Bartolini, Alessandro Pisano, and Elio Usai, 2002, Second-order sliding-mode
control of container cranes. Proc. of the 38th Automatica, pp. 1783-1790.
304
Adaptive Control
Kazunobu Yoshida, 1998, Nonlinear control design for a crane system with state
constraints. Proc. of the American Control Conference, Philadelphia, Pennsylvania, pp.
1277-1283.
M. J. Er, M. Zribi, and K. L. Lee, 1988, Variable Structure Control of Overhead Crane. Proc.
of the 1998 IEEE International Conference on Control Applications, Trieste, Italy,
pp.398-402.
Jianqiang Yi, Naoyoshi Yubazaki, and Kaoru Hirota, 2003, Anti-swing and positioning
control of overhead traveling crane. Proc. of the Information Sciences 155, pp. 19-42.
Diantong Liu, Jianqiang Yi, Dongbin Zhao, and Wei Wang, 2004, Swing-Free Transporting
of Two-Dimensional Overhead Crane Using Sliding Mode Fuzzy Control. Proc. of
the 2004 American Control Conference, Boston, Massachusetts, pp. 1764-1769.
J.A. Mendez, L. Acosta, L. Moreno, A. Hamilton, and G.N. Marichal, 1998, Design of a
Neural Network Based Self-Tuning Controller for an overhead crane. Proc. of the
1998 IEEE International Conference on Control Applications, pp. 168-171.
Ho-Hoon Lee, and Seung-Gap Choi, 2001, A Model-Based Anti-Swing Control of
Overhead Cranes with High Hoisting Speeds. Proc. of the 2001 IEEE International
Conference on Robotics and Automation, Seoul, Korea, pp. 2547-2552.
Wahyudi, and J. Jalani, 2006, Robust Fuzzy Logic Controller for an Intelligent Gantry
Crane System. First International Conference on Industrial and Information Systems,
pp. 497-502.
Wanlop Sridokbuap, Songmoung Nundrakwang, Taworn Benjanarasuth, Jongkol
Ngamwiwit and Noriyuki Komine, 2007, I-PD and PD Controllers Designed by
CRA for Overhead Crane System. International Conference on Control, Automation
and Systems, COEX, Seoul, Korea, pp. 326-330.
APPENDIX A
Mathematical Description of The Dynamic Model
The dynamic equation of the 3D overhead crane system can be derived by using
Largrange-Euler formula and shown in the following
M (q)q&& + C (q, q& ) + G (q) =

where
0
m1 + mc
0
m1 + m2 + mc
M (q) =
mc l cos mc l sin sin
0
mc l cos cos
mc l cos
mc l sin sin
mc l 2
0
mc l cos cos
mc l 2 cos 2
0
305
m c l & 2 sin
2
2
&
&
m c l (& + ) cos sin 2 m c l & sin cos
C ( q , q& ) =
m c l 2 & 2 sin cos
2 m c l 2& & sin cos
G (q) =
mc gl sin cos
mc gl cos sin
= [u x u y
q = [x
0 0
]T
To satisfy property P2 as stated in section 2 the vector

C (q, q& ) = B(q& , q)q& where
0
B ( q , q& ) =
0
0
0
C (q, q& )
can be re-arranged as
m c l & 2 sin
m c l & cos sin m c l & sin cos
m c l & sin cos

2
m c l & sin cos m c l & cos sin
m c l 2 sin cos
2
m c l & sin cos
It can be easily checked that
306
Adaptive Control
0
M& 2C =
mc l& sin
0
0
mc l& cos sin mc l& sin cos
0
mc l& sin
mc l& cos sin + mc l& sin cos
0
2 &
2mc l sin cos
2mc l 2 & sin cos
which is skew-symmetrical matrix.
14
Adaptive Inverse Optimal Control of
a Magnetic Levitation System
Yasuyuki Satoh1, Hisakazu Nakamura1,
Hitoshi Katayama2 and Hirokazu Nishitani1
1Nara
Institute of Science and Technology, 2Shizuoka University

Japan
1. Introduction
In recent years, control Lyapunov functions (CLFs) and CLF-based control designs have
attracted much attention in nonlinear control theory. Particularly, CLF-based inverse
optimal controllers are some of the most effective controllers for nonlinear systems [Sontag
(1989); Freeman & Kokotovi (1996); Sepulchre et al. (1997); Li & Krsti (1997); Krsti & Li
(1998)]. These controllers minimize a meaningful cost function and guarantee the optimality
and a stability margin. Moreover, we can obtain the optimal controller without solving the
Hamilton-Jacobi equation. An inverse optimal controller with input constraints has also
been proposed [Nakamura et al. (2007)]. On the other hand, these controllers assume that
the desired state of the controlled system is an equilibrium state. Then, if the controlled
system does not satisfy the assumption, we have to use a pre-feedback control design
method to the assumption is virtually satisfied. However, a pre-feedback control design
causes the luck of robustness. This implies that a stability margin of inverse optimal
controllers is lost. Hence the designed controller does not asymptotically stabilize the
system if there exists a parameter uncertainty in the system.
In this article, we study how to guarantee a stability margin when the pre-feedback
controller design is used. We consider a magnetic levitation system as an actual control
example and propose an adaptive inverse optimal controller which guarantees a gain
margin for the system. The proposed controller consists of a conventional inverse optimal
controller and a pre-feedback compensator with an adaptive control mechanism. By
introducing adaptive control law based on adaptive control Lyapunov functions (ACLFs),
we can successfully guarantee the gain margin for the closed loop system. Furthermore, we
apply the proposed method to the actual magnetic levitation system and confirm its
effectiveness by experiments.
This article is organized as follows. Section 2 introduces some mathematical notation and
definitions, and outlines the previous results of CLF-based inverse optimal control design.
Section 3 describes the experimental setup of the magnetic levitation system and its
mathematical model. In section 4, we design an inverse optimal controller with a prefeedback compensator for the magnetic levitation system. The problem with the designed
controller is demonstrated by the experiment in section 5. To deal with the problem, we
Adaptive Control
308
propose an adaptive inverse optimal controller in section 6. The effectiveness of the

proposed controller is confirmed by the experiment in section 7. Section 8 is devoted to
concluding remarks.
2. Preliminaries
In this section, we introduce some mathematical definitions and preliminary results of CLFbased inverse optimal control. We also refer to ACLF-based adaptive control techniques.
2.1 Mathematical notations and definitions
We use the notation R 0 := [0, ) .
Definition 1 A function sgn(y ) is defined for y R by the following equation:
1 ( y < 0)
sgn( y ) = 0 ( y = 0)
1 ( y > 0).
(1)
In this section, we consider the following input affine nonlinear system:
x& = f ( x) + g ( x)u,
(2)
where x R n is a state vector, u U R m is an input vector and U is a convex subspace

containing the origin u = 0 . We assume that f : R n R n and g : R n R n m are continuous
vector fields, and f (0) = 0 . Let L f V and LgV be the Lie derivative of f ( x) and g ( x)
respectively, which are defined by
L f V ( x) =
V
f ( x),
x
(3)
LgV ( x) =
V
g ( x).
x
(4)
For simplicity of notations, we shall drop (x) in the remaining of this article. We suppose
that a local control Lyapunov function is given for system (2).
Definition 2 A smooth proper positive-definite function V : X R0 defined on a
neighborhood of the origin X R n is said to be a local control Lyapunov function (local
CLF) for system (2) if the condition
inf {L f V + LgV u} < 0

u U
(5)
309
is satisfied for all x X \ {0} . Moreover, V (x) is said to be a control Lyapunov function
(CLF) for system (2) if V (x) is a function defined on entire R n and condition (3) is satisfied
for all x R n \ {0} .
If there exists no input constraint ( U = R m ), a smooth proper positive-definite function
V : R n R0 is a CLF if and only if
LgV = 0 L f V < 0, x 0.
(6)
In this article, we guarantee the robustness of controllers by sector margins and gain
margins.
Definition 3 A locally Lipschitz continuous mapping (u ) R m is said to be a sector
nonlinearity in ( , ) with respect to u R m if the following conditions are satisfied:
u T u < u T (u ) < u T u, u 0,
(0) = 0.
(7)
Definition 4 System (2) is said to have a sector margin ( , ) with respect to u R m if the
closed system
x& = f ( x) + g ( x) (u )
(8)
is asymptotically stable, where (u ) is any sector nonlinearity in ( , ) with respect to
u Rm .
Definition 5 System (2) is said to have a gain margin ( , ) with respect to u R m if the
closed system (8) is asymptotically stable, when (u ) is given as follows:
(u ) = u, ( , ).
(9)
By the definition, gain margins are the special case of sector margins. If system (2) has a
sector margin ( , ) , it also has a gain margin ( , ) .
2.2 Inverse optimal controller
We introduce the inverse optimal controller proposed by Nakamura et al [Nakamura et al.
(2007)]. The following results are obtained for system (2) with input constraint
1
m k k
m
U := u R u k = ui < C ( x),
i=1
C
k
(10)
Adaptive Control
310
where 1 < k < is a constant and C ( x) > 0 is continuous on R n .
Theorem 1 We consider system (2) with input constraint (10). Let V ( x) be a local CLF for
system (2) and a1 > 0 be the maximum number satisfying
inf {L f V + LgV u}< 0,
uU kC
(11)
x W1 := {x | V ( x) < a1 }.
Then, W1 is a domain in which the origin is asymptotically stabilizable. If V (x) is a CLF,

then a1 = and W1 = R n .
Theorem 2 We consider system (2) with input constraint (10) . Let V (x) be a local CLF for
system (2) , P(x) be a function defined by
P( x) =
LfV
C ( x) LgV
(12)
k
k 1
and ar (0, a1 ) be the maximum number such that the condition
k 1
inf L f V +
LgV u < 0,
k
x Wr := {x | V ( x) < a r }
uU kC
(13)
is satisfied, and d be a positive constant. Then, input

1
ui =
1
Lg V k 1 sgn( Lgi V )
R( x) i
(i = 1,K, m),
1
1
k 1
k 1
( 2 + q ( x ) Lg V k ) L g V k
k 1
k 1
1
k
R ( x) =
( P + P ) + q( x) LgV kk1
k 1
k 1
q( x)
(14)
( LgV 0) .
(15)
( LgV = 0),
q( x) = dC k 1 ( x)
(16)
311
asymptotically stabilizes the origin in Wr , and minimizes the cost function:
R k 1 ( x) k
J = l ( x) +
u k dt ,
k
l ( x) =
k 1 1
LgV
k R ( x)
k
k 1
k
k 1
(17)
LfV.
Moreover, it achieves at least a sector margin ( , ) in Wr .

2.3 Adaptive control problem
We consider an adaptive control problem for nonlinear systems. In this section, we
introduce some definitions and properties. We consider the following input affine nonlinear
system:
x& = f 0 ( x) + f1 ( x) + g ( x)u,
(18)
where x R n is a state vector, u R m is an input vector, and R p is a constant unknown

parameter vector. We assume that f 0 : R n R n , g : R n R nm and f1 : R n R n p are
continuous vector fields, and f 0 (0) = 0 . Note that there exists no input constraint.
The stabilizability of the system with unknown parameters is defined as the following.
Definition 6 Let be an estimate of . We say that (18) is globally adaptively stabilizable
if there exist a function ( x, ) continuous on R n \ {0} R p with (0, ) 0 , a continuous
function ( x, ) , and a positive definite symmetric p p matrix , such that the dynamic
controller
u = ( x,),
&
= ( x, )
(19)
(20)
guarantees that the solution ( x, ) is globally bounded, and x 0 as t for any value
of the unknown parameter R p .
For the stabilization problem, we introduce an adaptive control Lyapunov function (ACLF)
as the following.
Definition 7 We consider system (18) and assume that Va ( x, ) is a CLF for system (18).
Then, Va ( x, ) is called an adaptive control Lyapunov function (ACLF) for system (18) if
there exists a positive-definite symmetric matrix such that for each R p , Va is a CLF
for the modified system
Adaptive Control
312
T
V
x& = f 0 ( x) + f1 ( x) + a + g ( x)u.
(21)
Krsti et al. (1995) proved the following theorem.

Theorem 3 The following two statements are equivalent:
(1) There exists a triple ( , Va , ) such that ( x,) globally asymptotically stabilizes (21) at
x = 0 for each R p with respect to the Lyapunov function Va ( x, ) .

(2) There exists an ACLF Va ( x, ) for system (18).
3. Magnetic Levitation System

3.1 System configuration
We consider a stabilization problem of a magnetic levitation system shown in Fig.
1[Mizutani et al. (2004)]. The system consists of a magnet with a disk, a glass guide rod,
upper and lower magnetic drive coils that generate a magnetic field in response to a DC
current and two laser-based sensors that measure the magnetic position using the reflection
of the disk surface.
3.2 Mathematical model of the system
In this article, we control the position of the magnet using attractive force generated by the
upper drive magnetic coil. The force diagram is illustrated in Fig. 2. is the position of the
magnet from the upper coil, and Fu is an attractive force for the magnet generated by the
upper drive magnetic coil.
The dynamical equation for the magnet is described by
m&& = Fu m& mg 0 ,
(22)
where m is the mass of the magnet, is a friction constant. g 0 is the gravitational

acceleration.
Fig. 1. Magnetic levitation system
313
Fig. 2. Force diagram of the magnetic levitation system

Here, Fu is modeled by
Fu =
u
,
a( + b) 4
(23)
where a and b are constants determined by numerical modeling of the magnetic

configuration, and u is a control input voltage for the upper coil. The position is
measured by the upper laser sensor.
Let * be the desired position of the magnet, x1 = * , and x2 = x&1 . We set x = [ x1 , x2 ]T .
Then we obtain the following state equation:
x& = f ( x) + g ( x)u,
(24)
where f ( x) and g ( x) are defined as
x2
1
.
, g ( x) =
f ( x) =
ma( x * + b) 4
x2 g 0
1
(25)
The system parameters are shown in Table 1.
[kg]
[-]
[m/s2]
[V/Nxm4]
[m]
0.12
4.5
9.80665
40118.9
0.056464
g0
Table 1. Parameter values of the magnetic levitation system
Adaptive Control
314
There exists the following input constraint in system (24):
u 2 = u 2 < 5 [V].
(26)
By the above discussion, the control problem is reduced to the stabilization problem of
system (24) with the input constraint (26).
4. Pre-feedback Gravity Compensation

In system (2), we assume that f (0) = 0 . However, f (0) 0 in system (24). Therefore, we
cannot directly apply the inverse optimal controller (14) to system (24). To achieve f (0) = 0 ,
we design a controller to compensate for gravity by a pre-feedback input.
We consider the following gravity compensation input u c (x) as
uc ( x) = mg 0 a( x1 * + b) 4 .
(27)
Substituting (27) into (24), the gravitational acceleration g 0 is successfully canceled. Then,
we split the input u ( x) using u c (x) as
u ( x) = uc ( x) + u s ( x),
(28)
where u s is an asymptotic stabilizing input for system (24) when g 0 = 0 .

By using (26) and (28), the input constraint is rewritten to
u ( x) 2 = uc ( x) + u s ( x) 2 < 5.
(29)
To handle input constraint (29) as a norm constraint, we rewrite (29) as
u s ( x) < 5 uc ( x)
:= C ( x).
(30)
(30) represents a constraint depending on the state. Note that constraint (30) is more severe
than the original constraint (29). The problem of designing controller (28) is reduced to the
problem of designing controller u s (x) with input constraint u s ( x) < C ( x) .
2
To apply inverse optimal controller (14), we construct a CLF for system (24). In general, the
controller performance often depends on a CLF. However, it is unclear which CLF achieves
the best control performance. Hence, we construct a CLF with a design parameter. Using the
integrator backstepping method, a CLF V ( x) can be carried out as
V ( x) =
1 2
1 2
2
(r + 1) x1 + rx1 x2 + x2 ,
2
2
(31)
315
where r is a positive constant and also a design parameter.

Now, we construct input u s . Let f 0 ( x) be the function defined by
x
f 0 ( x) = f ( x) | g0 =0 = 2 .
x2
(32)
By using (31), we can calculate L f V and LgV as

0
L f 0 V = {(r 2 + 1) r}x1 x2 + (r ) x2 ,
(33)
rx1 + x2
.
ma( x1 * + b) 4
(34)
Lg V =
Substituting (33) and (34) into (14) and (15), we get the following input u s (x) .
u s ( x) =
( 2 + q ( x) L V ) L V
g
g
2
2( P + P ) + q( x) L V
1
g
R1 ( x) = 1
2
q( x)
P1 ( x) =
1
Lg V , ,
R1 ( x)
(35)
( LgV 0)
( LgV = 0),
L f0V
C ( x ) L gV
q( x) = dC ( x) = d 5 uc ( x)
(36)
(37)
.
2
(38)
According to Theorem 2, u s (x ) has a sector margin (1 / 2, ) .

Finally, the following controller u ( x) is obtained:
u ( x) = mg 0 a( x1 * + b) 4
1
LgV .
R1 ( x)
(39)
5. Experiment 1
We apply controller (39) to the magnetic levitation system. We set x(0) = [ 1.4, 0.0]T and
* = 2.0 [cm]. The controller is implemented by MATLAB/SIMULINK. The sampling
Adaptive Control
316
interval is 1 10 3 [sec] and control parameters are r = 8 and d = 1.25 10 4 , respectively.

The time response of the controlled system is shown in Fig. 3. Although the velocity x2
vibrates due to sensor noise, the input constraint (26) is satisfied. However, the position x1
does not converge to zero (an offset error remains). Then, the actual magnetic levitation
system is not asymptotically stabilized by the proposed controller (39).
The biggest reason for the offset error is the lack of robustness with respect to u c . If there
exists a parameter uncertainty in g ( x) , the gravitational acceleration g 0 is not completely
canceled by the pre-feedback u g ( x) . Therefore, the proposed controller u ( x) does not
guarantee the robustness for the system (24) even if the stabilizing input u s (x) guarantees
the sector margin (1 / 2, ) for the system
x& = f 0 ( x) + g ( x)u s .
0.2
30
20
offset error
-0.4
x2[cm/s]
x1[cm]
-0.2
-0.6
-0.8
10
0
-10
-1
-20
-1.2
-1.4
(40)
4
6
Time[sec]
10
4
6
Time[sec]
4
6
Time[sec]
10
10
u[V]
4
3.5
3
2.5
2
1.5
1
0.5
0
-0.5
0
-30
0
Fig. 3. Experimental result of controller (39)
6. Adaptive Inverse Optimal Controller Design

6.1 Robustness recovery via adaptive control
To solve the problem stated in section 5, we propose a controller that guarantees a gain
317
margin for u . We apply an adaptive control technique to achieve a gain margin for input u .
Before applying the adaptive controller, we rewrite the system (24) to
x& = f 0 ( x) + f1 ( x) g 0 + g ( x)u ,
(41)
where f 0 ( x) is defined by (31) and f1 ( x) = [0,1]T . Additionally, to consider a gain margin

for (41), we rewrite the system to
x& = f 0 ( x) + f1 ( x) + g ( x)u ,
(42)
where is an unknown constant and := g 0 / is a constant unknown parameter. Note

that the range of , in which the origin of the system (42) is asymptotically stable, is a gain
margin for input u . Furthermore, we consider the following input:
u ( x, ) = uc ( x, ) + us ( x, ),
(43)
where := g 0 / and is an estimate of . We suppose that input us ( x,) asymptotically

stabilizes the system (40) and guarantees the gain margin (1 / 2, ) . Let u g ( x, ) be a gravity
compensation input defined as follows:
u c ( x, ) = ma( x1 * + b) 4 .
(44)
Remark 1 In this section, we do not mention whether the input constraints exist or not.
&
Then, we construct an adaptive law such that the input (43) stabilizes the system (42) and
show the input (43) has a gain margin (1 / 2, ) .
In this section, we use an ACLF to construct an adaptive law. The following lemma is
available for constructing an ACLF.
Lemma 1 We consider system (42). Let V ( x) be a CLF for system (41). Then, V ( x) is an
ACLF for system (42).
Proof: If V ( x) is an ACLF for system (42), V ( x) is a CLF for the following system:
x& = f 0 ( x) + f1 ( x) +
+ g ( x)u,

(45)
where is a positive constant. Note that V / = 0 , the above system is rewritten to
x& = f 0 ( x) + f1 ( x) + g ( x)u.
(46)
Adaptive Control
318
Sytem (46) is asymptotically stabilized by the input
u ( x) = u c ( x) + u s ( x),
(47)
where u c (x) and u s (x) are defined by (27) and (35) respectively. This implies all CLFs for
system (41) are ACLFs for system (42).
By Lemma 1, CLF (31) is applicable to an ACLF for system (42).
Lemma 2 We consider system (42) and assume that an ACLF V ( x) for (42) is obtained. Let
V ( x, ) be a function defined by
~
V ( x,) = V ( x) + ( ) 2 = V ( x) + 2 ,
2
2
(48)
~
&
where 1 / 2 < < and := . Let the adaptive law be
&
= ( x) =
V
f1 ( x).
x
(49)
Then, V ( x, ) is a Lyapunov function for the closed loop system of (42).

Proof: Let the origin of system (42) be ( x,) = (0, ) . Then, V is a positive definite function.
&
~&
Assume u is input (43) and note that = . Then,
[
[
V
V& ( x,) =
f 0 ( x) + f1 ( x) + g ( x) uc ( x,) + us ( x,)
x
V
=
f 0 ( x) + g ( x)us ( x,) 0.
x
)}]
(50)
Since the input u s ( x, ) has a gain margin (1 / 2, ) , V& ( x, ) is less than or equal to zero.
Then V ( x, ) is a Lyapunov function for the closed loop system of (42) and the origin
( x,) = (0, ) is stable.

Remark 2 Lyapunov function (48) contains an unknown constant . However, it does not
become a problem because both input (43) and adaptive law (49) do not contain .
Lemma 3 We consider system (42) and assume that an ACLF V ( x) for (42) is obtained.
Then, if 1 / 2 < < , x 0(t ) and (t ) are achieved by input (43) and
adaptive law (49).
319
Proof: By Lemma 2, we can construct a Lyapunov function V ( x, ) (47) for system (41).
The input and the adaptive law are given by (42) and (48), respectively. Then, we obtain
V& ( x, ) 0( x 0) because the input us ( x, ) has a gain margin (1 / 2, ) . Let S be a set
defined by
S := {( x,) V& ( x,) = 0, x R n , R},

= {( x, ) x = 0, R}.
(51)
We show that the largest invariant set contained in S consists of only a point ( x,) = (0, ) .
Consider the following solution of (42) belonging to S :
x(t ) 0, t 0.
(52)
Note that us (0,) = 0 , we obtain the following equation for (42):
x& = f 0 (0) + f1 (0) + g (0)u (0,) ,

= f (0) + g (0)u ( x,) ,
(53)
= f1 ( ) 0,
where 0 and f1 (0) 0 , we obtain . On the other hand, if x = 0 and , we
obtain x& 0 by (50). Therefore, the largest invariant set contained in S is a set {(0, )} .
Finally, we obtain x 0 and when t by LaSalles invariance principle [Khalil
(2002)].
The following theorem is obtained by Lemmas 2 and 3.
Theorem 4 We consider system (42), controller (43) and adaptive law (49). Then, the
controller has a gain margin (1 / 2, ) .
6.2 Adaptive inverse optimal controller
&
We calculate of (49) by using CLF (31) as:
= [(r 2 + 1) x1 + rx2 rx1 + x2 ]

1
= (rx1 + x2 ).
&
0
(54)
Furthermore, taking into consideration the input constraint, we obtain the following
controller:
Adaptive Control
320
u ( x, ) = uc ( x, ) + us ( x, ) = ma( x1 * + b) 4 + us ( x, ),
us ( x,) =
1
LgV , ,
R2 ( x,)
(2 + q( x, ) L V ) L V
g
g
2
2
P
P
q
x
L
V
2
(
)
(
,
)
+
+
2
2
g
R 2 ( x, ) =
2
2
q( x, )
P2 ( x, ) =
L f0V
C ( x, ) LgV
(55)
(56)
( L g V 0)
( L g V = 0),
,
2
q( x, ) = dC ( x, ),
C ( x,) = 5 u c ( x,)
,
2
(57)
(58)
(59)
(60)
where we use u s ( x) given by (35) as u s ( x, ) . Then, note that the input constraint C ( x) is
rewritten to C ( x, ) given by (60). According to Lemma 2 and the result of [Nakamura et al.
(2007)], we can show the input u s ( x, ) minimizes the following cost function:
R ( x,) 2
J = l ( x,) + 2
us dt ,
2
0
(61)
where
l ( x,) =
1
LgV
2 R2 ( x,)
2
2
L f0V .
(62)
It is obvious that a gain margin (1 / 2, ) is guaranteed for controller (55) at least in the
neighborhood of the origin.
7. Experiment 2
In this section, we apply controller (55) to the magnetic levitation system and confirm its
effectiveness by the experiment. To consider the input constraint, we employ the following
adaptive law with projection instead of (54):
321
0
= 2 g 0 , rx1 + x2 > 0
&
=
0
= 0, rx1 + x2 < 0 .
(rx + x )
otherwise
1
2
(63)
We set the adaptation gain = 160 and the initial value of the estimate (0) = 820 . The
other experimental conditions and control parameters are the same as in section 5. The
experimental result is shown in Fig. 5. Position x1 converges to zero without any tuning of
control parameters. The gain margin guaranteed by the adaptive law seems quite effective.
We can observe that the input is larger than the non-adaptive controller (39), however, the
input constraint is satisfied. The parameter estimate also tends to converge to the true
value . As a result, the effectiveness of the proposed controller (55) is confirmed.
0.2
25
20
15
10
5
0
-5
-10
-15
-20
0
-0.4
x2[cm/s]
x1[cm]
-0.2
-0.6
-0.8
-1
-1.2
-1.4
4
6
Time[sec]
10
u[V]
4
3.5
3
2.5
2
1.5
1
0.5
0
-0.5
Time[sec]
10

4
6
Time[sec]
10
10
Time[sec]
Fig. 4. Experimental result of controller (55)
8. Conclusion
In this article, we proposed an adaptive inverse optimal controller for the magnetic
levitation system. First, we designed an inverse optimal controller with a pre-feedback
gravity compensator and applied it to the magnetic levitation system. However, this
controller cannot guarantee any stability margin. We demonstrated that the controller did
not work well (offset error remained) in the experiment. Hence, we proposed an improved
controller via an adaptive control technique to guarantee the stability margin. Finally, we
322
Adaptive Control
confirmed the effectiveness of the proposed adaptive inverse optimal controller by the
experiment. As a result, we achieved offset-free control performance.
9. References
Freeman, R.A. & Kokotovi, P.V. (1996). Robust Nonlinear Control Design. State-space and
Lyapunov Techniques. Birkhuser, Boston.
Khalil, H.K. (2007). Nonlinear systems, 3rd ed. Prentice Hall, Upper Saddle River, New Jersey.
Krsti, M.; Kanellakopoulos, I. & Kokotovi, P. (1995). Nonlinear and Adaptive Control
Design. Wiley-Interscience, New York.
Krsti, M. & Li, Z. (1998). Inverse optimal design of input-state-stabilizing nonlinear
controllers. IEEE Transaction on Automatic Control, 43, 3, 336-350.
Li, Z. & Krsti, M. (1997). Optimal design of adaptive tracking controllers for non-linear
systems. Automatica, 33, 8, 1459-1473
Mizutani, T.; Katayama, H. & Ichikawa A. (2004). Tracking control of a magnetic levitation
system by feedback linearization. Proceedings of SICE Annual Conference 2004, 121126.
Nakamura, N. ; Nakamura, H.; Yamashita, Y. & Nishitani, H. (2007). Inverse optimal control
for nonlinear systems with input constraints. Proceedings of the European Control
Conference 2007 (CD-ROM).
Sepulchre, R.; Jankovi, M. & Kokotovi, P.V. (1997). Constructive Nonlinear Control. Springer,
London.
Sontag, E.D. (1989). A universal construction of Artsteins theorem on nonlinear stabilization.
System & Control Letters, 13, 117-123.
15
Adaptive Precision Geolocation Algorithm with
Multiple Model Uncertainties
Wookjin Sung , Kwanho You
Sungkyunkwan University, Department of Electrical Engineering
Korea
1. Introduction
In the unmanned ground vehicle (UGV) case, the estimation of a future position with a
present one is one of the most important techniques (Madhavan & Schlenoff, 2004).
Generally, the famous global positioning system (GPS) has been widely used for position
tracking because of its good performance (Torrieri, 1984; Kim et al., 2006). However, there
exist some defects. For example, it needs a separate receiver and it must have at least three
available satellite signals. Moreover it is also vulnerable to the indoor case (Gleanson, 2006)
or the reflected signal fading.
There have been many researches to substitute or to assist the GPS. One of them is the
method of using the time difference of arrival (TDoA) which needs no special equipment
and can be operated in indoor multipath situation (Najar & Vidal, 2001). The TDoA means
an arrival time difference of signals transmitted from a mobile station to each base station. It
is the basic concept of estimation that the position of a mobile station can be obtained from
the crossing point of hyperbolic curves which are derived from the definition of TDoA.
Including some uncertainties, there have been several approaches to find the solution of
TDoA based geolocation problem using the least square method, for example, Tayler series
method (Xiong et al., 2003), Chans method (Ho & Chan, 1993), and WLS method (Liu et al.,
2006). However in case of a moving source, it demands a huge amount of computational
efforts each step, so it is required to use a method which demands less computational time.
As a breakthrough to this problem, the application of EKF can be reasonable.
The modeling errors happen in the procedure of linear approximation for system behaviors
to track the moving sources position. The divergence caused from the modeling errors is a
critical problem in Kalman filter applications (Julier & Uhlmann, 2004). The standard
Kalman filter cannot ensure completely the error convergence because of the limited
knowledge of the systems dynamical model and the measurement noise. In real
circumstances, there are uncertainties in the system modeling and the noise description, and
the assumptions on the statistics of disturbances could be restrictive since the availability of
a precisely known model is very limited in many practical situations. In practical tracking
filter designs, there exist model uncertainties which cannot be expressed by the linear statespace model. The linear model increases modeling errors since the actual mobile station
moves in a non-linear process. Especially even with a little priori knowledge it is quite
Adaptive Control
324
valuable concerning the strategy.

Hence, the compensation of model uncertainties is an important task in the navigation filter
design. In modeling or formulating the mathematical equations, the possible prediction
errors are approximated or assumed as a model uncertainty. The facts discussed above leads
to unexpected deterioration of the filtering performance. To prevent the divergence problem
due to modeling errors in the EKF approach, the adaptive filter algorithm can be one of the
good strategies for estimating the state vector. This chapter suggests the adaptive fading
Kalman filter (AFKF) (Levy, 1997; Xia et al., 1994) approach as a robust solution. The AFKF
essentially employs suboptimal fading factors to improve the tracking capability. In AFKF
method, the scaling factor is introduced to provide an improved state estimation. The
traditional AFKF approach for determining the scaling factors mainly depends on the
designers experience or computer simulation using a heuristic searching plan. In order to
resolve this defect, the fuzzy adaptive fading Kalman filter (FAFKF) is proposed and used as
an adaptive geolocation algorithm. The application of fuzzy logic to adaptive Kalman
filtering gains more interests. The fuzzy logic adaptive system is constructed so as to obtain
the suitable scaling factors related to the time-varying changes in dynamics. In the FAFKF,
the fuzzy logic adaptive system (FLAS) is used to adjust the scaling factor continuously so
as to improve the Kalman filter performance.
In this chapter, we also explain how to compose the FAFKF algorithm for TDoA based
position tracking system. Through the comparison using the simulation results from the
EKF and FAFKF solution under the model uncertainties, it shows the improved estimation
performance with more accurate tracking capability.
2. Geolocation with TDoA analytical methods

When the mobile station (MS: the unknown position) sends signals to each base station (BS:
the known position), there is a time difference because of the BSs isolated location from MS.
The fundamental principle of position estimation is to use the intersection of hyperbolas
according to the definition of TDoA as shown in Fig. 1.
The problem of geolocation can be formulated as
di = s - bi
di1 = cti1 = cti ct1
bi = col{xi , yi }, i = 1, 2, 3, L , m
(1)
s = col{x, y}
where bi is the known position of i-th signal receiver (BS), s is the unknown position of
signal source (MS), and c is the propagation speed of signal. In Eq. (1), di means the
distance between MS and i-th BS and ti is the time of signal arrival (ToA) (Schau &
Robinson, 1987) from MS to i-th BS. Hence ti1 becomes the time difference of arrival
(TDoA) which is the difference of ToA between ti (from MS to the i-th BS) and t1 (from
Adaptive Precision Geolocation Algorithm with Multiple Model Uncertainties
325
Fig. 1. Geometric method using hyperbolas.

MS to the first BS). The distance difference of di1 results from the multiplication of TDoA
and c.
Generally it is possible to estimate the source location if the values of ToA could be
provided exactly. However, it is required to be synchronized for all MS and BSs in this case.
To find the TDoA of acknowledgement signal from MS to BSs, the time delay estimation
can be used. As shown in Fig. 1, the estimation of geolocation can be obtained by solving the
nonlinear hyperbolic equation from the relation of TDoA. If there are three BSs as in Fig. 1,
we can draw three distinct hyperbolic curves using distance difference from TDoA signal. It
is the principle of geometric method that the cross point becomes the position estimation of
MS.
To find the position estimation (s) of the unknown MS in an analytical method, lets rewrite
the distance difference equation (1) as
di = d1 + di 1 , i = 2, 3, L , m.
(2)
By squaring Eq. (2) with the relation of ( d i )2 = s - bi , s - bi , the nonlinear equation for
positional vector of s can be formulated as following.
2
s 2 biT s + bi
= s 2 b1T s + b1
+ 2 d1 di 1 + ( di 1 )2 , i = 2,L , m
(3)
Adaptive Control
326
To represent the solution in linear matrix equality form, Eq. (3) can be simplified as
b1
bi
+ ( d i1 ) 2 = 2 b1 bi , s 2d1 d i1 , i = 2,L , m
(4)
Using the distance from MS to the first BS, (d1 ) = ( x x1 ) + ( y y1 ) , and with b1 as the
2
origin of coordinates, i.e., b1 = col{0, 0} , we can obtain the position estimation from the
following two nonlinear constraints.
1
2
( bi + (d i1 ) 2 ) = bi , s d1 d i1 , i = 2, 3,L , m
2
x 2 + y 2 (d1 ) 2 = 0
(5)
To find the solution of s, Eq. (5) is rewritten in linear matrix equation. Now the source vector
s can be acquired by solving the following MS geolocation problem.
Gs = h + d1
(d1 )2 = s, s
b 2 (d ) 2
2
21
1
1
M
h=
= (d )
2
2
2
2
bm (d m1 )
d 21
b2 , b2

= M , d = M
d m1
bm , bm

(6)
where G = [b2 L bm ] and means the Hadamar operator.

T
3. Geolocation with model uncertainty

This section describes the geolocation using the estimation filter in state-space. As stated in
the section 2, the conventional analytical methods are focused on solving the nonlinear
hyperbolic equations. In this section, we introduce the fuzzy adaptive fading Kalman filter
to get the precision estimation for multiple model uncertainties.
3.1 System modeling
In the real case, TDoA signal can be distorted by the timing error due to non-line-of sight or
by additive white Gaussian noise. To find the precision geolocation in real case, the system
modeling must include the model uncertainty. Let
to be the ideal TDoA signal and t is
327
the distorted amount by external noises. The real value of TDoA is changed as t = to + t .
If the real value of TDoA is used in Eq. (6), it becomes more complicated nonlinear equation
and this complexity may cause huge computational efforts in the real-time process. As a
breakthrough to this problem, the Kalman filter which needs relatively less computational
time can be an alternative solution.
Since the hyperbolic equation of TDoA is nonlinear, the extended Kalman Filter (EKF) can
be used as a nonlinear state estimator. The basic algorithm of EKF is shown as in Fig. 2.
Fig. 2. Flow chart of extended Kalman filter.

The first step is the time update in which it predicts the state of next steps from processing
model and it compares the real measurement with the prediction measurement of s
obtained by time update process. For TDoA based geolocation using extended Kalman filter,
the discrete state-equation of the processing and measurement model for MS can be
formulated as
sk +1 = Ask + Buk + wk
1
0
A=
0
0 0
0
0
1 0
, B=
1
0 0 0
0 0 0
0
0
0
0
(7)
Adaptive Control
328
where s (k ) = [ x y x& y& ] , u k is the known velocity of moving MS,
T
is the time interval
of sampling, wk is an additive white Gaussian noise (AWGN).

From the definition of TDoA, the measurement model can be written as
zk = h( sk , vk )
1
= (|| Msk bi || || Msk b j ||) + vk
c
1 0 0 0
M =
0 1 0 0
(8)
where k is the measurement noise in AWGN.

The output result of zk which is the TDoA signal provides the information of MS position.
As an accurate geolocation method, the frequency difference (FDoA) resulted from Doppler
shifts observation can be added in the state equation. However, to make the problem more
simple, we consider only the TDoA signal as an output measurement in this section.
Since the measurement model zk is nonlinear equation, the linear approximation using
partial differential method should be done for the use of EKF.
zk = zk 1 + H k sk + Vk
Hk
hk
h
, Vk k
sk
vk
(9)
3.2 Geolocation using fuzzy adaptive fading Kalman filter (FAFKF)

EKF is a very useful method for nonlinear state estimation. However, as EKF is based on the
linearization of nonlinear system using partial differential method, the modeling errors can
easily lead to the divergence problem. To solve this problem, an adaptive fading Kalman
filter (AFKF) with a fading factor can be applied. The application of AFKF to geolocation
estimation is given in the following mathematical expression.
Basically, the fading factor k is added in the error covariance projection during the time
update process.
Pk = k Ak Pk 1 AkT + Wk QkWkT
(10)
where k = diag (1 , 2 , L , m ) . In normal case of k = 1 , it means the general EKF. If the

estimated value approaches to the steady-state value, the fading factor k becomes less
than 1. If k is greater than 1, the divergence could happen. This iterative process is called
as the adaptive fading loop given as follows.
329
tr[ Fk ]
tr[ Ek ]
k +1 = max 1,
Fk = C0 Rk H k Qk H kT
Ek = H k Ak Pk AkT H kT
00T
, k =0
2
C0 =
T
k k k , k 1
1 + k
k = zk zk , zk = h( sk , vk )
(11)
where is a scaling factor and tr[] is the trace of a matrix.

Moreover the measurement estimation of zk is predicted through the estimation of sk .
That is, if we get more accurate sk , then the more accurate zk can be obtained. As the
error of output measurement is within the neighborhood i.e., || zk zk || k and
k k 1 0 , it is confirmed that the present estimation performance is guaranteed and the

fading factor becomes k +1 1 .
The fuzzy logic adaptive system (FLAS) offers an effective method when the problem is too
complicated or hard to be analyzed in mathematical way. The procedure of general fuzzy
system can be classified as three parts; fuzzification, fuzzy inference, and defuzzification.
The first step of fuzzification is to make linguistic variables from inputs and outputs. The
second step of fuzzy inference is to make rules using if-then expression. Finally the third step
of defuzzification is to decide the degree of the output value.
Using the scaling factor ( ) as an output from FLAS, the fading factor in FAFKF is updated
as k +1 = tr[ Fk ] tr[ Ek ] . According to the following two degree of divergence (DoD)
parameters from the innovation covariance matrix and the trace of innovation covariance
matrix, it is possible to identify the changing degree in dynamics of MS. The first DoD
parameter is defined as the ratio of the trace of innovation covariance matrix at present
state and the number of measurements used for estimating location.
k T k
m
(12)
where k = [1 2 L m ] , m is the number of measurements (number of TDoA signals).

T
The second DoD parameter is defined as the average of the absolute value of the
measurement error k .
1 m
| i |
m i =1
(13)
330
Adaptive Control
The fading factor k updated through the adaptive fading loop is used to change the error
covariance Pk .
Fig. 3. Flow chart of the fuzzy adaptive fading Kalman filter process.
Fig. 3 shows how the FAFKF works for TDoA geolocation problem. As a first step in the
process of FAFKF, the two DoD parameters ( , ) are obtained from measurement
difference between the real value ( zk ) and the estimation result ( zk ). These DoD
parameters are used as the inputs for the fuzzy system. Finally the FLAS is employed for
determining the scaling factor from the innovation information. According to the scaling
factor , the estimation accuracy is determined. Using the fuzzy logic system, we can
adjust the fading factors adaptively to improve the estimation performance.
4. Simulation results
The basic circumstance to be used in the simulation is shown in Fig. 4. There are two BSs
and the signal source of MS is supposed to move at a constant speed but changes its
direction every 2.5 sec. In Fig. 4, the dotted line is an ideal path of MS with no external forces.
The solid line is the real path which is affected by the multiple noises such as the
331
measurement noise vk and the process noise wk . The thick solid line is the path of MS
estimated by the standard EKF with no adaptive method. Fig. 4 shows that the performance
of EKF is restricted especially when the MS changes the direction. The accumulated position
error is increased as the MS changes its direction frequently.
To prove the effectiveness of the adaptive fading factor in TDoA gelocation, the simulation
parameters are set close to the real values. Table 1 shows the simulation parameters.
Fig. 4. Simulation circumstance for MS

The FLAS consists of the following 9 rules and is represented in the following if-then form.
The membership functions of input fuzzy variable (DoD parameters: and ) and
output (scaling factor: ) are shown in Fig. 5.
i.
ii.
if is n (negative) and is n,
if is z (zero) and is n,
then is nb (negative big).

then is ns (negative small).
EKF
AFKF
FAFKF
Constant
Constant
Constant.
(time interval)
0.1 sec
0.1 sec
0.1 sec
(scaling factor)
None
0.12
FLAS output
(fading factor)
None
Constant
Fuzzy based
Speed
Table 1. Parameters for the TDoA geolocation simulation
Adaptive Control
332
Fig. 5. Membership functions in FLAS

iii.
iv.
v.
vi.
vii.
viii.
ix.
if
if
if
if
if
if
if
is p (positive) and is n,
is n and is z,
is z and is z,
is p and is z,
is n and is p,
is z and is p,
is p and is p,
then
then
then
then
then
then
then
is z.
is ns.
is z.
is ps (positive small).
is z.
is ps.
is pb (positive big).
As the DoD parameter ( ) and the averaged magnitude ( ) of (k ) change within

0.003~0.007 and 0.03~0.1 respectively, we define those range as zero for and . The
output of the scaling factor ( ) is determined as 0.12 following that of AFKF in the
333
associated range. Other values can be determined from experiential way. The simulation
result of FLAS and adaptive fading loop is given in Fig. 6.
Fig. 6 shows the change of the scaling factor k and the fading factor k . The values of
k and k change very steeply to correct the position error from the beginning and the
estimate sk gets close to the real value within -neighborhood about after 1 sec since the
fading factor becomes small.
Fig. 6. Change of scaling factor ( ) and fading factor ( ).

Fig. 7 shows the performance of the proposed geolocation algorithm (FAFKF) through the
comparison with AFKF and EKF. The performance is measured in terms of the norm of
positioning error, i.e. sk sk . As shown in Fig. 7, the positioning error of FAFKF is much
smaller than that of EKF. It can be confirmed that the difference of position error between
EKF and FAFKF is increased as the MS changes its direction more frequently. It means that
the position estimation with FAFKF is tracking more precisely to the real value of sk than
334
Adaptive Control
Fig. 7. Comparison of error performance

AFKF or the standard EKF.
Fig. 8 indicates the path estimation performance of the proposed geolocation algorithm
through the comparison with AFKF and EKF under the situation of Fig. 4. As the adaptive
fading factor takes the sub-optimal value at each iteration, the error covariance has been
updated and is used to modify the Kalman filter gain adaptively. As shown in Fig. 8, the
trajectory estimation using FAFKF is close to the real value under noise added real
circumstance.
5. Conclusion
In this chapter, we introduced TDoA geolocation algorithm to reduce the position
estimation error. To be more similar to real circumstance, the MS is supposed to change its
direction periodically. The standard EKF which solves a huge computational problem of
TDoA based geolocaion can estimate the location of source through the linearization of
nonlinear measurement equation. However, the linearization from partial differentiation
causes a divergence problem which restricts the performance of the EKF.
To solve this problem, we applied FAFKF algorithm which changes the error covariance
using an adaptive fading factor ( ) from fuzzy logic. The scaling factor which is used
335
Fig. 8. Comparison of path estimation.

to update the fading factor has been decided by the fuzzy logic to minimize the estimation
error. Through the simulation results, it is confirmed that the trajectory estimation using
FAFKF follows the real one more precisely than EKF. The positioning error from FAFKF is
less than that performed by AFKF.
6. References
Gleason, C.P. (2006). Tracking human movement in an indoor environment using mobility profiles,
M.S. thesis, University of Nebraska-Lincoln, August, 2006.
Ho, K.C. & Chan, Y.T. (1993). Solution and performance analysis of geolocation by TDoA,
IEEE Tr. Aerospace & Electronic Systems, Vol. 29, No. 4, pp. 1311-1322, 1993.
Julier, S.J. & Uhlmann, J.K. (2004). Unscented filtering and nonlinear estimation, IEEE
Review, Vol. 92, No. 3, pp. 401-422, 2004.
Kim K.H., Lee, J.G. & Park, C.G. (2006). Adaptive two-stage EKF for INS-GPS loosely
coupled system with unknown fault bias, Jour. of Global Positioning System, Vol. 5,
pp. 62-69, 2006.
Levy, L.J. (1997). The Kalman filter: navigation's integration workhorse, Annual report in
Applied Physics Laboratory, Johns Hopkins University, 1997.
Liu, J.M., Zhang, C. & Liu, S. (2006). A TDOA location algorithm based on data fusion,
Frontiers of Electronical and Electronic Engineering in China, Vol. 1, No.3, pp. 330-333,
2006.
336
Adaptive Control
Madhavan, R. & Schlenoff, C. (2004). The effect of process models on short-term prediction
of moving objects for unmanned ground vehicles, International IEEE Conf. Intelligent
Transportation Systems, Vol. 1, pp. 471-476, 2004.
Najar, M. & Vidal, J. (2001). Kalman tracking based on TDOA for UMTS mobile location,
IEEE International Symp. Personal, Indoor and Mobile Radio Communications, Vol. 1,
pp. B45-B49, 2001.
Schau, H.C. & Robinson, A.Z. (1987). Passive source localization employing intersecting
spherical surfaces from Time-of-Arrival differences, IEEE Tr. Acoustics, Speech, &
Signal Processing, Vol. ASSP-35, No. 8, pp. 1223-1225, 1987.
Torrieri, D.J. (1984). Statistical theory of passive location systems, IEEE Tr. on Aerospace and
Electronic Systems, Vol. AES-20, No. 2, pp. 183-197, 1984.
Xia, Q., Rao, M., Ying, Y. & Shen, X. (1994). Adaptive fading Kalman filter with an
application, Automatica, Vol. 30, No. 8, pp. 1333-1338, 1994.
Xiong, J.Y., Wang, W. & Zhu, Z.L. (2003). An improved Taylor algorithm inTDOA
subscriber position location, Proc. of ICCT, Vol. 2, pp. 981-984, 2003.
16
Adaptive Control for a Class of Non-affine
Nonlinear Systems via Neural Networks
Zhao Tong
School of Automatization and Electronic Engineering, Qingdao University of Science and

Technology
China
1. Introduction
Adaptive control of highly uncertain nonlinear dynamic systems has been an important
research area in the past decades, and in the meantime neural networks control has found
extensive application for a wide variety of areas and has attracted the attention of many
control researches due to its strong approximation capability. Many significant results on
these topics have been published in the literatures (Lewis et al., 1996 ; Yu & Li 2002;
Yesidirek & Lewis 1995). It is proved to be successful that neural networks are used in
adaptive control. However, most of these works are applicable for a kind of affine systems
which can be linearly parameterized. Little has been found for the design of specific
controllers for the nonlinear systems, which are implicit functions with respect to control
input. We can find in literatures available there are mainly the results of Calise et al. (Calise
& Hovakimyan 2001) and Ge et al. (Ge et al. 1997). Calise et al. removed the affine in control
restriction by developing a dynamic inversion based control architecture with linearly
parameterized neural networks in the feedback path to compensate for the inversion error
introduced by an approximate inverse. However, the proposed scheme does not relate to the
properties of the functions, therefore, the special properties are not used in design. Ge, S.S.
et al., proposed the control schemes for a class of non-affine dynamic systems, using mean
value theorem, separate control signals from controlled plant functions, and apply neural
networks to approximate the control signal, therefore, obtain an adaptive control scheme.
Furthermore, when controlling large-scale and highly nonlinear systems, the presupposition
of centrality is violated due to either due to problems in data gathering when is spread out
or due to the lack of accurate mathematical models. To avoid the difficulties, the
decentralized control architecture has been tried in controller design. Decentralized control
systems often also arise from various complex situations where there exist physical
limitations on information exchange among several subsystems for which there is
insufficient capability to have a single central controller. Moreover, difficulty and
uncertainty in, measuring parameter values within a large-scale system may call for
adaptive techniques. Since these restrictions encompass a large group of applications, a
variety of decentralized adaptive techniques have been developed (Ioannou 1986).
338
Adaptive Control
Earlier literature on the decentralized control methods were focused on control of largescale linear systems. The pioneer work by Siljak (Siljak 1991) presents stability theorems of
interconnected linear systems based on the structure information only. Many works
consider subsystems which are linear in a set of unknown parameters (Ioannou 1986 ; Fu
1992 ; Sheikholeslam & Desor 1993 ; Wen 1994 ; Tang et al. 2000), and these results were
focused on systems with first order interconnections. When the subsystems has nonlinear
dynamics or the interconnected is entered in a nonlinear fashion, the analysis and design
problem becomes even challenging.
The use of neural networks learning ability avoids complex mathematical analysis in
solving control problems when plant dynamics are complex and highly nonlinear, which is
a distinct advantage over traditional control methods. As an alternative, intensive research
has been carried out on neural networks control of unknown nonlinear systems. This
motivates some researches on combining neural networks with adaptive control techniques
to develop decentralized control approaches for uncertain nonlinear systems with
restrictions on interconnections. For example, in (Spooner & Passino 1999), two
decentralized adaptive control schemes for uncertain nonlinear systems with radial basis
neural networks are proposed, which a direct adaptive approach approximates unknown
control laws required to stabilize each subsystem, while an indirect approach is provided
which identifies the isolated subsystem dynamics to produce a stabilizing controller. For a
class of large scale affine nonlinear systems with strong interconnections, two neural
networks are used to approximate the unknown subsystems and strong interconnections,
respectively (Huang & Tan 2003), and Huang & Tan (Huang & Tan 2006) introduce a
decomposition structure to obtain the solution to the problem of decentralized adaptive
tracking control a class of affine nonlinear systems with strong interconnections. Apparently,
most of these results are likewise applicable for affine systems described as above. For the
decentralized control research of non-affine nonlinear systems, many results can be found
from available literatures. Nardi et al. (Nardi & Hovakimyan 2006) extend the results in
Calise et al. (Calise & Hovakimyan 2001) to non-affine nonlinear dynamical systems with
first order interconnections. Huang (Huang & Tan 2005) apply the results in (Ge & Huang
1999) to a class of non-affine nonlinear systems with strong interconnections.
Inspired by the above researches, in this chapter, we propose a novel adaptive control
scheme for non-affine nonlinear dynamic systems. Although the class of nonlinear plant is
the same as that of Ge et al. (Ge et al. 1997), utilizing their nice reversibility, and invoking
the concept of pseudo-control and inverse function theorem, we find the equitation of error
dynamics to design adaptation laws. Using the property of approximation of two-layer
neural networks (NN), the control algorithm is gained. Then, the controlled plants are
extended to large-scale decentralized nonlinear systems, which the subsystems are
composed of the class of non-affine nonlinear functions. Two schemes are proposed,
respectively. The first scheme designs a RBFN-based (radial basis function neural networks)
adaptive control scheme with the assumption which the interconnections between
subsystems in entire system are bounded linearly by the norms of the tracking filtered error.
In the scheme, unlike most of other approaches in available literatures, the weight of BBFN
and center and width of Gaussian function are tuned adaptively. In another scheme, the
interconnection is assumed as stronger nonlinear function. Moreover, in the former, in every
subsystem, a RBFN is adopted which is used to approximate unknown function, and in the
latter, in every subsystem, two RBFNs are respectively utilized to approximate unknown
Adaptive Control for a Class of Non-affine Nonlinear Systems via Neural Networks
339
function and uncertain strong interconnection function. For those complicated large-scale
decentralized dynamic systems, in order to decrease discontinuous factors and make
systems run smooth, unlike most of control schemes, the hyperbolic tangent functions are
quoted in the design of robust control terms, instead of sign function. Otherwise, the citation
of the smooth function is necessary to satisfy the condition of those theorems.
The rest of the paper is organized as follows. Section 2 gives the normal form of a class of
non-affine nonlinear systems. Section 3 proposes a novel adaptive control algorithm, which
is strictly derived from some mathematical and Lyapunov stability theories, and the
effectiveness of the scheme is validated through simulation. Extending the above-mentioned
result, Section 4 discusses two schemes of decentralized adaptive neural network control for
the class of large-scale nonlinear systems with linear function interconnections and
nonlinear function interconnections, respectively. Finally, the Section 5 is concluding
remarks.
2. Problem Statement
We consider a general analytic system
& = g(, u ), R n , u R
y R.
y = h(),
(1)
where g (, ) is a smooth vector fields and h() is a scalar function. In practice, many
physical systems such as chemical reactions, PH neutralization and distillation columns are
inherently nonlinear, whose input variables may enter in the systems nonlinearly as
described by the above general form (Ge et al. 1998). Then, the Lie derivative (Tsinias &
Kalouptsidis 1983) of h ( ) with respect to g (, u ) is a scalar function defined
by Lg h = [h() ] g (, u ) . Repeated Lie derivatives can be defined recursively
for i = 1, 2L . The system (1) is said to have relative degree
as Lig h = Lg ( Lig1h),
( 0 , u0 )
at
if
there
exists
smallest
positive
integer
such
that L h u = 0 , Lg h u 0, i = 1, L , 1 .
i
g
Let
R n and u R be compact subsets containing 0 and u0 , respectively. System
(1) is said to have a strong relative degree

degree
in a compact set D = u , if it has relative
at every point ( 0 , u0 ) D . Therefore, system (1) is feedback linearizable and the
mapping () = [1 (), 2 (),Ln ()] , with
j () = Lgj 1h, j = 1, 2,L
has a Jacobian
matrix which is nonsingular for all x () , system (1) can be transformed into a normal
form
Adaptive Control
340
x&1 = x2
x& = x
3
2
M
x& = f ( x, u )
n
y = x1
where
(2)
T
f ( x, u ) = Lng h and x = 1 () with x = [ x1 , x2 ,L , xn ] . Define the domain of
normal system (2) as D
{( x, u) x ( ); u } .
3. Adaptive Control for a Class of Non-affine Nonlinear Systems via TwoLayer Neural Networks
Now we consider the n th order nonlinear systems of the described form as (2). For the
considered systems in the chapter, we may make the following assumptions.
Assumption 1.
f ( x, u ) / u 0 for all ( x, u ) R .
Assumption 2. f () : R n +1 R , is an unknown continuous function and f ( x, u ) a smooth

function with respect to control input u .
The control objective is: determine a control law, force the output, y , to follow a given
desired output,
xd
with an acceptable accuracy, while all signals involved must be
bounded.
Assumption 3. The desired signals xd (t ) = [ yd , yd(1) ,L , yd( n 1) ], and
bounded, with
X d = [ xdT , yd( n ) ]T are
X d X d , X d a known positive constant.
Define the tracking error vector as
e = x xd ,
(3)
= [T 1]e ,
(4)
and a filtered tracking error as
with
a gain parameter vector selected so that e(t ) 0 as 0. Differentiating (4), the
filtered tracking error can be written as
& = x&n xd( n ) + [0 T ]e.

Define a continuous function
(5)
341
= k + xd( n ) [0 T ]e.
k is a positive
[ f ( x, u ) ] u 0 .
where
constant. We know
(6)
f ( x, u ) u 0 (Assumption 1), thus,
Considering the fact that
u = 0 ,
we invoke the implicit
u in
function theorem (Lang 1983), there exists a continuous ideal control input
neighborhood of ( x, u ) R , such that
f ( x, u ) = 0 ,
= f ( x, u ) may represent ideal control inverse.

Adding and subtracting to the right-hand side of x&n = f ( x, u )
i.e.
= f ( x, u ) holds.
of (2), one obtains
x&n = f ( x, u ) k + xd( n ) [0 T ]e ,
(7)
and yields
& = k + f ( x, u ) .
Considering the following state dependent transformation
(8)
= x&n , where
is commonly
referred to as the pseudo-control (Calise & Hovakimyan 2001). Apparently, the pseudocontrol is not a function of the control u but rather a state dependent operator.
Then,
u = 0 , from Assumption 1, f ( x, u ) u 0 thus [ f ( x, u )] u 0 .
( x, u ) R , there exists a
function such that f ( x, u ) = 0 holds, i.e. = f ( x, u ) . Therefore, we have
With the implicit function theorem, for every
= f ( x, u ) .
implicit
(9)
[ f ( x, u )] u 0
and f ( x, u ) is a smooth with respect to control input, u , then, f ( x, u ) defines a local
Furthermore, using inverse function theorem, with the fact that
diffeomorphism (Slotine & Li 1991), such that, for a neighborhood of

smooth inverse function and
u = f ( x, ) holds.
1
u,
there exists a
If the inverse is available, the control
problem is easy. But this inverse is not known, we can generally use some techniques, such
as neural networks, to approximate it. Hence, we can obtain an estimated function,
u = f 1 ( x, ) . This result in the following equation holding:
= f ( x, u ) ,
where
(10)
may be referred to as approximation pseudo-control input which represents
actual dynamic approximation inverse.
Adaptive Control
342
Remark 1. According to the above-mentioned conditions, when one designs the pseudocontrol signal, , must be a smooth function. Therefore, in order to satisfy the condition,
we adopt hyperbolic tangent function, instead of sign function in design of input. This also
makes control signal tend smooth and system run easier. The hyperbolic tangent function
has a good property as follows (Polycarpou 1996) :
0 < tanh( ) ,
(11)
with = 0.2785 , any positive constant. Moreover, theoretically, is approximation

inverse, generally a nonlinear function, but it must be bounded and play a dynamic
approximation role and make system stable. Hence, it represents actual dynamic
approximation inverse.
Based on the above conditions, in order to control the system and make it be stable, we
design the approximation pseudo-control input as follows:
= f ( x, u ) + uad + vr ,
(12)
uad is output of a neural network controller, which adopts

network, vr is robustifying control term designed in stability analysis.
where
Adding and subtracting to the right-hand side of (8), with
= f ( x, u ) , we have
& = k + f ( x, u ) + f ( x, u ) uad vr
= k + % ( x, u, u ) + uad vr ,
where
a two-layer neural
(13)
% ( x, u, u ) = f ( x, u ) f ( x, u ) is error between nonlinear function and its ideal
control function, we can use the neural network to approximate it.

3.1 Neural network-based approximation
A two-layer NN consists of two layers of tunable weights, a hidden layer and an output
layer. Given a > 0 , there exists a set of bounded weights M and N such that the
% C () , with
nonlinear error
compact subset of R , can be approximated by a two-
layer neural network, i.e.
% = M T ( N T xnn ) + ( xnn ) ,
with
xnn = [1, xdT , eT , ] input vector of NN.
(14)
Assumption 4. The approximation error
is bounded as follows:
N ,
where
Let
343
(15)
N > 0 is an unknown constant.

and
be the estimates respectively of
and N . Based on these estimates, let
uad be the output of the NN

uad = M T ( N T xnn ).
M% = M M and N% = N N ,
Z% = diag[ M% , N% ] , Z = diag[ M , N ] for
Define
(16)
where we use notations:
Z = diag[ M , N ] ,
convenience. Then, the following inequality
holds:
tr ( Z% T Z ) Z%
The Taylor series expansion of
( N T xnn )
Z%
for a given
2
F
(17)
xnn can be written as:
( N T xnn ) = ( N T xnn ) + ( N T xnn ) N% T xnn + O ( N% T xnn ) 2 ,
:= ( N T xnn ) and denoting
O( N% T xnn ) 2 the
T
T
the following, we use notations: := ( N xnn ) , % := ( N% xnn ) .
with
its Jacobian,
(18)
term of order two. In
With the procedure as Appendix A, the approximation error of function can be written as
M T ( N T xnn ) M T ( N T xnn ) = M% T ( N T xnn ) + M T N% T xnn + ,

and the disturbance term
(19)
can be bounded as
xnn M T
+ M N T xnn + M 1 ,
(20)
where the subscript F denotes Frobenius norm, and the subscript 1 the 1-norm.
Redefine this bound as
( M , N , xnn ) ,
(21)
Adaptive Control
344
where
= max{ M , N
, M 1} and = xnn M T
+ N T xnn + 1 . Notice that
is an unknown coefficient, whereas is a known function.

3.2 Parameters update law and stability analysis
& = k + M T ( N T xnn ) M T ( N T xnn ) + vr + ( xnn ).
(22)
Using(19), the above equation can become
& = k + M% T ( N T xnn ) + M T N% T xnn + vr + + .
(23)
Theorem 1. Consider the nonlinear system represented by Eq. (2) and let Assumption 1-4
hold. If choose the approximation pseudo-control input as Eq.(12), use the following
adaptation laws and robust control law
&
M = F ( Nxnn ) k1M ,
&
N = R xnn M T k1 N ,
&
( + 1)
= ( + 1) tanh

( + 1)
vr = ( + 1) tanh
where F = F T > 0, R = RT > 0 are any constant matrices,

design parameters,
k1 > 0 and > 0 are
(24)
scalar
is the estimated value of the uncertain disturbance term
= max( , N ) , defining % =
with
error of , then, guarantee that all signals
in the system are uniformly bounded and that the tracking error converges to a
neighborhood of the origin.
Proof. Consider the following positive define Lyapunov function candidate as
1
1
1
1
L = 2 + tr ( M% T F 1M% ) + tr ( N% T R 1 N% ) + 1% 2
2
2
2
2
(25)
The time derivative of the above equation is given by
% %&
& + tr ( M% T F 1M%& ) + tr ( N% T R 1 N%& ) + 1
L& =
(26)
345
Substituting (23) and the anterior two terms of (24) into (26), after some straightforward
manipulations, we obtain
L& = k 2 + [ M% T ( N T xnn ) + M T N% T xnn + ( ) vr + + ]

&
&
% &%
+ tr ( M% T F 1M% ) + tr ( N% T R 1 N% ) + 1
% &% + k tr ( Z% T Z ).
= k 2 + ( ) vr + ( + ) + 1
1
2
1 % &
k + ( ) vr + ( + 1) + % + k1 tr ( Z% T Z ).
(27)
With (4),(6),(12),(16) and the last two equations of (24), the approximation error between
actual approximation inverse and ideal control inverse is bounded by
c1 + c2 + c3 Z%
where
(28)
c1 , c2 , c3 are positive constants.
Using (11) and the last two terms of (24), we obtain
( + 1)
L& k 2 + ( ) ( + 1) tanh
( + 1)
+ ( + 1) % ( + 1) tanh
+ k1 tr ( Z% T Z )
T
2
%
k + ( ) + + + k tr ( Z% Z )
(29)
%
Applying (17),(28) , and
% %
, after completing square, we have the following
inequality
2
L& (k c2 ) + D1 + D2
where D1
Let D3
then
= c1 +
c
k1
1
( Z M + 3 ) 2 , D2 = 2 + .
k1
4
4
= D12 + 4 D2 (k c2 ) + D1 ,
L& 0 holds.
Now define
(30)
thus, as long as
D3 [2(k c2 )] ,
and
k > c2 ,
Adaptive Control
346
= % % , Z = Z%
Z%
(k1Z M + c3 ) , =
k1
1
D3 .
2(k c2 )
(31)
Z M , k1 , k , D1 , D2 , D3 , c2 , c3 are positive constants, as long as k is chosen to be big

enough, such that k > c2 holds, we conclude that , Z and are compact sets.
& is negative outside these compacts set. According to a standard Lyapunov
Hence L
theorem, this demonstrates that % , Z% and are bounded and will converge
Since
to , Z and , respectively. Furthermore, this implies
is bounded and will converge
to a neighborhood of the origin and all signals in the system are uniformly bounded.
3.3 Simulation Study
In order to validate the performance of the proposed neural network-based adaptive control
scheme, we consider a nonlinear plant, which described by the differential equation
x&1 = x2
x&2 = 2 x1 0.02( + x12 ) x2 + u 3 + ( x12 + x22 ) (u ) + tanh(0.2u ) + d
= 0.4 , (u ) = (1 e u ) (1 + e u )
xd = 0.1 [sin(2t ) cos(t )] .
where
and
d = 0.2
(32)
. The desired trajectory
To show the effectiveness of the proposed method, two controllers are studied for
comparison. A fixed-gain PD control law is first used as Polycarpou, (Polycarpou 1996).
Then, the adaptive controller based on NN proposed is applied to the system.
= [1, xdT , eT , ] , and number of hidden layer nodes 25.

(0) = (0), N (0) = (0) . The initial condition of
network is M
Input vector of neural network is xnn

The initial weight of neural
x(0) = [0.1, 0.2]T . The other parameters are chosen as follows:

k1 = 0.01, = 0.1, = 0.01, = 10 , = 2, F = 8I M , R = 5I N , with I M , I N corresponding
controlled plant is
identity matrices.
Fig.1, 2, and 3 show the results of comparisons, the PD controller and the adaptive controller
based on NN proposed, of tracking errors, output tracking and control input, respectively.
These results indicate that the adaptive controller based on NN proposed presents better
control performance than that of the PD controller. Fig.4 depicts the results of output of NN,
norm values of
M , N and
M , N , respectively, to illustrate the boundedness of the estimates of
the control role of NN. From the results as figures, it can be seen that the
learning rate of neural network is rapid, and tracks objective in less than 2 seconds.
Moreover, as desired, all signals in system, including control signal, tend to be smooth.
0.05
0
-0.05
Trackingerror
-0.1
-0.15
-0.2
-0.25
-0.3
-0.35
-0.4
-0.45
0
10
15
20
time sec
Fig. 1. Tracking errors: PD(dot) and NN(solid).

0.6
O
utput tracking
0.4
0.2
-0.2
-0.4
10
15
20
time sec
Fig. 2. Output tracking: desired (dash), NN(solid) and PD(dot).

1.5
C
ontrolinput
0.5
-0.5
-1
-1.5
0
10
15
20
time sec
Fig. 3. Control input: PD (dash), NN(solid)

2.5
2
1.5
1
0.5
0
-0.5
-1
-1.5
0
10
time sec
Fig. 4. M (dash), N (dot), output of NN(solid)
15
20
347
Adaptive Control
348
4. Decentralized Adaptive Neural Network Control of a Class of Large-Scale

Nonlinear Systems with linear function interconnections
In the section, the above proposed scheme is extended to large-scale decentralized nonlinear
systems, which the subsystems are composed of the class of the above-mentioned non-affine
nonlinear functions. Two schemes are proposed, respectively. The first scheme designs a
RBFN-based adaptive control scheme with the assumption which the interconnections
between subsystems in entire system are bounded linearly by the norms of the tracking
filtered error. In another scheme, the interconnection is assumed as stronger nonlinear
function.
We consider the differential equations in the following form described, and assume the
large-scale system is composed of the nonlinear subsystems:
where
xi R li
x&i1 = xi 2
x&i 2 = xi 3
x& = f ( x , x ,L , x , u ) + g ( x , x ,L , x )
i
i1
i2
ili
i
i
1
2
n
ili
yi = xi1
i = 1, 2,L n,
(33)
xi = [ xi1 , xi 2 ,L , xili ]T , ui R
is the input and
is the state vector,
yi R is the output of the i th subsystem.

fi ( xi , ui ) : R li +1 R
is an unknown continuous function and
implicit and smooth
function with respect to control input ui .
fi ( xi , ui ) / ui 0 for all ( xi , ui ) i R .
gi ( x1 , x2 ,L , xn ) is the interconnection term. In according
Assumption 5.
to the distinctness of the
interconnection term, two schemes are respectively designed in the following.

4.1 RBFN-based decentralized adaptive control for the class of large-scale nonlinear
systems with linear function interconnections
Assumption 6. The interconnection effect is bounded by the following function:
n
gi ( x1 , x2 ,L , xn ) ij j
j =1
where
ij
are unknown coefficients, j is a filtered tracking error to be defined shortly .
(34)
The control objective is: determine a control law, force the output,
desired output,
xdi
yi
349
, to follow a given
, with an acceptable accuracy, while all signals involved must be
bounded.
Define the desired trajectory vector xdi
tracking error
= [ ydi , y& di ,L , ydili 1 ]T and X di = ydi , y& di ,L , ydi(li ) ,
ei = xi xdi = [ei1 , ei 2 ,L , eili ]T
, thus, the filter tracking error can be
written as
i = [Ti 1]ei = ki ,1ei + ki ,2e&i + L + ki ,l 1ei(l 2) + ei(l 1) ,

i
(35)
where the coefficients are chosen such that the polynomial
ki ,1 + ki ,2 s + L + ki ,li 1s (li 2)
+ s (li 1) is Hurwitz.
Assumption 7. The desired signal xdi (t ) is bounded, so that
X di X di ,
X di is
where
known constant.
For an isolated subsystem, without interconnection function, by differentiating (35), the
filtered tracking error can be rewritten as
&i = x&il xdi( l ) + [0 iT ]ei = fi ( xi , ui ) + Ydi
(36)
with
Ydi = xdi(li ) + [0 iT ]ei .
i = ki i Ydi
where
ki
thus, [ f
(37)
f ( xi , ui ) ui 0 ,
that i u i = 0 , we invoke the
is a positive constant. With Assumption 5, we know
( xi , ui ) i ] u i 0 .
Considering the fact
implicit function theorem, there exists a continuous ideal control input

neighborhood of ( xi , ui ) i R , such that
i = fi ( xi , ui )
ui
in a
f ( xi , ui ) i = 0 , i.e. i = fi ( xi , ui ) holds.
represents ideal control inverse.
Adding and subtracting
i to
the right-hand side of
x&ili = f i ( xi , ui ) + g i
of (33), one
obtains
x&ili = f i ( xi , ui ) + g i i ki i Ydi ,
and yields
(38)
Adaptive Control
350
&i = ki i + f i ( xi , ui ) + gi i
(39)
In the same the above-discussed manner as equations (9)-(10) , we can obtain the following
equation:
i = fi ( xi , ui ) .
(40)
design the approximation pseudo-control input i as follows:
i = ki i Ydi + uci + vri ,

where
uci
(41)
is output of a neural network controller, which adopts a RBFN,
vri
is
robustifying control term designed in stability analysis.
Adding and subtracting i to the right-hand side of (39), with
i = ki i Ydi = fi ( xi , ui ) ,
we have
&i = ki i + % i ( xi , ui , ui ) uci + i i vri + gi ,

where
(42)
% i ( xi , ui , ui ) = fi ( xi , ui ) f i ( xi , ui ) is error between nonlinear function and its
ideal control function, we can use the RBFN to approximate it.

4.1.1 Neural network-based approximation
n1i and m1i be node number of input layer and
Given a multi-input-single-output RBFN, let
hidden layer, respectively. The active function used in the RBFN is Gaussian
function, Sl ( x) = exp[ 0.5( zi lk ) / k2 ] , l = 1, , n1i , k = 1, , m1i where zi R n1i 1 is input
2
vector of the RBFN,
i R n1i m1i and i R m
1i 1
Based on the approximation property of RBFN,
are the center matrix and the width vector.
% i ( xi , ui , ui ) can be written as
% i ( xi , ui , ui ) = WiT Si ( zi , i , i ) + i ( zi ) ,
(43)
i ( zi ) is approximation error of RBFN, Wi R m 1 .

Assumption 8. The approximation error ( xnn ) is bounded by i Ni , with Ni > 0 is
where
1i
an unknown constant.
The input of RBFN is chosen as zi = [ xiT , i , i ]T . Moreover, output of RBFN is designed as
uci = WiT Si ( zi , i , i ).
Define
Wi , i , i
(44)
as estimates of ideal Wi , i , i , which are given by the RBFN tuning
algorithms.
Assumption 9. The ideal values of
Wi , i , i
Wi WiM ,
where
351
WiM , iM , iM
satisfy
are positive constants.
iM ,
and
i iM ,
(45)
denote Frobenius norm and 2-
norm, respectively. Define their estimation errors as
W%i = Wi Wi ,
Using the notations: Z i
% i = i i , % i = i i .
= diag[Wi , i , i ], Z%i = diag[W%i , % i , % i ], Zi = diag[Wi , i , i ] for
convenience.
The Taylor series expansion for a given
i and i is
Si ( zi , i , i ) = Si ( zi , i , i ) + S i % i + S i% i + O ( % i , % i ) 2
where S i
(46)
(47)
S i
S k ( zi , i , i ) i evaluated at i = i ,
i = i , O( %i , % i )2 denotes the terms of order two. We use notations: Si := Si ( zi , i , i ),
S% := S ( z , % , % ) , S := S ( z , , ) .
i
S k ( zi , i , i ) i ,
Following the procedure in Appendix B, it can be shown that the following operation. The
function approximation error can be written as
WiT Si WiT Si = W%iT ( Si S i i S i i ) + WiT ( S i % i + S i% i ) + i (t ),
(48)
The disturbance term i (t ) is given by
i (t ) = WiT ( Si Si ) + WiT ( S i i + S i i ) WiT ( S i i + S i i )
(49)
Then, the upper bound of i (t ) can be written as
i (t ) Wi ( S i i
+ S i i
) + Wi T S i
+ Wi T S i
i + 2 Wi 1 ii
(50)
Adaptive Control
352
where i = max( Wi , i
, i , 2 Wi 1 ) , i = S i i
with 1 1 norm. Notice that
+ S i i
+ WiT S i
+ Wi T S i
+1 ,
i is an unknown coefficient, whereas i is a known
function.
4.1.2 Controller design and stability analysis
&i = ki i + WiT Si WiT Si + i i vri + gi + i ( zi ) ,
(51)
using (48), the above equation can become
&i = ki i + W%iT ( Si S i i S i i ) + WiT ( S i % i + S i% i )

+ i i vri + gi + i ( zi ) + i (t ).
(52)
Theorem 2. Consider the nonlinear subsystems represented by Eq. (33) and let assumptions
hold. If choose the pseudo-control input i as Eq.(41), and use the following adaptation
laws and robust control law
&
Wi = Fi ( Si S i i S i i ) i WiWi i ,
(53)
& i = Gi S iT Wi i Wi i i ,
(54)
& i = H i S iT Wi i Wi i i ,
(55)
i*i
&
i = i ii tanh(
) ii i ,
i
&
di = di ( i2 di di i ) ,
vri = i*i tanh( i i ) + di i
(56)
(57)
(58)
where
*i = i + 1
matrices,
Fi = FiT > 0, Gi = GiT > 0, H i = H iT > 0
Wi , i , di , i , di and i are
positive design parameters,
value of the uncertain disturbance term i
%i
error,
di > 0 is
= max( i , Ni ) ,
353
are any constant
i is
defining
the estimated
%i = i i
with
used to estimate unknown positive number to shield interconnection
effect, di is its estimated value, with
d%i = di di
estimated error, then, guarantee that all
signals in the system are bounded and the tracking error
ei will converge to a neighborhood
of the origin.
1
1
Li = i2 + tr (W%iT Fi 1W%i ) + tr (%iT Gi 1%i ) + tr (%iT Hi 1%i ) + i1%i 2 + di1d%i2
2
2
(59)
&
% %& + 1d% d&%
L&i = i&i + tr (W%iT Fi 1W%i ) + tr(%iT Gi 1%&i ) + tr (%iT Hi 1%&i ) + i1
i i
di i i
(60)
Applying(52) to (60), we have
k + W%iT (Si S i i S ii ) + WiT (S i %i + S i%i )

L&i = i i i
+i i vri + gi + i + i
&
% %& + 1d% d%&
+ tr (W%iT Fi 1W%i ) + tr(%iT Gi 1&%i ) + tr (%iT Hi 1&%i ) + i1
i i
di i i
Substituting the adaptive laws (53), (54) and (55) into (61), and
(61)
( &% ) = ( & ) ,yields
% %& + 1d% d&%

L&i = i [ ki i + i i vri + gi + i + i ] + Wi i tr (Z%iT Zi ) + i1
i i
di i i
ki i2 + i ( i i ) vri i + i gi + i ( ii + Ni )
% &% + 1d% d%&
+ Wi i tr (Z%iT Zi ) + i1
i i
di i i
*
ki i2 + i ( i i ) vri i + i gi + i
i i
% &% + 1d% d&%
+ Wi i tr (Z%iT Zi ) + i1
i i
di i i
(62)
Adaptive Control
354
Inserting (56) and (58) into the above inequality, we obtain
L&i ki i2 + i ( i i ) + i g i + i i*i ii*i tanh( i i )

*
%i i*i tanh( i i ) ii i di i2
i
2
d% ( d ) + tr ( Z% T Z )
*
di
Wi
(63)
*
= ki i2 + i ( i i ) + i i *i i*i tanh( i i ) + i i %ii
i
d 2 + g + d% d + tr ( Z% T Z )
i i
di
Wi
Using (11), (63) becomes
L&i ki i2 + i ( i i ) + i i i d i i2 + i gi
(64)
+ i i%ii + di d%i di + Wi tr ( Z%iT Zi )
By completing square, we have
g2
L&i k i i2 + i ( i i ) + i i i + i
4di
(65)
+ i i%ii + di d%i di + Wi tr ( Z% iT Z i )
With (41), (44), (53)-(58), approximation error between actual approximation inverse and
ideal control inverse is bounded by
i i c1i + c2i i + c3i Z%i

where
(66)
c1i , c2i , c3i are positive constants.
L&i ( ki c2 i ) i 2 + i c1i + c3i Z% i
)+
i i
g2
+ i + i i%ii + di d%i di + Wi tr ( Z% iT Z i )
4d i
Since
tr ( Z%iT Zi ) Z% i
Zi
Z%i
2
F
(67)
2
2
&
, %ii %i i %i , d%i d%i d%i d i d%i hold, the
above inequality can be written as
L&i ( ki c2i ) i 2 + i c1i + c3i Z%i
) + + 4gd
355
i i
(68)
+ i i ( %i i %i ) + di ( d%i di d%i ) + Wi ( Z%i
Zi
Z%i )
F
By completing square for (68), we get
g2
L&i ( ki c2i ) i 2 + c5i i + i i i + i
4di
= c1i + c4i ,with c4i =
where c5i
i +
di
4
di
(
+
Wi
Zi
+ c3i )
(69)
For the overall system, it can be derived that the bound as

n
n
g2
L& = L&i ( ki c2i ) i 2 + c5i i + i i i + i
4d i
i =1
i =1
According
to
(34),
gi ij j = T i
define
j =1
(70)
= [ 1 , 2 ,L n ]T
i = [ i1 , i 2 ,L in ]T , K = diag[k1 c21 , k2 c22 ,L , kn c2 n ] , C = [c51 , c52 ,L , c5n ]T

, D=
(
i =1
i i ) , the above inequality can be rewritten as
1
L& T K
i iT + C T + D = T E + C T + D
4d i
(71)
min ( E ) + C + D
where E
= K (4di ) 1 i iT , min ( E )
as long as
&
the minimum singular value of E . Then L
ki > c2i and sufficiently large di , E would be positive definite, and
0,
Adaptive Control
356
+ Dmin ( E )
2
min
(E)
%i i , d%i d i , Z% i
2min ( E )
1
Wi
Wi
= A,
(72)
Zi
+ c3i )
Now, we define
Zi
A} , i = %i
d%i di , Zi = Z%i
di = d%i
Since
, i , di , Wi , c3i
Z%i
Wi
Wi
Zi
L&
, Z i
is negative outside these compacts set. According to a
standard Lyapunov theorem, this demonstrates that
Z%i , %i , d%i
and
are bounded and will
converge to , Z i , and d , respectively. Furthermore, this implies

i
(73)
+ c3i )
are positive constants, we conclude that
and d are compact sets. Hence

i
%i i ,
ei
is bounded
and will converge to a neighborhood of the origin and all signals in the system are bounded.
4.1.3 Simulation Study
In order to validate the effectiveness of the proposed scheme, we implement an example,
and assume that the large-scale system is composed of the following two subsystems
defined by
Subsystem
Subsystem
x&11 = x12
2
2
1 : x&12 = x11 + 0.02( x11 ) x12 + u1
+ ( x112 + x122 ) (u1 ) + 0.2 + sin(0.2 x21 )
(74)
x&21 = x22
2
2
2 : x&22 = x21 + 0.1(1 + x22 )u2 + tanh(0.1u2 )
+ 0.15u23 + tanh(0.1x11 )
(75)
where = 0.4 , (u1 ) = (1 eu1 ) (1 + eu1 ) . The desired trajectory xd 11 = 0.1 [sin(2t ) cos(t )] ,
xd 21 = 0.1 cos(2t ) .
357
Input vectors of neural networks are zi = [ xiT , i , i ]T , i = 1, 2 , and number of hidden layer
nodes both 8. The initial weight of neural network is Wi (0) = (0) . The center values and the
widths of Gaussian function are initialized as zeroes, and
5 , respectively. The initial
condition of controlled plant is x1 (0) = [0.1, 0.2] x2 (0) = [0, 0] . The other parameters are
T
chosen as follows:
i = 5, ki = 5 Wi = 0.001, i = 1, di = 1, i = 0.01, di = 0.01 , i = 10 , Fi = 10 IWi ,
G = 2 I i , H = 2 I i , with IWi , I i , I i corresponding identity matrices.

Fig.5 shows the results of comparisons of tracking errors of two subsystems. Fig.6 gives
control input of two subsystems, Fig.7 and Fig.8 the comparison of tracking of two
subsystems, respectively. Fig.9 and Fig.10 illustrate outputs of two RBFNs and the change of
norms of W , , , respectively. From these results, it can be seen that the effectiveness of the
proposed scheme is validated, and tracking errors converge to a neighborhood of the zeroes
and all signals in system are bounded. Furthermore, the learning rate of neural network
controller is rapid, and can track the desired trajectory in about 1 second. From the results of
control inputs, after shortly shocking, they tend to be smoother, and this is because neural
networks are unknown for objective in initial stages.
0.4
e11,e21
0.2
0
-0.2
-0.4
0
10
time sec
15
20
Fig. 5. Tracking error of two subsystems: 1(solid), 2(dot)

8
6
control input
4
2
0
-2
-4
-6
-8
0
10
time sec
Fig. 6. Control input of two subsystems: 1(solid), 2(dot)
15
20
Adaptive Control
358
0.6
x11,xd11
0.4
0.2
0
-0.2
-0.4
-0.6
-0.8
0
10
time sec
Fig. 7. Comparison of the tracking of subsystem 1:
15
20
x11 (solid) and xd 11 (dot)
0.6
x21,xd21
0.4
0.2
0
-0.2
-0.4
-0.6
-0.8
0
10
time sec
Fig. 8. Comparison of the tracking of subsystem 2:
15
20
x21 (solid) and xd 21 (dot)
15
10
5
0
-5
-10
0
10
time sec
Fig. 9. Subsystem 1: Output of RBFN (solid), norms of

dot)
15
20
W (dash), (dot), (dash-
359
15
10
-5
0
10
time sec
15
20
W (dash), (dot), (dash-
Fig. 10. Subsystem 2: Output of RBFN (solid), norms of

dot)
4.2 RBFN-based decentralized adaptive control for the class of large-scale nonlinear
systems with nonlinear function interconnections
Assumption 10. The interconnection effect is bounded by the following function:
gi ( x1 , x2 ,L , xn ) j =1 ij (| j |) ,
n
where
ij (| j |)
are unknown smooth nonlinear function,
(76)
j are filtered tracking errors to
be defined shortly .
The control objective is: determine a control law, force the output,
desired output,
xdi
yi
, to follow a given
, with an acceptable accuracy, while all signals involved must be
bounded.
X di = [ ydi , y& di ,L , ydi(li ) ]T and
Define the desired trajectory vector xdi = [ ydi , y& di ,L , ydili 1 ]T ,
tracking error ei = xi xdi = [ei1 , ei 2 ,L , eil ]T , thus, the filter tracking error can be written as
i
i = [Ti 1]ei = ki ,1ei + ki ,2 e&i + L + ki ,l 1ei( l 2) + ei(l 1) ,

i
(77)
where the coefficients are chosen such that the polynomial ki ,1 + ki ,2 s + L + ki ,l 1s (li 2) + s ( li 1)
i
is Hurwitz.
Assumption 11. The desired signal
xdi (t )
is bounded, so that
X di X di , with X di a
known constant.
For an isolated subsystem, without interconnection function, by differentiating (77), the
filtered tracking error can be rewritten as
Adaptive Control
360
&i = x&il ydi( l ) + [0 iT ]ei = fi ( xi , ui ) + Ydi
(78)
= ydi( li ) + [0 iT ]ei .
with Ydi
i = ki i + Ydi ,
(79)
ki is a positive constant. With Assumption 5, we know f ( xi , ui ) ui 0 ,

thus, [ f ( xi , ui ) i ] u i 0 . Considering the fact that i u i = 0 , with the implicit
where
function theorem, there exists a continuous ideal control input
a neighborhood
i = fi ( xi , ui ) holds.
Here, i = f i ( xi , ui ) represents an ideal control inverse. Adding and subtracting i to the

of
( xi , ui ) i R
right-hand side of
, such that
f ( xi , ui ) i = 0
ui in
x&ili = f i ( xi , ui ) + g i
, i.e.
of (33), one obtains
x&ili = fi ( xi , ui ) + gi + i Ydi ki i ,
(80)
&i = ki i + fi ( xi , ui ) + gi + i ,
(81)
and yields
Similar to the above-mentioned equation (40), i
= f i ( xi , ui )
holds.
design the approximation pseudo-control input i as follows:
i = ki i Ydi uci W giT S gi (| i |) i vri ,

where
uci
(82)
is output of a neural network controller, which adopts a RBFN,
robustifying control term designed in stability analysis,
W S gi (| i |)
T
gi
vri
is
is used to
compensate the interconnection nonlinearity (we will define later).

Adding and subtracting i to the right-hand side of (81), with i
= ki i + Ydi = f i ( xi , ui ) ,
we have
&i = ki i + % i ( xi , ui , ui ) uci W giT S gi (| i |) i + i i vri + gi ,
(83)
where
% i ( xi , ui , ui ) = fi ( xi , ui ) f i ( xi , ui ) is
361
error between the nonlinear function
and its ideal control function, we can use the RBFN to approximate it.
4.2.1 Neural network-based approximation
Based on the approximation property of RBFN,
% i ( xi , ui , ui ) can be written as
% i ( xi , ui , ui ) = WiT Si ( zi ) + i ( zi ) ,
where
Wi
is the weight vector,
S i ( zi )
is Gaussian basis function,
(84)
i ( zi )
is the
R , q the number of input node.

Assumption 12. The approximation error i ( zi ) is bounded by | i | Ni , with Ni > 0 is
approximation error and the input vector zi
an unknown constant. The input of the RBFN is chosen as zi
= [ xiT , i , i ]T . Moreover,
output of the RBFN is designed as
uci = WiT Si ( zi ).
Define
Wi
(85)
as estimates of ideal Wi , which are given by the RBFN tuning algorithms.
Assumption 13. The ideal value of
Wi satisfies
|| Wi || WiM ,
where
WiM is positive known constant, with estimation errors as W%i = Wi Wi
(86)
.
4.2.2 Controller design and stability analysis

&i = ki i + W%iT Si + i i vri + gi W giT S gi (| i |) i + i ( zi )
(87)
Theorem 3. Consider the nonlinear subsystems represented by Eq. (33) and let assumptions
hold. If choose the pseudo-control input i as Eq.(82), and use the following adaptation
laws and robust control law
&
Wi = Fi [ Si i WiWi | i |] ,
(88)
Adaptive Control
362
&
W gi = Gi [ S gi (| i |) i2 giW gi | i |] ,
(89)
i = i [ i (| i | +1) tanh( i i ) ii | i |] ,
(90)
&
vri = i (| i | +1) tanh( i i ) ,
(91)
= FiT > 0 , Gi = GiT > 0 are any constant matrices, i , Wi , gi , i and i are
positive design parameters, is the estimated value of the unknown approximation errors,
where Fi
which will be defined shortly, then, guarantee that all signals in the system are bounded and
the tracking error
ei will converge to a neighborhood of the origin.
2 Li = i2 + W%iT Fi 1W%i + W% giT Gi 1W% gi + i1%i 2
(92)
&
&
L&i = i&i + W%iT Fi 1W%i + W% giT Gi 1W% gi + i1%i%i
Applying (87) and(53) to (59) and ( &% )
(93)
= (& ) , we have
L&i = i [ki i + i i vri + gi W giT S gi (| i |) i + i ]

&
&
+ WiW%iT Wi | i | +W% giT Gi 1W% gi + i1%i%i
(94)
Using (76), (94) is rewritten as

n
L&i ki i2 + i ( i i ) vri i + i [ j =1 ij (| j |) W giT S gi (| i |) i ]
&
+ | i | Ni + % % + WiW%iT Wi i + W% giT Gi 1W%& gi
(95)
1
i i i
ij () is a smooth function, there exists a smooth function ij (| j |) , (1 i, j n)

such that ij (| j |) =| j | ij (| j |) hold. Thus, we have
Since
363
n
L&i ki i2 + i ( i i ) vri i + i2 [ j =1 ij (| j |) W giT S gi (| i |)]
(96)
&
&
+ | i | Ni + % % + WiW%iT Wi | i | +W% giT Gi 1W% gi
1
i i i
Since the function
di (| i |) can
d i (| i |) = i =1 ij (| i |) is smooth and i is on a compact set,

n
be approximated via a RBFN, i.e.,
bounded approximation error

boundedness || Wgi
gi , | gi | gNi
|| WgMi , WgMi > 0 a
di (| i |) = WgiT S gi (| i |) + gi ,
W is estimate of ideal W ,
gi
gi
with
with
known constant, and the estimation errors as
W% gi = Wgi W gi . Then, (96) becomes

n
L&i ki i2 + i ( i i ) vri i + i [ ij ( j ) W giT S gi ( i ) i ]

j =1
&
&
+ i Ni + i1%i%i + WiW%iT Wi i + W% giT Gi 1W% gi
(97)
ki i2 + i ( i i ) vri i + i2W% giT S gi (| i |) + gi i2 + | i | Ni

&
&
+ i1%i%i + WiW%iT Wi i + W% giT Gi 1W% gi
Substituting the adaptive law (89), we obtain
% &%
L&i ki i2 + i (i i ) vri i + gNi i2 + | i | Ni + i1
i i
+ WiW%iTWi | i | + giW%giTWgi | i |
Define i
(98)
= max( Ni , gNi ) , with i is its estimate, and %i = i i with %i error. (98) can
be rewritten as
L&i ki i2 + i ( i i ) vri i + i ( i2 + i )
&
+ i1%i%i + WiW%iTWi i + giW% giTW gi i
Applying the adaptive law (56) and robust control term (58), we have
(99)
Adaptive Control
364
L&i ki i2 + i ( i i ) i i ( i + 1) tanh( i i ) + i i ( i + 1)
%i i ( i + 1) tanh( i i ) + WiW%iT Wi i + giW% giT W gi i + i%ii i
= ki i2 + i ( i i ) + i i ( i + 1) i i ( i + 1) tanh( i i )
+ WiW%iT Wi i + giW% giT W gi i + i%ii i
(100)
= ki i2 + i ( i i ) + i ( i + 1) i i tanh( i i )
+ W% T W + W% T W + %
Wi
gi
gi
gi
i i i
Using (11), we get
L&i ki i2 + i (i i ) + i (| i | +1)i
% | |
+ WiW%iTWi | i | + giW%giTWgi | i | + i
i i
i
(101)
With (82), (85), and (88)-(91), the approximation error between the ideal control inverse and
the
actual
approximation
inverse
is
bounded
by
| i i | c1i + c2i | i |
+c3i || W%i || + c4i || W% gi ||, with c1i , c2i , c3i , c4i positive constants. Moreover, we utility the
facts, a%
a || a% |||| a || || a% ||2
, (101) can be rewritten as
L&i ( ki c2i ) i 2 + i c1i + c3i W%i + c4i W% gi
Wi W%i ( Wi W%i )
+ i + gi W% gi ( Wgi W% gi
+ % ( % )
i
i i i
W% + ( W + c ) W%
i
i
3i
Wi i
2
2
( ki c2i ) i + c1i i + i i i + i + gi W% gi + ( Wgi + c4i ) W% gi

% 2 + %
i
i
i i
2
) + i i i
(102)
Completing square for (102), we have
L&i ( ki c2i ) i 2 + c8i | i | +i i i
(103)
365
c5i = Wi || Wi || +c3i , c6i = gi || Wgi || +c4i , c7i = i 2 4i + c62i 4 gi + c52i 4 Wi
with
c8i = c1i + i i i + c7 i .
For the overall system, we have
n
n
L& = i =1 L&i i =1[ ( ki c2i ) i 2 + c8i | i | +i i i ]
Now, define
(104)
= [| 1 |,L | n |]T , K = diag[k1 c21 ,L , kn c2 n ] , C = [c81 , c82 ,L , c8 n ]T ,
D = i =1 (i i i ) . (104) can be rewritten as

n
L& T K + C T + D min ( K ) || ||2 + || C |||| || + D
(105)
By completing square, yields

2
C
C
L& min ( K )
+D
+
2min ( K ) 4min ( K )
Clearly,
(106)
L& 0 , as long as ki > c2i , and

A, %i i1 i , W%i c5i 1 Wi , W% gi c6i 1 Wgi
Wi
where
(107)
gi
2
A = [ C + Dmin ( K )] [4min
( K )] + C [2min ( K )] with min ( K ) the minimum
singular value of K .
Now, we define
Wi = W%i
Since
that
1
W%i c5i Wi
Wi , Wgi = W% gi
Wi , Wgi , i , i , Wi , Wgi , c5i , c6i
, i
A} , i = %i %i i1 i ,
W i
and
Wgi
are
W% gi c6i gi1 Wgi ,
positive
are compact sets. Hence
L&
constants,
we
(108)
conclude
is negative outside these
compacts set. According to a standard Lyapunov theorem, this demonstrates that
W%i , W% gi , %i
respectively.
and
are bounded and will converge to
, i
W i
and
Wgi
Adaptive Control
366
Furthermore, this implies
ei
is bounded and will converge to a neighborhood of the origin
and all signals in the system are bounded.

4.2.3 Simulation Study
In order to validate the effectiveness of the proposed scheme, we implement an example,
and assume that the large-scale system is composed of the following two subsystems
defined by
Subsystem
Subsystem
x&11 = x12
2
2
2
2
1 : x&12 = x11 + 0.02( x11 ) x12 + u1 + ( x11 + x12 ) (u1 )
+ 0.1|| x2 || exp(0.5 || x2 ||)
(109)
x&21 = x22
2
2
3
2 : x&22 = x21 + 0.1(1 + x22 )u2 + tanh(0.1u2 ) + 0.15u2
+ 0.2 || x2 || exp(0.1|| x2 ||)
(110)
The
desired
trajectory
(u1 ) = (1 e u ) (1 + e u ) .
xd 11 = 0.1 [sin(2t ) cos(t )] , xd 21 = 0.1 sin(2t ) . For the RBFNs as (84), input vectors are
where
= 0.4
chosen as zi
= [ xiT , i , i ]T , i = 1, 2
and number of hidden layer nodes both 8, the initial
weights Wi (0) = (0) and the center values and the widths of Gaussian function zero, and 2,
respectively. For the RBFNs, which used to compensate the interconnection nonlinearities,
both input vectors are
[ 1 , 2 ]T
, number of hidden layer nodes is 8, the initial
weights W gi (0) = (0) , and the center values and the widths of Gaussian function zero,
and
5 , respectively. The initial condition of controlled plant is
x1 (0) = [0.2, 0.2]T ,
x2 (0) = [0.3,0.2]T . The other parameters are chosen as follows: i = 1, ki = 2 ,
Wi = 0.001, i = 0.1, i = 0.01, i = 10
Fi = 10 IWi , G = 2 I gi , with IWi , I gi
corresponding identity matrices. Fig.11 and 12 show the results of comparisons of tracking
errors and control input of two subsystems, Fig.13 and 14 the comparison of tracking of two
subsystems, respectively. Fig.15 and Fig.16 illustrate the norm of the four weights in two
subsystems, respectively. From these results, it can be seen that the effectiveness of the
proposed scheme is validated, and tracking errors converge to a neighborhood of the zeroes
and all signals in system are bounded. Furthermore, the learning rate of neural network
controller is rapid, and can track the desired trajectory in less than 3 seconds. From the
results of control inputs, after shortly shocking, they tend to be smoother, and this is
because neural networks are unknown for objective in initial stages. As desired, though the
system is complex, the whole running process is well.
u1
u2
2.5
2
1.5
u1,u2
1
0.5
0
-0.5
-1
-1.5
-2
-2.5
0
10
time sec
15
20
Fig. 12. Control input of subsystem1: u1 , and subsystem 2: u2
xd11
x11
0.8
0.6
x11,xd11
0.4
0.2
0
-0.2
-0.4
10
time sec
Fig. 13. Comparion of tracking of subsystem 1
15
0.5
20
xd21
x21
0.4
0.3
x21,xd21
0.2
0.1
0
-0.1
-0.2
-0.3
-0.4
0
10
time sec
Fig. 14. Comparion of tracking of subsystem 2
15
20
367
Adaptive Control
368
1.2
NN1
||Wg1||
||W1||
1
0.8
0.6
0.4
0.2
0
-0.2
-0.4
-0.6
0
10
time sec
15
20
Fig. 15. The norms of weights and output of RBFNof subsystem1

0.6
0.4
0.2
0
-0.2
-0.4
||Wg2||
||W2||
NN2
-0.6
-0.8
0
10
time sec
15
20
Fig. 16. The norms of weights and output of RBFNof subsystem 2
5. Conclusion
In this chapter, first, a novel design ideal has been developed for a general class of nonlinear
systems, which the controlled plants are a class of non-affine nonlinear implicit function and
smooth with respect to control input. The control algorithm bases on some mathematical
theories and Lyapunov stability theory. In order to satisfy the smooth condition of these
theorems, hyperbolic tangent function is adopted, instead of sign function. This makes
control signal tend smoother and system running easier. Then, the proposed scheme is
extended to a class of large-scale interconnected nonlinear systems, which the subsystems
are composed of the above-mentioned class of non-affine nonlinear functions. For two
classes of interconnection function, two RBFN-based decentralized adaptive control schemes
are proposed, respectively. Using an on-line approximation approach, we have been able to
relax the linear in the parameter requirements of traditional nonlinear decentralized
adaptive control without considering the dynamic uncertainty as part of the
interconnections and disturbances. The theory and simulation results show that the neural
network plays an important role in systems. The overall adaptive schemes are proven to
369
guarantee uniform boundedness in the Lyapunov sense. The effectiveness of the proposed
control schemes are illustrated through simulations. As desired, all signals in systems,
including control signals, are tend to smooth.
6. Acknowledgments
This research is supported by the research fund granted by the Natural Science Foundation
of Shandong (Y2007G06) and the Doctoral Foundation of Qingdao University of Science and
Technology.
7. References
Lewis,F.L.; Yesildirek A. & Liu K.(1996). Multilayer neural-net robot controller with
guaranteed tracking performance. IEEE Trans.on Neural Networks, Vol. 7, No.2, Mar.
1996, pp.388-399, ISSN 1045-9227
Yu, W. & Li, X.(2002). Adaptive control with multiple neural networks. Proceeding of the
American Control Conference, pp. 1543-1549, ISBN 0-7803-7298-0, May 8-10,2002
Anchorage, AK
Yesidirek, A. and Lewis,F.L.(1995). Feedback linearization using neural networks.
Automatica, Vol.31, No.11, 1995, pp. 1659-1664, ISSN 0005-1098
Calise,A.J. & Hovakimyan, N. (2001). Adaptive output feedback control of nonlinear system
using neural networks. Automatica, Vol.37, 2001, pp.1201-1211, ISSN 0005-1098
Ge, S. S. ; Hang, C. C. & Zhang, T.(1997). Direct adaptive neural network control of
nonlinear systems. Proceedings of the American Control Conference, pp. 1568-1572,
ISBN 0-7803-3832-4, 1997, Albuqerque, New Mexico
Ioannou, P.A. (1986). Decentralized adaptive control of interconnected systems, IEEE Trans.
on Automatic Control, Vol. 31, Apr. 1986, pp. 291-298, ISSN 0018-9286
Siljak, D.D.(1991). Decentralized control of complex systems. Academic, 1991, ISBN-10:
0126434301, Boston
Fu, L.C. (1992). Robust adaptive decentralized control of robot manipulators. IEEE Trans. on
Automatic Control,Vol.37, 1992, pp.106110, ISSN 0018-9286
Sheikholeslam, S. & Desor, C.A. (1993). Indirect adaptive control of a class of interconnected
nonlinear dynamical systems. Int J Control, Vol. 57, No.3, 1993, pp.742765, ISSN
0020-7179
Wen, C. (1994). Decentralized adaptive regulation. IEEE Trans. on Automatic Control, Vol.39,
pp.21632166, ISSN 0018-9286
Tang, Y. ; Tomizuka, M. & Guerrero, G. (2000). Decentralized robust control of mechanical
systems. IEEE Trans. on Automatic Control, Vol.45, No.4, 2000, pp. 21632166, ISSN
0018-9286
Spooner, J.T. & Passino, K.M.(1999). Decentralized adaptive control of nonlinear systems
using radial basis neural networks, IEEE Trans. on Automatic Control, Vol. 44, No.11,
1999, pp.2050-2057, ISSN 0005-1098
Huang, S.; Tan, K.K. & Lee, T.H. (2003). Decentralized control design for large-scale systems
with strong interconnections using neural networks, IEEE Trans. on Automatic
Control, Vol.48, No.5, 2003, pp. 805-810, ISSN 0018-9286
Adaptive Control
370
Huang, S.N. & Tan, K.K. (2006). Nonlinear adaptive control of interconnected systems using
neural networks. IEEE Trans Neural Networks, Vol.17, No.1, 2006, pp.243246,
ISSN 1045-9227
Nardi, F. & Hovakimyan, N.(2006). Decentralized control of largescale systems using single
hidden layer neural networks. Proceedings of the American control conference,
pp.31233127, ISBN 0-7803-6495-3, June 2001,Arilington
Huang, S.N. & Tan, K.K. (2005). Decentralized control of a class of large-scale nonlinear
systems using neural networks. Automatica, Vol.41, 2005, pp.16451649, ISSN 00051098
Ge, S. S. ; Hang, C. C. & Zhang, T.(1998). Nonlinear adaptive control using neural networks
and its application to CSTR systems. Journal of Process Control, Vol.9, 1998, pp.313323, ISSN 0959-1524
Tsinias, J. & Kalouptsidis,N.(1983). Invertibility of nonlinear analytic single-input systems,
IEEE Trans. on Automatic Control, Vol.28, No. 9, 1983, pp. 931 933, ISSN 0018-9286
Lang, S. (1983). Real Analysis, Reading, ISBN-10: 0201141795, MA: Addison-Wesley, Reading
Slotine, J.-J. E. & Li, W.P. (1991). Applied Nonlinear Control, Englewood Cliffs, ISBN-10:
0130408905, NJ: Prentice Hall
Polycarpou, M.M.(1996). Stable adaptive neural control scheme for nonlinear system, IEEE
Trans. on Automatic Control, Vol. 41, No. 3, 1996, pp.447-451, ISSN 0018-9286
Appendix A
As Eq.(19), the approximation error of function can be written as
M T M T = M T M T + M T M T = M T ( ) + M% T
Substituting (18) into the above equation, we have
M T ( ) + M% T
= M% T [ + N% T xnn + O( N% T xnn )2 ] + M T [ N% T xnn + O( N% T xnn )2 ]
= M% T + M% T N% T x + M T N% T x + M T O( N% T x )2
nn
nn
nn
= M% T + M% T N T xnn M% T N T xnn + M T N% T xnn + M T O( N% T xnn )2

= M% T ( N T x ) + M T N% T x + M% T N T x + M T O( N% T x )2
nn
nn
Define that
= M% T NT xnn + M T O(N% T xnn )2
nn
nn
371
so that
M T M T = M% T ( N T xnn ) + M T N% T xnn +
Thus,
= M T M T M% T ( N T xnn ) M T N% T xnn
= M T M T + M% T N T xnn M T N% T xnn
= M T ( ) + M T N T x M T N T x M T N% T x
nn
nn
nn
= M ( ) + M N xnn M N xnn
T
Appendix B
Using (46) and (47), the function approximation error can be written as
WiT Si WiT Si = WiT Si WiT Si + WiT Si WiT Si = W%iT Si + WiT S%i

= W%iT [Si + S i %i + S i%i + O(%i ,%i )2 ] + WiT [Si + S i %i + S i%i + O(%i ,%i )2 Si ]
= W%iT Si + W%iT (S i %i + S i%i ) + W%iT O(%i ,%i )2 + WiT (S i %i + S i%i ) + WiT O(%i ,%i )2
= W%iT Si + W%iT [S i (i i ) + S i ( i i )] + WiT (S i %i + S i%i ) + WiT O(%i ,%i )2
= W%iT (Si S i i S ii ) + W%iT (S i i + S i i ) + WiT (S i %i + S i%i ) + WiT O(%i ,%i )2
= W%iT (Si S i i S ii ) + WiT (S i %i + S i%i ) + i (t ).
define as
i (t) = W%iT (S i i + S i i ) +WiT O(%i ,%i )2

Thus,
Adaptive Control
372
i (t ) = W%iT Si + WiT S%i W%iT (Si S i i S ii ) WiT (S i %i + S i%i )

= W%iT S%i + WiT S%i + W%iT (S i i + S ii ) WiT (S i %i + S i%i )
= WiT S%i + W%iT (S i i + S ii ) WiT (S i %i + S i%i )
= WiT S%i + WiT (S i i + S ii ) WiT (S i i + S i i )
This book was distributed courtesy of:
For your own Unlimited Reading and FREE eBooks today, visit:
http://www.Free-eBooks.net
Share this eBook with anyone and everyone automatically by selecting any of the
options below:
To show your appreciation to the author and help others have

wonderful reading experiences and find helpful information too,
we'd be very grateful if you'd kindly
post your comments for this book here.
COPYRIGHT INFORMATION
Free-eBooks.net respects the intellectual property of others. When a book's copyright owner submits their work to Free-eBooks.net, they are granting us permission to distribute such material. Unless
otherwise stated in this book, this permission is not passed onto others. As such, redistributing this book without the copyright owner's permission can constitute copyright infringement. If you
believe that your work has been used in a manner that constitutes copyright infringement, please follow our Notice and Procedure for Making Claims of Copyright Infringement as seen in our Terms
of Service here:
http://www.free-ebooks.net/tos.html

Adaptive Control

Uploaded by

Copyright:

Available Formats

Adaptive Control

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Adaptive Control

Uploaded by

Copyright:

Available Formats

Adaptive Control

Adaptive Control, Edited by Kwanho You

for Adaptive Control Strategies, by R. Ribeiro and K.

Sungkyunkwan University, Korea

Automatic 3D Model Generation based on a Matching of Adaptive Control

Na-Young Lee, Joong-Jae Lee, Gye-Young Kim and Hyung-Il Choi

Adaptive Estimation and Control for Systems with Parametric and

Hongbin Ma and Kai-Yew Lum

Adaptive output regulation of unknown linear systems with unknown

Ikuro Mizumoto and Zenta Iwai

Output Feedback Direct Adaptive Control for a Two-Link Flexible Robot

Selahattin Ozcelik and Elroy Miranda

S. Alonso-Quesada and M. De la Sen

Hybrid Schemes for Adaptive Control Strategies

Ricardo Ribeiro and Kurios Queiroz

Adaptive Control for Systems with Randomly Missing Measurements in a

Yang Shi and Huazhen Fang

Adaptive Control Based On Neural Network

Sun Wei, Zhang Lujin, Zou Jinhai and Miao Siyi

Krzysztof Szabat and Teresa Orlowska-Kowalska

10. Adaptive Control of Dynamic Systems with Sandwiched Hysteresis Based

11. High-Speed Adaptive Control Technique Based on Steepest Descent

Ken Tanizawa and Akira Hirose

12. Adaptive Control of Piezoelectric Actuators with Unknown Hysteresis

Wen-Fang Xie, Jun Fu, Han Yao and C.-Y. Su

13. On the Adaptive Tracking Control of 3-D Overhead Crane Systems

Yang, Jung Hua

Adaptive Inverse Optimal Control of a Magnetic Levitation System

Yasuyuki Satoh, Hisakazu Nakamura, Hitoshi Katayama and Hirokazu Nishitani

Research Division, Korea Atomic Energy Research Institute

Fig. 1. Overall flow of the system configuration

Automatic 3D Model Generation based on a Matching of Adaptive Control Points

distal length Os distal length

Table 1. Database of the coronary artery

AP0 LAO60 LAO60

Table 2. Measured angles of the vessel bifurcation from six angiographies

Automatic 3D Model Generation based on a Matching of Adaptive Control Points

Fig. 3. 3D model generation of the standard vessel from six angiographies

Fig. 4. Projection result for 2D image of standard vessel

3. Matching of the Adaptive Control Points

Automatic 3D Model Generation based on a Matching of Adaptive Control Points

Fig. 5. Extracted 9 control points in individual vessel

(a) Segmented vessel

representation, it is natural to describe the coronary tree by a graph structure (Chalopin et

I thin = { v po int , bif }

v po int = { vstart _ po int , vend _ po int 1, vend _ po int 2 }

Fig. 7. Primitives of a vascular net

Automatic 3D Model Generation based on a Matching of Adaptive Control Points

(a) Individual vessel

Fig. 9. Result of an adaptive interpolation of the corresponding control points

K (r ) is the solution of the biharmonic equation (2 K = 0) that satisfies the condition of a

(a) Individual vessel

(c) Warped vessel

5. Results of the Vessel Transformation

Automatic 3D Model Generation based on a Matching of Adaptive Control Points

Fig. 12. Result of standard vessel warping

Automatic 3D Model Generation based on a Matching of Adaptive Control Points

A, Wahle.; E, Wellnhofer.; I, Mugaragu.;H.U, Sauer.; H, Oswald. & E, Fleck. (1994). Accurate

Temasek Laboratories, National University of Singapore

appeared in previous study on capability and limitations of the feedback