Position Control of A Permanent Magnet DC Motor by Model Reference Adaptive Control

POSITION CONTROL OF A PERMANENT MAGNET DC MOTOR BY
MODEL REFERENCE ADAPTIVE CONTROL
by
Fedai YENİCİ
July 2006
v
by
Fedai YENİCİ
A thesis submitted to
the Graduate Institute of Sciences and Engineering
of
Fatih University
in partial fulfillment of the requirements for the degree of
Master of Science
in
Electronics Engineering
July 2006
Istanbul, Turkey
v
APPROVAL PAGE
I certify that this thesis satisfies all the requirements as a thesis for the degree of Master of
Science.
Prof. Dr. Muhammet KÖKSAL

Chairperson
Electronics Engineering Department
This is to certify that I have read this thesis and that in my opinion it is fully adequate, in
scope and quality, as a thesis for the degree of Master of Science.
Prof. Dr. Muhammet KÖKSAL

Supervisor
Examining Committee Members
Prof. Dr. Muhammet KÖKSAL ...............................................
Assoc. Prof. Dr. Erkan İMAL ...............................................
Assist. Prof. Dr. Atakan KURT ...............................................
It is approved that this thesis has been written in compliance with the formatting rules laid
down by the Graduate Institute of Sciences and Engineering.
Assist. Prof. Dr. Nurullah ASLAN

Director
Date
July 2006
v
iii

Fedai YENİCİ
M. S. Thesis - Electronics Engineering
July 2006
Supervisor: Prof. Dr. Muhammet KÖKSAL
ABSTRACT
Model reference adaptive control is one of the various techniques of solving the control
problem when the parameters of the controlled process are poorly known or vary during normal
operation. To understand the dynamic behavior of a dc motor it is required to know its parameters.
Armature inductance, armature resistance, inertia of the rotor, motor constants and friction
coefficient are the main parameters of a dc motor. To identify all these parameters, some
experiments should be performed. However, motor parameters change during the operation
according to several conditions. Therefore, the performance of the controller, which has been
designed considering constant motor parameters, becomes poorer due to parameter variations. For
this reason, a model reference adaptive control method is proposed to control the position of a dc
motor without requiring any fixed motor parameter. Experimental results show how well this
method controls the position of the motor.
Keywords: Model Reference Adaptive Control, Lyapunov’s Direct Method, Dc Motor.

iv
MODEL REFERANS UYARLAMALI KONTROL İLE KUTUPLARI SABİT

MIKNATISLI BİR DC MOTORUN KONUM KONTROLU
Fedai YENİCİ
Yüksek Lisans Tezi – Elektronik Mühendisliği
Temmuz 2006
Tez Yöneticisi: Prof. Dr. Muhammet KÖKSAL
ÖZ
Prosesin parametrelerinin tam olarak bilinmemesi veya parametrelerin çalışma esnasında

değişmesi durumunda kontrol probleminin çözümünde kullanılabilecek çeşitli tekniklerden biri
de model referans uyarlamalı denetimdir. Doğru akım motorunun dinamiğinin anlaşılması için
motor parametrelerinin bilinmesi gerekmektedir. Bu parametrelerden başlıcaları armatür
endüktansı, armatür direnci, rotorun dönel eylemsizliği, motor sabitleri ve rotor yatağındaki
sürtünmedir. Tüm bu parametrelerin belirlenmesi için bazı deneylerin yapılması gerekir. Ancak
motor parametreleri, motorun çalışması sırasındaki koşullara bağlı olarak değişmektedir.
Dolayısıyla, parametrelerin değişmediği varsayılarak tasarlanmış olan denetleyicinin performansı,
çalışma sırasındaki değişimlerden dolayı düşer. Bu nedenden dolayı bu çalışmada, sabit motor
parametrelerine gerek kalmadan, bir dc motorun konumunun kontrol edildiği bir uyarlamalı
denetim yöntemi sunulmaktadır. Deney sonuçları motor konumunun önerilen yöntem ile
denetlenebildiğini göstermiştir.
Anahtar Kelimeler: Model Referans Uyarlamalı Kontrol, Lyapunov’un direkt metodu, Dc

Motor.
v
ACKNOWLEDGEMENT
I would like to thank my advisor, Prof. Dr. Muhammet Köksal, for his knowledge and
support in helping me throughout my research. I appreciate his enlightening guidance and
advice to me in completing this study. Especially his serious attitude on research and his
pursuit for the perfect work will help me in a long run.
I also sincerely thank my committee member, Assoc. Prof. Dr. Erkan İmal, for his
patience and kind support for completing this thesis. Their lectures helped me fulfilling this
research.
vi
TABLE OF CONTENTS
ABSTRACT ....................................................................................................................iii
ÖZ.................................................................................................................................... iv
ACKNOWLEDGMENT.................................................................................................. v
TABLE OF CONTENTS ................................................................................................ vi
LIST OF FIGURES.......................................................................................................viii
LIST OF SYMBOLS AND ABBREVIATIONS............................................................. x
CHAPTER 1 INTRODUCTION................................................................................. 1
CHAPTER 2 MODEL REFERENCE ADAPTIVE CONTROL.............................. 10
2.1 General Control System Design Steps. .............................................................. 10
2.2 Adaptive Control. ............................................................................................... 12
2.2.1 Direct and Indirect Adaptive Control ..................................................... 13
2.2.2 Model Reference Adaptive Control .........................................................16
CHAPTER 3 MRAC FOR SYSTEMS WITHOUT FINITE TRANSMISSION
ZEROS .............................................................................................19
3.1 The structure of the adaptive control system.......................................................19
3.2 Derivation of the adaptive laws...........................................................................21
CHAPTER 4 MRAC FOR DC MOTOR....................................................................28
4.1. Mathematical Modeling of PM DC motors........................................................28
4.2 Parameters of DC Motor to be controlled..........................................................31
4.2.1 Usual Empirical Method..........................................................................32
4.2.2 Parameter Estimation with Matlab’s System Identification Toolbox......35
4.3 Approximation of High Order Systems by Low Order System.........................39
4.3.1 Order Reduction with Negligence of L ................................................39
a
4.3.2 Order Reduction with Matlab’s “MODRED” and “BALREAL”
commands................................................................................................40
vii
4.3.3 Order Reduction with Matlab’s System Identification Toolbox.............40

4.3.4 Comparison Between Third Order Transfer Function with
Second Order Transfer Function..............................................................42
4.4 MRAC of PM DC Motors.....................................................................................43
4.4.1 Experimental Setup for MRAC ...............................................................44
4.4.2 Review of Adaptive Laws for the Control of Tested Motors..................45
CHAPTER 5 SIMULATION RESULTS ................................................................ 48
5.1 Effect of Adaptive Gains α on the Performance..............................................48
5.2 Effects of Positive Definite Matrices P on the Performance.............................52
5.3 Effects of Simulation Step Sizes T on the Performance......................................55
5.4 Effects of Reference Models on the Performance...............................................57
5.5 Effects of Disturbances on the Performance.......................................................59
5.5.1 Disturbances with Changeable Armature Resistance...............................59
5.5.2Disturbances with Changeable Load Torque............................................61
CHAPTER 6 CONCLUSIONS ..............................................................................64
APPENDIX A LYAPUNOV STABILITY..................................................................66
APPENDIX B THE FEATURES OF APPLIANCES................................................. 69
APPENDIX C THE MATLAB PROGRAM CODES ................................................71
REFERENCES............................................................................................................... 73
viii
LIST OF FIGURES
FIGURE
2.1 General control system design steps………………………………………….........10

2.2 Controller building with adjustable controller gains ………………………............13
2.3 Indirect adaptive control........................................................................................... 14
2.4 Direct adaptive control. ............................................................................................ 15
2.5 Model reference control. .......................................................................................... 16
2.6 Indirect MRAC......................................................................................................... 17
2.7 Direct MRAC. .......................................................................................................... 18
3.1 The structure of the adaptive control system ........................................................... 19
4.1 Model of a separately excited PM DC motor........................................................... 28
4.2 State diagram of a dc motor ..................................................................................... 30
4.3 Block diagram of a dc motor system........................................................................ 31
4.4 Connection scheme for empirical method................................................................ 32
4.5 Velocity alteration of motor ..................................................................................... 34
4.6 Simulink scheme for getting measured input-output data........................................ 35
4.7 Parameter identification window related with third order estimation…...................37
4.8 Comparison of measured and simulated outputs related with third
order estimation ……………………………………………………………………37
4.9 Parameter identification window related with second order estimation................41
4.10 Comparison of measured and simulated outputs related with second order
estimation …………………………………………………………………….......41
4.11 Bode Diagrams of Transfer Functions…………………………….....……...........42
4.12 The principle MRAC scheme of a dc motor displacement……………..……..…43
4.13 Physical scheme of MRAC of a dc motor displacement........................................44
4.14 State transition signal flow graph of the motor model............................................45
5.1 Simulation results for α = 10 −4 , b ≅ c , T =1 ms ................................................... ..51
ix
5.2 Simulation results for α = 10 −3 , b ≅ c , T =1 ms .................................................... .51
5.3 Simulation results for α = 10 −2 , b ≅ c , T =1 ms .................................................... 52
5.4 Simulation results for α = 10 −2 , b > c , T =1 ms. ................................................. ..53
5.5 Simulation results for α = 10 −2 , b < c ,T =1 ms. ..................................................... 54
5.6 Simulation results for α = 10 −2 , b=0, T =1 ms........................................................ 55
5.7 Simulation results for α = 10 −2 , b ≅ c , T =10 ms.....................................................56
5.8 Simulation results for α = 10 −2 , b ≅ c , T =100 ms...................................................56
5.9 Simulation results with respect to faster and less stable model...............................59
5.10 Simulation results with respect to armature resistance variations...........................60
5.11 Simulation results with respect to abrupt armature resistance variations
(0 to 110 Ω and 110 to 0 Ω ).....................................................................................61
5.12 Simulation results with respect to torque disturbances............................................62
5.13 Simulation results with respect to abrupt torque disturbances.................................63

x
LIST OF SYMBOLS AND ABBREVIATIONS
SYMBOL
Δ Unmodeled plant phenomena

P (θ * ) Plant model which is parameterized with respect to some unknown
parameter vector θ *
θ* Coefficients of the numerator and denominator of the plant transfer
function
θ (t) Estimate of θ * at each time t
Pˆ (θ (t) ) Estimated plant model

θ C (t) Controller parameter
C (θC ) Control law

WM(s) Transfer function of the reference model
r Reference signal
u The input of plant
g Feed forward gain of reference signal
g* Optimal values of feed forward gain of reference signal
A Coefficient matrix of x
Am Coefficient matrix of z
b Coefficient matrix of u
bm Coefficient matrix of r
T
F= [F1 F2......Fn] Feedback gains of plant’s states
F* Optimal values of feedback gains of plant’s states
x Controllable canonical form of plant states
z Controllable canonical form of model’s states
e Error signal between model and plant states
α Adaptive gain
xi
sT Abbreviated form of bmT P

T Adaptation period or simulation step size
Tpi Time constants of plant
Tmi Time constants of reference model
Tc Computation time
f Adaptation frequency (inverse of T )
P Positive definite matrices for Lyapunov equation
Q Positive definite matrices for Lyapunov equation
ξ Damping ratio
wn Natural undamped frequency
K Numerator of plant’s transfer function
K* Numerator of reference model’s transfer function
aj Denominator coefficients of plant’s transfer function
*
aj Denominator coefficients of reference model’s transfer function
ΦF Difference between F and F *
Φg Difference between g * and g
V Lyapunov function
ia (t ) armature current
Ra armature resistance
eb (t ) back electromotive force (emf)
TL (t ) load torque
Tm (t ) motor torque
θ m (t ) rotor displacement
Ki torque constant
Kb back emf constant
La armature inductance
xii
ea (t ) applied voltage
Φ magnetic flux in the air gap
wm (t ) rotor angular velocity
J m (t ) rotor inertia
Bm viscous- friction coefficient
td the delay time
tr the rise time
ts the settling time
ABBREVIATION
DC Direct Current
DSP Digital Signal Processing
DMRAC Decentralized Model Reference Adaptive Control
FDD Fast Detection and Diagnosis
HFG High Frequency Gain
I/O Input/Output
LOS Line Of Sight
LTI Linear Time Invariant.
MRAC Model Reference Adaptive Control
MRC Model Reference Control
MIMO Multi Input Multi Ouput
PM Permanent Magnet
SISO Single Input Single Output
SOC Self Organizing Control
SPR Strictly Positive Real
CHAPTER I
INTRODUCTION
In spoken language, “to adapt” shows to change a behavior to become similar to new
circumstances. Instinctively, an adaptive controller is so a controller that can modify its
behavior in response to changes in the dynamics of the process and character of the
disturbances. Since ordinary feedback also tries to reduce the effects of disturbances and plant
uncertainty, the question of the distinction between feedback control and adaptive control
without delay arises. Through the years there have been many efforts to explain adaptive
control formally. At an early symposium in 1961 a long discussion ended with the following
suggestion: “An adaptive system is any physical system that has been designed with an adaptive viewpoint.”
A renewed try was made by an IEEE committee in 1973. It submitted a new vocabulary based
on concepts like self-organizing control (SOC), and learning control system. However, these
tries were not widely accepted. A meaningful definition of adaptive control, which would
make it possible to look at a controller hardware and software and decide whether or not it is
adaptive, is still lacking. However, there seems to be a general agreement that a constant –
gain feedback system is not an adaptive system (Aström, 1995).
Adaptive systems have two advantages according to non adaptive systems. First of all,
if plant parameters change during the operation according to the several conditions an
adaptive system adjusts itself, and the performance of the plant becomes as desired. In the
non adaptive systems, controller is designed using constant plant parameters. When the plant
parameters change, performance of the controller decreases.
Second advantage of adaptive systems is the information of plant’s parameters is not

required for a controller design. Control can be made partly or without any plant parameters.
1
2
Plant parameters must be known in the non adaptive systems. For this reason, research of the
plant parameters brings us extra difficulties.
Generally, the aim of adaptive systems is to control the plant with unknown parameters.
There are two techniques for this purpose direct and indirect adaptive control respectively.
The plant parameters are estimated on-line in indirect adaptive control. Direct adaptive
control doesn’t need on line parameter estimation. Direct adaptive control is considered in this
thesis.
This method requires knowing plant’s zeros. Therefore, we consider plants without
finite transmission zeros. For this reason, the method is proper for position control of dc
motor. In this thesis the position control of a dc motor is considered by using model reference
adaptive control (MRAC). Adaptation mechanism which adjusts recursively plant’s feed-
forward and feedback gains, tries that equalize the coefficients of closed loop plant to model’s
coefficients. Hence if we choose proper model, plant output converges to the model output
with time.
The parameters in a dc motor are armature inductance, armature resistance, rotor inertia,
friction coefficient, motor moment, load moment, armature voltage, motor speed, speed
coefficient and moment coefficient. These parameters change with external effects and
working conditions. These alterations effect motor dynamic. We will investigate the
performance of adaptive tracking in spite of parameters variation.
The solution of problem against to the parameters variations is investigated by many

scientists. These are summarized below.
(Zeng at all, 1999) controls of flexible spacecraft with using output feedback and
variable structure model reference adaptive control theory. For the derivation of control law, it
is assumed that the parameters and the structure of the nonlinear functions in the model are
unknown. It is shown that in the closed-loop system including the variable structure model
reference adaptive control system designed using bounds on uncertain functions, the pitch
angle tracks given reference trajectory and the vibration is suppressed.
The speed which is needed for motor driver of induction motor is estimated from motor
current without using sensor. The parallel MRAC is used for speed estimation. The error
signal between reference model output and adjustable system output is driven to zero through
3
an adaptive law. The speed of motor is estimated very fast with using large adaptive gains
(Kojabadi, 2005), (Park and Kwon, 2004).
(Zhou and Wang, 2005) control of a permanent magnet synchronous motor with using
techniques of MRAC and back stepping control. In the controller design, the input output
feedback linearization is first of all used to compensate the nonlinearities in the nominal
system. Then, adaptive back stepping control approach is adopted in order to derive the
control scheme, which is strong to the parameter uncertainties and load torque disturbance.
(Marino at all, 1998) present an adaptive nonlinear control algorithm for current fed
induction motors which is adaptive with regards to both load torque and rotor resistance. The
eighth order adaptive controller supplies reference signals for stator currents on the basis of:
measurements of rotor speed, stator currents and stator voltages; estimates of rotor fluxes,
which are the unmeasured state variables; estimates of torque load and rotor resistance which
may vary significantly during operations. The dynamic controller assures speed tracking and
bounded signals for every initial condition of the motor. When persistency of excitation
conditions are satisfied, the rotor flux tracking error goes asymptotically to zero so that motor
power efficiency may be enhanced. In addition, in this case, the estimates of rotor fluxes,
torque load and rotor resistance tend asymptotically to their true values. Results show that
persistency of excitation conditions are satisfied in physical operating conditions and that all
estimation errors go quickly to zero so that high tracking performances are obtained both for
speed and rotor flux.
(Lee at all, 2000) present and compare possible intelligent control designs for precision
motion control applications which are established upon the use of linear actuators. The control
of linear motor is realized three different ways which include adaptive control, composite
control using a radial-basis function for nonlinear compensation and an iterative learning
control. Experimental results show that first two are more successful than last technique.
(Eldeeb and Elmaraghy, 1998) present a new optimal controller designed for rigid-body
robots containing motor dynamics. The new optimal adaptive controller developed in this
work is established upon feedback linearization, it does not need acceleration feedback, and it
does not assume full state is available for measurement but it needs an observer. Of course, it
does not assume exact knowledge of either robot or actuator parameters. The optimality is on
4
the basis of the minimization of a performance index which turns out to be possible if we
could find a solution to the Hamilton Jacobi equation.
(Aiko and Kimura, 2002) aim to establish control theoretical validity of the feedback
error learning scheme suggested as an architecture of brain motor control with deep
physiological root in computational neuroscience. The feedback error learning method is
formulated as a two-degree of freedom adaptive control. The stability of the adaptive control
law is shown clearly based on the strict positive realness, under the supposition that the plant
is stable and stably invertible. Results prove the effectiveness of the method.
(McLain and Henson, 2000) present a nonlinear adaptive control strategy established upon
radial basis function networks and principal component analysis. The suggested method is well
suited for low dimensional nonlinear systems that are difficult to model and control via
conventional means. The effective system dimension is decreased by applying nonlinear principal
component analysis to state variable data obtained from open-loop tests. This permits the radial
basis functions to be placed in a lower dimensional space than the original state space. The total
number of basis functions is clearly described a priori, and an algorithm which adjusts the place of
the basis function centers to encompass the current operating point is presented. The basis
function weights are adapted on-line such that the plant output asymptotically follows a linear
reference model. A highly nonlinear polymerization reactor is used to compare the nonlinear
adaptive controller to a linear state feedback controller that takes advantage of the same amount
of plant information.
(Makoudi and Radouane 2000) present a distributed model reference adaptive control
for interconnected subsystems in the sense that no information exchange occurs between the
subsystems. The approach is established upon the interconnection output estimation using the
polynomial series which suggests a general solution for interconnected subsystems. The
parameter estimation scheme is an integrated adaptive data filtering with a recursive least-
squares algorithm with parameter projection and normalization. The problem of minimum
phased subsystems is handled by an adaptive input output data filtering. Hence the zeros of
each subsystem estimated model are replaced inside the unit circle. This estimated model
which is minimum phased is then used for the control synthesis. It is shown that the stability
conditions established upon weak interconnections are relaxed. Also the robustness of the
5
suggested adaptive control against unmodeled dynamics is expressed. At last, the results are
illustrated by numerical examples.
(Tsai and Lin, 1997) present a model reference adaptive control approximation, it is
formed in the modal space; it is applied for flutter control of a cantilever pipe conveying fluid.
The control input is supplied by a pair of surface mounted piezoelectric actuators which are
driven 180 o out of phase to provide an equivalent bending moment acting on the controlled
system. Comparison of performance of the model reference adaptive control with that of the
optimal independent modal space control shows that the former is more robust than the latter
in terms of flow speed variations, which are unknown in the control system designed ; that is,
the adaptive approach can compensate a larger range of flow speed uncertainties without
resulting in an unstable control system, hence successful flutter suppression of the fluid
conveying cantilever pipe with high flow speed can be performed.
(Lee at all, 1998 ) present that an adaptive neural network full state feedback controller
has been designed and applied to the passive line of sight (LOS) stabilization system. Model
reference adaptive control (MRAC) is well founded for linear systems. However, this method
cannot be utilized directly because the LOS system is nonlinear in nature. Utilizing the
universal approximation property of neural networks, an adaptive neural network controller is
presented by generalizing the model reference adaptive control technique, in which the gains
of the controller are approached by neural networks. This ejects the requirement of linearizing
the dynamics of the system, and the stability properties of the closed loop system can be
satisfied.
(Taware at all, 2003) present that friction correction for a benchmark system with load
friction plus joint flexibility and damping is addressed. This is a difficulty of controlling a
sandwich dynamic system with a non-smooth nonlinearity. Few non-adaptive and adaptive
compensation designs are analyzed, established upon a state feedback output tracking model
reference adaptive control scheme. Adequate output matching conditions are derived for
friction compensation. Approximate linear parameterizations of nonlinear friction are built for
adaptive friction compensator designs. Simulation results confirm the desired system
performance.
(Zhong, 2005) presents that model reference adaptive control problem for single input
single output time invariant continuous time plants with input saturation is taking into account
6
with main attention focused on global properties. A sufficient condition is presented and a
new design method of adaptive control systems is suggested. If a priori knowledge about the
plant is available to choose the reference model and the reference input so that the sufficient
condition holds, the closed loop adaptive control system designed by the suggested method
can have global stability and globally output tracking property. It is shown that the sufficient
condition is necessary in some cases.
(Chien and Yao , 2004) present that a model reference adaptive control strategy is used
to design an iterative learning controller for a class of repeatable nonlinear systems with
uncertain parameters, high relative degree, initial output resetting error, input disturbance and
output noise. The class of nonlinear systems should gratify some differential geometric
conditions such that the plant can be transformed via a state transformation into an output
feedback canonical form. An appropriate error model is derived based on signals filtered from
plant input and output. The learning controller compensates for the unknown parameters,
uncertainties and nonlinearity by means of projection type adaptation laws which update
control parameters along the iteration domain. It is shown that the internal signals stay
bounded for all iterations. The output tracking error will converge to a profile which can be
adjusted by design parameters and the learning speed is increased if the learning gain is large.
(Costa at all, 2003) present that the design of Model Reference Adaptive Control for
Multi Input Multi Output (MIMO) linear systems has not yet achieved, despite significant
efforts, the completeness and simplicity of its Single Input Single Output (SISO) counterpart.
One of the main obstructions has been the generalization of the SISO assumption that the sign
of the high frequency gain (HFG) is known. Here they overcome this obstacle and present a
more complete MIMO analog to the renowned Lyapunov based SISO design which is
significantly less restrictive than the existing analogs. Their algorithm makes use of a new
control parameterization derived from a factorization of the HFG matrix Kp = SDU, where S
is positive symmetric definite, D is diagonal, and U is unity upper triangular. Only the signs
of the entries of D or, equally, the signs of the leading principal minors of Kp, are assumed to
be known.
(Krstic and Banaszuk, 2005) consider a class of Multi Input Multi Output (MIMO) LTI
models with uncertain resonant modes and time delays, which are common in control of
instabilities arising in jet engines. With uncertain delays preventing the use of model
reference adaptive control, they develop an adaptive MIMO pole placement scheme for the
7
system. They use indirect adaptation, estimating a small number of physical parameters from
a nonlinearly parameterized plant. To address the highly noisy environment in jet engines
they introduce the dead zone in the adaptation law and present simulations that successfully
stabilize the system in the presence of noise and serious actuator saturation.
(WANG at all, 1997) present novel approach for the fault detection and diagnosis
(FDD) of faults in actuators and sensors by way of the use of adaptive updating rules. The
system considered is linear time invariant and is subjected to an unknown input that
represents either model uncertainty or immeasurable disturbances. Firstly, fault detection and
diagnosis for linear actuators and sensors are considered, where a fixed observer is used to
detect the fault while an adaptive diagnostic observer is built to diagnose the fault. Utilizing
the augmented error technique from mode1 reference adaptive control, an observation error
model is formulated and used to establish an adaptive diagnostic algorithm that produces an
estimate of the gains of actuator and the sensor. An extension to the fault detection and
diagnosis to include nonlinear actuators is also made, where a similar augmented error model
to that used for linear actuators and sensors is acquired. As a result, a convergent adaptive
diagnostic algorithm for estimating the parameters in the nonlinear actuators is improved.
(Sinha and Pechev, 1999) present a model reference adaptive controller (MRAC) for
magnetically suspended vehicles (maglev) using the criterion of stable maximum descent. The
adaptation algorithm is forced to reduce the air gap error between the reference model and the
actual system. The explicit relationship between the parameters of the performance criterion
(function of the air gap error and its derivative) and the state feedback adaptation rule is
produced for a single degree of freedom suspension system. Experimental results from a small
representative test rig are presented to illustrate the efficiency of the suggested non linear
controller in the presence of variations in pay load (suspended mass), disturbance force and
air gap set point. Hardware aspects of the transporter and Digital Signal Processing (DSP)
based real time controller are briefly discussed to emphasize some of the practical issues
related to digital execution of the air gap adaptive control law.
(Makoudi and Radouane, 1999) present a decentralized model reference adaptive
control (DMRAC) for interconnected subsystems with unknown or time-varying time delay.
The decentralization approximation is established upon the interconnection output estimation
using the polynomial series which suggests a general solution for interconnected subsystems.
The parameter estimation scheme is a combined adaptive data filtering with a recursive least
squares algorithm with parameter projection and signal normalization. A “good data” model
8
is determined by an adaptive filtering of the input and output signals. The acquired model
allows dealing with non-minimum phased subsystems with unknown or time-varying dead
time and at the same time to relax the hypothesis of weak interconnections for decentralized
control.
(Tian and Hoo, 2003) present that transition control is determined as a type of control
method that is operated when the plant transitions from one steady state to another as a result
of a set point change. Recent approaches have depended on multiple models and centralized
or decentralized controller designs to address this issue. This work presents and improves a
transition control framework that consists of multiple fixed and adaptive models within a
state-shared non minimal realization and an H ∞ controller design. The effectiveness of this
transition control framework is presented on two nonlinear single-input single-output reactors
in the face of modeling errors, parameter uncertainties and disturbances.
(Chaoa and Neoub, 2000) present that model reference adaptive control of air-lubricated
capstan drive for precision positioning. Because friction-induced nonlinearities in positioning
systems are mostly range of motion-dependent, dual-model or dual-stage strategies are
frequently adopted to deal with the incompatibility encountered when a system moves from
submicrometer steps (micro mode) to larger scale strokes (macro mode). Despite the fact that
good performance is usually acquired when each stage functions in its designed range of
motion, a system often performs less sufficiently when operating near the switching point
between models or stages. An air-lubricated capstan drive was used in this work to minimize
the inconsistency between macro and micro modes, and a single mode MRAC was designed
to control the capstan drive system for precision positioning. Accuracy better than 615 nm
with no overshooting was obtained in all conditions tested (including 50 nm, 500 nm and 10
mm steps).
(Mirkin and Gutman, 2005) present that two new output feedback adaptive control
schemes established upon Model Reference Adaptive Control (MRAC) and adaptive laws for
updating the controller parameters are improved for a class of linear MIMO systems with
state delay. An effective controller structure established upon a new error equation
parameterization is suggested to achieve tracking with the error tending to zero
asymptotically. To accomplish exact asymptotical tracking, they introduce, in the standard
MRAC structure for plants without delay, a new supplemental adaptive feedforward control
component as an output of a dynamical system driven by the reference signal. Adaptive laws
are improved using the SPR-Lyapunov design approximation and two suppositions regarding
the previous knowledge of the high frequency matrix Kp. This study is the first asymptotic
9
exact zero tracking results for this class of systems in the framework of the certainty
equivalence approximation.
In this thesis the standard model reference adaptive control is used to control the
position of a permanent magnet dc motor, some of its parameters can not be measured
directly. Further, some of the other parameters are not constant and vary as the motor
operates. A second order system is used as the reference model.
This thesis is organized as follows: Fundamentals of MRAC is summarized in Chapter

2. Model reference adaptive control for systems without finite transmission zeros is
introduced in Chapter 3. MRAC for dc motor is given in Chapter 4. Simulation results are
presented in Chapter 5. Finally conclusions are made in Chapter 6.
CHAPTER II
2.1 GENERAL CONTROL SYSTEM DESIGN STEPS
The steps in a general control design problem are shown in Figure 2.1. Each step is
explained below (Ioannou, 1994).
Figure 2.1 General control system design steps.
10
11
Step 1. Modeling
A plant model may be built by using physical laws or by processing the plant input
output (I/O) data obtained by performing various experiments. But, this model may be
complex for the controller design and additional simplifications may be necessary. Some of
the approaches in many cases used to obtain a simplified model are
(i) Linearization around functioning points,
(ii) Model order decreasing techniques.
In approximation (i) the plant is approximated by a linear model that is acceptable

around a given operating point. Different functioning points may lead to some different linear
models that are used as plant models. Linearization is performed by using Taylor's series
expansion and approximation, fitting of empirical data to a linear model, etc.
In approach (ii) small impacts and phenomena outside the frequency range of interest
are disregarded leading to a lower order and uncomplicated plant model.
Step 2. Controller Design
The controller is planned to meet the performance necessities for the plant model. Δ
symbolizes most of the unmodeled plant phenomena. The control engineer may be able to
modify or redesign the controller to be more robust with respect to Δ . This robustness
analysis and redesign increases the ability for a successful execution in Step 3.
Step 3. Implementation
The implementation can be done using a digital computer. The type of computer, the
type of interface devices between the computer and the plant, software tools are considered
priority matters. Computer speed and preciseness limitations may constrain on the complexity
of the controller. It may force the control engineer to go back to Step 2 or even Step 1. Other
important aspect of implementation is the final adjustment, or as often called the tuning which
is often done by trial and error and depends very much on the experience and intuition of the
control engineer.
12
2.2 ADAPTIVE CONTROL
The words “adaptive systems" and “adaptive control" have been employed as early as
1950 (Ioannou, 1994). The design of autopilots for high-performance airplane was one of the
first incentives for active investigation on adaptive control in the early 1950s. Airplane
function over a broad range of speeds and heights, and its dynamics are nonlinear and
conceptually time varying. For a given functioning point, described by the airplane speed and
height, the complex airplane dynamics can be approached by a linear model of the similar
shape as (2.1).
x& = Ax + Bu ; x( 0 ) = x0 , y = C T x + Du . (2.1)
For instance, for a functioning point i , the linear airplane model has the following shape:
T
x& = Ai x + Bi u ; x(0 ) = x0 , y = C i x + Di u (2.2)
where x , u , y are the state, input, output vectors, respectively; and Ai , Bi ,C i and Di are
coefficient matrix functions of the functioning point i. As the airplane moves through
different flight situations, the functioning point changes goes to different values
for Ai , Bi , C i and Di . Since the output response y (t) transports information about the state x
as well as the parameters, one may dispute that in principle, a complicated feedback controller
should be able to learn about parameter alterations by processing y (t) and utilize the suitable
gains to adapt them. This quarrel goes to a feedback control structure on which adaptive
control is established. The controller building comprise of a feedback loop and a controller
with adjustable gains as shown in Figure 2.2.
13
Figure 2.2 Controller building with adjustable controller gains.
2.2.1 Direct and Indirect Adaptive Control

An adaptive controller is shaped by combining an on-line parameter estimator, which
gives estimates of unknown parameters at each moment, with a control law that is prompted
from the known parameter case. The method of the parameter estimator, also referred to as
adaptive law, is joined with the control law lead to two different approximations. In the first
approximation, mentioned as indirect adaptive control, the plant parameters are estimated on-
line and used to compute the controller parameters. This approximation has also been
mentioned as explicit adaptive control, due to the fact that the design is on the foundation of
an explicit plant model.
In the second approximation, mentioned as direct adaptive control, the plant model is
parameterized regarding the controller parameters that are estimated directly without
intermediate computations involving plant parameter estimates. This approximation has also
been mentioned as implicit adaptive control due to the fact that the design is established upon
the estimation of an implicit plant model.
In indirect adaptive control, the plant model P( θ * ) is parameterized with regards to
some unknown parameter vector θ * . For instance, for a linear time invariant (LTI) SISO
plant model, θ * may symbolize the unknown coefficients of the numerator and denominator
of the plant model transfer function. An on-line parameter estimator generates an estimate
θ (t) of θ * at each time t by working the plant input u and output y. The parameter estimate
θ (t) clearly describes an estimated plant model characterized by P̂(θ(t) ) that for control
design aims is treated as the “true" plant model and is used to calculate the controller
parameter θ C (t) by solving a specific algebraic equation θ C (t) = F (θ (t) ) at every time t. The
14
shape of the control law C (θ C ) and algebraic equation θ C = F (θ ) is selected to be identical to
( ) and equation θ
that of the control law C θ C
*
C
*
= F (θ * ) that could be used to encounter the
performance necessities for the plant model P (θ * ) if θ * was known. It is, hence, clear that
with this approximation, C (θ C (t) ) is designed at every time t to gratify the performance
necessities for the estimated plant model Pˆ (θ (t) ) , which may not be the same from the
unknown plant model P (θ * ). And so, the head problem in indirect adaptive control is to select
the class of control laws C( θ C ) and the class of parameter estimators that produce θ (t) and
likewise the algebraic equation θ C (t) = F(θ (t )) so that C (θ C (t) ) encounters the
performance necessities for the plant model P (θ * ) with unknown θ * . The block diagram of
an indirect adaptive control scheme is seen in Figure 2.3.
Figure 2.3 Indirect adaptive control.
In direct adaptive control, the plant model P (θ * ) is parameterized in the sense of the
unknown controller parameter vector θ C , for which C θ C

*
( ) *
encounters the performance
necessities, to acquire the plant model PC (θ C* ) with precisely the similar input output
characteristics as P( θ * ). The on-line parameter estimator is designed based on PC (θ C* ) in
place of P ( θ * ) to supply direct estimates θ C (t) of θ C at every time t by working the plant
*
input u and output y. The estimate θ C (t) is then used to update the controller parameter
vector θ C without intermediate computations. The selection of the class of control laws
C (θ C ) and parameter estimators producing θ C (t) for which C (θ C (t) ) encounters the
15
performance necessities for the plant model P( θ * ) is the basic problem in direct adaptive
control. The characteristics of the plant model P (θ * ) are very important in acquiring the
parameterized plant model PC (θ C* ) that is useful for on-line estimation. Due to that, direct
adaptive control is constrained to a specific class of plant models. A class of plant models that
is appropriate for direct adaptive control comprise of all SISO plant models that are
minimum-phase, i.e., their zeros are placed in Re [s] < 0. The block diagram of direct
adaptive control is seen in Figure 2.4.
Figure 2.4 Direct adaptive control.
The foundation behind the design of direct and indirect adaptive control seen in
Figures 2.3 and 2.4 is conceptually simple. The design of C (θ C ) acts the estimates θ C (t )
(regarding direct adaptive control) or the estimates θ (t ) (regarding adaptive control) as if they
were the true parameters. This design approximation is named certainty equivalence and can
be used to produce a broad class of adaptive control schemes by joining different on-line
parameter estimators with different control laws.
The idea in back of the certainty equivalence approximation is that as the parameter
estimates, θ C (t ) and θ (t ) converge to the true ones θ C and θ * , respectively, the performance
*
of the adaptive controller C (θ C ) be inclined to that accomplished by C (θ C* ) concerning

known parameters.
16
2.2.2 Model Reference Adaptive Control
Model reference adaptive control (MRAC) comes from the model following problem or
model reference control (MRC) problem. In MRC, a good comprehension of the plant and the
performance necessities it has to encounter permit the designer to invent a model, mentioned
as the reference model, that depicts the desired I/O characteristics of the closed-loop plant.
The purpose of MRC is to discover the feedback control law that alters the structure and
dynamics of the plant so that its I/O characteristics are precisely the same as those of the
reference model. The building of an MRC diagram for a LTI, SISO plant is shown in Figure
2.5. The transfer function W M (s ) of the reference model is designed so that for a given
reference input signal r (t ) the output y m (t ) of the reference model depicts the wanted
response the plant output y (t ) must follow.
Figure 2.5 Model reference control.
( )
The feedback controller symbolized by C θ C* is designed so that all signals are delimited and
the closed-loop plant transfer function from r to y is equal to W M (s ) . This transfer function
matching assures that for any given reference input r (t ) , the tracking error e1 ≅ y − y m , which
symbolizes the divergence of the plant output from the desired trajectory y m , approaches to
zero with time. The transfer function matching is accomplished by canceling the poles of the
plant transfer function G (s ) and substituting them with those of W M (s ) through the use of the
feedback controller ( )
C θ C* . The cancellation of the plant poles brings a limitation on the
plant to be minimum phase, in other words, have stable poles. When the plant poles are not
stable, its cancellation may easily go to uncontrolled signals.
17
( )
The design of C θ C* necessitates the learning of the coefficients of the plant transfer
function G (s) . If θ * is a vector including all the coefficients of G ( s ) = G ( s ; θ * ) , then the
parameter vector θ C may be calculated by solving an algebraic equation of the shape

*
θ C* = F ( θ * ) . (2.3)
Finally for the realization of MRC, the plant model has to be minimum phase and its
parameter vector θ * has to be known precisely.
When θ * is not known the MRC scheme of Figure 2.5 can not be executed because
θ C* can not be computed applying Eq. 2.3. One method of referring to the unknown parameter
matter is to utilize the certainty equivalence approximation to substitute the unknown θ C* in
the control law with its estimate θ C (t ) acquired using the direct or the indirect approach. The
deriving control schemes are accepted as MRAC and can be classified as indirect MRAC seen
in Figure 2.6 and direct MRAC seen in Figure 2.7.
Figure 2.6 Indirect MRAC.

18
Figure 2.7 Direct MRAC.
In this thesis, direct MRAC approach is used for the control of a permanent magnet dc
motor which does not have finite transmission zeros and the method for this is discussed in
the following chapter.
CHAPTER III
MRAC FOR SYSTEMS WITHOUT

FINITE TRANSMISSION ZEROS
The aim of this method realizes model reference adaptive control for the plant that has
single input single output and without finite transmission zeros. Adaptation mechanism
adjusts feedback gains of states and the feed-forward gain of reference signal. The
coefficients which are in the transfer function of closed loop plant are equalized to model’s.
3.1 THE STRUCTURE OF THE ADAPTIVE CONTROL SYSTEM
The structure of the adaptive control system is given Figure 3.1 (Karadeniz et al, 2004).
In this figure r is reference signal , u is plant’s input , g is feed forward gain , F=[F1
T
F2......Fn] is feedback gains of plant’s states, x and z are controllable canonical form of
plant and model’s states respectively, e is error signal between model and plant states. Plant
follow model via F (t ) and g (t ) which are updated every time step.
Figure 3.1 The structure of the adaptive control system.
19
20
The transfer function of one input one output continuous time plant without finite
transmission zeros is
K
n −1
. (3.1)
s + an s
n
+ .... + a 2 s + a1
The controllable canonical form of (3.1) is
⎡ x1 ⎤ ⎡ 0 1 0 K 0 ⎤ ⎡x ⎤
⎡0 ⎤
⎢ ⎥ ⎢ 0 0 1 K 0 ⎥⎥ ⎢ x ⎥ ⎢ ⎥
1
d ⎢ x2 ⎥ ⎢ ⎢ 2⎥ 0
=⎢ K K K K K ⎥ ⎢ ⎥ + ⎢ ⎥ u (t ) , (3.2 -a)
dt ⎢⎢.....⎥ ⎢ ⎥ ⎢K⎥
⎥ 0 0 0 K 1 ⎥ ⎢.....⎥ ⎢ ⎥
⎢x ⎥ ⎢ ⎢ ⎥ K
⎣ n ⎦ ⎢⎣− a1 − a 2 − a 3 K − a n ⎥⎦ ⎣ x n ⎦ ⎣ ⎦
y = [1 0 K 0 ] x , (3.2 -b)
which is in the form
x& = Ax + bu , (3.3 -a)

y = x1 . (3.3 -b)
As shown in this figure, the following feedback control law used is
u = gr − F T x . (3.4)
We can show the controlled plant’s state space equation is
x& = A x + b (g r − F T x ) = (A − bF T )x + bgr , (3.5-a)
Or most explicitly,
⎡ x1 ⎤ ⎡ 0 1 0 K 0 ⎤ ⎡ x1 ⎤ ⎡0 ⎤
⎢x ⎥ ⎢ ⎥ ⎢x ⎥ ⎢ ⎥
0 0 1 K 0 ⎢ 2 ⎥ ⎢0 ⎥
d ⎢ 2⎥ ⎢ ⎥
= ⎢ ⎥ ⎢ ⎥ + ⎢K ⎥ r (t ) .
dt ⎢.....⎥ ⎢
K K K K K
⎢ ⎥ ⎥ ⎢.....⎥ ⎢ ⎥
0 0 0 K 1 ⎥
⎢x ⎥ ⎢ ⎢ x ⎥ ⎣ Kg ⎦
⎣ n ⎦ ⎢⎣− a1 − KF1 − a 2 − KF1 − a3 − KF1 K − a n − KF1 ⎥⎦ ⎣ n⎦
(3.5-b)
The reference model transfer function and state space equation are
K*
* * * , (3.6-a)
s n + a n s n −1 + .... + a2 s + a1
21
⎡ ⎤
⎡ z1 ⎤ ⎢ 0 1 0 K 0
⎥ ⎡ z1 ⎤ ⎡ 0 ⎤
⎢ ⎥ ⎢ 0 0 1 K 0 ⎥ ⎢z ⎥ ⎢ ⎥
d ⎢z2 ⎥ ⎢ ⎢ 2 ⎥ ⎢0 ⎥
⎢ ⎥ + ⎢K ⎥ u (t )
= K K K K K ⎥ (3.6-b)
dt ⎢⎢.....⎥ ⎢ ⎥
⎢.....⎥ ⎢ ⎥
⎥ ⎢ 0 0 0 K 1 ⎥
⎢z ⎥ ⎢ ⎥ ⎢ z ⎥ ⎣⎢ K * ⎥⎦
⎣ n ⎦ −a * *
− a 2 − a3 K − a n
* * ⎣ n⎦
⎣ 1 ⎦
which is in the form

z& = Am z + bm r . (3.6-c)
Assume the model output
y m = z1 . (3.6-d)
3.2 DERIVATION OF THE ADAPTIVE LAWS
Similar adaptive laws used in (Karadeniz et al, 2004) are followed in the sequel. From
Eqs. 3.5-a and 3.6-c, for the equivalence of the controlled plant and the reference model
Am = A − b F * T , (3.7-a)
bm = b g * , (3.7-b)
or explicitly
*
aj − aj
Fj = Fj =
*
, ( j = 1,2,...., n ) , (3.8-a)
K
K*
g = g* = . (3.8-b)
K
*
Normally, a j and K are not known. Resultantly, Fj and g * are not known.
*
Therefore F j and g may be different from F j and g * . The purpose of the adaptive law is to
*
converge the present values of F j and g to F j and g*, respectively, i.e,
lim F (t ) ⎯⎯→ F , (3.9-a)

*
t ⎯⎯→ ∞
lim g (t ) ⎯⎯→ g
*
. (3.9-b)
t⎯
⎯→ ∞
Define the error by

e = z − x. (3.10-a)
Then, derivative of the error is
22
e& = z& − x& . (3.10-b)

We manipulate the plant and model equations in (3.4), (3.5) and (3.6), respectively and obtain
e& = ( Am z + bm r ) − ( A x + b u )
(
= ( Am z + bm r ) − A x + b rg − F T x ( ))
= Am z + bm r − A x − brg + bF T x
(
= Am z + bm r − Am + bF *T x − brg + bF T x )
(
= Am (z − x ) + bm r + b F T − F *T x − brg )
= Am e +
1
( ) b rg
b F T − F * T x + bm r − m *
* m
g g
= Am e +
1
( )1
b F T − F * T x + bm r * g * − g .
* m
( ) (3.11)
g g
Define:
ΦF = F − F * , (3.12-a)
Φ g = g* − g . (3.12-b)
Then,
Φ& F = F& , (3.12-c)
Φ& g = − g& . (3.12-d)
We renew the derivative of the error equation in (3.11),

1 1
e& = Am e + b Φ FT x + * rΦ g bm ,
* m
g g
(3.13)
1 1
= Am e + * bm x T Φ F + * rΦ g bm .
g g
1
We suggest Lyapunov’s method for updating of F and g. For this we define the Lyapunov
function
1 1
V = eT P e + Φ FT Φ F + Φ g2 , (3.14)
αg *
αg *
where P is strictly positive real matrix to be found, α is a positive real coefficient.
1
Extra knowledge with relating to Lyapunov Theorem is placed in Appendix A.
23
Assuming g* > 0, Lyapunov function satisfies V ≥ 0 .
Derivative of the Lyapunov’s function can be found from Eq. 3.14 and using Eq. 3.13
V& =
d T
(
e Pe + )
1
Φ FT Φ& F +
1 &T
ΦF ΦF +
2 &
ΦgΦg . (3.15-a)
dt αg *
αg *
α g*
Where, first term of Eq.3.15-a can be rewritten as shown in Eq.3.15-b
d T
dt
( )
e P e = e&T P e + eT P e&
T
⎛ 1 1 ⎞ ⎛ 1 1 ⎞
= ⎜⎜ Ame + * bm xT ΦF + * rΦg bm ⎟⎟ P e + eT P ⎜⎜ Ame + * bm xT ΦF + * rΦg bm ⎟⎟
⎝ g g ⎠ ⎝ g g ⎠
⎛ 1 1 ⎞ ⎛ 1 1 ⎞
= ⎜⎜eT AmT + * ΦFT x bmT + * rΦg bmT ⎟⎟P e + eT P ⎜⎜ Ame + * bm xT ΦF + * rΦg bm ⎟⎟
⎝ g g ⎠ ⎝ g g ⎠
1 1 1 1
= eT AmT Pe + * ΦFT x bmT P e + * rΦg bmT P e + eT P Ame + * eT P bm xT ΦF + * rΦg eT P bm
g g g g
( ) 1 1 1 1
= eT AmT P + PAm e + * ΦFT x bmT P e + * eT P bm xT ΦF + * rΦg bmT P e + * rΦg eT P bm
g g g g
3.15 − b
Eq. 3.15-b is replaced in to Eq. 3.15-a. So that, the derivative of Lyapunov’s function is
acquired as
( ) 1 1
V& = eT AmT P + PAm e + * ΦFT x bmT P e + * eT P bm xT ΦF
g g
1 1 1 & + 1 Φ & T Φ + 2 Φ& Φ .
+ * rΦg bmT P e + * rΦg eT P bm + * ΦFT Φ F F F g g (3.15−c)
g g αg α g* α g*
The matrix of the first quadratic form in Eq. 3.15-c is chosen as
AmT P + PAm = −Q (3.16)
which is known as Lyapunov Equation where, the eigenvalues of Am are negative for a stable
reference model and the matrix Q is chosen as positive definite symmetric matrix. So, the
solution of Lyapunov Equation gives us positive definite symmetric matrix P . Thus the first
term of Eq. 3.15-c e T (AmT P + PAm )e is made negative definite. If, the remaining terms of Eq.
24
3.15-c cancel each other, V& < 0 and stability is acquired. Therefore, the following equations
must be realized for cancellation :
Φ& FT ΦF = −α eT P bm xT ΦF , (3.17-a)
Φ FT Φ& F = − α Φ FT x b mT P e , (3.17-b)
2Φ& gΦ g = −α rΦ g bmT P e − α rΦ g e T P bm . (3.17-c)
Eqs. (3.17-a) and (3.17-b) are equivalent. To satisfy these equations, the parameter Φ& F can
be chosen
Φ& F = −α x (t ) bmT P e (t ) . (3.18-a) To
satisfy Eq. (3.17-c), the parameter Φ& g can be set
Φ& g = −α r (t )bmT P e(t ) . (3.18-b)
The adaptive laws to update F and g are finally obtained from Eqs. 3.18-a, b together with
Eqs. 3.12 c, d as follows
F& = −α x (t ) bmT P e (t ) , (3.18-c)
g& = α r (t )bmT P e (t ) . (3.18-d)

Hence Eqs 3.18-c and 3.18-d are the adaptive laws which are chosen. So, with these choices,
the derivative of Lyapunov function in Eq. 3.15 c is negative definite, that is
V& = −e T Q e < 0 (3.19)
for the nonzero error vector e.
This implies
lim V (t ) = 0 . (3.19-a)
t→∞
From Eq. 3.14, Eq. 3.19-a implies

lim Φ F = 0 , (3.19-b)
t →∞
lim Φ g = 0 . (3.19-c)
t→ ∞
Consequently, Eq. 3.12 a and b imply
lim F → F * , (3.19-d)
t→∞
lim g → g *
, (3.19-e)
t→ ∞
respectively. With these equations, Eq. 3.11 becomes

e& = Am e (3.19-f)
25
Therefore,
lim e → 0 , (3.19-h)
t →∞
since Am is a stable matrix. Finally, Eq. 3.19-h leads to
lim x (t ) → z (t ) (3.19-i)
t→ ∞
since e = z − x . Thus the adaptive laws in Eqs. 3.18-c and 3.18-d approach the plant states to
the model states. These are the desired requirements for the MRAC. Hence, the controlled
closed loop plant is globally asymptotically stable. Even if, the plant is unstable, it can be
controlled by at least the same order stable reference model.
In the adaptive laws, α > 0 is an arbitrary design constant referred to as the adaptive
gain. The use of a larger value of the adaptive gain α led to a faster convergence of plant
output to model’s. The adaptive gain α is usually chosen by trial and error using simulations
in order to achieve a good rate of convergence. Small α may result in slow convergence rate
whereas large α may make the differential equations stiff and difficult to solve numerically
on a digital computer (Ioannou, 1994).
For digital computations, we can convert Eq. (3.18-c) and (3.18-d) to (3.20-a) and
(3.20-b), respectively:
F ( k + 1 ) = F ( k ) − α x( k )bmT Pe( k ) , (3.20-a)
g( k + 1 ) = g( k ) + α r( k )bmT Pe( k ) , (3. 20-b)

which are discrete time versions and k denotes kT, k=0,1,..2 and T is the sampling period,
which will be called simulation step size or adaptation period in the sequel.
There are two important matters with respect to choosing matrix P. These issues are
explained with related to the second order plant and reference model. The matrices used for
the changes of F and g are explicitly written as
⎡x ⎤ ⎡0 ⎤ ⎡a b⎤ ⎡ e1 ⎤
x = ⎢ 1 ⎥ , bm = ⎢ ⎥ , P = ⎢ ⎥ , e=⎢ ⎥ . (3.21)
⎣ x2 ⎦ ⎣1⎦ ⎣b c ⎦ ⎣e 2 ⎦
i. With these values, the adaptive laws in Eqs. 3.20-a, b are written as
⎡ F1 ( k + 1 )⎤ ⎡ F1 ( k )⎤ ⎡ x1 ( k ) [b e1 ( k ) + c e2 ( k )]⎤
=
⎢F ( k + 1 )⎥ ⎢F ( k )⎥ − α ⎢ x ( k ) [b e ( k ) + c e ( k )]⎥ , (3.22-a)
⎣ 2 ⎦ ⎣ 2 ⎦ ⎣ 2 1 2 ⎦
g( k + 1 ) = g( k ) + α r( k ) [b e1 ( k ) + c e2 ( k )] . (3.22-b)
26
Obviously bm has negative property for the adaptive laws. Because, except only last element,
its remaining element(s) is (are) zero. For this reason and as shown in Eqs. 3.22, except last
row elements, other rows elements of P are ineffective on the adaptive laws. Advise that, the
absolute values of last row elements of P are not very different from each other, otherwise the
effect of elements of e on the adaptive laws are different ratios (Karadeniz et al, 2004).
ii. As shown below in Eqs. 3.23-a and 3.23-b, if except diagonal elements, remaining
elements of P are zero (b=0), only e2 is effective on the adaptive laws. In these
circumstances, adaptation may not occur. In order not to meet with such a result advise that,
all the elements of matrix Q which is used in the Lyapunov equation (Eq. 3.16) are chosen
different from zero. In fact with b=0 Eqs. 3.22-a, b reduces to
⎡ F1 ( k + 1 )⎤ ⎡ F1 ( k )⎤ ⎡ x1 ( k )(c e2 ( k ))⎤
⎢F ( k + 1 )⎥ = ⎢F ( k )⎥ − α ⎢ x ( k )(c e ( k ))⎥ , 3.23-a
⎣ 2 ⎦ ⎣ 2 ⎦ ⎣ 2 2 ⎦
g ( k + 1 ) = g ( k ) + α r( k )(c e2 ( k )) . 3.23-b
There are two important matters with respect to choosing of adaptation period T or
simulation step size. The inverse of T is called as adaptation frequency f.
i. The selected simulation step size T must be large enough to permit the computations
to be performed in real time. But the large simulation step size T may tend to result in an
inferior performance in the closed loop plant and it is bad for disturbances rejection in
adaptive law (CHAK at all, 1997). For this reason simulation step size T must be bigger than
computation time Tc ( T > Tc ) , otherwise simulation program does not work properly and it
gives error message. Computation time Tc involves the evaluation of Eqs. 3.20-a, b and it
depends on hardware features. It is not measured precisely or properly and already it is not
unnecessary, because Simulink gives error message automatically for ( T ≤ Tc ) .
ii. Simulation step size T must be smaller than plant’s Tpi and model’s Tmi time
constants ( T < TPi and T < Tmi ) . Otherwise adaptation mechanism is insufficient for
perception and correction to the parameters variations of plant.
27
For simplicity and shortening the computation time bmT P is abbreviated by s T in the
adaptive laws, i.e .,
s T = b mT P
Now, adaptive laws in Eqs. 3.20 a, b are defined as follow

F ( k + 1 ) = F ( k ) − α x( k ) s T e( k ) , (3.24-a)
g( k + 1 ) = g( k ) + α r( k )s T e( k ) . (3.24-b)
Theoretically, the initial values of adaptation gains F ( 0 ) and g (0 ) may be any values.
But, the desired performance results with a delay due to choosing the initial values very far
from the steady stead values or optimal values F * and g * . The initial values of adaptation
gains may be chosen zero or small values due to unknown optimal values. The initial values
are chosen zero in this thesis.
In the next chapter, the plant which is a permanent magnet DC motor will be introduced
and its MRAC is achieved as desired in this chapter.
CHAPTER 4
MRAC FOR DC MOTOR
Mathematical modeling of permanent magnet (PM) DC motor is presented.

Parameters of DC motor to be controlled and order reduction of motor transfer function are
investigated in this chapter. Then model reference adaptive control of the mentioned motor is
presented.
4.1 MATHEMATICAL MODELLING OF PM DC MOTORS
Model of a PM DC motor is shown Figure 4.1 (Kuo, 1995).
Figure 4.1 Model of a separately excited PM DC motor.
The motor variables and parameters in this figure are explained as follows:
ia (t ) =armature current,
Ra =armature resistance,
eb (t ) = back electromotive force (emf),
28
29
TL (t ) =load torque,
Tm (t ) = motor torque,
θ m (t ) =rotor displacement,
K i =torque constant,
K b =back emf constant,
La =armature inductance,
ea (t ) =applied voltage,
Φ =magnetic flux in the air gap,
wm (t ) =rotor angular velocity,
J m (t ) =rotor inertia,
Bm =viscous- friction coefficient.
The mathematical modeling of PM DC motor is given as follows:
dia (t ) 1 R 1
= e a (t ) − a ia (t ) − eb (t ) ,
d (t ) La La La (4.1)
Tm (t ) = K i ia (t ) ,
(4.2)
dθ (t )
eb (t ) = K b m = K b wm (t ) ,
dt (4.3)
d θ m (t ) 1
2
1 B dθ m (t )
2
= Tm (t ) − TL (t ) − m . (4.4)
dt Jm Jm J m dt
By direct substitution and eliminating all the nonstate variables from Eq. (4.1) through (4.4),
the state equations of dc motor are presented in vector matrix form:
30
⎡ dia (t ) ⎤ ⎡ Ra Kb ⎤
⎢ d (t ) ⎥ ⎢ − − 0⎥ ⎡ i ⎤ ⎡ 1 ⎤
L La a ⎡ 0⎤
⎢ ⎥ ⎢ a ⎥ ⎢ ⎥ ⎢ La ⎥ ⎢ ⎥
⎢ dwm (t ) ⎥ ⎢ K i Bm ⎥⎢ ⎥ ⎢ ⎥ 1
⎢ dt ⎥ = ⎢ J − 0 ⎥ ⎢ wm ⎥ + ⎢0 ⎥ e a (t ) − ⎢ ⎥TL (t ) . (4.5)
Jm ⎢ Jm ⎥
⎢ ⎥ ⎢ m ⎥ ⎢θ ⎥ ⎢0 ⎥ ⎢ ⎥
⎢ dθ m (t ) ⎥ ⎢0 0 ⎥⎢ ⎥ ⎢ ⎥
m
1 ⎣ 0 ⎦
⎢ dt ⎥ ⎢ ⎥ ⎣ ⎦ ⎢⎣ ⎥⎦
⎣ ⎦ ⎣ ⎦
Notice that in this case, TL (t ) is taken as second input in the state equations.
The state diagram of the system is depicted as shown in Fig. 4.2, using Eq. (4.5).
Figure 4.2 State diagram of a dc motor.
The transfer function between the motor displacement and the applied voltage is
acquired from the state diagram as
Θ m (s ) Ki
= (4.6)
E a (s ) La J m s + (Ra J m + Bm La )s 2 + (K b K i + Ra Bm )s
3
Here TL (t ) is accepted zero. The importance of the transfer function Θ m (s ) E a (s ) is that the
dc motor is basically an integrating device between these two variables due to the
transmission pole at s = 0.
31
Figure 4.3 displays a block-diagram representation of the dc-motor. The benefit of

utilizing block diagram is that it gives a comprehensible drawing of the transfer function
relations between the internal and / or external variables of the system.
Figure 4.3 Block diagram of a dc motor system.
The values of K b and K i are identical if K b is represented in V/(rad/sec) and K i is
in (Nm)/A. If K b and K i are replaced by K , Eq. 4.6 reduces to
Θ m (s ) K
= 4.7
E a (s ) (
La J m s + (Ra J m + Bm La )s 2 + K 2 + Ra Bm s
3
)
Dividing both numerator and denominator by La J m , Eq. 4.7 is normalized as
Θ m (s ) K ( La J m )
= . 4.8
E a (s ) ( )
s + (Ra / La + Bm / J m )s 2 + K 2 + Ra Bm s /( La J m )
3
4.2 PARAMETERS OF DC MOTOR TO BE CONTROLLED
The knowledge of plant’s order is enough for MRAC design. In our application plant’s
order is 3 which is seen in Eqs. 4.6-8. For simplicity and shortening the number of operation,
plant’s order is reduced to 2. The exact knowledge of dc motor parameters is needed for using
more formal order reduction methods. For finding of the parameters, two methods are utilized
which are called usual empirical method and matlab’s system identification toolbox.
32
4.2.1 Usual Empirical Method
The transfer function between applied voltage E a ( s) and rotor angular velocity
Wm ( s ) and its parameters are researched using experimental method. The transfer function
Wm (s ) E a (s ) is easily obtained by substituting of Θ m (s ) with Wm ( s ) s −1 in Eq. 4.7 or in Eq.

4.8 to yield
W m (s ) K
= , 4.9
(
E a (s ) La J m s + (Ra J m + Bm La )s + K 2 + Ra Bm
2
)
W m (s ) K ( La J m )
= 2 , 4.10
( )
E a (s ) s + (Ra / La + Bm / J m )s + K 2 + Ra Bm /( La J m )
respectively.
Ra and La are easily measured by RLC meter to be Ra = 15.36 Ω , La = 0.42 mH.

Other parameters are found by applying the following experimental scheme 2 shown in Fig.
4.4. Where, E a is applied voltage, S is on-off switch, A is ammeter, M1 is controlled motor,
E is encoder, wm is rotor angular velocity (rad/s). Encoder output which is connected to the
data acquisition board is displacement. The derivative of the displacement which is called
angular velocity is realized by matlab’s simulink.
Figure 4.4 Connection scheme for empirical method.
2
For a detailed version of the experimental setup refer to Fig. 4.13 where many of the subsytems are made
functionless to obtain experimental test implied by Fig. 4.4.
33
The applied voltage Ea is 8.00 V. After the switch is closed, steady-stead armature
current is measured as I a =120 mA and velocity alteration becomes as in Fig. 4.5. Steady-
stead velocity is 668 rad/s. Dc motor has two time constants. One of them is very large with
respect to the other. The big time constant is observable, the smaller time constant is not
observable on this figure, because its mode goes rapidly to zero as soon as motor starts. As
can be obtained in Fig. 4.5, motor reaches 63 % of its steady-stead velocity value at
τ 1 = 0.0638 second where τ1 is the larger time constant of motor. Hence, (s+ τ 1−1 ) is one
the factor of the denominator polynomial of the motor transfer function Wm (s ) Ea (s ) . The
smaller time constant or pole is to be found by computing.
Under steady-state conditions, i.e. for constant ia and eb , Eq. 4.1 can be written as
E b = E a − I a Ra , 4.11
where, Ea = 8.00 V, I a = 120 mA and Ra = 15.36 Ω . (Although the result are valid with in 2
digits due to measurement errors in the value of La , and 3 digits of I a , Wm , in order not to
cause the additional accumulative errors due to computation at least 4 digits are carried in the
following.) This equation yields
Eb = K b Wm = 8.00 − 0.120 × 15.36 = 6.157 V . Since, Wm is measured to be 668 rad/s

the torque or back emf constant is found to be
Eb 6.157
Kb = Ki = K = = = 92.17 × 10 -4 Vs/r . 4.12
Wm 668
34
Figure 4.5 Velocity alteration of motor
For computation of Bm , consider the dc gain
Wm K
= 2
E a K + Ra Bm 4.13
computed from Eq. 4.10 replacing s by zero with Wm =668r/s, Ea =8.00V,
K = 92.17 × 10 -4 Vs/r , Ra = 15.36 Ω ; this equation yields Bm = 1.656 × 10 −6 Nms/r.
For computation of J m , the denominator polynomial of the transfer function in Eq. 4.10
−1
is divided by (s+ τ 1 ) and the remainder is equated to zero. This operation leads to
τ 1 (K 2 + Ra Bm ) = J m (Ra − La τ 1 ) + La Bm . 4.14
With the measured values of Ra , La and the already computed values of τ 1 , K , Bm
( Ra = 15.36 Ω , La =0.42 mH), τ 1 = 0.0638s, K = 92.17 × 10 -4 Vs/r , Bm = 1.656 × 10 −6 Nms/r )
this equation yields J m = 4.587 × 10 -7 Kg m 2 .

35
Thus all parameters of motor are found by a resistance measurement ( Ra ), an inductance
measurement ( La ), the steady-stead test ( Wm ) and the transient test ( τ 1 ). In fact, the last
two tests are combined in the step response of the motor shown in Fig. 4.5. Hence, the
transfer function H EMP which is found by empirical method between applied voltage Ea and
displacement Θ m is obtained by inserting the found parameters in Eq. 4.8. The result is as in
Eq. (4.15), where, the second time constant τ 2 is 2.735 × 10 -5 s .
Θ (s) 4.784× 107 4.784× 107

H EMP = = =
E a (s ) s 3 + 3.656× 10 4 s 2 + 5.729× 105 s s (s + 3.656× 10 4 ) (s + 15.67)
83.51
= 4.15
( )
s 1 + 2.735× 10-5 s (1 + 0.0638 s )
4.2.2 Parameter Estimation with Matlab’ s System Identification Toolbox
System Identification Toolbox of Matlab builds mathematical models from measured

input-output data. The transfer function of plant is estimated using Prediction Error Method
from the measured dataset. In our application, measured dataset is acquired from executing
the simulink scheme shown in Fig.4.6.
Figure 4.6 Simulink scheme for getting measured input-output data
The motor input signal Ea is a swept sinusoid (a chirp signal) whose frequency range is
selected as between 0.05 and 10 Hz. The bottom level of the input signal Ea is chosen as 2
Volt, because the motor has dead zone between -1 and 1 V which produce bad effect for
estimation of proper transfer function. Upper level of Ea selected as 10 V which is the
36
allowed maximum input voltage of motor. The output signal of motor is the shaft
displacement Θ m which is obtained from output of the encoder. These input output data are
saved to matlab’s workspace while the simulink program is executing. For estimation
operation, Ident which is a graphical user interface to the System Identification Toolbox is
started by writing to command line ‘ident’. The measured dataset import to Ident.
The transfer functions are estimated within certain classes of candidate descriptions (model or
template transfer functions). We select model transfer function K [s ( 1 + s τ 1 )(1 + s τ 2 )]
which is shown in left pane of Fig. 4.7. It contains one integral element, two poles and one
gain. At first estimation, the fit rate between measured output and simulated output is very
high (99.9 %) but τ 1 , τ 2 and K are dissimilar from those in Eq.4.15. Before the second
estimation, τ 1 , τ 2 which are put in to the template transfer function are accepted as known
values from Eq. 4.15. As shown Fig.4.7, K is found as 84.23 whilst the experimental value is
83.51. Hence, the transfer function H IDE 3 which is found by the identification toolbox
becomes as in Eq. 4.16. As shown left pane of Fig.4.8, measured output and simulated model
output are seem one line. The fit rate between them is (99.9 %). It is a very good estimate.
84.23 4.827 × 10 7
H IDE 3 = =
s(1 + 2.735 × 10 −5 s )(1 + 0.0638 s ) s 3 + 3.656 × 10 4 s 2 + 5.729 × 10 5 s
4.827 × 10 7
= 4.16
s (s + 3.656 × 10 4 ) (s + 15.67)
37
Figure 4.7 Parameter identification window related with third order estimation.
Figure 4.8 Comparison of measured and simulated outputs related with third
order estimation.
38
To check the accuracy of the experimental results for the motor parameters, we now want to
find motor parameters using Eq. 4.16. So Eq. 4.16 and Eq. 4.8 are equalized each other. Three
equations which are shown below are obtained from this.
K ( La J m ) = 4.827 × 10 7 (r s ) /V
3
4.17
Ra / La + Bm / J m = 3.656 × 10 4 r/s , 4.18
(K 2
)
+ Ra Bm /( La J m ) = 5.729 × 10 5 (r s )
2
4.19
There are 5 unknown parameters in these equations, for solution of the equations; we must
accept that, the values of whichever at least two are known. As mentioned in Section 4.2.1
that, Ra La and K are known values with high confidence.
Suppose that, Ra = 15.36 Ω and K = 92.17 × 10 -4 Vs/r are known and Bm , J m , La

are desired parameters. The solution of the equations for the remaining 3 parameters are
Bm = 1.591 × 10 -6 Nms/r,
J m = 4.534 × 10 −7 Kg m 2
La = 42.01 × 10 −5 H.
The consistency of these results with the experimental ones are correct within errors 96.07%,
98.84%, and 99.97%, respectively.
For the second case assume La = 42 × 10 −5 H and K = 92.17 × 10 -4 Vs/r are precisely
determined experimentally and Bm , J m , Ra are desired parameters computed from the result
of Identification Toolbox. The solution of equations in 4.17-19 for the remaining
variables Bm , J m , Ra become
Bm = 1.592 × 10 -6 Nms/r,
J m = 4.546 × 10 -7 Kg m 2 ,
Ra = 15.35 Ω .
These are consistent with the experimental results within coherency 96.13%, 99.11% and
99.93%, respectively.
39
Consequently, these results based on the System Identification Toolbox are in good
agreement with experimental ones of the previous section.
4.3 APPROXIMATION OF HIGH ORDER SYSTEMS BY LOW ORDER

SYSTEM
A high order system often contains less important poles that have little effect on the
system response. Thus, given a high order system, it is desirable to find a low order
approximating system, so that, the number of operation is reduced for the solution of control
problem. Generally, in many applications which are utilizing dc motor, plant’s order is
accepted 2 due to neglecting of small La . In addition to negligence of La , we will offer extra
two methods for order reduction.
4.3.1 Order Reduction with Negligence of L

a
In this method, La is accepted zero. Replacement of La = 0 in Eq. 4.7 and using the
first set of parameter values obtained by System Identification Toolbox found in Section
4.2.2, namely
Ra = 15.36 Ω ,
K = 92.17 × 10 -4 Vs/r ,
La = 42.01 × 10 −5 H ≅ 0 ,
Bm = 1.591 × 10 -6 Nms/r,
J m = 4.534 × 10 −7 (Kg m 2 ),
the reduced transfer function
1323
H L neg = 4.20
s (s + 15.71)
is obtained.
40
4.3.2 Order Reduction with Matlab’s “MODRED” and “BALREAL” Commands
The order of the transfer function in Eq.4.16 is reduced to 2 using “modred” (model
reduction) and “balreal” (balanced realization) commands 3. Reduced transfer function
H mod red is
1319
H mod red = 4.21
s (s + 15.65)
which is almost the same within an error less than 0.2 % in the coefficients as in Eq. 4.20
obtained by simply replacing La by zero.
4.3.3 Order Reduction with Matlab’s System Identification Toolbox
Second order transfer function of motor is estimated from measured dataset which is
the same dataset used in estimation of Eq.4.16. This time, we select template transfer function
K [s ( 1 + s τ 1 )] as shown in left pane of Fig.4.9. This transfer function contains one integral
element, one pole and one gain. Hence, the transfer function H IDE 2 which is predicted as
second order becomes

1319
H IDE 2 = 4.22
s (s + 15.66 )
This result is also in good coherence with those in Eqs. 4.20 and 4.21. In fact, having the most
similarity to the remaining two others, this model will be taken as reference in the sequel
when discussing the MRAC.
As shown in the left pane of Fig. 4.10, measured output and simulated model output are
coincident. The fit rate between them is 99.9 % which is the same fit rate of Eq. 4.16.
3
The matlab program code with concerning these commands is placed in to the C.2
41
Figure 4.9 Parameter identification window related with second

order estimation.
Figure 4.10 Comparison of measured and simulated outputs related with second
order estimation.
42
4.3.4 Comparison Between Third Order Transfer Function and Second Order
Transfer Function
In the previous section, it has already been concluded that the results of different
reduction techniques are in good agreement. The transfer functions in Eqs. 4.16, 4.20, 4.21,
4.22 are now compared with each other by plotting bode diagrams. All bode diagrams are
shown in Fig. 4.11.
Fig. 4.11 Bode Diagrams of Transfer Functions
In this figure bode diagrams of the transfer functions in Eqs. 4.16, 4.20, 4.21, 4.22 are
remarked with . (point), o (circle), x (multiplication), + (plus) symbols, respectively. All of
the bode curves concerning third and second order transfer functions seem identical over a 3
decade frequency range strating from 0.1 r/s and extending to 100r/s. Hence, the order of dc
motor’s transfer function Θ m (s ) E a (s ) may be well accepted as second order, and all the
reduced second order transfer functions are almost equally well.
43
4.4 MRAC OF PM DC MOTORS
Since, the motor can be accepted as second order, the order of reference model and the
number of states and its feedback gains of dc motor should be two accordingly. The principle
MRAC scheme of a dc motor displacement is shown in Fig.4.12. Where, θ r (t ) is reference
signal input which is chosen as square wave. ea (t ) is the input voltage of motor. The
displacement θ (t ) is the first state of motor and it is assigned to be output. wm (t ) is the
second state or velocity of the motor which is gotten by first derivative of displacement. F1
and F2 are feedback gains of first and second states. g is feed forward gain of reference
signal. e(t) is identified as difference or error signal between reference model output and dc
motor output. Adaptation mechanism whose input are first state, second state, reference
signal, error signal and its derivative adjust feedback gains F1 , F2 and feed forward gain g .
The aim of the adaptation is that, dc motor output converges to reference model output and
error signal goes to zero.
Figure 4.12 The principle MRAC scheme of a dc motor displacement.

44
4.4.1 Experimental Setup for MRAC

Physical scheme of MRAC of dc motor displacement is shown in Fig. 4.13. Where,
M1 is the controlled motor which is a permanent magnet dc motor whose control voltage is
got from female connector X1 of data acquisition board. E is dual channel encoder which
measures displacement of motor shaft. The phases of the two-output (A and B) pulse trains
are 90 degrees apart electrically. Each output produce 100 square wave per rotation. Encoder
outputs are connected to female connector X2 of data acquisition board. M D which is
identical with M1 is permanent magnet dc motor. It acts as load to for M1 and applies
disturbance effect for displacement. Disturbance torque or load of M1 ( TL ) is adjusted by the
armature current I D of M D .
Rext is 200 Ω external resistance which provides parameter variations for armature resistance
so it applies disturbance effect for displacement. The maximum output voltage and maximum
output current of motor driver are limited 10 V and 0.5 A, respectively. The shafts of M1, E
and M D are coupled to each other for experiment 4.
Figure 4.13 Physical scheme of MRAC of a dc motor displacement.
4
The features of appliances and program codes are in appendix B and appendix C respectively.
45
4.4.2 Review of Adaptive Laws for the Control of Tested Motors
The second order mathematical model in Eq. 4.22 is assumed as reference for the plant
of this section. Adapting the notation in Fig. 4.12, we write this transfer function again for
convenience
Θ 1
= 4.23
1319 E a s (s + 15.66)
Direct realization procedure yields the following steps:
Θ 1 s −2 s −2 F
= = =
1319 E a s 2 + 15.66 s 1 + 15.66 s −1 F + 15.66 s −1 F . 4.24-a
Equating the numerators and denominators, we obtain

Θ = s −2 F , 4.24-b
F = 1319 E a − 15.66 s −1 F , 4.24-c

respectively. The following state transition signal flow graph in Fig. 4.14 follows directly.
Fig. 4.14 State transition signal flow graph of the motor model.
Designing Θ and w = dΘ dt as the first and second state equations, we obtain the
controllable state equations
d ⎡Θ ⎤ ⎡0 1 ⎤ ⎡Θ ⎤ ⎡ 0 ⎤
=⎢ + ea , 4.25-a
dt ⎣ w ⎦ ⎣0 − 15.66⎥⎦ ⎢⎣ w ⎥⎦ ⎢⎣1319⎥⎦
⎢ ⎥
⎡Θ ⎤
Θ = [1 0] ⎢ ⎥ . 4.25-b
⎣ w⎦
These equations constituting the plant model correspond to Eqs. 3.2-a and 3.2-b respectively.
Hence
⎡0 1 ⎤ ⎡ 0 ⎤
A=⎢ ⎥ , B=⎢ ⎥.
⎣0 − 15.66⎦ ⎣1319⎦
46
Feedback control law in Eq. 3.4 becomes

⎡Θ ⎤
ea = g r − [F1 F2 ] ⎢ ⎥ . 4.26
⎣ w⎦
For the reference model, we assume a prototype second order transfer function with
undamped natural frequency wn and damping coefficient ξ (Kuo, B. C., 1995). Hence,
Θ m (s ) wn
2
= . 4.27
Θ r (s ) s 2 + 2ξwn s + wn 2
Following Eqs. 3.6-a, 3.6-b, 3.6-c and 3.6-d, we have
d ⎡ z1 ⎤ ⎡ 0 1 ⎤ ⎡ z1 ⎤ ⎡ 0 ⎤
= + Θr , 4.28
dt ⎢⎣ z 2 ⎥⎦ ⎢⎣− wn2 − 2ξ wn ⎥⎦ ⎢⎣ z 2 ⎥⎦ ⎢⎣ wn2 ⎥⎦
⎡z ⎤
Θ m = [1 0 ] ⎢ 1 ⎥ , 4.29
⎣z2 ⎦
⎡ 0 1 ⎤ ⎡0 ⎤
Am = ⎢ ⎥ , Bm = ⎢ 2 ⎥ , 4.30
⎣ − wn
2
− 2ξ wn ⎦ ⎣ wn ⎦
for the reference model.
Finally adaptation laws are obtained explicitly using Eq. 3.20-a and b,
⎡ F1 ( k + 1 )⎤ ⎡ F1 ( k )⎤ p12 ⎤ ⎡ z 1 − Θ ⎤
=
⎢F ( k + 1 )⎥ ⎢F ( k )⎥ − α
⎡ x1 ( k ) ⎤
⎢ x ( k )⎥ 0 [
w 2 ⎡ p 11
n ⎢p ] p 22 ⎥⎦ ⎢⎣ z 2 − w⎥⎦
, 4.31
⎣ 2 ⎦ ⎣ 2 ⎦ ⎣ 2 ⎦ ⎣ 21
p12 ⎤ ⎡ z 1 − Θ ⎤
[ ⎡p
g ( k + 1 ) = g ( k ) + α Θ r ( k ) 0 wn2 ⎢ 11 ] p 22 ⎥⎦ ⎢⎣ z 2 − w⎥⎦
. 4.32
⎣ p 21
The explicit form of the Lyapunov Equation in (3.16) becomes
⎡0 − wn2 ⎤ ⎡ p11 p12 ⎤ ⎡ p11 p12 ⎤ ⎡ 0 1 ⎤ ⎡Q11 Q12 ⎤

⎢ ⎥⎢ + = − 4.33
⎣1 − 2ξ wn ⎦ ⎣ p21 p22 ⎥⎦ ⎢⎣ p21 p22 ⎥⎦ ⎢⎣− wn2 − 2ξ wn ⎥⎦ ⎢Q ⎥
⎣ 21 Q22 ⎦
We find it convenient to remind that the rate of decrease of the Lyapunov function V is
determined by Q due to Eq. 3.19 i.e,
47
V& = −e T Q e .
The error decays to zero by the time constants of the reference model if Φ F and Φ g in Eqs.
3.13 become zero shortly so that this equation reduces to e& = Am e .
On the base of the specific MRAC equations associated with the plant, the reference
model the MRAC method given in this chapter, the obtained simulation result will be
presented in the next chapter.
CHAPTER 5
SIMULATION RESULTS
In this chapter, we investigate the performance of adaptive tracking for various adaptive
gains α , positive definite matrices P, simulation step sizes T, reference models and various
disturbance effects.
5.1 EFFECTS OF ADAPTIVE GAIN α ON THE PERFORMANCE
We investigate in this section, the effect of adaptive gain α on the adaptive tracking
performance for a proper positive definite matrix P, for fixed simulation step size T and a
chosen reference model. We can suggest a reference model having the transfer function
Θ m (s )
2
wn
= 2 via choosing proper values for the ξ and wn parameters. As
Θ r (s ) s + 2ζwn s + wn 2
shown in Eq. 5.1 and Eq. 5.2, percent maximum overshoot and quality factor of the
denominator polynomial is adjusted by only ζ , as
1−ζ 2
Percent maximum overshoot = 100 e −π ζ / , 5.1
Quality factor of denominator polynomial = 1 (2ζ ) . 5.2
The time delay t d , rise time t r and settling time t s are approximately adjusted by both
ς and wn (Kuo, 1995) as in the following formulas, respectively;
( )
t d ≅ 1.1 + 0.125ζ + 0.469ζ 2 / wn , 5.3
( )
t r ≅ 1 − 0.4167ζ + 2.917ζ 2 / wn , 5.4
48
49
t s ≅ 3.2 / (ζwn ) for 0 < ζ < 0.69 , 5.5-a
t s ≅ (4.5ζ ) / wn for ζ > 0.69 . 5.5-b
We choose ζ =1 and wn = 4 for the reference model. We get the following results by
using the formulas 5.1 through 5.5;
Percent maximum overshoot = 0,

Quality factor of denominator polynomial = 0.5 < 1,
t d ≅ 0.4235 s,
t r ≅ 0.875 s,
t s ≅ 1.125 s.
From Eq. 4.30, Am and B m are
⎡ 0 1 ⎤ ⎡0⎤
Am = ⎢ ⎥ , Bm = ⎢ ⎥ .
⎣− 16 − 8⎦ ⎣16 ⎦
⎡2 1⎤
Q is chosen as Q = ⎢ ⎥.
⎣1 1⎦
P is solved from the Lyapunov Equation 4.33 which becomes
⎡0 − 16 ⎤ ⎡ p11 p 12 ⎤ ⎡ p11 p 12 ⎤ ⎡ 0 1 ⎤ ⎡2 1⎤
⎢1 − 8 ⎥ ⎢ p + = −⎢ ⎥ ,
⎣ ⎦ ⎣ 21 p 22 ⎥⎦ ⎢⎣ p 21 p 22 ⎥⎦ ⎢⎣− 16 ⎥
− 8⎦ ⎣1 1⎦
Its solution for P is
⎡ 0.625 0.0625⎤
P=⎢ ⎥.
⎣0.0625 0.0703⎦
P satisfies criterions that are mentioned at the end of Section 3.2. Namely, all elements of P
are different from zero and the elements of the bottom row be close each other ( b ≅ c ).
From Eqs. 4.31 and 4.32, adaptive laws are

50
⎡ F1 ( k + 1 ) ⎤ ⎡ F1 ( k ) ⎤ ⎡ x1 ( k ) ⎤ ⎡ 0 . 625 0 . 0625 ⎤ ⎡ z 1 − Θ ⎤
⎢ F ( k + 1 ) ⎥ = ⎢ F ( k ) ⎥ − α ⎢ x ( k ) ⎥ [0 16 ] ⎢
0 . 0703 ⎥⎦ ⎢⎣ z 2 − w ⎥⎦
,
⎣ 2 ⎦ ⎣ 2 ⎦ ⎣ 2 ⎦ ⎣ 0 . 0625
⎡ 0.625 0.0625⎤ ⎡z1 − Θ ⎤

g( k + 1 ) = g( k ) + α Θr ( k )[0 16] ⎢ ⎥⎢ ⎥.
⎣0.0625 0.0703⎦ ⎣ z2 − w⎦
As used in Eqs. 3.24a-b, bmT P is abbreviated by s = [1 1.125 ] .

T
So adaptive laws are written by simple way
⎡ F1 ( k + 1 )⎤ ⎡ F1 ( k )⎤ ⎡ x1 ( k )⎤ ⎡ z1 − Θ ⎤
=
⎢F ( k + 1 )⎥ ⎢F ( k )⎥ − α ⎢ x ( k )⎥ [1 1.125] ⎢ z − w⎥ ,
⎣ 2 ⎦ ⎣ 2 ⎦ ⎣ 2 ⎦ ⎣ 2 ⎦
⎡z −Θ ⎤
g ( k + 1 ) = g ( k ) + α Θ r ( k ) [1 1 . 125 ] ⎢ 1 ⎥.
⎣ z2 − w ⎦
Adaptation period T is chosen as 1 ms. The time constant of the plant (Eq. 4.22)
is T p = 1 15.66 = 6.38 × 10 −2 s ≅ 64 ms and the time constants of model are Tm1 = Tm 2 =250 ms.
Simulation step size T is sufficiently smaller than the plant’s and model’s time constants
(T<Tpi and T<Tmi). It satisfies criterions that mentioned at the end of Section 3.2.
Simulation step size T is bigger than the computation time Tc because computer is
forced and gives error message due to using simulation step size T smaller than 1ms, i.e; Tc
≈ 1 ms. In fact T =1 ms is almost the minimum allowable value and the effects of other values
are also investigated in the sequel.
Reference signal is a square wave, its frequency is chosen 0.1 Hz (period is 10 s) and its
amplitude is chosen between π 2 and π in all simulations.
As mentioned at the end of Section 3.2, adaptive gain α will be chosen by trial and
error using simulations in order to achieve a good rate of convergence. Simulation results for
the rotor angular position are shown in Figures 5.1, 5.2, 5.3 for adaptive
gains α = 10 −4 , α = 10 −3 , α = 10 −2 respectively. All the figures indicate satisfactory
convergence of the plant output to that of the model. But the use of a larger value of the
adaptive gain α led to a faster convergence of plant output to model’s. The biggest adaptive
51
gain is chosen as α = 10 −2 because adaptive laws become stiff and difficult to solve
numerically on the computer and simulink gives application error for α > 10 −2 . The solid
lines show model output, dotted lines show plant output, solid-dotted lines if exist show error
signal in all simulations.
Figure 5.1 Simulation results for α = 10 −4 , b ≅ c , T =1 ms
Figure 5.2 Simulation results for α = 10 −3 , b ≅ c , T =1 ms

52
Figure 5.3 Simulation results for α = 10 −2 , b ≅ c , T =1 ms.
5.2 EFFECTS OF POSITIVE DEFINITE MATRIX P ON THE PERFORMANCE
Now, we investigate how adaptive tracking performance is affected, if P matrix is

chosen in improper form. We consider three different P matrices whose bottom elements are
not close each other for the first two cases, the first bottom element is equal to zero for the
last case.
⎡ 2.126 0.3125⎤ ⎡10 1 ⎤

i. P = ⎢ ⎥ for Q = ⎢ ⎥,
⎣0.3125 0.0391⎦ ⎣ 1 0.001⎦
As seen in Eqs. 3.22-a and 3.22-b, the first bottom element (b) of P is the multiplier of e1
( z1 − Θ ) and the second bottom element (c) of P is the multiplier of e2 (z 2 − w) . This
situation may be stated as, e1 is 0.3125/0.0391=8 times more effective than e2 on the
adaptive laws. The simulation results are shown in Fig. 5.4 with respect to this situation. The
error signal e1 changes between +0.15 rad and -0.15 rad. The error signal in Fig. 5.4 is 3
times bigger than in Fig. 5.3. Shortly, the case of ( b > > c ) is less successful than ( b ≅ c ).
53
Figure 5.4 Simulation results for α = 10 −2 , b > > c , T =1 ms.
⎡ 7.625 0.0625 ⎤ ⎡2 1⎤
ii. P = ⎢ ⎥ is for Q = ⎢ ⎥ . Contrary of previous case, this time the
⎣0.0625 0.5078 ⎦ ⎣1 8⎦
coefficient of e2 is approximately 0.5078/0.0625 ≈ 8 times bigger than e1. As shown in
Fig. 5.5 adaptive tracking is much worse than the previous one. The plant doesn’t follow the
model properly, especially around π / 2 rad under steady-state condition. Shortly, the case of
( b << c ) is more terrible than ( b > > c ).
54
Figure 5.5 Simulation results for α = 10 −2 , b << c , T =1 ms.
⎡1 0 ⎤ ⎡0 0 ⎤
iii. P = ⎢ ⎥ for Q = ⎢ ⎥
⎣0 0.0625⎦ ⎣0 1 ⎦
As shown and mentioned in Eqs. 3.23-a and 3.23-b, only e2 is effective on the adaptive
laws. As shown in Fig. 5.6, adaptation can not occur and the plant output goes to away from
the model output little by little. In order not to meet such a result it is advise that all the
elements of matrix Q are chosen different from zero.
55
Figure 5.6 Simulation results for α = 10 −2 , b=0, T =1 ms.
5.3 EFFECTS OF SIMULATION STEP SIZE T ON THE PERFORMANCE
Now, we investigate how adaptive tracking performance is affected if simulation step

size (T ) is chosen larger than 1ms. We consider two different T values 10 ms and 100 ms.
The simulation results with respect to the these values are shown in Fig. 5.7 and Fig. 5.8
respectively. Although the simulation step size T (10 ms) is smaller than Tp (64 ms) and Tmi
(250 ms), adaptive tracking performance shown in Fig. 5.7 is less successful than the one in
Fig. 5.3 whose T value is (1ms). Hence, even if T<Tpi and T<Tmi, the larger simulation step
size T, the lower performance we get.
Because, simulation step size T (100 ms) is not smaller than Tp (64 ms) adaptive
tracking doesn’t occur as shown in Fig. 5.8.

56

57
5.4 EFFECTS OF REFERENCE MODEL ON THE PERFORMANCE
Now, we investigate the success of adaptive tracking for a faster and less stable
reference model. ζ and wn are chosen as 0.2 and 8 respectively.
We get the following results by using the formulas 5.1 through 5.5.
Percent maximum overshoot = 52.66,

Quality factor of denominator polynomial = 2.5 > 1,
t d ≅ 0.01787 s,
t r ≅ 0.01615 s,
t s ≅ 2 s for 0 < ζ < 0.69 .
From Eq. 4.30, Am and B m are
⎡ 0 1 ⎤ ⎡0 ⎤
Am = ⎢ ⎥ , Bm = ⎢ ⎥ .
⎣− 64 − 3.2⎦ ⎣64 ⎦
⎡29 1⎤
Q is chosen as Q = ⎢ ⎥.
⎣ 1 1⎦
P is solved from the Lyapunov Equation (Eq. 4.33)
⎡0 − 64 ⎤ ⎡ p11 p12 ⎤ ⎡ p11 p12 ⎤ ⎡ 0 1 ⎤ ⎡29 1⎤

⎢1 − 3.2⎥ ⎢ p + ⎢− 64 − 3.2⎥ = − ⎢ 1 1⎥
⎣ ⎦ ⎣ 21 p 22 ⎥⎦ ⎢⎣ p 21 p 22 ⎥⎦ ⎣ ⎦ ⎣ ⎦
of which solution is
⎡14.2563 0.2266 ⎤
P=⎢ ⎥.
⎣ 0.2266 0.2271⎦
P satisfies criterions that are mentioned at the end of Section 3.2. Namely, the elements of
bottom row be close each other.
From Eqs. 4.31 and 4.32, adaptive laws are

58
⎡ F1 ( k + 1 )⎤ ⎡ F1 ( k )⎤ ⎡ x1 ( k ) ⎤ ⎡14.2563 0.2266 ⎤ ⎡ z 1 − Θ ⎤
=
⎢F ( k + 1 )⎥ ⎢F ( k )⎥ − α ⎢ x ( k )⎥ [0 64 ] ⎢ 0.2266 0.2271 ⎥ ⎢ z − w⎥ ,
⎣ 2 ⎦ ⎣ 2 ⎦ ⎣ 2 ⎦ ⎣ ⎦⎣ 2 ⎦
⎡14.2563 0.2266 ⎤ ⎡ z 1 − Θ ⎤
g( k + 1 ) = g ( k ) + α Θ r ( k ) [0 64 ] ⎢ ⎥⎢ ⎥.
⎣ 0.2266 0.2271 ⎦ ⎣ z 2 − w⎦
As used in Eqs. 3.21a-b, bmT P is abbreviated by s T = [14.5 14.5313] .
So adaptive laws are simply written as
⎡ F1 ( k + 1 )⎤ ⎡ F1 ( k )⎤ ⎡ x1 ( k ) ⎤ ⎡ z1 − Θ ⎤
⎢ F ( k + 1 )⎥ = ⎢F ( k )⎥ − α ⎢ x ( k )⎥[14.5 14.5313] ⎢ z − w⎥ ,
⎣ 2 ⎦ ⎣ 2 ⎦ ⎣ 2 ⎦ ⎣ 2 ⎦
⎡z −Θ ⎤
g ( k + 1 ) = g ( k ) + α Θ r ( k ) [14 . 5 14 . 5313 ]⎢ 1 ⎥.
⎣ z2 − w⎦
Adaptation period T is chosen 1ms in this simulation. The time constant of the plant is
still T p = 1 15.66 = 6.38 × 10 −2 s ≅ 64 ms , and the time constants of the model yield Tm1 =625
ms for the exponential decaying and Tm2 =125 ms for the oscillations period. In fact this
amount of oscillation period corresponds to a much smaller (approximately 4 times) time
constant. Simulation step size T is much smaller than plant’s and model’s time constants
(T<<Tpi and T<<Tmi). It satisfies criterions that mentioned at the end of Section 3.2.
Reference signal is a square wave, its frequency is chosen as 0.1 Hz and its amplitude
is chosen between π 2 and π in this simulation. Simulation result with respect to the
mentioned model is shown Fig. 5.9. Because model trajectory or desired trajectory changes
very rapidly due to faster behavior of the model, the plant does not follow the model properly
especially around π / 2 rad.
59
Figure 5.9 Simulation result with respect to faster and less stable model.
5.5 EFFECTS OF DISTURBANCES ON THE PERFORMANCE
We will investigate the success of the adaptation mechanism against to the disturbance
effects. We apply disturbance effects with two ways which are changeable armature
resistance (Rext) and load torque ( TL ).
5.5.1 Disturbances with Changeable Armature Resistance
To begin with, we will apply disturbance effect by means of the disturbance

potentiometer (Rext) which is connected in series with the armature resistance Ra of dc motor.
The experimental scheme is shown in Fig. 4.13. The reference model and the values of other
parameters which are used for this section are same as those used for Fig. 5.3 namely,
α = 10 −2 , b = 0.0625, c = 0.0703, T =1 ms.
60
Figure 5.10 Simulation result with respect to armature resistance variations
As shown in Fig. 5.10 the value of disturbance potentiometer (Rext) is increased by

gradually within specified time intervals. No appreciable deviation occurs between resistance
levels up to 110 Ω , plant follows model very well. At time 25 second, disturbance resistance
is increased to the value of 130 Ω and adaptive tracking becomes poor. At time 39 second
disturbance resistance is increased to the value of 140 Ω and motor shaft stops. At first
glance, this situation may be perceived as weakness of adaptation mechanism. As mentioned
in the Section 4.2.2, dc motor has dead zone between -1 and 1 V levels of applied
voltages Ea . Ra is already known as 15.36 Ω from Section 4.2.1, under these circumstances,
from ohm law, dead zone armature current I a interval is between -65mA and +65mA. In
other words motor torque becomes not sufficient to overcome the static resistance and motor
shaft doesn’t rotate between these levels. Total armature resistance is
RT = 140 + 15.36 = 155.36 Ω for Rext = 140 Ω . Then maximum armature current is I a = 10 /
155.36 =64.36 mA for maximum allowed input voltage E a = 10 V . This value is inside of
dead zone level. Hence, the stopping of the shaft for Rext = 140 Ω is not an abnormal
situation. Armature current I a is 68.79 mA for Rext = 130 Ω. This value is close to dead
zone, for this reason adaptive tracking is poor. Figure 5.10 shows that the value of sudden
61
change of 20 Ω in Rext does not spoil the control mechanism since there is not an perceptible
difference on the values of the models and plant outputs (See subsequent intervals from 0 to
25 ms). To view the effect of abrupt change of Rext within the region where the adaptation
mechanism work well ( 0 ≤ R ext ≤ 110 ), we cary Rext with parallel switch from 0 to 110 Ω and
110 to 0 Ω suddenly. Simulation results are shown in Fig. 5.11.
Figure 5.11 Simulation results with respect to abrupt armature resistance variations
(0 to 110 Ω and 110 to 0 Ω )
The switch wich is connected in parallel with disturbance potentiometer (Rext) is on-off
within specified time intervals. Although small deviations occur after switching time,
adaptation mechanism adjust itself very rapidly in almost 1s.
5.5.2 Disturbances with Changeable Load Torque
Now, we will apply disturbance effect by means of disturbance motor M D which is

identical with M1. M D acts as load for M1 and applies disturbance effect for displacement.
Disturbance torque or load of M1 ( TL ) is adjusted by the armature current I D of M D . As
shown in Figure 5.12, disturbance current I D is increased by gradually within specified time
intervals. No deviations occur between disturbance current levels up to 0.4 A, plant follows
62
model very well. At time 33 second, disturbance current is increased to the value of 0.5 A
and, after 2 seconds plant output leave the model output when the reference input reversed.
As mentioned in Section 4.4.1, maximum output current of driver of M1 is limited with 0.5 A.
At time 33 second, the currents of identical motors are equal, therefore their torqueses are
same, and then the directions of torques are converse of each other, so motor shaft doesn’t
rotate.
Figure 5.12 Simulation result with respect to torque disturbances.
In reality, the motor may be subject to two types of saturation or limitations. One
limitation is that as the armature current increases due to the increase in Ea , the magnetic
circuit will saturate, so that the motor torque cannot exceed a certain maximum value. The
second limitation is due to the maximum current that the motor can handle due to the heat
dissipation rating of the motor. Amplifier gain or motor driver gain is subject to magnetic
saturation and heat dissipation. For these reasons, the maximum output voltage and maximum
output current of motor driver are limited 10V and 0.5A respectively. Unless these
limitations are not to be for motor and its driver, MRAC overcomes almost all excessive load
parameter variations and adaptive tracking is continuous.
63
Figure 5.12 shows that the value of sudden change of 0.1mA in disturbance current of
M D does not spoil the control mechanism since there is not an perceptible difference on the
values of the models and plant outputs (See subsequent intervals from 0 to 0.4mA). To view
the effect of abrupt change of TL within the region where the adaptation mechanism work well
( 0 ≤ I D ≤ 0.4 mA ), we cary I D from 0 to 0.4mA and 0.4 to 0 mA suddenly. Simulation results
shown in Fig. 5.13. The switch wich is connected in series with disturbance current source is
made on-off within specified time intervals. Although small deviations occur just after the
switching time, adaptation mechanism adjust itself very rapidly in almost 1s.
Figure 5.13 Simulation result with respect to abrupt torque disturbances.

CHAPTER 6
CONCLUSIONS
Model Reference adaptive control of a permanent DC motor is investigated in this

study. After an introduction of different control problems and MRAC, the general theory
about the method and its application to the considered motor (plant) is treated. Identification
of motor parameters by experimental methods are evaluated to know the motor for handling
the proper choice of the reference model and the correct judgement of the MRAC method.
Other than gaining the theoretical and practical knowledge about MRAC, identification
problem, matlab tools, this work has also been beneficial in getting acquaintance with the
interactive study involving computer and real life.
The main results of this study may be summarized as in the following bullets;
• Model reference adaptive control is one of the various techniques of solving control problem
when the parameters of the controlled process are poorly known or vary during normal operation.
For this reason, a model reference adaptive control method is proposed to control the position of a
dc motor without requiring any fixed motor parameter. Experimental results show how well this
method controls the position of the motor.
• This method is only suitable for plants without finite transmission zeros.
• This method adjusts feedback gains of states and feed forward gain of closed loop plant,
equalizes them to model’s.
• If plant dynamics can be expressed with lower order transfer function, reference model can
be chosen as lower order.
• The displacement of motor is measured by encoder. Encoder output has insignificant noise,
hence, we don’t use filter for plant output.
• To choice of small adaptive gain α may result in slow convergent rate whereas large α
may make the differential equations stiff and difficult to solve numerically on the computer.
• Simulation step size T must be bigger than the computation time Tc ( T > Tc ) , otherwise
simulation program does not work properly and it gives error message. But the large
64
65
simulation step size T reduces the adaptive tracking performance. Computation time Tc
involves the evaluation of adaptive laws and it depends on hardware features. It is not
measured precisely or properly and already it is not unnecessary, because Simulink gives
error message automatically for ( T ≤ Tc ) .
• Simulation step size T must be smaller than plant’s and model’s time constants
( T < TPi and T < Tmi ) otherwise adaptation mechanism is insufficient for perception and
correction to the parameters variations of plant.
• The absolute values of bottom row elements of P must be close each other ( b ≅ c ) , these
elements are coefficients of error signals e (e1 and e2) on the adaptive laws. Otherwise the
effect of elements of e on the adaptive laws are different ratios and adaptive tracking
performance decreases. That is adaptation is more efficient when all entries of the error vector
are almost equally weighted for adaptation.
• If except diagonal elements, remaining elements of P are zero (b=0), only e2 is effective on
the adaptive laws. In these circumstances, adaptation doesn’t occur. In order not to meet with
such a result advise that all the elements of matrix Q which is used in the Lyapunov equation
(Eq. 3.16) are chosen different from zero.
• In spite of change of motor parameters during working time, MRAC adjusts plant’s input so
that output of the plant becomes as model output.
• The motor is subjected to two types of limitations, one is magnetic saturation and second is
heat dissipation. For these reasons, the maximum output voltage and maximum output current
of motor driver are limited within specific intervals. Unless these limitations are not exceeded
for motor and its driver, MRAC overcomes all excessive load parameter variations and
adaptive tracking is continuous.
• The parameter variations with time which are faster as compared to simulation step size T
are not perceived and corrected by MRAC such as white noise and high frequency
disturbances.
• However, the method might be improved to apply the plants with finite transmission zeros
and this is the subject of the future work.
APPENDIX A
LYAPUNOV STABILITY
Consider the nonlinear, time-varying system (Butler, 1992):
x& = f ( x, t ) . (A.1)
Here, f is an n x 1 vector function and x is the n x 1 state vector. An equilibrium point
( )
x * of system (A.1) is characterized by f x * = 0 . If the state x of system (A.1) is situated at
an equilibrium point at t = 0 , it will stay on the equilibrium point for all t > 0 . If the initial
state isn’t exactly equal to the equilibrium state ( x(0 ) = x * + δ ), there are four possibilities.
1. The state x can be made to remain in any specified vicinity ε of x * by choosing a

sufficiently small deviation δ of x * at t = 0 . However, x isn’t guaranteed to converge to x * .
This behavior is called stable.
2. The system isn’t stable. It is than said to be unstable. Note that if a system is unstable,
this doesn’t imply divergence from x * . It merely states that it is possible to specify an ε
which doesn’t allow any δ . A well known example of this is a harmonic oscillator which is
oscillating with a specific frequency and amplitude: x(t ) = A * sin( wt ) .While x always lies
between -A and +A, specifying ε > A allows any initial deviation δ < ε . However, no value
of δ can be found that allows an ε < A because x will always exceed ε for some time t .
3. The state x will converge to x * as long as the initial deviation δ of x * is smaller than
some boundary R. More formally, if δ < R , for an arbitrarily close vicinity μ of x * a time t *
exists such that for all t > t * , x − x * < μ .This notion is called asymptotic stability. Note that
in this definition, starting within a boundary R from x * doesn’t imply that x will remain
within R for all t > 0 , but only that x − x * → 0 as t → ∞ .
4. If the asymptotic stability is guaranteed for any initial deviation δ , the asymptotic
property is said to be global. To check to which stability class a given system of the form
66
67
(A.1) belongs, Lyapunov’s direct method can be applied. Lyapunov’s method can only
investigate stability properties of an equilibrium point x * = 0 . Hence, other equilibrium points
have to be transformed to x * = 0 by a transformation x ′ = x − x * . In applying Lyapunov’s
method, first a Lyapunov function V ( x ) is defined. This function can be considered as a sort
of energy function, while it has similar properties to the energy stored in the system. The
Lyapunov function itself must satisfy
V ( x ) >0, ∀x ≠ 0 .
In addition to be comparable to energy function V ( x ) should be monotonically increasing,
and go to infinity as x → ∞;
0 < α x < V (x ) ; α > 0
V ( x ) → ∞ as x → ∞.
Now, it can be felt that if the stored energy in a system decreases as time passes, all energy
will eventually leave the system and equilibrium x = 0 will be reached. Similarly, if the time
derivative of the Lyapunov function is always negative and hence V is decreasing with time, V
will eventually become zero because V is monotonous. As V=0 implies x =0, also due to the
monotonous character of V, a negative definite V& guarantees asymptotic stability;
V& < 0 , ∀x ≠ 0 .
To calculate V& , partial derivatives of V with respect to the elements of x are needed;
dx ∂ V ∂V
V& ( x, t ) = ⋅ = f (x ) ⋅ .
dt ∂ x ∂x
∂V
Hence, the partial derivative must be continuous. The above requirements on V
∂x
guarantee global asymptotic stability. By using less strict requirements, other forms of
stability are obtained. For example, if V& is negative semi-definite, which implies that V& = 0
for some x ≠ 0 , the stability is no longer asymptotic. If V isn’t monotonously increasing
with x → ∞ , the stability isn’t global. Note that Lyapunov’s method provides a stability
guarantee: if the requirements mentioned are met, the system is guaranteed to be globally
asymptotically stable. However, if the requirements aren’t satisfied, the system may still be
68
stable. In addition, the choice of the Lyapunov function V is crucial in the stability check.
Different Lyapunov functions may give different stability results.
For a linear system x& = Ax , let us consider a quadratic Lyapunov function:
V = x T P x , in which P is a symmetric positive definite matrix. Then;
V& = x& P x + x T P x& = − x T Q x

in which
− Q = AT P + PA .
According to Lyapunov’s theorem, a positive definite symmetric matrix Q always yields a

positive definite symmetric matrix P if the system x& = Ax is asymptotically stable.
APPENDIX B
THE FEATURES OF APPLIANCES
B.1 MOTOR SPECIFICATIONS
• M1 and M D are identical permanent magnet dc motor.

• Maximum armature current is 0.5 A
• Maximum armature voltage is 10 V
• Nominal speed is 5700 RPM.
• Armature resistance is 15.36 Ω
• Armature inductance is 0.42 mH
B.2 ENCODER SPECIFICATIONS
• Typical dual-channel encoder

• It produces 100 square waves per rotation.
B.3 DATA ACQUISITION BOARD SPECIFICATIONS
The MF 614 multifunction I/O card is designed for the need of connecting PC
compatible computers to real world signals. The MF 614 contains a 100 kHz throughput 12
bit A/D converter with sample/hold circuit, four software selectable input ranges and 8
channel input multiplexer, 4 independent 12 bit D/A converters, 8 bit digital input port and 8
bit digital output port, 4 quadrature encoder inputs with single-ended or differential interface
and 5 timers/counters. The card is designed for standard data acquisition and control
applications and optimized for use with Real Time Toolbox for MATLAB®.
69
70
B.4 HARDWARE SPECIFICATIONS
• Microprocessor is Pentium 4, 3 GHz

• RAM is 512MB
• Hard disk is 20GB
B.5 SOFTWARE SPECIFICATIONS
• Matlab version is 7.1

• Simulink version is 6.3
• System Identification Toolbox version is 6.1.2
APPENDIX C
THE MATLAB PROGRAM CODES
C.1 SIMULINK SCHEME OF MRAC OF A DC MOTOR DISPLACEMENT
Figure C.1 Simulink scheme of MRAC
71
72
C.1.1 Inside of Model Block
Figure C.2 Interior of Model Block
C.1.2 Inside of Plant Block
Figure C.3 Interior of Plant Block.
C.1.3 Inside of S-Function Builder Block (Adaptation Mechanism)
%this code is written with C++ program

Static double F1=0.0, F2=0.0, g=0.0, a=.01, s1=1, s2= 1.1250;
If (reset [0]>0.5)
F1=F2=g=0.0;
F1 -= a*(s1*x1[0]*e1 [0] + s2*x1[0]*e2 [0]);
F2 -= a*(s1*x2[0]*e1 [0] + s2*x2[0]*e2 [0]);
g += a*r [0]*(s1*e1 [0] +s2*e2 [0]);
U [0] = r [0]*g - F1*x1[0] - F2*x2[0];
C.2 ORDER REDUCTION of MOTOR TRANSFER FUNCTION with MATLAB
%ORDER REDUCTİON WITH MATLAB'S BALREAL AND MODRED COMMANDS

TFIDE3=TF ([4.827E7], [1 3.656E4 5.729E5 0])
[NUM, DEN] = TFDATA (TFIDE3,'V');
SYS=TF([NUM(1,4)],[DEN(1,1) DEN(1,2) DEN(1,3)]);
[SYSB, G] = BALREAL (SYS);
SYSR = MODRED (SYSB, [2],'DEL');
SYSR=TF (SYSR);
[NUM, DEN] = (TFDATA (SYSR,'V'));
FPRINTF ('ORDER REDUCTION WITH MATLAB’ S BALREAL COMMAND =%5.15E\N')
TFBAL=ZPK(TF([NUM(1,2)],[DEN(1,1) DEN(1,2) 0]))%END OF PROGRAM
REFERENCES
Aiko Miyamura , Hidenori Kimura, April 2002, “Stability of feedback error learning
scheme”, Systems & Control Letters, Automatica, Volume 45, No.4, Pages 303-316.
Aström, K. J. and Wittenmark,B, Adaptive Control, Addison Wesley, New York, 1995.
Butler, H., Model Reference Adaptive Control, Prentice Hall, New York, 1992.
Chak C. K., Gang Feng and T. Hesketh, March 1997, “Multirate adaptive optimal control
with application to dc motor” computers elect. engng vol. 23, no. 2, pp. 65-79.
Chaoa C.L., J. Neoub, October 2000, “Model reference adaptive control of air-lubricated
capstan drive for precision positioning”,Precision Engineering, Vol. 24, Issue 4 , pp.
285–290.
Chien Chiang-Ju and Chia-Yu Yao , May 2004, “Iterative learning of model reference
adaptive controller for uncertain nonlinear systems with only output measurement”,
Automatica, Volume 40, Issue 5, Pages 855-864.
Costa Ramon R., Liu Hsu , Alvaro K. Imai , Petar Kokotovic, July 2003, “Lyapunov based
adaptive control of MIMO systems”, Automatica, Volume 39, Issue 7, Pages 1251-1257.
Ioannou, P. A., Robust Adaptive Control, Prentice Hall, New York, 1994.
Karadeniz, M. et al, 2004 “Adaptive Pole Placement Control of a DC Motor Speed”, J.

Fac.Eng. Arch. Gazi Univ. Vol 19, No 3, 327-334.
Krstic Miroslav and Andrzej Banaszuk, June 2005, “Multivariable adaptive control of
instabilities arising in jet engines”, Control Engineering Practice, In Press, Corrected
Proof, Available online 13 June 2005
Kojabadi H. M., 2005, “Simulation and experimental studies of model reference adaptive
system for sensorless induction motor drive”, Simulation Modelling Practice and Theory,
Volume 13, Pages 451-464
Kuo, B. C., Automatic Control Systems, Prentice Hall, New Jersey, 1995.
73
74
Lee T. H, S. S. Ge, C. P. Wong , December 1998, “Adaptive neural network feedback

control of a passive line of sight stabilization system”, Mechatronics, Volume 8, Issue 8,
Pages 887-903.
Lee T.H., K.K. Tan, S.N. Huang, H.F. Dou, December 2000, “Intelligent control of precision
linear actuators”, Engineering Applications of Artificial Intelligence, Volume 13, Issue 6,
Pages 671-684.
Makoudi M., L. Radouane, July 2000 , “Robust decentralized adaptive control for non-
minimum phase systems with unknown or time varying delay”, Automatica, Vol. 35, No.
2, pp. 1417-1426.
Makoudi M. and L. Radouane, August 1999, “A robust decentralized model reference

adaptive control for non-minimum-phase interconnected systems”, Automatica, Volume
36, Issue 7, Pages 1057-1065.
Marino R, S. Peresada and P. Tomei, May 1998, “Adaptive Output Feedback Control of
Current-Fed Induction Motors with Uncertain Rotor Resistance and Load Torque”,
Automatica, Volume 34, Issue 5, Pages 617-624.
McLain Richard B., Michael A. Henson, April 2000, “Principal component analysis for
nonlinear model reference adaptive control”, Computers and Chemical Engineering,
Volume 24, Pages 99-110.
Mirkin Boris M., Olof Gutman, October 2005, “Output feedback model reference adaptive
control for multi-input–multi-output plants with state delay”, Systems & Control Letters,
Vol. 54, Issue 10, pp. 961–972.
Park C., Kwon H., 2004, “Simple and robust speed sensorless vector control of induction
motor”, Electric Power Systems Research, Volume 71, Pages 257-266.
Sinha P.K., A.N. Pechev, August 1999, “Model reference adaptive control of a maglev
system with stable maximum descent criterion”, Automatica, Vol. 35, No. 8, pp. 1457-
1465.
Taware A., Gang Taoa, Nilesh Pradhana, Carole Teolisb, March 2003, “Friction
compensation for a sandwich dynamic system”, Automatica, Volume 39, Issue 3, Pages
481-488.
Tian Z., K.A. Hoo, November 2003, “Transition control using a state shared model
approach”, Computers and Chemical Engineering, Vol. 27, Issue 11, pp. 1641-1656.
Tsai Y . K . and Y . H . Lin, July 1997,” Adaptive modal vibratıon control of a fluid-
conveying cantilever pipe” Journal of Fluids and Structures,Volume 11, Issue 5, Pages
535-547.
Wang Hong, F Zhen J. Huangs And Steve Daley, February 1997, “On the Use of Adaptive
Updating Rules for Actuator and Sensor Fault Diagnosis”, Automatica, Volume 33, Issue
2,, Pages 217-225
75
Yasser Eldeeb and W H. Elmaraghy, July 1998, “Optimal control of a single link manipulator
including motor dynamics”, Computer Integrated Manufacturing Systems, Volume 11,
Issue 3, Pages 199-205
Zeng Y., A.D. Araujo and S.N. Singh, 1999 “Output feedback variable structure adaptive
control of a flexible spacecraft”, Acta Astronautica Vol. 44, No. 1, pp. 11±22,.
Zhou J.and Y. Wang, October 2005, “Real-time nonlinear adaptive backstepping speed
control for a PM synchronous motor, Control Engineering Practice, Volume 13, Issue 10,
pp 1259-1269
Zhong Y. S., September 2005, “Globally stable adaptive system design for minimum phase
SISO plants with input saturation”, Automatica, Volume 41, Issue 9, Pages 1539-1547.

Position Control of A Permanent Magnet DC Motor by Model Reference Adaptive Control

Uploaded by

Copyright:

Available Formats

Position Control of A Permanent Magnet DC Motor by Model Reference Adaptive Control

Uploaded by

Document Information

Original Title

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

Position Control of A Permanent Magnet DC Motor by Model Reference Adaptive Control

Uploaded by

Copyright:

Available Formats

POSITION CONTROL OF A PERMANENT MAGNET DC MOTOR BY

MODEL REFERENCE ADAPTIVE CONTROL

the Graduate Institute of Sciences and Engineering

in partial fulfillment of the requirements for the degree of

Prof. Dr. Muhammet KÖKSAL

Prof. Dr. Muhammet KÖKSAL

Examining Committee Members

Prof. Dr. Muhammet KÖKSAL ...............................................

Assoc. Prof. Dr. Erkan İMAL ...............................................

Assist. Prof. Dr. Atakan KURT ...............................................

Assist. Prof. Dr. Nurullah ASLAN

POSITION CONTROL OF A PERMANENT MAGNET DC MOTOR BY

M. S. Thesis - Electronics Engineering

Supervisor: Prof. Dr. Muhammet KÖKSAL

Keywords: Model Reference Adaptive Control, Lyapunov’s Direct Method, Dc Motor.

MODEL REFERANS UYARLAMALI KONTROL İLE KUTUPLARI SABİT

Yüksek Lisans Tezi – Elektronik Mühendisliği

Tez Yöneticisi: Prof. Dr. Muhammet KÖKSAL

Prosesin parametrelerinin tam olarak bilinmemesi veya parametrelerin çalışma esnasında

Anahtar Kelimeler: Model Referans Uyarlamalı Kontrol, Lyapunov’un direkt metodu, Dc

4.3.3 Order Reduction with Matlab’s System Identification Toolbox.............40

5.5.2Disturbances with Changeable Load Torque............................................61

CHAPTER 6 CONCLUSIONS ..............................................................................64

APPENDIX A LYAPUNOV STABILITY..................................................................66

APPENDIX B THE FEATURES OF APPLIANCES................................................. 69

APPENDIX C THE MATLAB PROGRAM CODES ................................................71

2.1 General control system design steps………………………………………….........10

5.2 Simulation results for α = 10 −3 , b ≅ c , T =1 ms .................................................... .51

5.3 Simulation results for α = 10 −2 , b ≅ c , T =1 ms .................................................... 52

5.4 Simulation results for α = 10 −2 , b > c , T =1 ms. ................................................. ..53

5.5 Simulation results for α = 10 −2 , b < c ,T =1 ms. ..................................................... 54

5.6 Simulation results for α = 10 −2 , b=0, T =1 ms........................................................ 55

5.7 Simulation results for α = 10 −2 , b ≅ c , T =10 ms.....................................................56

5.8 Simulation results for α = 10 −2 , b ≅ c , T =100 ms...................................................56

5.10 Simulation results with respect to armature resistance variations...........................60

5.11 Simulation results with respect to abrupt armature resistance variations

(0 to 110 Ω and 110 to 0 Ω ).....................................................................................61

5.12 Simulation results with respect to torque disturbances............................................62

5.13 Simulation results with respect to abrupt torque disturbances.................................63

LIST OF SYMBOLS AND ABBREVIATIONS

Δ Unmodeled plant phenomena

Pˆ (θ (t) ) Estimated plant model

C (θC ) Control law

sT Abbreviated form of bmT P

ΦF Difference between F and F *

Φg Difference between g * and g

eb (t ) back electromotive force (emf)

Kb back emf constant

Φ magnetic flux in the air gap

wm (t ) rotor angular velocity

Bm viscous- friction coefficient

td the delay time

tr the rise time

ts the settling time

Second advantage of adaptive systems is the information of plant’s parameters is not

The solution of problem against to the parameters variations is investigated by many

This thesis is organized as follows: Fundamentals of MRAC is summarized in Chapter

MODEL REFERENCE ADAPTIVE CONTROL

2.1 GENERAL CONTROL SYSTEM DESIGN STEPS