Adaptive Control of Nonaffine Systems With Applications To Flight Control

Adaptive Control of Nonane Systems with Applications to Flight
Control
Amanda Young
Thesis submitted to the Faculty of the
Virginia Polytechnic Institute and State University
in partial fulllment of the requirements for the degree of
Master of Science
in
Aerospace Engineering
Dr. Naira Hovakimyan, Committee Chair
Dr. Daniel Stillwell, Committee Member
Dr. Craig Woolsey, Committee Member
May 5, 2006
Blacksburg, Virginia
Keywords: Flight Control, Adaptive Control, Nonane Systems
Copyright 2006, Amanda Young
Adaptive Control of Nonane Systems with Applications to Flight Control
Amanda Young
(ABSTRACT)
Traditional ight control design is based on linearization of the equations of motion around
a set of trim points and scheduling gains of linear (optimal) controllers around each of these
points to meet performance specications. For high angle of attack maneuvers and other
aggressive ight regimes (required for ghter aircraft for example), the dynamic nonlinearities
are dependent not only on the states of the system, but also on the control inputs. Hence,
the conventional linearization-based logic cannot be straightforwardly extended to these
ight regimes, and non-conventional approaches are required to extend the ight envelope
beyond the one achievable by gain-scheduled controllers. Due to the nonlinear-in-control
nature of the dynamical system in aggressive ight maneuvers, well-known dynamic inversion
methods cannot be applied to determine the explicit form of the control law. Additionally,
the aerodynamic uncertainties, typical for such regimes, are poorly modelled, and therefore
there is a great need for adaptive control methods to compensate for dynamic instabilities.
In this thesis, we present an adaptive control design method for both short-period and
lateral/directional control of a ghter aircraft. The approach uses a specialized set of radial
basis function approximators and Lyapunov-based adaptive laws to estimate the unknown
nonlinearities. The adaptive controller is dened as a solution of fast dynamics, which veries
the assumptions of Tikhonovs theorem from singular perturbations theory. Simulations
illustrate the theoretical ndings.
Acknowledgments
First and foremost, I would like to thank God for giving me all the blessings in the world.
I would like to thank my advisor, Naira Hovakimyan, for believing in me and investing so
much of her time and eort. If it were not for her passion and dedication, I would not be as
successful as I am today.
Many thanks to Dr. Chengyu Cao for his patience, Dr. Eugene Lavretsky from the Boeing
company for his suggestions and insights, Dr. Vijay Patel for sharing his guidance and
experience, and Dr. Konda Reddy for his wisdom. I would like to thank my committee
members Dr. Daniel Stillwell and Dr. Craig Woolsey for their time and commitment.
Additionally, I thank my friends Imraan Faruque, Lili Ma, Nina Mahmoudian, Vahram
Stepanyan, Laszlo Techy, and Jiang Wang for their help and support. I would also like to
thank the department of Aerospace and Ocean Engineering and its faculty for transforming
me into an engineer.
All my thanks to my best friend, Sam for his support and love. You mean everything to
me. Thanks to my little pup, Kane, for big slobbery kisses and for putting up with my long
hours and not messing in the house. Finally, I would like to thank my mom, Katherine for
all her hard work and sacrices. My opportunities in life are made possible because of you.
iii
Contents
1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Problem Denition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.4 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2 Literature Review 4
2.1 Short Period Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2 Dutch-Roll Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
3 Mathematical Preliminaries 7
3.1 Preliminaries on Nonlinear Systems Theory . . . . . . . . . . . . . . . . . . . 7
3.2 Preliminaries on Approximation Theory . . . . . . . . . . . . . . . . . . . . 9
3.3 Preliminaries on Singular Perturbations Theory . . . . . . . . . . . . . . . . 10
4 Adaptive Controller for Nonane Single Input System 14
4.1 Problem Formulation for Short Period Dynamics . . . . . . . . . . . . . . . . 14
4.2 Ideal Reference Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4.3 RBF Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.4 State Predictor and Adaptive Law . . . . . . . . . . . . . . . . . . . . . . . . 19
iv
4.5 Nonane Control Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.6 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
5 Adaptive Controller for Nonane Multi-Input System 31
5.1 Problem Formulation for Dutch-Roll Dynamics . . . . . . . . . . . . . . . . . 31
5.2 Nominal Dutch-Roll Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
5.3 Multi-Input RBF approximation . . . . . . . . . . . . . . . . . . . . . . . . . 36
5.4 Roll-Yaw Dynamics State Predictor . . . . . . . . . . . . . . . . . . . . . . . 38
5.5 MIMO Control Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
5.6 Dutch-Roll Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
6 Conclusion 54
6.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
6.2 Recommendations For Future Work . . . . . . . . . . . . . . . . . . . . . . . 55
v
List of Figures
4.1 Intersection of the functions
e
ad
(t) and

W
(t)((t),
e
nom
(t)
e
ad
(t)). . . . 23
4.2 Gaussian as a function of with an angle of attack range of 60
. . . . . . . . 26
4.3 f(,
e
) as a function of elevator control
e
. . . . . . . . . . . . . . . . . . . . 27
4.4 Short-period nominal system tracking with respect to dened commanded
input r
cmd
(t). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.5 LQR performance violated in the presence of uncertainties. . . . . . . . . . . 28
4.6 Loss of tracking of and q caused by nonlinearity f(,
e
) of order O(0.9). . 28
4.7 States (t),
r
(t), (t) of short-period system. . . . . . . . . . . . . . . . . . 29
4.8 States q(t), q
r
(t), q(t) of short-period system. . . . . . . . . . . . . . . . . . . 29
4.9 Adaptive elevator control,
e
ad
(t)of short-period system versus unknown non-
linearity f((t),
e
(t)). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5.1 Dutch-roll nominal system
r
(t) and
r
(t) tracking of commanded input R
cmd
(t). 49
5.2 LQR performance violated for angle
r
,
r
tracking. . . . . . . . . . . . . . . 50
5.3 Loss of tracking of and caused by nonlinearities f
1
(,
a
) of order O(1)
and f
2
(,
r
) of order O(1.1). . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
5.4 Curve-tting for f
l
(, p
s
, r
s
,
a
,
r
). . . . . . . . . . . . . . . . . . . . . . . . . 51
5.5 Curve-tting for f
2
(, p
s
, r
s
,
a
,
r
). . . . . . . . . . . . . . . . . . . . . . . . 51
5.6 States

(t),
r
(t), (t) of Dutch-Roll system. . . . . . . . . . . . . . . . . . . 52
5.7 States

(t),
r
(t), (t) of Dutch-Roll system. . . . . . . . . . . . . . . . . . . 52
vi
5.8 States p
s
(t), p
s
r
(t), p
s
(t), r
s
(t), r
s
r
(t), r
s
(t) of Dutch-Roll system. . . . . . . . 53
5.9 Adaptive control eort of Dutch-Roll System compared with uncertainties f
1
and f
2
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
vii
List of Tables
5.1 Wind Tunnel Data of Aerodynamic and Control Derivatives at = 0
. . . . 47
5.2 Wind Tunnel Data of Aerodynamic and Control Derivatives at = 5
. . . . 47
5.3 Averaged Wind Tunnel Data of Aerodynamic and Control Derivatives to Ap-
proximate 2.11
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
viii
Chapter 1
Introduction
For straight and level ight, the dynamic modes for conventional airplanes decouple into two
independent sets: 1) longitudinal modes, called the short period and phugoid, and 2) lateral
modes, called the roll, Dutch-Roll, and spiral. In this thesis, we are interested in examining
the behavior of the short-period and Dutch-Roll systems.
1.1 Motivation
Research in adaptive control of nonane systems, in general, is motivated by the many
emerging applications that employ novel actuation devices, like piezo-electric lms or syn-
thetic jets, which are typically nonlinearly coupled to the dynamics of the processes they
are intended to control. Modelling for these applications varies from having accurate low
frequency models in the case of structural control problems, to having no reasonable set of
model equations in the case of active control of ows and combustion processes. From the
perspective of ight control design, development of adaptive control methods for nonane
systems has the potential of expanding the ight envelope to near-stall angles of attack ight
regimes. This is especially relevant for design of ight control systems for ghter aircraft that
are very much needed for military operations and homeland security. This thesis presents
an adaptive augmentation of a baseline linear optimal (LQR) controller to encounter for
uncertain control-dependent nonlinearities in aggressive ight regimes. Two specic systems
are considered: short-period dynamics and Dutch-Roll dynamics.
1
Amanda Young Chapter 1. Introduction 2
1.2 Problem Denition
Consider the following system:
x(t) = Ax(t) + B(u(t) + f(x(t), u(t))), x(0) = x
0
y(t) = C
x(t), (1.1)
where x R
n
is the state of the system, A R
nn
is the known system matrix, B R
nm
is the input matrix, C
is the output matrix, y R

m
is the regulated output, u R
m
is the
control input, and f R
m
is an unknown Lipshitz continuous nonlinear function dependent
on both x and u. The control objective is to determine a control input u, such that y(t)
tracks a given set of bounded smooth reference inputs r = [r
1
, . . . , r
m
]
, while all other error

signals remain bounded.
1.3 Approach
The goal of this thesis is to design a controller for two nonlinear aircraft systems with
uncertain dynamics that are nonlinear-in-control or nonane. First, a nominal system is
designed using standard linear quadratic regulator (LQR) methods for the linearized system
in the absence of uncertainties. This controller is then augmented with an adaptive element
to compensate for modelling uncertainties. The adaptive laws for approximating system
uncertainties are derived from Lyapunov-like analysis and the incremental adaptive control
signal is sought as a solution of fast dynamics which veries the assumptions of Tikhonovs
theorem from singular perturbations theory for the closed-loop system.
1.4 Overview
This thesis is presented as follows. In chapter 3, we review some preliminaries from approx-
imation theory and Tikhonovs theorem from singular perturbations theory, which are the
key elements in the synthesis and analysis. In chapter 4, we formulate the single input short-
period ghter aircraft dynamics at high angle of attack and propose a dynamic-inversion
based controller to achieve the control objective. Simulations are given for this system that
show the desired tracking performance. Next, we extend the methodology to the multi-input
Amanda Young Chapter 1. Introduction 3
multi-output roll-yaw system and show simulations. A brief summary and some recommen-
dations for future work are given in conclusion.
Chapter 2
Literature Review
2.1 Short Period Dynamics
The continuing development of ight control design methods in parallel to advances in non-
linear control theory have made controllability at high angles of attack near and beyond stall
increasingly relevant. The main challenge encountered at high angles of attack, is that the
short-period aircraft dynamics appear to be highly uncertain and nonlinear both in the states
and in the control signal [6]. Hence, traditional linearization and gain scheduling methods
are insucient for modelling and controlling the system at high-angles of attack. To obtain
a better control design and model, various methods have been explored and discussed. One
such method is to re-write the system linearly by expanding the output and control vectors
into a truncated Taylor series expansion in order to apply optimal control design [7]. Non-
linear control methods for ane-in-control systems using feedback linearization via dynamic
inversion are found in [8] - [11]. The authors of [8] have utilized the natural time-scale sepa-
ration between the faster pitch rate dynamics and the slower angle-of-attack dynamics and
assumed that the control eectors have no eect on so that can be pseudo-controlled
through the pitch rate. Without any modelling uncertainties, this controller was valid for an
angle of attack range of up to 55
. In [9], two nonlinear controllers are implemented by sep-

arating the 6 DOF system dynamics into two time scales: angular rates which are fast, and
angles of attack and sideslip, which are slow. The controller for the fast dynamics are found
by inversion of the momentum equations, and the slow dynamics controller is obtained by in-
verting the force equations. The controller performance is tested in high-delity simulations,
valid up to angles of attack of 30
. The problem with feedback linearization is that it has

4
Amanda Young Chapter 2. Literature Review 5
a low tolerance twoards modelling errors. Neural network approximations are introduced in
[10] and [11] in order to compensate for modelling uncertainties. [10] uses an o-line neural
network conguration to make an initial approximation of the nonlinear system dynamics
and a second on-line neural network approximation to estimate the inversion error. In [11],
the neural network approximation captures the nonlinear dynamics of the system for an
angle of attack range of [5
, 21
], and the inverse dynamics are computed by dieren-

tiating the output with respect to time until a term appears that contains the input u. The
authors of [12] dierentiate the nonane-in-control system with respect to time to obtain a
higher-order system, in which the derivative of the control input appears linearly. They then
apply linear control methods and integrate to obtain the structure of the controller. This
method results in a large control signal and would be susceptible to system saturation in
realistic systems. Direct projection-based adaptive control is then easily implemented and
the system is shown to be stable using a standard Lyapunov argument.
In the rst part of this thesis, we apply the theory developed in [13] to a nonane ghter
aircraft like the F-16 performing at or near the stall angle in the presence of system uncer-
tainties. The method in [13] introduces a specialized set of radial basis functions (RBF) that
retains the monotonic property with respect to control eectiveness of the original aircraft
dynamics. It further denes the adaptive controller via fast dynamics and achieves time-
scale separation between the system dynamics and the controller dynamics. The stability
and tracking results follow from Tikhonovs theorem in singular perturbations theory.
2.2 Dutch-Roll Dynamics
The Dutch-Roll is a type of lightly damped oscillatory lateral/directional motion between
sideslip, roll, and yaw that consists of an out-of-phase combination of tail-wagging and
rocking from side to side that occurs when an aircraft is disturbed laterally from equilibrium
ight. This type of motion occurs when a statically stable aircraft tries to re-establish lateral
and directional equilibrium. Dutch roll is usually the most troublesome of the natural modes
associated with the dynamics of an aircraft in free ight due to its sensitive stability balance,
which could easily result in instability [16]. Hence, it is important to model the Dutch-Roll
system as accurately as possible. Traditional Dutch-Roll approximations are made by lin-
earizing the full nonlinear equations of motion so that the dynamic modes can be decoupled
and evaluated [17], [18]. Additionally, the aerodynamic stability and control derivatives are
taken from wind-tunnel data and the eigenvalues of the linearized equations can be numeri-
Amanda Young Chapter 2. Literature Review 6
cally derived, so that a linear controller can be designed [15], [19] - [24]. Typically, these types
of approximations neglect the rolling rate, bank angle, rolling momentum and aerodynamic
coupling, which decreases the accuracy of the models [25]. An improved closed-loop form
is derived in [16] using Taylor series expansion to approximate the roll-damping derivative.
The eects of inertial, gyroscopic, and aerodynamic coupling are included for a linear model
in [26]. The problem with linearization approaches, is that although they are accurate for
low angles of attack, they becomes less than ideal for aggressive aircraft maneuvers and high
angle of attack ight regimes when the nonlinear eects become signicant enough to drive
the system to instability.
In the second part of this thesis, we extend the method developed in [14] to uncertain
nonane-in-control multi-input multi-output (MIMO) systems, which describe the lateral/directional
dynamics of an F-16 at dierent angles of sideslip in the presence of system uncertainties.
Chapter 3
Mathematical Preliminaries
In this chapter, we recall some facts that will be used for derivation of main results. First,
some results from nonlinear systems theory is given. Next, we summarize the main result
from [2] on approximation of monotonic functions. Lastly, Tikhonovs theorem is stated
from [1], which enables denition of the adaptive controller via fast dynamics.
3.1 Preliminaries on Nonlinear Systems Theory
Lemma 1 ([3], p.116) A time-varying linear system
x = A(t)x, (3.1)
is globally exponentially stable if
1. for any time t 0, the eigenvalues of A(t) have negative real parts, that is,
> 0, t 0, (A(t)) , (3.2)
where (A(t)) denotes the eigenvalues of A(t),
2. the matrix A(t) remains bounded:
_

0
A
(t)A(t)dt < . (3.3)

7
Amanda Young Chapter 3. Mathematical Preliminaries 8
Lemma 2 (Comparison Lemma [1], p.102) Consider the scalar dierential equation
x = f(t, x), x(t
0
) = x
0
, (3.4)
where f(t, u) is continuous in t and locally Lipshitz in x, for all t 0. Let [t
0
, T) (where
T could be innity) be the maximal interval of existence of the solution x(t). Let v(t) be a
continuous function which satises the dierential inequality
v(t) f(t, v(t)), v(t
0
) x
0
. (3.5)
Then, v(t) x(t) for all t [t
0
, T).
Lemma 3 ([1], p.368) Consider the system
z = f(z + h(), ) g(z, ) (3.6)
and suppose g(z, ) is continuously dierentiable and the Jacobian matrices
g
z
and
g
satisfy
_
_
_
_
g
z
(z, )
_
_
_
_
L
1
,
_
_
_
_
g
(z, )
_
_
_
_
L
2
|z| (3.7)
for all (z, ) D , where R
m
and D = z R
n
[ |z| < r. Let k
1
,
1
, and
0
be positive constants with r
0
<
r
k
1
, and dene D
0
= z R
n
[ |z| < r
0
. Assume that the
trajectories of the system satisfy
|z(t)| k
1
|z(0)|e
1
t
, z(0) D
0
, , t 0. (3.8)
Then, there is a function V : D
0
R that satises
c
1
|z|
2
V (z, ) c
2
|z|
2
(3.9)
V
z
g(z, ) c
3
|z|
2
(3.10)
_
_
_
_
V
z
_
_
_
_
c
4
|z| (3.11)
_
_
_
_
V
_
_
_
_
c
5
|z|
2
(3.12)
for positive constants c
1
, c
2
, c
3
, c
4
, c
5
and all z D, . Moreover, if all the assumptions
hold globally (in z), then V (z, ) is dened and satises (3.9) through (3.12) on R
n
.
3.2 Preliminaries on Approximation Theory
In this section, we recall the main result from [2]. To follow the notations in [2], let ^,
R, and R
r
denote the set of natural numbers, the set of real numbers, and the set of real
r-vectors, respectively. Let L
p
(R
r
), L
(R
r
), ((R
r
), (
c
(R
r
), denote the usual spaces of R-
valued maps f : R
r
R such that f is p
th
power integrable, essentially bounded, continuous,
and continuous with compact support. Let further L
p
+
(R
r
), L
+
(R
r
), (
+
(R
r
), (
c+
(R
r
) be the
spaces of positive valued maps f : R
r
R
+
respectively. Since (
c
(R
r
) is dense in L
p
(R
r
)
( [4], p. 69), then for every f L
p
(R
r
) there exists an f
c
(
c
(R
r
) such that [[f
c
f[[
p
,
where | |
p
denotes the p-norm. The main result from [2] is summarized via the following
theorem:
Theorem 1 Let K : R
r
R be an integrable bounded function such that K is continuous
almost everywhere and
_
R
r
K(x)dx ,= 0. Then the family o
K
of functions q : R
r
R dened
as
q(x) =
M
i=1
w
i
K
_
x z
i
_
, (3.13)
where M ^, > 0, w
i
R, and z
i
R
r
, is dense in L
p
(R
r
) for every p [1, ).
The proof of this theorem in [2] gives an explicit expression for the coecients w
i
in (3.13)
as:
w
i
=
1
r
f
c
(
i
)
_
2T
n
_
r
1
_
R
r
K(x)dx
, (3.14)
where the set
i
R
r
: i = 1, ..., n
r
consists of all points in [T, T]
r
of the form
[T + (2i
1
T/n), ...., T + (2i
r
T/n)], i
1
, i
2
, .., i
r
= 1, 2..., n,
in which T is a number such that sup(f
c
) [T, T]
r
, where sup() is dened as the
supremum or least upper bound. This consequently implies that if f L
p
+
(R
r
), then f
c

(
c+
(R
r
), and consequently the coecients w
i
in (3.14) are positive, i.e. w
i
> 0. Let o
G
be
the subclass of functions from o
K
with positive w
i
s. Then o
G
is dense in L
p
+
(R
r
). We also
notice that the Gaussians given by
_
x x
c
i
i
_
= e
(xx
c
i
)
2
2
i
,
where x
c
i
is the center, while
i
is the width parameter, represent one particular choice of
K.
3.3 Preliminaries on Singular Perturbations Theory
Consider the problem of solving the system
0
:
_
x(t) = f(t, x(t), u(t), ), x(0) = ()
u(t) = g(t, x(t), u(t), ), u(0) = (),
(3.15)
where : () and : () are smooth. Additionally, assume that f and g are
continuously dierentiable in their arguments for (t, x, u, ) [0, ] D
x
D
u
[0,
0
],
where D
x
R
n
and D
u
R
m
are domains,
0
> 0. In addition, let
0
be in standard form,
that is,
0 = g(t, x, u, 0) (3.16)
has k 1 isolated real roots u = h
i
(t, x), i 1, . . . , k for each (t, x) [0, ] D
x
.
Choosing a particular i and keeping it xed, the subscript i is dropped. Let (t, x) =
u h(t, x).
In singular perturbations theory, the system given by
00
: x(t) = f(t, x(t), h(t, x(t)), 0), x(0) = (0), (3.17)
is called the reduced system, and the system given by
b
:
d
d
= g(t, x, + h(t, x), 0), (0) =
0
h(0,
0
) (3.18)
is called the boundary layer system, where
0
= (0) and
0
= (0), (t, x) [0, ) D
x
are
treated as xed parameters. The new time scale is related to the original time t via the
relationship =
t
. The following result is due to Tikhonov [1], [5]:

Theorem 2 Consider the singular perturbation system
0
given in (3.15) and let u = h(t, x)
be an isolated root of (3.16). Assume that the following conditions are satised for all [t, x, u
h(t, x), ] [0, )D
x
D
v
[0,
0
] for some domains D
x
R
n
and D
v
R
m
, which contain
their respective origins:
A1. On any compact subset of D
x
D
v
, the functions f, g, their rst partial derivatives with
respect to (x, u, ), and the rst partial derivative of g with respect to t are continuous
and bounded, h(t, x) and
_
g
u
(t, x, u, 0)
have bounded rst derivatives with respect to

their arguments,
_
f
x
(t, x, h(t, x))
is Lipschitz in x, uniformly in t, and the initial data

given by and are smooth functions of .
A2. The origin is an exponentially stable equilibrium point of the reduced system
00
given
by equation (3.17). There exists a Lyapunov function V : [0, ) D
x
[0, ) that
satises
W
1
(x) V (t, x) W
2
(x)
V
t
(t, x) +
V
x
(t, x)f(t, x, h(t, x), 0) W
3
(x)
for all (t, x) [0, )D
x
, where W
1
, W
2
, W
3
are continuous positive denite functions
on D
x
, and let c be a nonnegative number such that x D
x
[ W
1
(x) c is a compact
subset of D
x
.
A3. The origin is an equilibrium point of the boundary layer system
b
given by equation
(3.18), which is exponentially stable uniformly in (t, x).
Let R
v
D
v
denote the region of attraction of the autonomous system
dv
d
= g(0,
0
, v +
h(0,
0
), 0), and let
v
be a compact subset of R
v
. Then for each compact set
x
x
D
x
[ W
2
(x) c, 0 < < 1, there exists a positive constant
such that for all t 0,
0

x
,
0
h(0,
0
)
v
and 0 < <
,
0
has a unique solution x
on [0, ) and
x
(t) x
00
(t) = O()
holds uniformly for t [0, ), where x
00
(t) denotes the solution of the reduced system
00
in (3.17).
The following Remarks of [1] will be useful for verifying Assumption A3:
Remark 1 Verication of Assumption A3 can be done via a Lyapunov argument. If there
exists a Lyapunov function V : [0, ) D
x
D
v
that satises
c
1
||
2
V (t, x, v) c
2
||
2
(3.19)
V
g(t, x, + h(t, x)) c

3
||
2
, (3.20)
1
, c
2
, c
3
, c
4
and for all (t, x, v) [0, ) D
x
D
v
, then Assumption
A3 is satised.
Remark 2 Assumption A3 can be locally veried by linearization. Let denote the map
g(t, , + h(t, ), ). It can be shown that if there exists
0
> 0 such that the Jacobian
matrix

satises the eigenvalue condition

Re
_
(t, x, h(t, x), 0)

__

0
< 0 (3.21)
for all (t, x) [0, ) D
x
, then Assumption A3 is satised.
Proof. The boundary layer system (3.18) can be written as a perturbation of its linearization
at the equilibrium = 0,
d
d
= A(t, x) + (g(t, x, + h(t, x), 0) A(t, x))
. .
(t,x,)
, (3.22)
where
A(t, x) =
g
(t, x, h(t, x), 0). (3.23)

From Assumption A1 of Tikhonovs theorem,
g
(t, x, +h(t, x), 0) and g(t, x, +h(t, x), 0)

are Lipshitz in so that the perturbation term (t, x, ) is also Lipshitz and satises
|(t, x, )|
p
k||
2
p
, (t, x) [0, ) D
x
, (3.24)
where | |
p
denotes the p-norm. Assumption A1 implies that A(t, x) is bounded for all
time and (3.21) implies that the eigenvalues of A(t, x) have negative real parts, so it can be
concluded from Lemma 1 that the nominal system
d
d
= A(t, x) (3.25)
is globally exponentially stable for all (t, x) [0, ) D
x
and the trajectories of the linear
time-varying system (3.25) satisfy
|v(t)|
p
k
1
|v(0)|
p
e
1
t
, v(0) D
, x D
x
, t 0 (3.26)
for positive constants k
1
,
1
.
The partial derivatives of A(t, x) with respect to its arguments are bounded as a result
of Assumption A1 of Tikhonovs theorem and the time-varying nominal system (3.25) is
globally exponentially stable. Hence, it follows then from applying Lemma 3, that there
exists a Lyapunov function V : [0, ) D
x
D
R that satises the inequalities

c
1
||
2
p
V (t, x, ) c
2
||
2
p
V
A(t, x) c
3
||
2
p
(3.27)
_
_
_
_
V
(t, x, )
_
_
_
_
p
c
4
||
p
for some positive constants c
1
, c
2
, c
3
, and c
4
. Taking the derivative of the Lyapunov function
V along the trajectory of (3.18) yields
V (t, x, ) =
_
V
(t, x, )
_
g(t, x, + h(t, x), 0)
=
_
V
(t, x, )
_
(A(t, x) + (t, x, ))
c
3
||
2
p
+ c
4
k||
3
p
c
3
||
2
p
for ||
p

c
3
c
4
k
. This in turn implies
V (t, x, )
c
3
c
2
V (t, x, ), (3.28)
and, consequently,
V () V (0) e
c
3
c
2
. (3.29)
Using Lemma 2 (Comparison Lemma), it follows that
|()|
p

_
c
2
c
1
|(0)|
p
e
c
3
2c
2
, |(0)|
c
3
c
4
k
_
c
1
c
2
, x D
x
, 0 (3.30)
1
, c
2
, c
3
, c
4
, k. Thus, it can be concluded that Assumption A3 of
Tikhonovs theorem on exponential stability of the time-varying boundary layer system at
the origin can be veried locally by linearization.
Chapter 4
Adaptive Controller for Nonane
Single Input System
4.1 Problem Formulation for Short Period Dynamics
Neglecting the inuence of gravity and thrust, the short-period dynamics of a rigid aircraft
performing high angle of attack maneuvers can be given as
_
_
=
L
(
0
)
V
+ q
L
e
(
0
)
V

e
q = M
0
() + M
q
()q + M
e
(,
e
),
(4.1)
where q is the pitch rate, is the angle of attack,
0
is the trim angle,
e
is the incre-
mental elevator deection, L
(
0
) is the known lift curve slope at
0
, L
e
(
0
) is the known
lift eectiveness due to elevator deection, V is the trimmed airspeed, M
q
() is the pitch
damping, and M
0
() and M
e
are the pitching moment components. From (4.1), it fol-
lows that the pitch dynamics depend on M
e
, which is nonlinear with respect to and
e
.
Hence, the short-period dynamics of an aircraft at high angles of attack are nonlinear and
nonane-in-control.
In general, the terms M
0
(), M
q
(), and M
e
(,
e
) are unknown. However, some partial
knowledge of aerodynamic stability and control derivatives are usually available from wind-
tunnel experiments and theoretical predictions. Incorporating prior known data, the short-
14
Amanda Young Chapter 4. Adaptive Controller for Nonane Single Input System 15
period dynamics of (4.1) can be rewritten as:
_
_
=
L
(
0
)
V
+ q
L
e
(
0
)
V

e
q = M
(
0
) + M
q
(
0
)q + M
e
(
0
, 0)
e
+
(M
0
() M
(
0
))
. .
M
0
()
+ (M
q
() M
q
(
0
))
. .
M
q
()
q + (M
e
(,
e
) M
e
(
0
, 0),
e
)
. .
M
e
(,
e
)
,
(4.2)
where it is assumed that M
(
0
), M
q
(
0
), and M
e
(
0
, 0) are the known constant stability
and control derivatives. Additionally, the lift derivative L
e
(
0
) is known to be small with
respect to airspeed V , so that for control design purposes, one can assume
L
e
(
0
)
V
0, (4.3)
which leads to the following model description:
_
_
=
L
V
+ q
q = M
+ M
q
q + M
e
+ M
e
(M
0
() + M
q
()q + M
e
(,
e
))
. .
f(,
e
)
.
(4.4)
This can be rewritten in state space form
_
(t)
q(t)
_
=
_

L
V
1
M
M
q
_
. .
A
_
(t)
q(t)
_
. .
x(t)
+
_
0
M
e
_
. .
b
(
e
(t) + f((t),
e
(t))), (4.5)
y(t) = [1 0]
. .
c
x(t),
where f(,
e
) is unknown to the controller. Based on wind-tunnel experiments and using
curve-tting methods, f(,
e
) is known to have the structure
f(,
e
) =
_
(1 C
0
) e

2
2
2
+ C
0
_
(tanh (
e
+ h) + tanh (
e
h) + 0.01
e
) , (4.6)
which is well-dened for all ,
e
R and continuously dierentiable, where e

2
2
2
is a
Gaussian function with width and C
0
, h are positive constants. A scaled plot demonstrating
the structure of f(,
e
) is shown in Figure 4.3. When the angle of attack is at the trim angle,
then full elevator control eectiveness is obtained. If control is small, the control surface is
within a boundary layer and does not produce an aerodynamic moment, so this is portrayed
in the small elevator control induced moment term 0.01
e
. As the control eort gets large, the
surface stalls as the system reaches saturation and stops producing any additional moment
which is approximated by the hyperbolic tangent functions. The control objective is to design
elevator surface deection
e
(t) =
e
nom
(t)
e
ad
(t) (4.7)
so that the angle of attack (t) tracks a commanded input r
cmd
(t), while all other signals in
the system remain bounded. In (4.7),
e
nom
(t) is the nominal control elevator deection for
the nominal (known) linear model, while
e
ad
(t) is the incremental adaptive augmentation
to compensate for unknown f((t),
e
(t)).
4.2 Ideal Reference Model
To nd the nominal control elevator for the short period aircraft dynamics (4.1), consider
the nominal linear model of the longitudinal dynamics (4.5) in the absence of uncertainties
(f((t),
e
(t)) = 0,
e
ad
= 0) and design an LQR controller to achieve the tracking objective:
(t) r
cmd
(t) as t for a smooth r
cmd
(t). Towards that end, let
A
r
=
_

L
V
1
M
M
q
_
. .
A
b [k
lqr
k
lqr
q
]
. .
k
lqr
(4.8)
=
_
V
1
M
e
k
lqr
. .
a
r
2,1
M
q
M
e
k
lqr
q
. .
a
r
2,2
_
_
,
where the gain k
lqr
= R
1
b
P is found by solving the corresponding Riccati equation for

the unique positive denite symmetric matrix P:
Q + PA + A
P PbR
1
b
P = 0, (4.9)
in which Q is a positive denite matrix and R is a positive scalar control weight. This leads
to the following form of the closed-loop reference system:
_

r
(t)
q
r
(t)
_
= A
r
_

r
(t)
q
r
(t)
_
+ br(t),
r
(0) =
r
0
, q
r
(0) = q
r
0
, (4.10)
y
r
(t) = c
x
r
(t).
Due to the time-varying nature of the commanded reference input r
cmd
(t) and the denition
of the reference model (4.10), setting r(t) = r
cmd
(t) will cause y
r
(t) to track r
cmd
(t) with
bounded errors. To force y
r
(t) to track the commanded signal r
cmd
(t) asymptotically and
guarantee zero steady state tracking error, dene a new reference input r
cmd
(t) as
r
cmd
(t) =
1
M
e
_
r
cmd
(t) r
cmd
(t)
_
a
r
2,2
V
_
r
cmd
(t)
_
a
r
2,2
L
V
+ a
r
2,1
__
. (4.11)
The proof of this can be obtained by noticing that for the system (4.10), one can compute
the transfer function
y
r
(s)
r(s)
= c(sI A
r
)
1
b, where I is the 2 2 identity matrix, s is the
Laplace variable, and y
r
(s) and r(s) are the Laplace transforms of y
r
(t) and r(t) respectively,
as
y
r
(s)
r(s)
= M
e
_
s
2
s
_
a
r
2,2
V
_
_
a
r
2,2
L
V
+ a
r
2,1
__
1
.
It is easy then to dene
r
cmd
(s) =
1
M
e
_
s
2
s
_
a
r
2,2
V
_
_
a
r
2,2
L
V
+ a
r
2,1
__
r
cmd
(s), (4.12)
whose representation in the time domain is given by (4.11), so that y
r
(s) = r
cmd
(s), which
results in asymptotic tracking r
cmd
(t). It follows then that the nominal system for the
short-period aircraft dynamics (4.5) is
_

r
(t)
q
r
(t)
_
= A
r
_

r
(t)
q
r
(t)
_
. .
x
r
(t)
+b r
cmd
(t),
r
(0) =
r
0
, q
r
(0) = q
r
0
, (4.13)
y
r
(t) = c
x
r
(t).
The nominal elevator control input is then
e
nom
(t) = k
lqr
_
(t)
q(t)
_
+ r
cmd
(t), (4.14)
which can be applied to the dynamics in (4.5) to yield
_
(t)
q(t)
_
= A
r
_
(t)
q(t)
_
+ b r
cmd
(t) + b
_
e
ad
(t) + f( (t),
e
nom
(t)
e
ad
(t))
_
. (4.15)
4.3 RBF Approximation
In this section, we address approximation of monotonic nonlinear functions using the result
from [2]. Towards that end, dene
g(,
e
) =
f
e
(,
e
) (4.16)
=
_
(1 C
0
) e

2
2
2
+ C
0
_
_
2.001 tanh
2
(
e
+ h) tanh
2
(
e
h)
_
0.01C
0
> 0 for all C
0
> 0.
Using Leibnitz formula, f(,
e
) can be expressed as
f(,
e
) = f(, 0) +
_

e
0
g(, ) d, (4.17)
where the integral term is strictly positive due to the positivity of g(,
e
) dened in (4.16).
Because f(, 0) (
c
(R) and g(,
e
) (
c+
(R
2
), Theorem 1 can be applied to conclude
that each component can be approximated by a linear combination of radial basis functions
(RBFs) arbitrarily closely on a compact set. Consider a set of radial basis functions (RBF)
: R R R given by:
(,
e
) =
1
() +
2
(,
e
), (4.18)
where
1
()
1
() =
N
i=1
i
e
(
c
i
)
(
c
i
)
2
i
,
i
> 0, (4.19)
2
(,
e
) w
2
(,
e
) =
M
j=1
w
j
_
_
_

e
0
e
(
c
j
)
(
c
j
)
2
j
d
_
_
, (4.20)
[
,
j
> 0,
and M, N are positive constants,
1
() is the vector of Gaussians dependent only on , and
2
(,
e
) is the vector of the integrals of the Gaussians dependent on both and
e
. Since
g(,
e
) > 0, then following (3.14), the unknown coecients w
i
are positive. The vectors
c
i
, i = 1, ..., N,
c
j
= [
c
j

e
c
j
]
, j = 1, ..., M represent the xed centers of the RBFs,

i
,
j
are the xed widths of the basis functions respectively, and the unknown parameters
are
i
R and w
j
R
+
. It is easy to see from (4.18) - (4.20) that
sgn
_
e
(,
e
)
_
= sgn
_
e
(,
e
)
_
> 0. (4.21)
Then,
f(,
e
) = f(, 0) +
_

e
0
g(, ) d
= W
(,
e
) + (,
e
), for (,
e
)
e
, [(,
e
)[
,(4.22)
where g(,
e
) is given in (4.16), W
= [
, (,
e
) = [
1
()
2
(,
e
)]
, w
i
> 0,
(,
e
) is the uniformly bounded approximation error,
e
are compact sets of R, and
is a constant.
4.4 State Predictor and Adaptive Law
A state predictor for the system (4.15) can be constructed as:
_

(t)
q(t)
_
= A
r
_
(t)
q(t)
_
+ b r
cmd
(t) + b
_
e
ad
(t) +

W
(t)((t),
e
(t))
_
, (4.23)
(0) =
0
, q(0) = q
0
,
where A
r
is dened in (4.8), r
cmd
is given in (4.11), and

W(t) is used to estimate the unknown
constant vector W. The prediction error signal e
s
(t) is dened as
e
s
(t) =
_
(t) (t)
q(t) q(t)
_
. (4.24)
Prediction error dynamics can be written as
e
s
(t) = A
r
e
s
(t) + b
_

W
(t)((t),
e
(t)) ((t),
e
(t))
_
, e
s
(0) = e
s
0
, (4.25)
and

W(t) =

W(t) W. The following adaptive law
W(t) = Proj
_

W(t), ((t),
e
(t) e
s
(t) P
0
B
_
,

W(0) = W
0
, (4.26)
where is a positive denite matrix of adaptation rates, Proj ( , ) denotes the projection
operator, and P
0
is the solution to the Lyapunov equation A
r
P
0
+P
0
A
r
= Q
0
for a positive
denite matrix Q
0
, ensures that the prediction error dynamics in (4.25) and the parameter
estimation error

W(t) are bounded independent of
e
(t). Indeed, consider the following
Lyapunov function candidate:
V (e
s
(t),

W(t)) = e
s
(t)P
0
e
s
(t) +

W
(t)
1

W(t). (4.27)
Taking the time derivative along the trajectories (4.25), (4.26) gives
V (t) = e
s
(t)Q
0
e
s
(t) + 2e
s
(t)P
0
B(t)
|e
s
|
min
(Q
0
)|e
s
| + 2|e
s
| |P
0
B|
,
where
min
(Q
0
) denotes the minimum eigenvalue of Q
0
. The Projection Operator ensures
boundedness of parameter errors so that outside the compact set
_
|e
s
| 2
|P
0
B|
min
(Q
0
)
_
_
|

W| W
_
, (4.28)
where W
is the maximum allowable upper-bound selected by the Projection operator and

| | indicates the 2-norm, the derivative of the Lyapunov function (4.27) is negative denite.
Following standard invariant set arguments, it can be concluded that if the initial errors are
within the largest level set of the Lyapunov function for which the RBF approximation has
been dened, then the error dynamics (4.25) are ultimately bounded with respect to e
s
(t)
and

W(t). Furthermore, this ultimate bound can be any number larger than the value of
the Lyapunov function on the minimum level set embracing the compact set in (4.28). That
is,
|| >
_
_
_
_
V
_
2
|P
0
B|
min
(Q
0
)
, W
__
_
_
_
= 4
(
)
2
|P
0
B|
2
max
(P
0
)
2
min
(Q
0
)
+ (W
)
2
|
1
|,
where V is the Lyapunov function (4.27) and
max
(P
0
) denotes the maximum eigenvalue of
P
0
.
Remark 3 Notice that since w
i
are positive, the compact set in the application of the
Projection operator can be selected so that w
i
(t) remain positive for all t 0, that is,
w
i
(t) > w
i
0
> 0 for all i = 1, ..., M.
4.5 Nonane Control Design
Dene the tracking error between the predictor of (4.23) and the reference system (4.13) as
e(t) =
_
(t)
r
(t)
q(t) q
r
(t)
_
. (4.29)
The error dynamics are then:
e(t) = A
r
e(t) + b
_
e
ad
(t) +

W
(t)((t),
e
(t))
_
, e(0) = e
0
.
Dynamic inversion based control is to be determined from the solution of
e
ad
(t) =

W
(t)((t),
e
nom
(t)
e
ad
(t), (4.30)
so that the resulting closed-loop error dynamics are asymptotically stable, i.e. e(t) = A
r
e(t).
Due to the nonane nature of the longitudinal dynamics, (4.30) cannot be solved explicitly
for
e
ad
. We construct fast dynamics to determine the solution for
e
ad
:
e
ad
= sign
_
f
e
ad
_
f (t, e,
e
ad
),
e
ad
(0) =
e
ad
0
, (4.31)
where 0 < 1 and
f (t, e,
e
ad
) =
e
ad

W
(t)((t),
e
nom
(t)
e
ad
).
As an extension of Tikhonovs theorem, recall the main theorem of [13]:
Theorem 3 Assume that the following conditions are satised for all [t, e, u h(t, e), ]
[0, ) R
2
R [0,
0
]:
B1. On any compact subset of R
2
R, the function f and its rst partial derivative with
respect to (t, e, u) are continuous and bounded, h(t, e) and
f
u
(t, e, u) have bounded rst
derivatives with respect to their arguments, and
f
e
as a function of (t, e, h(t, e)) is
Lipshitz in e, uniformly in t.
B2. (t, e, )
f
u
(t, e, + h(t, e)) is bounded below by some positive number for all (t, e)
[0, ) R
2
.
Then the origin of (4.31) is exponentially stable. Moreover, let
be a compact subset of
R
, where R
R denotes the region of attraction of the autonomous system

d
d
= sign
_
f
e
ad
_
f (0, e
0
, + h(0, e
0
)). (4.32)
The for each compact subset
e
R
2
, there exists a positive constant, and a T > 0 such
that for all t 0, e
0

e
, u
0
h(t, e
0
)
nu
and 0 < <
, the system of equations (4.23),

(4.31 - 4.32) has a unique solution
(t), q
(t) on [0, ), and

_
(t)
q(t)
_
=
_

r
(t)
q
r
(t)
_
+ O().
Let
e
ad
(t) = h(t, e) be an isolated root of f (t, e,
e
ad
) = 0, (t, ) =
e
ad
(t) h(t, ),
and =
t
. The existence of such a root is guaranteed by noting that there must exist
a point of intersection between

W
(t)((t),
e
nom
(t)
e
ad
(t)), a set of Gaussians that
changes amplitude depending on the value of the positive multipliers

W
(t), and
e
ad
(t). An
illustration of this is shown in Figure 4.1.
The reduced system for (4.29) is
e(t) = A
r
e(t), e(0) = e
0
,
and the boundary layer system is
d
d
= sign
_
f
e
ad
_
f (t, e, + h(t, e)), (0) =
e
ad
0
h(t, e
0
). (4.33)
To apply Tikhonovs theorem of (Theorem 2), it is necessary to show that all of its assump-
tions are satised. A1 is satised by noting that the function f and its partial derivatives
with respect to (t, e,
e
ad
) are continuous and bounded, h(t, e)
f
e
ad
(t, e,
e
ad
) has bounded
rst derivatives with respect to its arguments, and
f
e
(t, e, h(t, e)) is Lipshitz in e, uniformly
in t. Because the matrix A
r
is Hurwitz by choice, the origin of the reduced system (4.33) is
an exponentially stable equilibrium point and there exists a Lyapunov function that satises
20 15 10 5 0 5 10 15 20
20
15
10
5
0
5
10
15
20
e
ad
Existence of Isolated Roots
e
ad
hat W
T
(,
e
nom

e
ad
)
Figure 4.1: Intersection of the functions
e
ad
(t) and

W
(t)((t),
e
nom
(t)
e
ad
(t)).
the inequalities of A2. To satisfy A3, note from (4.18) - (4.21) that
f
e
ad
= 1

W
(,
e
nom

e
ad
)
e
ad
= 1
w
2
(,
e
nom

e
ad
)
e
ad
= 1
_
_
M
j=1
w
j
_
_
_
e
nom
e
ad
0
e
(
c
j
)
(
c
j
)
2
j
d
_
_
_
_
e
ad
= 1 +
M
j=1
w
j
_
_
e
(
c
j
)
(
c
j
)
2
j
_
_
> b
0
> 0, (4.34)
where j = 1, . . . , M, w
j
s are positive (Remark 3), and M,
j
, ,
c
j
are introduced in (4.20).
Hence, linearization of the boundary layer system (4.33) with respect to implies that
the boundary layer system has locally exponentially stable origin. Following Remark 2,
Assumption A3 is satised. The complete controller consists of (4.23), (4.26), (4.30), (4.31)
and Tikhonovs theorem ensures that for the system of equations given by (4.23), (4.31),
there exists a unique solution , q that tracks
r
, q
r
, and
_
(t)
q(t)
_
=
_

r
(t)
q
r
(t)
_
+ O().
4.6 Simulations
Consider the short-period dynamics of an F-16 model trimmed at an airspeed of 502 ft/sec
and angle of attack = 2.11
. From [15], the nominal system data are:

A =
_
1.0190 1
0.8223 1.0774
_
, b =
_
0
0.1756
_
, (4.35)
which leads to open-loop system eigenvalues
1
= 1.9554,
2
= 0.1409. The commanded
reference input of interest to track is
r
cmd
(t) =
0.5
1 + e
t8
+
1
1 + e
t30
0.5.
The Ricatti equation is solved with the following weighting matrices:
Q =
_
8 0
0 0.5
_
, R = 0.01, (4.36)
which leads to the following LQR gains: k
lqr
= [16.8696 10.5911]. The new reference
input following (4.11) is dened as
r(t) = 5.6948 r
cmd
(t) 22.5291 r
cmd
(t) 29.2296r
cmd
(t), (4.37)
so that the reference model for the nominal values of the aircraft system is
x
r
(t) =
_
1.0190 1
2.1400 2.9372
_
. .
A
r
x
r
(t) + b r(t),
having its closed loop eigenvalues at
lqr
= 1.9781 1.1045i. The plot in Figure 4.4 shows
the tracking result of the ane system in the absence of uncertainties for initial conditions
r
0
= 2.11
, q
r
0
= 0
. Next, consider when f((t),

e
(t)) cannot be neglected and is unknown.
For simulation purposes, the parameter is chosen as 0.15 so that the Gaussian has a width
of 60
(see Figure 4.2), a range that contains both high and low angles of attack. The constant
C
0
is set to 0.3 so that at high angles of attack, a maximum of 30% control eectiveness
can be achieved, and the constant h = 3. Figure 4.3 demonstrates the complete structure
of f(,
e
). As seen from Figure 4.5, the performance of LQR is violated in the presence of
uncertainties and adaptive control must be implemented to restore desired system behavior.
The nonlinearity [f(,
e
)[ 0.9 causes a 6
loss of tracking in and an 8
loss of tracking
in q seen in Figure 4.6.
The state predictor is designed with 12 RBFs. Of these, 4 are
1
()-type Gaussians dis-
tributed over [30
, 30
] with the step size equal to 60
and the width = 1. The remain-

ing 8
2
(,
e
)-type basis functions were distributed over [30
, 30
],
e
ad
[30
, 30
]
with step size 60
and width = 5. The norm upper bound is W
= 10, the lower bound

for the positive widths w is 0.01, and the adaptation gain is = 0.2. The following state
predictor is implemented:
_

(t)
q(t)
_
= A
r
_
(t)
q(t)
_
+ b
_
e
ad
(t) +

(t)
1
((t)) + w
(t)
2
((t),
e
(t))
_
.
Then:
e(t) = A
r
e(t) + b
_
e
ad
(t) +

(t)
1
((t)) + w
(t)
2
((t),
e
(t))
_
.
To achieve the desired performance, the following equation needs to be solved for adaptive
elevator control,
e
ad
:
e
ad
(t) =

(t)
1
((t)) + w
(t)
2
((t),
e
nom
(t)
e
ad
(t)),
50 0 50
0.4
0.5
0.6
0.7
0.8
0.9
1
Gaussian as a function of angle of attack, C
0
= 0.3, = 0.15 radians

0
in degrees
(
1

C
0
)
e
0
)
2
/
2
2

+

C
0
Figure 4.2: Gaussian as a function of with an angle of attack range of 60
.
where
nom
(t) is dened in (4.14). Obviously, this cannot be solved in terms of analytical
functions.
The fast dynamics are designed as:
0.05
e
ad
(t) =
e
ad
(t)

(t)
1
((t)) w
(t)
2
((t),
e
nom
(t)
e
ad
(t)).
Initial conditions used for the simulations are:
[(0) q(0)] = [2.11
]
[ (0) q(0)] = [2.11
].
The plots are given in in Figures 4.7,4.8,4.9. Figure 4.7 shows the closed-loop tracking
performance of the commanded reference input r
cmd
(t) by the estimator state (t), the
reference state
r
(t) and the actual state (t). Figure 4.8 shows the closed-loop tracking
performance of the reference state q
r
(t) by the estimator state q(t) and the actual state q(t),
and Figure 4.9 shows the adaptive control eort
e
ad
(t) which closely matches the unknown
nonlinearity, f((t),
e
(t)) as a function of time. We note the convergence ability of the
predictor to the system states and the asymptotic tracking of r
cmd
(t) by . All the states
and control input remain in the domain of the RBF approximation.
40 30 20 10 0 10 20 30 40
2
1.5
1
0.5
0
0.5
1
1.5
2
f(,
e
) as a function of control effort
e
, C
0
= 0.3, = 0.15
control effort
e
(degrees)
f
(
e
)
Figure 4.3: f(,
e
) as a function of elevator control
e
.
0 5 10 15 20 25 30
30
20
10
0
10
20
30
Reference System Tracking of Commanded Input
time, t in seconds
r
cmd
(t) in degrees
r
(t in degrees)
q
r
(t) in degrees/sec
Figure 4.4: Short-period nominal system tracking with respect to dened commanded input
r
cmd
(t).
0 5 10 15 20 25 30
40
30
20
10
0
10
20
30
40
LQR Performance Violated in the Presence of Uncertainties
time, t in seconds
r
cmd
(t) in degrees
r
(t in degrees)
q
r
(t) in degrees/sec
Figure 4.5: LQR performance violated in the presence of uncertainties.
0 5 10 15 20 25 30
2
1
0
1
2
3
4
5
6
7
8
Loss of Tracking
time, t in seconds
d
e
g
r
e
e
s

q
Figure 4.6: Loss of tracking of and q caused by nonlinearity f(,
e
) of order O(0.9).
0 5 10 15 20 25 30
30
20
10
0
10
20
30
Tracking Performance: = 0.05
time, t in seconds
(
t
)
,

d
e
g
r
e
e
s
r
cmd
(t)
r
(t)
(t)
hat (t)
Figure 4.7: States (t),
r
(t), (t) of short-period system.
0 5 10 15 20 25 30
30
20
10
0
10
20
30
time, t in seconds
q
(
t
)
,

d
e
g
r
e
e
s
q
r
(t)
q(t)
hat q(t)
Figure 4.8: States q(t), q
r
(t), q(t) of short-period system.
0 5 10 15 20 25 30
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Adaptive Control Effort
e
ad
(t)
time, t in seconds
e
ad
(t) in degrees
f((t),
e
(t))
Figure 4.9: Adaptive elevator control,
e
ad
(t)of short-period system versus unknown nonlin-
earity f((t),
e
(t)).
Chapter 5
Adaptive Controller for Nonane
Multi-Input System
5.1 Problem Formulation for Dutch-Roll Dynamics
From the six degree-of-freedom at-earth, body-axis aircraft model, the kinematic Euler roll
rate equation is known to be
= p + tan (q sin + r cos ), (5.1)

where p, q, r are the body axis roll, pitch and yaw rates respectively, is the pitch angle,
and is the roll angle. Denote the trimmed pitch angle as
0
. For small roll angle , the
following holds:
= p + r tan
0
. (5.2)
For a trimmed angle of attack
0
, one can write
p
s
= p cos
0
+ r sin
0
r
s
= r cos
0
p sin
0
,
(5.3)
where p
s
, r
s
are the stability axis roll and yaw rates. Solving (5.3) for p, r and directly
substituting into (5.2) yields
= (cos
0
+ sin
0
tan
0
)p
s
+ (cos
0
tan
0
sin
0
)r
s
. (5.4)
31
Amanda Young Chapter 5. Adaptive Controller for Nonane Multi-Input System 32
An additional relationship between ight path angle , pitch angle, and angle of attack at
zero bank angle and zero angle of sideslip (AOS) is
0
=
0
0
, (5.5)
which can be substituted into (5.4) to get
=
cos
0
cos
0
p
s
+
sin
0
cos
0
r
s
. (5.6)
The angle of sideslip dynamics in stability axis assuming small angles has the form
=
1
V
(Y
+ Y
r
r
s
+ Y
a
+ Y
r
) +
_
g cos
0
V
_
r
s
, (5.7)
where V is the true airspeed, g is the gravitational constant, and
a
,
r
are aileron and rudder
control. The terms Y
, Y
r
Y
a
, Y
r
are aerodynamic stability and control derivatives that
change slowly with time and can be adequately approximated by constants. Additionally,
Y
a
, Y
r
are known to be small with respect to true airspeed V , so that for control design
purposes,
Y
a
V

Y
r
V
0. (5.8)
The full roll-yaw dynamics expressed in stability axis are
_
=
cos
0
cos
0
p
s
+
sin
0
cos
0
r
s
=
Y
V
+
Y
r
V
r
s
+
g cos
0
V
r
s
p
s
= L
+ L
p
p
s
+ L
r
r
s
+
l
(p
s
, r
s
) + L
a
(,
a
) + L
r
(,
r
)
r
s
= N
+ N
p
p
s
+ N
r
r
s
+
n
(p
s
, r
s
) + N
a
(,
a
) + N
r
(,
r
),
(5.9)
where
l
and
n
are incremental rolling moment and yawing moment that are generally
unknown functions of the rates p
s
, q
s
. Additionally, L
a
, L
r
, N
a
, N
r
are rolling and yaw
moments due to aileron and rudder deection that are also unknown. However, some partial
knowledge is assumed from ight tests so that L
a
, L
r
, N
a
, N
r
can be decomposed into a
known (nominal) linear part and an unknown nonlinear part as follows
L
a
(,
a
) = L
a
0
a
+ L
a
(,
a
)
L
r
(,
r
) = L
r
0
r
+ L
r
(,
r
)
N
a
(,
a
) = N
a
0
a
+ N
a
(,
a
)
N
r
(,
r
) = N
r
0
r
+ N
r
(,
r
).
(5.10)
Additionally, the nonlinear expressions
l
(p
s
, r
s
) + L
a
(,
a
) + L
r
(,
r
) and
n
(p
s
, r
s
) +
N
a
(,
a
) + N
r
(,
r
) can be expressed as a linear combination of two unknown functions
of , p
s
, r
s
,
a
,
r
. Substituting (5.10), we get that
l
(p
s
, r
s
) + L
a
(,
a
) + L
r
(,
r
) = L
a
0
(
a
+ f
1
(, p
s
, r
s
,
a
)) + L
r
0
(
r
+ f
2
(, p
s
, r
s
,
r
))
n
(p
s
, r
s
) + N
a
(,
a
) + N
r
(,
r
) = N
a
0
(
a
+ f
1
(, p
s
, r
s
,
a
)) + N
r
0
(
r
+ f
2
(, p
s
, r
s
,
r
)).
Substituting this expression into (5.9) and re-writing the dynamics in state space form gives
_
(t)
(t)
p
s
(t)
r
s
(t)
_
_
=
_
_
Y
V
g cos
0
V
Y
p
V
(1
Y
r
V
)
0 0
cos
0
cos
0
sin
0
cos
0
L
0 L
p
L
r
N
0 N
p
N
r
_
_
. .
A
_
_
(t)
(t)
p
s
(t)
r
s
(t)
_
_
. .
x(t)
+
_
_
0 0
0 0
L
a
0
L
r
0
N
a
0
N
r
0
_
_
. .
B
_

a
(t) + f
1
((t), p
s
(t), r
s
(t),
a
(t))
r
(t) + f
2
((t), p
s
(t), r
s
(t),
r
(t))
_
,
y(t) =
_
1 0 0 0
0 1 0 0
_
. .
C
_
(t)
(t)
p
s
(t)
r
s
(t)
_
_
,
(0) =
0
, (0) =
0
p
s
(0) = p
s
0
, r
s
(0) = r
s
0
,
(5.11)
where C is the output matrix that denes the outputs of interest for tracking. The con-
trollability matrix has rank = 4, so the system is controllable. Additionally, the A matrix
is Hurwitz but has slow eigenvalues that do not achieve desired system properties such as
settling time and overshoot. Wind tunnel data shows that the nonlinear terms f
1
and f
2
have the structure
f
1
(, p
s
, r
s
,
a
) =
_
(1 C
1
)e

2
2
2
1
+ C
1
_
(tanh (
a
+ h
1
) + tanh (
a
h
1
) + 0.001
a
) +
D
1
cos (A
1
p
s
1
) sin (A
2
r
s
2
) + D
2
(5.12)
f
2
(, p
s
, r
s
,
r
) =
_
(1 C
2
)e

2
2
2
2
+ C
2
_
(tanh (
r
+ h
2
) + tanh (
r
h
2
) + 0.001
r
) +
D
3
cos (A
3
p
s
3
) sin (A
4
r
s
4
) + D
4
. (5.13)
The positive constants C
1
, C
2
represent the percentage of control eectiveness available at
high angles of sideslip and are always less than unity. When the system is at trim sideslip
angle, full aileron and rudder eectiveness is obtained. The parameters
1
,
2
dene the
width of the Gaussian functions e

2
2
2
1
, e

2
2
2
2
. An additional small aileron and rudder control
induced moment term is introduced so that for small control eorts, the control surface is
within a boundary layer and does not produce an aerodynamic moment. As the control
eort increases, the surface stalls and fails to produce any additional moment as the system
reaches saturation, which is approximated by the hyperbolic tangent functions where h
1
, h
2
are constants. Lastly, the rolling and yaw moments can be locally approximated by sine
functions where the amplitude and phase are designed based on wind-tunnel data analysis.
The control objective is to design a control signal
(t) =
_

a
(t)
r
(t)
_
=
_

a
nom
(t) +
a
ad
(t)
r
nom
(t) +
r
ad
(t)
_
(5.14)
such that the angles , track a a smooth commanded reference input R
cmd
(t). First,
a nominal controller is designed to achieve desired performance for the linearized system
in the absence of uncertainties. Then, an adaptive element is augmented to the nominal
controller to achieve the tracking objective in the presence of uncertainties.
5.2 Nominal Dutch-Roll Model
The linear ane system for the Dutch-Roll dynamics of (5.11) in the absence of uncertainties
is
_
r
(t)
r
(t)
p
s
r
(t)
r
s
r
(t)
_
_
= A
r
_
r
(t)
r
(t)
p
s
r
(t)
r
s
r
(t)
_
_
. .
x
r
(t)
+B

R(t), (5.15)
y
r
(t) = C
r
(t)
r
(t)
p
s
r
(t)
r
s
r
(t)
_
_
,
r
(0) =
r
0
,
r
(0) =
r
0
, p
s
r
(0) = p
s
r
0
, r
s
r
(0) = r
s
r
0
,
where
A
r
= A B
_
K
lqr
1,1
K
lqr
1,2
K
lqr
1,3
K
lqr
1,4
K
lqr
2,1
K
lqr
2,2
K
lqr
2,3
K
lqr
2,4
_
,
. .
K
lqr
(5.16)
and the LQR gain K
lqr
= R
1
B
P is found by solving the matrix Riccati equation for the

unique 4 4 positive denite symmetric matrix P:
Q + PA + A
P PBR
1
B
P = 0 (5.17)
given positive denite symmetric control weights Q R
4
and R R
2
.
Consider
R(t) = k
g
R
cmd
(t), (5.18)
where k
g
is the feedforward gain chosen as:
k
g
=
1
C
A
1
r
B
(5.19)
to achieve a unity DC gain between the commanded signal R
cmd
(t) and the system output
y
r
(t) so that y
r
(t) R
cmd
(t) as t . Thus, it follows that the nominal controller is given
by
_

a
nom
(t)
r
nom
(t)
_
= K
lqr
_
_
(t)
(t)
p
s
(t)
r
s
(t)
_
_
+

R(t). (5.20)
Substituting the nominal controller into the roll-yaw dynamics (5.11) gives
_
(t)
(t)
p
s
(t)
r
s
(t)
_
_
= A
r
_
_
(t)
(t)
p
s
(t)
r
s
(t)
_
_
+ B

R(t) + B
_

a
ad
(t) + f
1
((t), p
s
(t), r
s
(t),
a
(t))
r
ad
(t) + f
2
((t), p
s
(t), r
s
(t),
r
(t))
_
,
y(t) = C
_
(t)
(t)
p
s
(t)
r
s
(t)
_
_
, (0) =
0
, (0) =
0
, p
s
(0) = p
s
0
, r
s
(0) = r
s
0
. (5.21)
5.3 Multi-Input RBF approximation
From (5.21), the unknown nonlinearities of the system are
_
f
1
((t), p
s
(t), r
s
(t),
a
(t))
f
2
((t), p
s
(t), r
s
(t),
r
(t))
_
, (5.22)
which will be approximated using neural networks. Using the structure of f
1
and f
2
given
in (5.4), (5.5), dene the Jacobian matrix of (5.22) as
P(,
a
,
r
) =
_
_
f
1
a
(,
a
)
f
1
r
(,
a
)
f
2
a
(,
r
)
f
2
r
(,
r
)
_
_
=
_
C
1
()
1
(
a
)
. .
P
1
(,
a
)
0
0

C
2
()
2
(
r
)
. .
P
2
(,
r
)
_
_
, (5.23)
where
C
1
() =
_
(1 C
1
)e

2
2
2
1
+ C
1
_
C
1
> 0,
C
2
() =
_
(1 C
2
)e

2
2
2
2
+ C
2
_
C
2
> 0, (5.24)
1
(
a
) = 2.001 tanh
2
(
a
+ h
1
) tanh
2
(
a
h
1
) 0.001,
2
(
r
) = 2.001 tanh
2
(
r
+ h
2
) tanh
2
(
r
h
2
) 0.001,
and C
1
, C
2
, h
1
, h
2
are positive constants with C
1
, C
2
less than one. For the roll-yaw dynamical
system, the diagonal elements of the Jacobian matrix (5.23) are strictly positive and P is
non-singular for every t 0 and (t),
a
(t),
r
(t) R. Leibnitz Formula allows the functions
f
1
, f
2
to be represented as
f
1
(,
a
) = f
1
(, 0) +
_

a
0
P
1
(,
1
) d
1
f
2
(,
r
) = f
2
(, 0) +
_

r
0
P
2
(,
2
) d
2
where the integrals must be positive by positivity of P
1
and P
2
. Each component of f
1
, f
2
can be approximated by a linear combination of radial basis functions (RBF) over a compact
set with arbitrarily small approximation error. In the same manner as in the single-input
short-period dynamics case, consider two sets of radial basis functions
1
,
2
: R R R
given by
1
(,
a
) =
1,1
() +
1,2
(,
a
) (5.25)
2
(,
r
) =
2,1
() +
2,2,
(,
r
) (5.26)
with the following denitions:
1,1
()
1

1,1
() =
M
1
i=1
1,i
e
(
c
i
)
(
c
i
)
2
1,i
,
1,i
> 0 (5.27)
1,2
(,
a
) w
1

1,2
(,
a
) =
N
1
j=1
w
1,j
_
_
_

a
0
e
(
1
1,c
j
)
(
1
1,c
j
)
2
1,j
d
1
_
_
, (5.28)
1
[
1
]
,
1,j
> 0,
2,1
()
2

2,1
() =
M
2
i=1
2,i
e
(
c
i
)
(
c
i
)
2
2,i
,
2,i
> 0 (5.29)
2,2
(,
r
) w
2

2,2
(,
r
) =
N
2
j=1
w
2,j
_
_
_

r
0
e
(
2
2,c
j
)
(
2
2,c
j
)
2
2,j
d
2
_
_
, (5.30)
2
[
2
]
,
2,j
> 0,
where M
1
, M
2
, N
1
, N
2
, are positive constants, the vectors
1,c
k
[
c
k

a
c
k
]
k 1, . . . , N
1
2,c
l
[
c
l

r
c
l
]
l 1, . . . , N
2
represent the xed centers of the basis,

1,i
,
2,j
,
1,k
,
2,l
are xed width parameters, while
1,i
,
2,k
R, w
1,k
, w
2,l
R
+
are the unknown constant parameters. It is straightforward to
verify from (5.27) - (5.30) that
sgn
_
a
(,
a
)
_
= sgn
_
1,2
a
(,
a
)
_
> 0
sgn
_
r
(,
r
)
_
= sgn
_
2,2
r
(,
r
)
_
> 0. (5.31)
Thus, over a compact set of initial conditions of interest,
f
1
(,
a
) = W
1

1
(,
a
) +
1
(,
a
), [
1
(,
a
)[
1
, (5.32)
f
2
(,
r
) = W
2

2
(,
r
) +
2
(,
r
), [
2
(,
r
)[
2
, (5.33)
where
W
1
= [
1
w
1
]
= [
1,1
. . .
1,M
1
w
1,1
. . . w
1,N
1
]
,
W
2
= [
2
w
2
]
= [
2,1
. . .
2,M
2
w
2,1
. . . w
2,N
2
]
1
(,
a
) = [
1,1
()
1,2
(,
a
)]
2
(,
r
) = [
2,1
()
2,2
(,
r
)]
with constant coecients w

1,1
, . . . , w
1,N
1
, w
2,1
, . . . , w
2,N
2
> 0 and
1
(,
a
),
2
(,
r
) are the
bounded approximation errors over the sets (,
a
)
a
, and (,
r
)
r
respectively for compact sets
a
,
r
R.
5.4 Roll-Yaw Dynamics State Predictor
Consider the following state predictor using a series parallel model for the dynamics of (5.11):
_
(t)
(t)
p
s
(t)
r
s
(t)
_
_
=A
_
(t)
(t)
p
s
(t)
r
s
(t)
_
_
. .
x(t)
+B

R(t) + B
_

a
ad
+

W
1
(t)
1
((t),
a
(t))
r
ad
+

W
2
(t)
2
((t),
r
(t))
_
, (5.34)
(0) =

0
, ,

(0) =

0
, p
s
(0) = p
s
0
, r
s
(0) = r
s
0
,
where the Hurwitz matrix A
r
is introduced in (5.16),

R is given in (5.18), and

W
1
,

W
2
are
the estimates of the unknown constant vectors W
1
, W
2
. Dene the prediction error signal
e
s
(t) =
_
(t) (t)
(t) (t)
p
s
(t) p
s
(t)
r
s
(t) r
s
(t)
_
_
. (5.35)
The prediction error dynamics are:
e
s
(t) = A
r
e
s
(t) + B
_

W
1
(t)
1
((t),
a
(t))
1
((t),
a
(t))
2
(t)
2
((t),
r
(t))
2
((t),
r
(t))
_
,
= A
r
e
s
(t) + B
1
_

W
1
(t)
1
((t),
a
(t))
1
((t),
a
(t))
_
+
B
2
_

W
2
(t)
2
((t),
r
(t))
2
((t),
r
(t))
_
, (5.36)
with initial conditions e
s,1
(0) =

(0)
0
, e
s,2
(0) =

(0)
0
, e
s,3
(0) = p
s
(0) p
s
0
, e
s,4
(0) =
r
s
(0) r
s
0
, the parameter errors

W
1
(t) =

W
1
(t) W
1
(t),

W
2
(t) =

W
2
(t) W
2
(t), and B
1
, B
2
denoting the rst and second columns of B respectively. That is,
B
1
=
_
_
0
0
L
a
0
N
a
0
_
_
, B
2
=
_
_
0
0
L
r
0
N
r
0
_
_
. (5.37)
Theorem 4 The adaptive law
W
1
(t) =
1
Proj
_

W
1
(t),
1
((t),
a
(t))e
s
(t)P
0
B
1
_
,

W
1
(0) =

W
1
0
W
2
(t) =
2
Proj
_

W
2
(t),
2
((t),
r
(t))e
s
(t)P
0
B
2
_
,

W
2
(0) =

W
2
0
, (5.38)
where Proj (, ) denotes the Projection operator [2], P
0
= P
0
> 0 solves the Lyapunov
equation A
r
P
0
+ P
0
A
r
= Q
0
for a matrix Q
0
> 0, and
1
=
1
> 0,
2
=
2
> 0 are
matrices of adaptation gains, ensures that the prediction error dynamics (5.36) are ultimately
bounded with respect to e
s
(t),

W
1
(t), and

W
2
(t).
Proof. Consider the following positive denite Lyapunov function candidate
V (e
s
(t),

W
1
(t),

W
2
(t)) = e
s
(t)P
0
e
s
(t) +

W
1
(t)
1
1
W
1
(t) +

W
2
(t)
1
2
W
2
(t). (5.39)
Taking the time derivative yields
V (e
s
(t),

W
1
(t),

W
2
(t)) = e
s
(t)P
0
e
s
(t) + e
s
(t)P
0
e
s
(t) +

1
(t)
1
1
W
1
(t) +

W
1
(t)
1
1
1
W(t) +
2
(t)
1
2
W
2
(t) +

W
2
(t)
1
2
2
W(t)
= e
s
(t)A
r
P
0
e
s
(t) + e
s
(t)P
0
A
r
e
s
(t) +
_
1
((t),
a
(t))

W
1
(t)
1
((t),
a
(t))
_
B
1
P
0
e
s
(t) +
_
2
((t),
r
(t))

W
2
(t)
2
((t),
r
(t))
_
B
2
P
0
e
s
(t) +
e
s
(t)P
0
B
1
_

W
1
(t)
1
((t),
a
(t))
1
((t),
a
(t))
_
+
e
s
(t)P
0
B
2
_

W
2
(t)
2
((t),
r
(t))
2
((t),
r
(t))
_
+
_
1
Proj(

W
1
(t),
1
((t),
a
(t))e
s
(t)P
0
B
1
)
_
1
1
W
1
(t) +
1
(t)
1
1
_
1
Proj(

W
1
(t),
1
((t),
a
(t))e
s
(t)P
0
B
1
)
_
+
_
2
Proj(

W
2
(t),
2
((t),
r
(t))e
s
(t)P
0
B
2
)
_
1
2
W
2
(t) +
2
(t)
1
2
_
2
Proj(

W
2
(t),
2
((t),
r
(t))e
s
(t)P
0
B
2
)
_
e
s
(t)Q
0
e
s
(t) 2e
s
(t)P
0
B
1

1
((t),
a
(t))
2e
s
(t)P
0
B
2

2
((t),
r
(t))
|e
s
(t)|
2
min
(Q
0
) + 2|e
s
(t)| |P
0
B
1
|
1
+ 2|e
s
(t)| |P
0
B
2
|
2
= |e
s
(t)|
_
min
(Q
0
)|e
s
(t)| 2|P
0
B
1
|
1
2|P
0
B
2
|
2
_
,
where
min
(Q
0
) denotes the minimum eigenvalue of Q
0
. It follows that outside the compact
set
_
|e
s
| 2
_
1
|P
0
B
1
| +
2
|P
0
B
2
|
min
(Q
0
)
__
_
|

W
1
| W
1
_
_
|

W
2
| W
2
_
, (5.40)
where W
1
, W
2
are determined by the Projection operator and | | denotes the 2-norm,
the derivative of the Lyapunov function candidate,

V (t) is negative. Applying well-known
invariant set arguments, it can be concluded that if the initial errors are within the largest
level set of the Lyapunov function (5.39), for which the RBF approximation has been dened,
then the signals e
s
(t),

W
1
(t),

W
2
(t) are ultimately bounded and the ultimate bound can
be any number greater than the value of the Lyapunov function on the minimum level set
embracing the compact set (5.40). That is,
|| >
_
_
_
_
V
_
2
_
1
|P
0
B
1
| +
2
|P
0
B
2
|
min
(Q
0
)
_
, W
1
, W
2
__
_
_
_
(5.41)
= 4
_
(
1
|P
0
B
1
| +
2
|P
0
B
2
|)
2
2
min
(Q
0
)
_
|
max
(P
0
)| + (W
1
)
2
|
1
1
|, (W
2
)
2
|
1
2
|,
where V is the Lyapunov function (5.39).
Remark 4 Notice that boundedness of the prediction error e

s
(t) and the parameter errors
W
1
(t),

W
2
(t) does not imply stability of the overall system. It must be proven additionally
that the state predictor dynamics (5.34) achieves the tracking objective in the presence of
feedback, implying that

(t),

(t), p
s
(t), r
s
(t) remain bounded to lead to the system states
(t), (t), p
s
(t), r
s
(t) boundededness.
Remark 5 Notice that since w
1
and w
2
are positive, the compact set in the application of
the Projection operator can be selected so that w
1
(t), w
2
(t) remain positive for all t 0, that
is, w
1
(t) > w
1
0
> 0, w
2
(t) > w
2
0
> 0.
5.5 MIMO Control Design
Let the tracking error signal between the predictor and the reference system be
e(t) =
_
(t)
r
(t)
(t)
r
(t)
p
s
(t) p
s
r
(t)
r
s
(t) r
s
r
(t)
_
_
. (5.42)
The open loop time-varying tracking error dynamics are:
e(t) = A
r
e(t) + B
_

a
ad
(t) +

W
1
(t)
1
((t),
a
(t))
r
ad
(t) +

W
2
(t)
2
((t),
r
(t))
_
,
e
1
(0) =

r
0
, e
2
(0) =

r
0
, e
3
(0) = p
s
0
p
s
r
0
, e
4
(0) = r
s
0
r
s
r
0
. (5.43)
Dynamic inversion based controller is to be determined from the solution of the following
system of equations
_

a
ad
r
ad
_
=
_

W
1

1
(,
a
nom
+
a
ad
)
2

2
(,
r
nom
+
r
ad
)
_
, (5.44)
which would result in the stable closed-loop tracking error dynamics e(t) = A
r
e(t). In
general, (5.44) cannot be solved explicitly for
a
ad
,
r
ad
, so an approximation of the dynamic
inversion controller is constructed via fast dynamics:
a
ad
(t)
r
ad
(t)
_
=
P(t, e(t),
a
ad
(t),
r
ad
(t))f (t, e(t),
a
ad
(t),
r
ad
(t)), 0 < 1,
a
ad
(0) =
a
ad
0
,
r
ad
(0) =
r
ad
0
, (5.45)
where
f (t, e,
a
ad
,
r
ad
) =
_

a
ad
+

W
1
(t)
1
((t),
a
nom
(t) +
a
ad
)
r
ad
+

W
2
(t)
2
((t),
r
nom
(t) +
r
ad
)
_
(5.46)
and
P(t, e,
a
ad
,
r
ad
) =
_
1
(t)

1
a
ad
((t),
a
ad
) + 1 0
0

W
2
(t)

2
r
ad
((t),
r
ad
) + 1
_
_
(5.47)
is the Jacobian matrix of f (t, e,
a
ad
,
r
ad
). To prove existence of a root for (5.46), recall
Implicit Function Theorem of [1]:
Implicit Function Theorem Assume that f : R
n
R
m
R
n
is continuously dierentiable
at each point (x, y) of an open set S R
n
R
m
. Let (x
0
, y
0
) be a point in S for which
f(x
0
, y
0
) = 0 and for which the Jacobian matrix
f
x
(x
0
, y
0
) is nonsingular. Then there exist
neighborhoods u R
n
of x
0
and V R
m
of y
0
such that for each y V , the equation
f(x, y) = 0 has a unique solution x U. Moreover, this solution can be given as x = g(y),
where g is continuously dierentiable at y = y
0
.
It is straightforward to verify from the denition of the RBF approximations (5.27) - (5.30)
that the determinant of

P is
det(
P(t,
a
ad
,
r
ad
)) =
_

W
1
(t)

1
a
ad
((t),
a
ad
) + 1
__

W
2
(t)

2
r
ad
((t),
r
ad
) + 1
_
=
_
w
1
(t)
1,2
a
ad
((t),
a
ad
) + 1
__
w
2
(t)
2,2
r
ad
((t),
r
ad
) + 1
_
. (5.48)
In the light of Remark 5, all the elements of w
1
and w
2
are positive and

1,2
a
ad
(t, e,
a
ad
),
2,2
r
ad
(t, e,
r
ad
) are positive-valued Gaussians, so the determinant of

P is always positive.
Furthermore, the matrix

P
(t, e,
a
ad
1
,
r
ad
1
)
P(t, e,
a
ad
2
,
r
ad
2
) is strictly positive denite, i.e.
there exists c
1
> 0 such that
(t, e,
a
ad
1
,
r
ad
1
)
P(t, e,
a
ad
2
,
r
ad
2
) 2c
1
||
2
, R
2
. (5.49)
The proof is straightforward by computing the left-hand side of (5.49) to obtain
(t, e,
a
ad
1
,
r
ad
1
)
P(t, e,
a
ad
2
,
r
ad
2
) =
_
w
1
(t)

1,2
a
ad
1
((t),
a
ad
1
) + 1
__
w
1
(t)

1,2
a
ad
2
((t),
a
ad
2
) + 1
_
2
1
+
_
w
2
(t)
2,2
r
ad
1
((t),
r
ad
1
) + 1
__
w
2
(t)
2,2
r
ad
2
((t),
r
ad
2
) + 1
_
2
2

2
1
+
2
2
,
and it is easy to see that c
1
=
1
2
is one choice of c
1
that satises the inequality of (5.49).
Hence, we apply Implicit Function Theorem to guarantee the existence of a unique root
within a neighborhood. Let h(t, e,
a
,
r
) denote this isolated root of f (t, e,
a
,
r
) and dene
=
t
, v(t, ) =
_

a
ad
(t)
r
ad
(t)
_
h(t, ). (5.50)
The reduced system for (5.42), (5.43), (5.45) is given by:
e(t) = A
r
e(t), e(0) = e
0
, (5.51)
and the boundary layer system is
d
d
=
P(t, e, + h(t, e))f (t, e, + h(t, e)), v(0)

0
=
_

a
0
r
0
_
h(t, e
0
). (5.52)
The main result for the roll-yaw dynamics is stated as follows:
Theorem 5 For the dynamics of (5.11), the boundary layer system (5.52) has exponentially
stable origin. Moreover, the system given by (5.34), (5.45) has a unique solution
_
(t)
(t)
p
s
(t)
r
s
(t)
_
_
=
_
r
(t)
r
(t)
p
s
r
(t)
r
s
r
(t)
_
_
+ O(). (5.53)
Proof. We must verify that the assumption in Tikhonovs theorem are satised. Verication
of A1 is straightforward, since the function f and its rst partial derivatives with respect
to t are continuous and bounded, h(t, e) and
f
a
(t, e, + h(t, e)),
f
r
(t, e, + h(t, e)) have
bounded rst derivatives with respect to their arguments, and
f
e
(t, e, +h(t, e)) is Lipshitz
in e, uniformly in t. Since A
r
is Hurwitz by denition, the origin is an exponentially stable
equilibrium point for the reduced system given in (5.51), and, hence, A2 is satised. For
verication of A3, consider the following Lyapunov function candidate for the boundary layer
system (5.52):
V (t, e, ()) =
1
2
f
(t, e, () + h(t, e, z))f (t, e, () + h(t, e)). (5.54)

Since f (t, e, h(t, e)) = 0, then V (t, e, 0) = 0. Therefore
V (t, e, ) =
_
L
V (t, e, s)ds, (5.55)
where
_
L
assumes integration along the pass L from zero to and
V (t, e, ) = f
(t, e, + h(t, e, z))
P(t, e, + h(t, e)). (5.56)

Since
f (t, e, + h(t, e))
=

P(t, e, + h(t, e)), (5.57)
then
f (t, e, + h(t, e)) =
_
L
P(t, e, s + h(t, e))ds. (5.58)

Hence, it follows from (5.55) that
V (t, e, ) =
_
L
__
L
1
P(t, e, s
1
+ h(t, e))ds
1
_
P(t, e, s + h(t, e)) ds,

and consequently,
V (t, e, ) =
_
L
_
L
1
(ds
1
)
(t, e, s
1
+ h(t, e))
P(t, e, s + h(t, e)) ds, (5.59)

where
_
L
,
_
L
1
assume integration along the pass L and L
1
from zero to , s, respectively.
From (5.49), it follows then that
V (t, e, )
_
L
_
L
1
2c
1
(ds
1
)
ds (5.60)
=
_
||||
0
_

0
2c
1
dd.
Therefore,
V (t, e, ) c
1
[[[[
2
, (5.61)
where c
1
=
1
2
. Condition A1 implies that there exists positive c
2
such that V (t, e, ) c
2
||
2
,
which along with (5.61), veries (3.19). Using the relationships in (5.52) and (5.54), one
obtains
V (t, e, ) = f
(t, e, + h(t, e))
P(t, e, + h(t, e))
(t, e, + h(t, e))f (t, e, + h(t, e)),

which implies that

V (t, e, ) 2c
1
V (t, e, ) 2c
2
1
||
2
, and this proves that (3.20) holds
with c
3
= c
2
1
. Hence, A3 holds, and the boundary layer (5.52) has exponentially stable
origin. Following Tikhonovs theorem, we can conclude that there exists a unique solution,
,

, p
s
, r
s
that tracks
r
,
r
, p
s
r
, r
s
r
with an error of order O().

5.6 Dutch-Roll Simulations
For our simulation we use the lateral/directional model of F-16 from [15] ying at sea level
with an airspeed of 502 ft/s and angle of attack = 2.11
. The model matrices are given as

A =
_
_
0.3220 0.0640 0.0364 0.9917
0 0 1 0.0393
30.6490 0 3.6784 0.6646
8.5395 0 0.0254 0.4764
_
_
, B =
_
_
0 0
0 0
0.7331 0.1315
0.0319 0.0620
_
_
,
and its open-loop eigenvalues are:
1
= 3.6153
2,3
= 0.4237 3.0635i
4
= 0.0142.
Note that the Dutch-Roll system is stable, but that its eigenvalues are slow. Next we select
our reference model by solving the Riccati equation with the following weighted matrices:
Q =
_
_
10 0 0 0
0 100 0 0
0 0 0 0
0 0 0 100
_
_
, R =
_
1 0
0 0.1
_
, (5.62)
so that
k
lqr
=
_
10.6901 9.5824 2.0328 6.1944
0.3982 0.2043 0.4170 27.0142
_
(5.63)
and
A
m
=
_
_
0.3220 0.0640 0.0364 0.9917
0 0 1 0.0393
22.7599 6.9980 5.1138 0.3242
8.8560 0.3183 0.1161 2.3489
_
_
.
The eigenvalues of the reference model are:
1,2
= 2.6518 0.6710i
3,4
= 1.2405 2.9696i.
The commanded reference input of interest to track is
R
cmd
(t) =
_
0.2
_
0.5
1+e
t8
+
1
1+e
t30
0.5
_
0.2
_
0.5
1+e
t8
+
1
1+e
t30
0.2
_
_
.
In the absence of uncertainties, the nominal system behaves optimally using the LQR con-
troller. Setting
k
g
=
_
2.9031 9.9924
156.5907 2.4300
_
, (5.64)
the angles , track R
cmd
5.1. Initial conditions used for the nominal system are
r
0
=
0
,
r
0
= 0
, p
s
r
0
= 0
/sec, r
s
r
0
= 0
/sec. To insert uncertainties into the roll-yaw dy-

namics, explicit expressions must be found for the nonlinearities f
1
(, p
s
, r
s
,
a
,
r
) and
f
2
(, p
s
, r
s
,
a
,
r
) introduced in (5.4) - (5.5) for simulations purposes. Towards that end,
the values of the constants C
1
, C
2
,
1
,
2
and D
i
, A
i
,
i
, for i = 1, 2, 3, 4, are found by
analyzing wind-tunnel data from [15] for an F-16 at angle of attack = 0
(Table 5.1) and

= 5
(Table 5.2). The two data sets are averaged to obtain an approximation at 2.11
in order to remain consistent with the rest of the aircraft data. Using this averaged data set,
curve-tting methods can be applied to nd approximations that are of the form (5.4), (5.5).
From Figures 5.4 - 5.5, the approximations on f
1
(, p
s
, r
s
,
a
,
r
) and f
2
(, p
s
, r
s
,
a
,
r
) are
shown along with the wind-tunnel data with the constant parameters chosen as
A
1
= 0.1 A
2
= 0.1 A
3
= 0.1 A
4
= 0.1
D
1
= 0.075 D
2
= 0.0016 D
3
= 0.45 D
4
= 0
1
= 1.5
2
= 0
3
= 1.5
4
= 0
C
1
= 0.3 C
3
= 0.3 h
1
= 7 h
2
= 4
1
= 0.15
2
= 0.15.
(5.65)
Physically, this means that at high sideslip angle, only 30% of the control is available to the
system. The Gaussian parameters
1
,
2
are chosen so that the Gaussian e

2
2
2
has a width
of 60
, a reasonable range to assume for high angle manuevers. The curve-t approximation
assumes [30
, 30
], p
s
[360
, 360
], r
s
[90
, 90
],
a
[21.5
, 21.5
], and
r

[30
, 30
], which is seen to be most accurate for sideslip angle range [20
, 20
].
Figures 5.2 shows that the performance of the LQR controller is violated when uncertainties
are present in the system. The uncertainties f
1
, f
2
of order O(1) cause to vary only slightly,
but tracking is o by almost 30
(Figure 5.3). Hence, it is necessary to use adaptive control

to compensate for these uncertainties to recover the optimal performance.

l

n
L
a
L
r
N
a
N
r
30
0 0 -0.053 0.014 -0.017 -0.052

20
-0.017 0.042 -0.053 0.014 -0.016 -0.043

10
-0.022 0.076 -0.052 0.011 -0.006 -0.038

0
-0.023 0.106 -0.051 0.015 -0.010 -0.045

10
-0.022 0.076 -0.048 0.013 -0.014 -0.041

20
-0.017 0.042 -0.048 0.010 -0.004 -0.036

30
0 0 -0.047 0.011 -0.003 -0.027

Table 5.1: Wind Tunnel Data of Aerodynamic and Control Derivatives at = 0

l

n
L
a
L
r
N
a
N
r
30
0 0 -0.056 0.010 -0.013 -0.052

20
-0.024 0.042 -0.053 0.014 -0.016 -0.046

10
-0.034 0.074 -0.051 0.012 -0.006 -0.040

0
-0.037 0.106 -0.052 0.014 -0.009 -0.045

10
-0.034 0.074 -0.049 0.013 -0.012 -0.041

20
-0.024 0.042 -0.047 0.011 -0.002 -0.036

30
0 0 -0.045 0.011 -0.005 -0.028

Table 5.2: Wind Tunnel Data of Aerodynamic and Control Derivatives at = 5

l

n
L
a
L
r
N
a
N
r
30
0 0 -0.055 0.012 -0.015 -0.052

20
-0.021 0.042 -0.053 0.014 -0.016 -0.045

10
-0.028 0.075 -0.052 0.012 -0.006 -0.039

0
-0.030 0.106 -0.052 0.015 -0.001 -0.045

10
-0.028 0.075 -0.049 0.013 -0.013 -0.041

20
-0.021 0.042 -0.048 0.011 -0.003 -0.036

30
0 0 -0.046 0.011 -0.004 -0.028

Table 5.3: Averaged Wind Tunnel Data of Aerodynamic and Control Derivatives to Approx-
imate 2.11

The state predictor is designed with two sets of 24 RBFs. For each of the sets, 8 are
i,1
()-
type Gaussians distributed over [30
, 30
] with the step size equal to 60
and width
1
,
2
= 1 for i = 1, 2. The remaining 16 were
1,2
(,
a
),
2,2
(,
r
)-type basis functions
distributed over [30
, 30
],
a
ad
[21.5
, 21.5
], and
r
ad
[30
, 30
] with width
1
=
2
= 5. The norm upper bound is W
1
= W
2
= 10, the lower bound for the positive
widths w
1
, w
2
is 0.01, and the adaptation gain
1
=
2
= 0.2. The state predictor is
_
(t)
(t)
p
s
(t)
r
s
(t)
_
_
= A
r
_
(t)
(t)
p
s
(t)
r
s
(t)
_
_
+ B
_

a
(t) +

1
(t)
1,1
((t)) + w
1
(t)
1,2
((t),
a
(t))
r
(t) +

2
(t)
2,1
((t)) + w
2
(t)
2,2
((t),
r
(t))
_
(5.66)
and the error dynamics are
e(t) = A
r
e(t) + B
_

a
ad
(t) +

1
(t)
1,1
((t)) + w
1
(t)
1,2
((t),
a
(t))
r
ad
(t) +

2
(t)
2,1
((t)) + w
2
(t)
2,2
((t),
r
(t))
_
. (5.67)
For desired performance, the following equation needs to be solved for adaptive aileron and
rudder control,
a
ad
,
r
ad
:
_

a
ad
(t)
r
ad
(t)
_
=
_

1
(t)
1,1
((t)) + w
1
(t)
1,2
((t),
a
nom
(t) +
a
ad
)
2
(t)
2,1
((t)) + w
2
(t)
2,2
((t),
r
nom
(t) +
r
ad
)
_
, (5.68)
where
a
nom
,
r
nom
has been dened in (5.20). The fast dynamics are designed as
0.05
_

a
ad
(t)
r
ad
(t)
_
=

P(t, e,
a
ad
(t),
r
ad
(t))f (t, e,
a
ad
(t),
r
ad
(t)), (5.69)
where
P(t, e,
a
ad
,
r
ad
) =
_
_
w
1
(t)
1,2
((t),
a
nom
(t) +
a
ad
) + 1 0
0 w
2
(t)
2,2
((t),
r
nom
(t) +
r
ad
)
_
_
and
f (t, e,
a
ad
,
r
ad
) =
_

a
ad
(t) +

1
(t)
1,1
((t)) + w
1
(t)
1,2
((t),
a
nom
(t) +
a
ad
)
r
ad
(t) +

2
(t)
2,1
((t)) + w
2
(t)
2,2
((t),
r
nom
(t) +
r
ad
)
_
.
Initial conditions used for the simulations are
_
0
p
s
0
r
s
0
_
_
=
_
_
0
/sec
0
/sec
_
_
,
_
0
p
s
0
r
s
0
_
_
=
_
_
0
/sec
0
/sec
_
_
.
0 5 10 15 20 25 30
6
4
2
0
2
4
6
8
10
DutchRoll Angle
r
and
r
Tracking of Commanded Input
time, t in seconds
d
e
g
r
e
e
s
R
cmd
1
(t)
R
cmd
2
(t)
r
(t)
r
(t)
Figure 5.1: Dutch-roll nominal system
r
(t) and
r
(t) tracking of commanded input R
cmd
(t).
The plots in Figures 5.6 - 5.8 show the tracking performance with the augmented adaptive
controller. Figure 5.9 shows the time history of the adaptive control signal and the unknown
nonlinearities f
1
((t),
a
(t)) and f
2
((t),
r
(t)) as functions of time. Asymptotic tracking
is achieved of the angles , while the other states of the system p
s
, r
s
, remain bounded.
Additionally, the approximations stay within their respective domains.
0 5 10 15 20 25 30
30
25
20
15
10
5
0
5
10
LQR Performance Violated in the Presence of Uncertainties
time, t in seconds
d
e
g
r
e
e
s
/
s
e
c
R
cmd
1
(t)
R
cmd
2
(t)
r
(t)
r
(t)
Figure 5.2: LQR performance violated for angle
r
,
r
tracking.
0 5 10 15 20 25 30
5
0
5
10
15
20
25
30
Loss of Tracking of ,
time, t in seconds
d
e
g
r
e
e
s
Figure 5.3: Loss of tracking of and caused by nonlinearities f

1
(,
a
) of order O(1) and
f
2
(,
r
) of order O(1.1).
30 20 10 0 10 20 30
0.075
0.07
0.065
0.06
0.055
0.05
0.045
0.04
0.035
f
1
Approximation Using CurveFitting
in degrees
f
1
(
,

p
s
,

r
s
,

a
,

r
)
WindTunnel Data
CurveFit
Figure 5.4: Curve-tting for f
l
(, p
s
, r
s
,
a
,
r
).
30 20 10 0 10 20 30
0.1
0.08
0.06
0.04
0.02
0
0.02
0.04
0.06
0.08
f
2
Approximation Using CurveFitting
in degrees
f
2
(
,

p
s
,

r
s
,

a
,

r
)
WindTunnel Data
CurveFit
Figure 5.5: Curve-tting for f
2
(, p
s
, r
s
,
a
,
r
).
0 5 10 15 20 25 30
6
4
2
0
2
4
6
time, t in seconds
(
t
)
,

d
e
g
r
e
e
s
R
cmd
1
(t)
r
(t)
(t)
hat (t)
Figure 5.6: States

(t),
r
(t), (t) of Dutch-Roll system.
0 5 10 15 20 25 30
4
2
0
2
4
6
8
10
time, t in seconds
(
t
)
,

d
e
g
r
e
e
s
R
cmd
2
(t)
r
(t)
(t)
hat (t)
Figure 5.7: States

(t),
r
(t), (t) of Dutch-Roll system.
0 5 10 15 20 25 30
3
2
1
0
1
2
3
4
5
6
7
8
time, t in seconds
d
e
g
r
e
e
s
/
s
e
c
p
s
r
(t)
p
s
hat p
s
(t)
r
s
r
(t)
r
s
(t)
hat r
s
(t)
Figure 5.8: States p
s
(t), p
s
r
(t), p
s
(t), r
s
(t), r
s
r
(t), r
s
(t) of Dutch-Roll system.
0 5 10 15 20 25 30
0
0.2
0.4
0.6
0.8
1
1.2
1.4
Control Effort
time, t in seconds
a
ad
(t) in degrees
r
ad
(t) in degrees
f
1
((t),
a
(t))
f
2
((t),
r
(t))
Figure 5.9: Adaptive control eort of Dutch-Roll System compared with uncertainties f
1
and f
2
.
Chapter 6
Conclusion
6.1 Summary
For aircraft systems, uncertainties often arise within the system dynamics due to external
and internal forces that are unknown to the controller. Additionally, for high-performance
aircraft, the nonlinearities that result from the equations of motion cannot be neglected or
linearized as in straight and level ight. Performance of traditional linear controllers are
violated in the presence of these nonlinearities that are unknown unstructured functions of
control input and states. This can lead to loss of tracking and instability. In order to account
for unpredictable changes and unknown nonlinear dependencies in aircraft dynamics, func-
tional approximations using basis functions can be used over a compact set to approximate
these unknowns nonlinearities. An approximate dynamic inversion methodology has been
presented for the nonane-in-control short-period and lateral/directional aircraft dynamics
via time-scale separation where the control signal was dened as a solution of fast dy-
namics. For both systems, the form of the unknown functions allowed for the existence of
an isolated root in control. Moreover, the derivatives with respect to control were strictly
positive denite. Hence, we were able to verify the assumptions of Tikhonovs theorem for
the closed-loop system.
54
Amanda Young Bibliography 55
6.2 Recommendations For Future Work
Development of general theory for adaptive control of multi-input nonane systems remains
still to be solved. The particular case of Dutch-Roll dynamics, assumed in the paper, had
a block-diagonal decoupled Jacobian with respect to two control inputs. This simplied
verication of assumptions, required in [28], for multi-input nonane systems. In case of
fully coupled dynamics, such verication for adaptive control design is not obvious and
requires further investigation.
Bibliography
[1] H.K. Khalil. Nonlinear Systems. Prentice Hall, 2002.
[2] J. Park, I. Sandberg. Universal Approximation Using Radial-Basis-Function Networks.
Neural Computation, 3:246-257, 1991.
[3] J. Slotine and W. Li. Applied Nonlinear Control. Pearson Education, 1990.
[4] W. Rudin. Real and Complex Analysis. McGraw Hill, NY, 1986.
[5] P.V. Kokotovi, H.K. Khalil, and J. OReilly. Singular Perturbation Methods in Control:
Analysis and Design. Academic Press, 1986.
[6] S. Lane and R. Stengel. Flight Control Design Using Nonlinear Inverse Dynamics. Au-
tomatica, 24(4):471-483, 1988.
[7] N. Slegers and M. Costello. Nonlinear Model Predictive Control of a 6 DOF Air Vehicle.
In Proceedings of AIAA Atmospheric Flight Mechanics Conference and Exhibit, San
Francisco, California, 2005.
[8] J. Reiner, G. Balas and W. Garrard. Flight Control Design Using Robust Dynamics
Inversion and Time-Scale Seperation. Automatica, 32(11):1493-1504, 1996.
[9] Q. Wang and R. Stengel. Robust Nonlinear Flight Control of a High-Performance Air-
craft. IEEE Transactions on Control Systems Technology, 13(1):15-26, January, 2005
[10] B. Kim and A. Calise. Nonlinear Flight Control Using Neural Networks. AIAA Journal
of Guidance, Control, and Dynamics, 20(1):26-33, 1997.
[11] R. Somakumar and J. Chandrasekhar. Neural Network Based Nonlinear Inverse Dynam-
ics for Flight Controller Design. In Proceedings of the IEEE International Conference
on Control Applications, Trieste, Italy, 1998.
56
[12] J.D. Boskovic, L. Chen, and R. Mehra. Adaptive Control Design for Nonane Models
Arising in Flight Control. Journal of Guidance, Control, and Dynamics, 27(2):209-217,
2004.
[13] N. Hovakimyan, E. Lavretsky, and C. Cao. Adaptive Dynamic Inversion for Nonane-in-
Control Systems via Time-Scale Seperation. Submitted to invited issue of IEEE Trans-
actions on Neural Networks, 2005.
[14] N. Hovakimyan, E. Lavretsky, and C. Cao. Dnyamic Inversion for Multivariable
Nonane-in-Control Systems via Time-Scale Seperation. In Proceedings of American
Control Conference, Minnapolis, MN, 2006.
[15] B. Stevens and F. Lewis. Aircraft Control and Simulation. John Wiley and Sons,1992.
[16] W. Phillips. An Accurate Closed Form Approximation for Dutch Roll. In Proceedings
of 38th Aerospace Sciences Meeting and Exhibit, Reno, NC, 2000.
[17] G. Bryan. Stability in Aviation. Macmillan, London, 1911.
[18] C. Perkins and R. Hage. Airplane Performance Stability and Control. Wiley, New York,
1949.
[19] B. Etkin and L. Reid. Dynamics of Flight: Stability and Control. Wiley, New York,
1996.
[20] B. MrCormick. Aerodynamics, Aeronautics, and Flight Mechanics. Wiley, New York,
1995
[21] N. Ananthkrishnan and Suraj Unnikrishnan. Literal Approximations to Aircraft Dy-
namic Modes. Journal of Guidance, Control and Dynamics, 24(6) 1196:1203, 2001.
[22] NG. Miller and J. Franklin. Lateral-Directional Flying Qualities For Power Approach. In
Proceedings of AIAA Guidance, Control, and Flight Dynamics Conference, Huntsville,
AL 1967.
[23] R. Stengel and G. Miller. Flying Qualities of an Aircraft With Strong Lateral-Directional
Coupling. In Proceeding of Atmospheric Flight Mechanics Conference, Palo Alto, CA,
1978.
[24] R. Stengel. Eect of Combined Roll Rate and Sideslip Angle on Aircraft Flight Stability.
Journal of Aircraft, 12(8) 683:685, 1975.
[25] C. Zahringer, M. Heller and G. Sachs. Lateral Seperation Dynamics and Stability of a
Two-Stage Hypersonic Vehicle. In Proceedings of AIAA International Space Planes and
Hypersonic Systems and Technologies, Norfolk, VA, 2003.
[26] W. Phillips and B. Santana. Aircraft Small-Disturbance Theory with Longitudinal-
Lateral Coupling. Journal of Aircraft, 39(6) 973:980, 2002.
[27] M. Rauw. FDC 1.2 - A Simulink Toolbox for Flight Dynamics and Control Analysis.
2001
[28] N. Hovakimyan, E. Lavretsky and C. Cao. Dynamic Inversion of Multi-Input Nonane
Systems Via Time-Scale separation. In Proceedings of American Control Conference,
Minneapolis, MN, 2006.

Adaptive Control of Nonaffine Systems With Applications To Flight Control

Uploaded by

Copyright:

Available Formats

Adaptive Control of Nonaffine Systems With Applications To Flight 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

Adaptive Control of Nonaffine Systems With Applications To Flight Control

Uploaded by

Copyright:

Available Formats

Adaptive Control of Nonane Systems with Applications to Flight

is the output matrix, y R

, while all other error

. In [9], two nonlinear controllers are implemented by sep-

. The problem with feedback linearization is that it has

], and the inverse dynamics are computed by dieren-

(t)A(t)dt < . (3.3)

. The following result is due to Tikhonov [1], [5]:

have bounded rst derivatives with respect to

is Lipschitz in x, uniformly in t, and the initial data

such that for all t 0,

g(t, x, + h(t, x)) c

satises the eigenvalue condition

(t, x, h(t, x), 0)

(t, x, h(t, x), 0). (3.23)

(t, x, +h(t, x), 0) and g(t, x, +h(t, x), 0)

R that satises the inequalities

P is found by solving the corresponding Riccati equation for

, j = 1, ..., M represent the xed centers of the RBFs,

is the maximum allowable upper-bound selected by the Projection operator and

R denotes the region of attraction of the autonomous system

, the system of equations (4.23),

(t) on [0, ), and

. From [15], the nominal system data are:

. Next, consider when f((t),

loss of tracking in and an 8

] with the step size equal to 60

and the width = 1. The remain-

and width = 5. The norm upper bound is W

= 10, the lower bound

= p + tan (q sin + r cos ), (5.1)

P is found by solving the matrix Riccati equation for the

represent the xed centers of the basis,

with constant coecients w

Remark 4 Notice that boundedness of the prediction error e

P(t, e, + h(t, e))f (t, e, + h(t, e)), v(0)

(t, e, () + h(t, e, z))f (t, e, () + h(t, e)). (5.54)

(t, e, + h(t, e, z))

P(t, e, + h(t, e)). (5.56)

P(t, e, s + h(t, e))ds. (5.58)

P(t, e, s + h(t, e)) ds,

P(t, e, s + h(t, e)) ds, (5.59)

(t, e, + h(t, e))

P(t, e, + h(t, e))

(t, e, + h(t, e))f (t, e, + h(t, e)),

Amanda Young Chapter 5. Adaptive Controller for Nonane Multi-Input System 45

. The model matrices are given as

/sec. To insert uncertainties into the roll-yaw dy-

(Table 5.1) and

], which is seen to be most accurate for sideslip angle range [20

(Figure 5.3). Hence, it is necessary to use adaptive control

0 0 -0.053 0.014 -0.017 -0.052

-0.017 0.042 -0.053 0.014 -0.016 -0.043

-0.022 0.076 -0.052 0.011 -0.006 -0.038

-0.023 0.106 -0.051 0.015 -0.010 -0.045

-0.022 0.076 -0.048 0.013 -0.014 -0.041

-0.017 0.042 -0.048 0.010 -0.004 -0.036

0 0 -0.047 0.011 -0.003 -0.027

0 0 -0.056 0.010 -0.013 -0.052