Nonlinear and Adaptive Control

Zhengtao Ding

Control Systems Centre

School of Electrical and Electronic Engineering
University of Manchester, P O Box 88
Manchester M60 1QD, United Kingdom

January 24, 2007

1
Aims

To introduce techniques for analysis and design of nonlinear and adaptive

control systems.

Learning Outcomes

On completion of the course, students should be able to

• Demonstrate a knowledge of effects of nonlinearities on the operation of
control systems.
• Show an understanding of methods for reducing nonlinear effects in control
systems.
• Apply phase-plane method to second-order nonlinear systems.
• Obtain locally linearized models for nonlinear dynamic systems.
2
• Use the describing function method for predicting nonlinear feedback sys-
tems behaviour.
• Investigate the stability of nonlinear systems by using Lyapunov functions.
• Use backstepping technique for control design of a special class of non-
linear systems.
• Design and analysis adaptive control of uncertain linear systems.
• Understand robustness of adaptive schemes.
• Design adaptive control for special class of uncertain nonlinear systems.

3
1. Introduction to Nonlinear Systems

1.1 Nonlinear functions and nonlinearities

A function f : R 7→ R is linear, if for any x, y, k ∈ R,
f (kx) = kf (x) (1)
f (x + y) = f (x) + f (y) (2)
A function does not satisfy these conditions is a nonlinear functions. Similar
conditions can be imposed for linear maps between Rm and Rn for any
positive integers m and n.
Linear dynamic systems
ẋ = Ax + Bu (3)
y = Cx + Du (4)
for state x ∈ Rn, u ∈ Rm, and u ∈ Rp, and matrices are in suitable
dimensions.
4
 Single-Input-Single-Output (SISO) systems with input saturation
ẋ = Ax + Bσ(u) (5)
y = Cx (6)
where σ : R 7→ R is a saturation 
function defined as









−1 for u < −1
σ(u) =  u otherwise (7)


 1 for u > 1






The saturation function σ is a nonlinear function and this system is a non-

linear system.
A general nonlinear system is often described by
ẋ = f (x, u) (8)
y = h(x, u) (9)
where functions f and h are often assumed to be smooth enough (ie, having
continuous derivatives to certain orders).
5
 System with unknown parameters
ẋ = ax + u (10)
where a is an unknown parameter. How to design a control system to ensure
the stability of the system? If a range a− < a < a+ is known, then we can
design a control law as
u = −cx − a+x (11)
with c > 0, which results the closed-loop system
ẋ = −cx + (a − a+)x (12)

6
 Adaptive control can be used in the case of complete unknown a.
u = −cx − âx (13)
â˙ = x2 (14)
If we let ã = a − â, the closed-loop system is described by
ẋ = −cx + ãx (15)
ã˙ = −x2 (16)
This adaptive system is nonlinear, even though the original uncertain system
is linear. This adaptive system is stable, but how to show it?

7
1.2 Common nonlinearities in control systems
• Nonlinear components, such springs, resistors etc.
• Dead zone and saturation in actuators
• Switching devices, relay and hysteresis relay
• Backlash

8
1.3 Common nonlinear systems behaviours
• Dynamics (of linearized systems) depends on operating point (hence the
input).
• There may be more than one equilibrium point.
• Limit cycles (persistent oscillations) may occur, and also chaotic motion.
• For a pure sinusoidal input at a fixed frequency, the output waveforms is
generally distorted, and the output amplitude is a nonlinear function of
the input amplitude.

9
1.4 Compensating nonlinearities and nonlinear control design
• Gain Scheduling switching controller parameters in different operating
regions.
• Feedback Linearisation By using nonlinear control laws.
• Nonlinear Function Approximations using fuzzy logic and neural net-
works.
• Adaptive Control such self-tuning control model reference adaptive con-
trol, and nonlinear adaptive control.
• Bang-bang Control Switching between operating limits (time-optimal
control).
• Variable Structure Control, sliding mode control.
• Backstepping Design design control input in an iterative way by moving
down the different virtual control until the control input is design.
10
• Forwarding Control an iterative design working in a way opposite to
backstepping.
• Semiglobal Control Design achieving stability in arbitrarily large re-
gions, often by using high gain control.

11
2. State Space Models

2.1 Nonlinear systems and linearisation around an operating point

Consider
ẋ = f (x, u) (17)
y = h(x, u) (18)
An operating point at (xe, ue) is taken with x = xe and u = ue being
constants such that f (xe, ue) = 0. A linerized model around the operation
point can then be obtained. Let
x̄ = x − xe (19)
ū = u − ue (20)
ȳ = h(x, u) − h(xe, ue) (21)
then the linearized model is given by
x̄˙ = Ax̄ + B ū (22)
12
 ȳ = C x̄ + Dū (23)
where
∂fi
ai,j = (xe, ue) (24)
∂xj
∂fi
bi,j = (xe, ue) (25)
∂uj
∂hi
ci,j = (xe, ue) (26)
∂xj
∂hi
di,j = (xe, ue) (27)
∂uj

13
2.2 Some special forms
Strict feedback from
ẋ1 = x2 + φ1(x1)
ẋ2 = x3 + φ2(x1, x2)
..
ẋn−1 = xn + φn−1(x1, x2, . . . , xn−1)
ẋn = u + φn(x1, x2, . . . , xn) (28)

14
Output feedback from
ẋ1 = x2 + φ1(y)
ẋ2 = x3 + φ2(y)
..
ẋr = xr+1 + φr (y) + br u
..
ẋn = φn(y) + bnu
y = x1 (29)

15
Normal from
ż = f0(z, ξ1, . . . , ξr )
ξ˙1 = ξ2
ξ˙2 = ξ3
..
ξ˙r = fr (z, ξ1, . . . , ξr ) + br u (30)
with z ∈ Rn−r .
Special normal from
ż = f0(z, ξ1)
ξ˙1 = ξ2
ξ˙2 = ξ3
..
ξ˙r = fr (z, ξ1, . . . , ξr ) + br u (31)
16
2.3 Autonomous systems
Autonomous systems: a dynamic system which does not explicitly de-
pend on time. Often it can be expressed as
ẋ = f (x) (32)
For an autonomous system, the set of all the trajectories provides a com-
plete geometrical representation of the dynamic behaviour. This is often
referred to as the phase portrait, especially for second order systems in the
format:
ẋ1 = x2
ẋ2 = φ(x1, x2) (33)
Singular point: the point xe such that f (xe) = 0.

17
2.4 Second order systems
Also referred to as phase plane.

ẋ1 = f1(x1, x2)

ẋ2 = f2(x1, x2) (34)
Classification of singular points (based of eigenvalues)

λ1 < 0, λ2 < 0 Stable Node

λ1 > 0, λ2 > 0 Unstable Node
λ1 < 0, λ2 > 0 Saddle Point
λ1,2 = µ ± iν(µ < 0) Stable Focus
λ1,2 = µ ± iν(µ > 0) Unstable Focus
λ1,2 = ±iν(ν > 0) Centre

18
Slope of trajectory

dx2 f1(x1, x2)

= (35)
dx1 f2(x1, x2)
An Isocline is a curve on which ff1(x
(x1,x2)
is constant.
2 1,x2)

19
Example 2.1 The swing equation of a synchronous machine
H δ̈ = Pm − Pe sin δ (36)
where H is the inertia, δ, the rotor angle, Pm, mechanical power, Pe the
maximum electrical power generated.
The state space model is obtained by let x1 = δ and x2 = δ̇ as
ẋ1 = x2
u − Pe sin x1
ẋ2 =
H
y = Pe sin x1 (37)
where u = Pm as the input.

20
Singular points
x1e = arcsin(ue/Pe) or π − arcsin(ue/Pe)
x2e = 0 (38)
Linearised model
   
 0 1  0 
A = 
Pe cos x1e  B=  

− H1
  

− H
  

0  

C = [ Pe cos x1e 0 ]
Stable or unstable?

21
2.5 Limit cycles
A special trajectory which takes the form of a closed curve, and is intrinsic
to the system, and not imposed from outside.
Limits cycles are periodic solutions.
A limit cycle is stable is nearby trajectories converge to it asymptotically,
unstable is move away.
A general condition for the existence of limit cycles in the phase plane is
the Poincare-Bendixson theorem:
If a trajectory of the second-order autonomous system remains in a finite
region, then one of the following is true:
(a) the trajectory goes to an equilibrium point
(b) the trajectory tends to an asymptotically stable limit cycle
(c) the trajectory is itself a stable limit cycle

22
Example 2.2 The Van der Pol oscillator
This is one of the best known models in nonlinear systems theory, originally
developed to describe the operation of an electronic oscillator, which depends
on the existence of a region with effective negative resistance.
ẋ1 = x2
ẋ2 = −x1 + (x2 − x32) (39)

23
2.6 Strange attractors and chaos
For high order nonlinear systems, there are more complicated features if the
trajectories remain in a bounded region.
Positive limit set of a trajectory the set of all the points for which the
trajectory converge to.
Strange limit sets are those limit sets which may or may not be asymp-
totically attractive to the neighbouring trajectories. The trajectories they
contain may be locally divergent from each other, within the attracting set.
Such struectures are associated with the quasi-random behaviour of solutions
called chaos.

24
Example 2.3 The Lorenz attractor
This is one the most widely studied examples of strange behavour in or-
dinary differential equations, which is originated from studies of turbulent
convection by Lorenz. The equation is in the form
ẋ1 = σ(x2 − x1)
ẋ2 = (1 + λ − x3)x1 − x2
ẋ3 = x1x2 − bx3 (40)
where σ, λ and b are positive constants.

25
3. Describing Functions

3.1 Describing function fundamentals

If a nonlinear system contains only one nonlinear component and the other
part is linear, then this kind of systems may be analyzed using describing
functions. Some genuinely nonlinear systems can also be arranged in the
form of with only one nonlinear element.
The basic assumption for describing function analysis are:
• there is only a single nonlinear component
• the nonlinear component is not time-varying
• corresponding to a sinusoidal input x = sin(ωt), only the fundamental
component has to be considered
• the nonlinearity is odd

26
Basic definition
For a nonlinear component described by single-value nonlinear function f (),
its output w(t) = f (A sin(ωt)) to a sinusoidal input A sin(ωt) is a periodical
function, although it may not be sinusoidal in general. A periodical function
can be expanded in Fourier series
a0 ∞
w(t) = + [an cos(nωt) + bn sin(nωt)] (41)
X

2 n=1
where
1 Zπ
a0 = −π w(t)d(ωt)
π
1 Zπ
an = −π w(t) cos(nωt)d(ωt)
π
1 Zπ
bn = −π w(t) sin(nωt)d(ωt)
π
Deu to the assumption that f is an odd function, we have a0 = 0, and the

27
fundamental component is given by
w1 = a1 cos(ωt) + b1 sin(ωt) = M sin(ωt + φ) (42)
where
v

M (A, ω) = a21 + b21

u
u
t

φ(A, ω) = arctan(a1/b1)
In complex expression, we have w1 = M ej(ωt+φ) = (b1 + ja1)ejωt.
Describing function is defined, similar to frequency response, as the com-
plex ratio of the fundamental component of the nonlinear element against
the input by
M ejωt+φ b1 + ja1
N (A, ω) = jωt = (43)
Ae A

28
Example 3.1 Describing function of hardening spring
The characteristics of a hardening spring are given by
w = x + x3/2 (44)
Given the input A sin(ωt), the output is
A3 3
w = A sin(ωt) + sin (ωt) (45)
2
Since w is an odd function, we have a1 = 0. The coefficient b1 is given by
1 Zπ A3 3
b1 = −π [A sin(ωt) + sin (ωt)] sin(ωt)d(ωt)
π 2
3 3
= A+ A (46)
8
Therefore the describing function is
3 2
N (A, ω) = N (A) = 1 + A (47)
8
29
3.3 Describing functions for common nonlinear components
Saturation A saturation function is described by




 kx for |x| < a
f (x) =  (48)


 sign(x)ka otherwise
The output to the input A sin(ωt), for A > a, is symmetric over quarters of
a period, and in the first quarter,




 kA sin(ωt) 0 ≤ ωt ≤ γ
w1(x) =  (49)


 ka γ < ωt ≤ π/2
where γ = sin−1(a/A). The function is odd, hence we have a1 = 0, and
the symmetry of w1(t) implies that

30
 4 Z π/4
b1 = 0 w1 sin(ωt)d(ωt)
π
4 Zγ 2 4 Z π/2
= 0 kA sin (ωt)d(ωt) + γ ka sin(ωt)d(ωt)
π v π
a2
u
2kA a uuuu
u

= [γ + t1 − 2 ] (50)
π A A
Therefore the describing function is given by
 v 

a2 
u
b1 2k  −1 a a
u
u

+ 1 − 2 
u
N (A) = = sin
u
u
t (51)
A π 
 A A A

31
Ideal relay The output from the ideal relay (sign function) is give by, with
M > 0,




 −M −π ≤ ωt < 0
w1(x) =  (52)


 M 0 ≤ ωt < π
It is again an odd function, hence we have a1 = 0. The coefficient b1 is
given by
Z π/2 4M
b1 = 0 M sin(ωt)(.ωt) = (53)
π
and therefore the describing function is given by
4M
N (A) = (54)
πA

32
3.4 Describing function analysis of nonlinear systems
Extended Nyquist criterion: Consider a linear feedback system with for-
ward transfer function G(s) and feedback control gain H(s). The charac-
teristic equation of this system is given by
G(s)H(s) + 1 = 0 or G(s)H(s) = −1 (55)
The Nyquist criterion can then be analyze the stability of the system based
on the Nyquist plot the open-loop transfer function G(s)H(s) with the en-
circlement of the point (−1, 0). An extension to this criterion for the case
of the forward transfer function is KG(s), characteristic equation is
KG(s)H(s) + 1 = 0 or G(s)H(s) = −1/K (56)
The same argument used in the derivation of Nyquist can be applied again,
with (−1, 0) being replace by (1/K, 0). Note that K can be a complex
number.
33
Existence of limit cycles
Consider a unit feedback system with N (A, ω) and G(s) in the forward
path. The equations describing this system can be put as
x = −y
w = N (A, ω)x
y = G(jω)w
with y as the output. From the above equations, it can be obtained that
y = G(jω)N (A, ω)(−y) (57)
and it can be arranged as
[G(jω)N (A, ω) + 1]y = 0 (58)
If there exists a limit cycle, then y 6= 0, which implies that
G(jω)N (A, ω) + 1 = 0 (59)
34
or
G(jω) = −1/N (A, ω) (60)
Therefore the amplitude A and frequency ω of the limit cycle must satisfy
the above equation. The equation 60 is difficult to solve in general. Graphic
solutions can be found by plotting G(jω) and −1/N (A, ω) to see if they
intersect each other. The intersection points are the solutions. It is easy to
plot −1/N if it is independent of the frequency.
Stability of limit cycles Discussions based on the extended Nyquist cri-
terion.

35
 K
Example 3.1 Consider a linear transfer function G(s) = s(s+1)(s+2) and
an ideal relay with M = 1 in a closed loop. Determine if there exists a limit
cycle. If√so, determine the amplitude and frequency of the limit cycle.
(ω = 2, A = 3π 2)

36
4. Stability Theory

4.1 Definitions
Consider the system
ẋ = f (x) (61)
with x = 0 as an equilibrium point.
Lyapunov Stability The equilibrium point x = 0 is said to be stable if for
any R > 0, there exists r > 0 such that if kx(0)k < r, then kx(t)k < R
for all t ≥ 0. Otherwise the equilibrium is unstable.
Critically stable linear system is Lyapunov stable.

37
Asymptotic Stability An equilibrium point 0 is asymptotically stable if it
is stable (Lyapunov) and if in addition there exists some r > 0 such that
kx(0)k < r implies that limt→∞ x(t) = 0.
Linear systems with poles in the left side of complex plane are asymptoti-
cally stable.
Consider ẋ = −x3

38
Exponential Stability An equilibrium point 0 is exponentially stable if there
exist two positive real numbers α and λ such that
∀t > 0, kx(t)k < αkx(0)ke−λt (62)
in some ball Br around the origin.
For linear systems, asymptotic stability implies exponential stability.
Global Stability If asymptotic (exponential) stability holds for any initial
state, the equilibrium point is said to be globally asymptotically (exponen-
tially) stable.

39
4.2 Linearization and local stability
Lyapunov’s linearisation method
• If the linearized system is strictly stable (ie, with system’s poles in the
open left-half complex plane), then the equilibrium point is asymptotically
stable for the actually nonlinear system.
• If the linearized system is unstable (ie, with system’s poles in the open
right-half complex plane), then the equilibrium point is unstable.
• If the linearized system is marginally stable (with poles on the imaginary
axis), then the stability of the original system cannot be concluded using
the linearised model.
Examples

40
4.3 Lyapunov’s direct method
Positive definite functions A scalar function V (x) is said to be locally
positive definite if V (0) = 0 and in a ball Br , x 6= 0 implies V (x) > 0. If
the above properties hold for the entire space, the V (x) is said to be globally
positive definite.
Lyapunov function If in a ball BR, the function V (x) is positive definite
and has continuous partial derivatives, and if its time derivative along any
state trajectory of system (61) is negative semidefinite, ie,
V̇ (x) ≤ 0 (63)
then V (x) is a Lyapunov function.

41
Lyapunov Theorem for Local Stability If in a ball BR, there exists a
scalar function V (x) with continuous first order derivatives such that
• V (x) is positive definite (locally in BR)
• V̇ (x) is negative semidefinite (locally in BR)
then the equilibrium point 0 is stable. Furthermore, if V̇ (x) is locally negative
definite in BR, then the stability is asymptotic.

42
Example 4.1 S simple pendulum is described by
θ̈ + θ̇ + sin θ = 0 (64)
Consider the scalar function
θ̇2
V (x) = (1 − cos θ) + (65)
2

43
Lyapunov Theorem for Global Stability Assume that there exists a
scalar function V (x) with continuous first order derivatives such that
• V (x) is positive definite
• V̇ (x) is negative definite
• V (x) → ∞ as kxk → ∞
then the equilibrium at the origin is asymptotically stable.

44
4.4 Lyapunov analysis of linear-time-invariant systems
Consider LTI system ẋ = Ax. If a Lyapunov function candidate is give by
V (x) = xT P x (66)
then direct evaluation gives
V̇ = xT AT P x + xT P Ax := −xT Qx (67)
where
AT P + P A = −Q (68)
If Q is positive definite, then the system is asymptotically stable.

45
Theorem of stability of LTI systems A necessary and sufficient condition
for a LTI system ẋ = Ax to be strictly stable is that, for any positive
definite matrix Q, the unique solution P of Lyapunov equation (68) is positive
definite.
Proof: Briefly, choose a positive definite Q, and define
P = 0∞ exp(AT t)Q exp(At)dt
Z
(69)
and note
−Q = 0∞ d[exp(AT t)Q exp(At)]
Z
(70)

46
5. Advanced Stability Theory

5.1 Positive Real Systems

Consider a SISO dynamic system
ẋ = Ax + bu
y = cT x (71)
and its transfer function matrix is given by
G(s) = cT (sI − A)−1b (72)
Positive Reals Systems A system described in (75) is said to be positive
real if
<(G(s)) ≥ 0 for all <(s) ≥ 0 (73)
It is strictly positive real if G(s − ) is positive real for some  > 0.
Examples: G(s) = s+λ 1 with λ > 0.

47
Theorem 5.1 A transfer function G(s) is strictly positive real (SPR) if and
only if
• G(s) is a strictly stable transfer function
• the real part of G(s) is strictly positive along the jω axis, ie.,
∀ω ≥ 0, <[G(jω)] > 0 (74)
Some necessary conditions for SPR
• G(s) is strictly stable
• The Nyquist plot of G(jω) lies entirely in the right half of complex plane
• G(s) has a relative degree of 0 or 1
• G(s) is strictly minimum phase
s−1 , G = s+1 , G =
Examples: G1 = s2+as+b 1 , G = s+1
2 s2−s+1 3 s2+as+b 4 s2+s+1

48
Theorem 5.2 A transfer function G(s) is positive real (PR) if and only if
• G(s) is a stable transfer function
• The poles of G(s) on the jω axis are simple and associated residues are
real and non-negative
• <[G(jω)] ≥ 0 for any ω such that jω is not a pole of G(s)

49
Kalman-Yakubovic Lemma Consider a controllable LTI system
ẋ = Ax + bu
y = cT x (75)
Its transfer function matrix is given by
G(s) = cT (sI − A)−1b (76)
is strictly positive real if and only if there exist positive definite matrices P
and Q such that
AT P + P A = −Q (77)
Pb = c (78)

50
5.2 Stability of feedback system
Consider the closed-loop system
ẋ = Ax + bu
y = cT x
u = −F (y)y (79)
If cT (sI − A)−1b is strictly positive real, then what is the condition on f (y)
for the stability if the closed-loop system?

51
 From K-Y lemma, there exist positive definite P and Q. Define V =
xT P x, its derivative is given by
V̇ = −xT Qx + 2xT P Bu
= −xT Qx − 2xT P BF (y)y
= −xT Qx − 2xT cF (y)y
= −xT Qx − 2F (y)y 2 (80)
Therefore if F (y) > 0 then system is exponentially stable. The condition is
only a sufficient condition.

52
5.3 Circle Criterion
Consider the case α < F (y) < β. Under what condition of cT (sI − A)−1b,
the closed-loop system (79) is stable?
Consider the transform
F −α
F̃ = (81)
β−F
and obviously we have F̃ > 0. How to use this transform for analysis of
systems stability?
Consider the characteristic equation of (79). We can write
G(s)F + 1 = 0 (82)
which gives
G(s)(F − α) = −αG − 1 (83)
G(s)(β − F ) = βG + 1 (84)
53
With (83) divided by (84), we can obtain
1 + βG F − α
· +1=0 (85)
1 + αG β − F
Let
1 + βG
G̃ := (86)
1 + αG
Based on the analysis of feedback stability of SPR system, we can conclude
that the closed-loop system is stable if G̃ is SPR.
What is the condition of G if G̃ is SPR? From (86), it can be obtained
that
G̃ − 1
G= (87)
β − αG̃
For the case of 0 < α < β, G̃ is SPR is equivalent to that Nyquist plot of
G does not encircle the circle centered as (− 12 (1/α + 1/β), 0) with radius of
1 (1/α − 1/β). That is why it is referred as circle criterion.
2
54
6. Feedback Linearization

6.1 Introduction
Consider the following nonlinear systems
x = f (x) + g(x)u
y = h(x) (88)
Two Examples:
ẋ1 = x2 + y 3
ẋ2 = y 2 + u
y = x1 (89)
ẋ1 = x2 + y 3 + u
ẋ2 = y 2 + u
y = x1 (90)
55
6.2 Input-Output Linearization
Lie Derivative A Lie derivative of a function h(x) along the vector field
f (x) is defined as
∂h(x)
Lf h(x) = f (x) (91)
∂x
Notation
∂L k−1h(x)
f
Lkf h(x) = Lf (Lk−1
f h(x)) = f (x) (92)
∂x
Relative Degree A system has relative degree ρ at a point x if the following
conditions are satisfied:
Lg Lkf h(x) = 0 for k = 0, . . . , ρ − 2 (93)
ρ−1
Lg Lf h(x) 6= 0 (94)
Examples. Use (89) and linear systems to demonstrate the concepts.
56
 It can be obtained that
y (k) = Lkf h(x) for k = 0, . . . , ρ − 1 (95)
(ρ) ρ ρ−1
y = Lf h(x) + Lg Lf h(x)u (96)
If we define the input as
1 ρ
u= ρ−1 [−Lf h(x) + v] (97)
Lg Lf h(x)
which results at
y (ρ) = v (98)

57
 Therefore the system is input-output linearized by defining
ξi = Li−1
f h(x) for i = 1, . . . , ρ (99)
and the linearized part of the system is described by
ξ˙1 = ξ2 (100)
.. (101)
ξ˙ρ−1 = ξρ (102)
ξ˙ρ = v (103)
The other part of system dynamics are characterized by the zero dynamics.
Examples

58
6.3 Full State Linearization
Consider
ẋ = f (x) + g(x)u (104)
where x ∈ rn, and u ∈ R. The full state linearization can be achieved if
there exists a function h(x) such that the relative degree regrading h(x) as
the output is n, the dimension of the system.
Lie Bracket For two vector fields f (x) and g(x), their Lie Bracket is defined
as
∂g ∂f
[f, g](x) = f (x) − g(x) (105)
∂x ∂x
Notations:
ad0f g(x) = g(x) (106)
ad1f g(x) = [f, g](x) (107)
adkf g(x) = [f, adk−1
f g](x) (108)
59
Example
   
 x2  0 

f (x) =

 ,


g(x) = 

(109)
− sin x1 − x2  x1 
  
 

Example
f (x) = Ax, g(x) = b (110)

60
Distribution The collection of vector spaces
∆ = span{f1(x), . . . , fn(x)} (111)
The dimension of distribution is defined as
dim(∆(x)) = rank[f1(x), . . . , fn(x)] (112)
Involutive A distribution is involutive if
g1 ∈ ∆ and g2 ∈ ∆ ⇒ [g1, g2] ∈ ∆

61
Example Let ∆ = span{f1, f2} where
   

2x2 



1 




f1(x) = 1  , f2(x) = 0  (113)

   
 
 
 
 

0  x2 
   
 

62
Theorem 6.1 The system (104) is feedback linearizable if and only if
• the matrix [g(x), adf g(x), . . . , adn−1
f g(x)] has full rank;
• the distribution span{g(x), adf g(x), . . . , adn−2
f g(x)} is involutive.

63
Example
ẋ1 = a sin x2
ẋ2 = −x21 + u

64
7. Backstepping Design

7.1 Integrator Backstepping

Consider
ẋ = f (x) + g(x)ξ
ξ˙ = u (114)
with x ∈ Rn and ξ ∈ R. There exist a function φ(x) with φ(0) = 0 and a
positive definite function V (x) such that
∂V
[f (x) + g(x)φ(x)] ≤ −W (x) (115)
∂x
where W (x) is positive definite.
Condition (115) implies that the system ẋ = f (x) + g(x)φ(x) is asymptot-
ically stable. How to design the control input so that the system is stable?

65
 Consider
ẋ = f (x) + g(x)φ(x) + g(x)(ξ − φ(x)) (116)
Let z = ξ − φ(x). It is easy to obtain that
ẋ = f (x) + g(x)φ(x) + g(x)z (117)
∂φ
ż = u − φ̇ = u − (f (x) + g(x)ξ) (118)
∂x
Consider a Lyapunov function candidate
1 2
Vc(x, ξ) = V (x) + z (119)
2
Its derivative is given by
∂V ∂V
V̇c = [f (x) + g(x)φ(x)] + g(x)z
∂x ∂x
∂φ
+z[u − (f (x) + g(x)ξ)] (120)
∂x
66
 Let
∂V ∂φ
u = −cz − g(x) + (f (x) + g(x)ξ) (121)
∂x ∂x
with c > 0 which results at
V̇c = −W (x) − cz 2 (122)
Theorem 7.1 For a system described in (114), the control design given in
(121) ensures the global asymptotic stability of the closed-loop system.

67
Example 7.1
ẋ1 = x21 + x2
ẋ2 = u (123)
Example 7.2
ẋ1 = −x1 + x21x2
ẋ2 = u (124)
Example 7.3
ẋ1 = x21 − x1x2
ẋ2 = u (125)

68
7.2 Iterative Backstepping
Consider a nonlinear system in the strict feedback form
ẋ1 = x2 + φ1(x1)
ẋ2 = x3 + φ2(x1, x2)
..
ẋn−1 = xn + φn−1(x1, x2, . . . , xn−1)
ẋn = u + φn(x1, x2, . . . , xn) (126)
Backstepping technique introduced in the previous section can be applied
iteratively. Define the notations
z1 = x1 (127)
zi = xi − αi−1(x1, . . . , xi−1), for i = 2, . . . , n (128)
where αi−1 for i = 2, . . . , n are stabilizing functions obtained in the iterative
beackstepping design.
69
Step 1:
ż1 = [x2 − α1] + α1 + φ1(x1)
= z2 + α1 + φ1(x1) (129)
Let
α1 = −c1z1 − φ1(x1) (130)
The resultant dynamics of z1 is
ż1 = −c1z1 + z2 (131)

70
Step 2: The dynamics of z2 is
ż2 = ẋ2 − α̇1
∂α1
= x3 + φ2(x1, x2) − (x2 + φ1(x1))
∂x1
∂α1
= z3 + α2 + φ2(x1, x2) − (x2 + φ1(x1)) (132)
∂x1
Design α2 as
∂α1
α2 = −z1 − c2z2 − φ2(x1, x2) + (x2 + φ1(x1)) (133)
∂x1
The resultant dynamics of z2 is given by
ż2 = −z1 − c2z2 + z3 (134)

71
Step i : For 2 < i < n, the dynamics of zi is given by
żi = ẋi − α̇i−1(x1, . . . , xi−1)
= xi+1 + φi(x1, . . . , xi)
i−1 ∂αi−1
− (xj+1 + φj (x1, . . . , xj ))
X

j=1 ∂xj
= zi+1 + αi + φi(x1, . . . , xi)
i−1 ∂αi−1
− (xj+1 + φj (x1, . . . , xj )) (135)
X

j=1 ∂xj
Design αi as
αi = −zi−1 − cizi − φi(x1, . . . , xi)
i−1 ∂αi−1
+ (xj+1 + φj (x1, . . . , xj )) (136)
X

j=1 ∂xj
The resultant dynamics of zi is given by
żi = −zi−1 − cizi + zi+1 (137)
72
Step n: At the final step, we have
żn = ẋn − α̇n−1(x1, . . . , xn−1)
= u + φn(x1, . . . , xn)
n−1 ∂αn−1
− (xj+1 + φj (x1, . . . , xj )) (138)
X

j=1 ∂xj
Design the final control input as
u = −zn−1 − cnzn − φn(x1, . . . , xn)
n−1 ∂αn−1
+ (xj+1 + φj (x1, . . . , xj )) (139)
X

j=1 ∂xj
The resultant dynamics of zn is given by
żn = −zn−1 − cnzn (140)

73
Stability Analysis Define a Lyapunov function candidate
1 Xn 2
V = zi (141)
2 i=1
It is easy to obtain that
V̇ = z1[−c1z1 + z2] + z2[−z1 − c2z2 + z3] + . . .
+zn−1[−zn−2 − cn−1zn−1 + zn] + zn[−zn−1 − cnz2]
n
= − cizi2 (142)
X

i=1
Therefore the closed loop system under the proposed control is asymptotically
stable.
Theorem 7.2 For a system in the strict feedback form (126), the control
input (139) renders the closed-loop system asymptotically stable.

74
Example 7.4
ẋ1 = x2 + x21
ẋ2 = x3 + x1x2
ẋ3 = u + x23 (143)
Example 7.5
ẋ1 = −x1 + x21x2
ẋ2 = x3
ẋ3 = u (144)

75
8. Adaptive Control I

8.1 Introduction
The basic objective of adaptive control is to maintain consistent performance
of a system in the presence of uncertainty or unknown variation in plant
parameters.
MRAC Model Reference Adaptive Control consists of reference model which
produce the desired output, and difference between the plant output and
the reference output is then used to adjust the control parameter and the
control input directly. MRAC is often in continuous-time domain, and for
deterministic plants.
STC Self-Tuning Control estimates systems parameters and then compute
the control input from the estimated parameters. STC is often in discrete-
time and and for stochastic plants.
We will focus on MRAC in this course.
76
Design of adaptive controllers Compared with the conventional con-
trol design, adaptive control is more involved, with the need to design the
adaptation law. Adaptive control design usually involves the following three
steps:
• Choose a control law containing variable parameters
• choose an adaptation law for adjusting those parameters
• analysis the convergence properties of the resulting control system

77
8.2 Adaptive control of first-order systems
Consider a first order system
ẏ + apy = bpu (145)
The output y is to follow the output of the reference model
ẏm + amym = bmr (146)
The reference model is stable, ie., am > 0. The signal r is the reference
input. The design objective is to make the tracking error e = y−ym converge
to 0.

78
Model Reference Control Rearrange the system model as
ap − am
ẏ + amy = bp[u − y] (147)
bp
and therefor the we obtain
ap − am bm
ė + ame = bp[u − y − r]
bp bp
:= bp[u − ay y − ar r] (148)
a −a
where ay = p bp m , ar = bbmp .
If all the parameters are known, the control law is designed as
u = ar r + ay y (149)

79
Adaptive control law With parameters unknown, let âr and ây denote
their estimates of ar and ay . The control law under the certainty equivalence
principle is given by
u = âr r + ây y (150)
Adaptive laws The estimates are updated by
â˙ r = −sign(bp)γr er (151)
â˙ y = −sign(bp)γy ey (152)
were γr and γy are positive real design parameters, and are often referred to
as adaptive gains.
The closed loop system under the adaptive control law is given by
ė + ame = bp[−ãy y − ãr r] (153)
where ãr = ar − âr and ãy = ay − ây .

80
Stability analysis Consider the Lyapunov function candidate
1 2 |bp| 2 |bp| 2
V = e + ãr + ãy (154)
2 2γr 2γy
Its derivative along the trajectory (153) and adaptive laws (151) and (151)
is given by
V̇ = −ame2 + ãr [|bp|ã˙ r − ebpãr r] + ãy [|bp|ã˙ y − ebpãy y]
= −ame2 (155)
Therefore the system is Lyapunov stable with all the variables e, âr and ây
are all bounded. Furthermore, we can show that
Z
∞ e2(t)dt = V (0) − V (∞) < ∞ (156)
o
am
Therefore we have established that e ∈ L2 ∩ L∞ and ė ∈ L∞. To conclude
the stability analysis, we need Babalat’s Lemma.
81
Barbalat’s Lemma If a function f (t) is uniformly continuous for t ∈
[0, ∞), and R0∞ f (t)dt exists, then limt→∞ f (t) = 0.
Since ė and e are bounded , e2 is uniformly continuous. Therefore we
can conclude from Barbalat’s Lemma that limt→∞ e2(t) = 0 and hence
limt→∞ e(t) = 0.

82
Simulation study of first order system

2
Gp = (157)
s−1
1
Gm = (158)
s+2

83
9. Adaptive Control II

9.1 Model reference control of high order systems

Consider a nth-order system
Zp(s)
y(s) = kp u(s) (159)
Rp(s)
where kp is the high frequency gain, Zp and Rp are monic polynomials with
orders of n−ρ and n respectively with ρ as the relative degree. The reference
model is given by
Zm(s)
y(s)m = km r(s) (160)
Rm(s)
where km > 0 and Zm and Rm are monic Hurwitz polynomials.

84
MRC of systems with ρ = 1 Following a similar manipulation to the first
order system, we have
y(s)Rm(s) = kpZp(s)u(s) − (Rp(s) − Rm(s))y(s)
Zp(s) Rm(s) − Rp(s)
= kpZm(s)[ u(s) + y(s)]
Zm(s) Zm(s)
θ1T α(s) θ2T α(s)
= kpZm(s)[u(s) − u(s) − y(s) − θ3y(s)]
(161)
Zm(s) Zm(s)
where θ1 ∈ Rn−1, θ2 ∈ Rn−1, θ3 ∈ R and
α(s) = [sn−2, . . . , 1]T (162)
Hence, we have
Zm(s) θ1T α(s) θ2T α(s)
y = kp [u(s) − u(s) − y(s) − θ3y(s)] (163)
Rm(s) Zm(s) Zm(s)

85
and
Zm(s) θ1T α(s) θ2T α(s)
e1 = kp [u(s) − u(s) − y(s) − θ3y(s) − θ4r](164)
Rm(s) Zm(s) Zm(s)
where e1 = y − ym and θ4 = kkmp . The control input for the model reference
control is given by
θ1T α(s) θ2T α(s)
u = u+ y + θ3 y + θ4 r
Zm(s) Zm(s)
= θT ω (165)
where
θT = [θ1T , θ4T , θ3, θ4]
ω = [ω1T , ω2T , y, r]T
α(s) α(s)
ω1 = u; ω2 = y
Zm(s) Zm(s)
86
Example 9.1 Design MRC for the system
s+1
y(s) = 2 u (166)
s − 2s + 1
with the reference model
s+3
ym(s) = 2 r (167)
s + 2s + 3

87
MRC of systems with ρ > 1 For a system with ρ > 1, the input in the
same format as (165) can be obtained. The only difference is that Zm is
of order n − ρ < n − 1. In this case, we let P (s) be a monic polynomial
with order ρ − 1 so that Zm(s)P (s) is of order n − 1. We adopt a slightly
different approach from the case of ρ = 1. Consider the identity
Zm R m P
y = [ y]
R m Zm P
Zm QRp + ∆1
= [ y]
R m Zm P
Zm QZpu + kp−1∆1y
= kp [ ]
Rm Zm P
Zm ∆2 kp−1∆1
= kp [u + u+ y] (168)
Rm Zm P Zm P
where Q is a monic polynomial of order ρ − 1, ∆1 and ∆2 are polynomials

88
of order n − 1 and n − 2 respectively. Therefore we can write
Zm θ1T α(s) θ2T α(s)
y = kp [u − u− y − θ3y] (169)
Rm Zm P Zm P
and
Zm θ1T α(s) θ2T α(s)
e1 = kp [u − u− y − θ3y − θ4r] (170)
Rm Zm P Zm P
Hence the control input is designed as
θ1T α(s) θ2T α(s)
u = u+ y + θ3y + θ4r]
Zm(s)P (s) Zm(s)P (s)
= θT ω (171)
with the same format as (165) but
α(s) α(s)
ω1 = u; ω2 = y
Zm(s)P (s) Zm(s)P (s)
89
Example 9.2 Design MRC for the system
1
y(s) = 2 u (172)
s − 2s + 1
with the reference model
1
ym(s) = 2 r (173)
s + 2s + 3

90
9.2 Adaptive Control (ρ = 1)
Basic Assumptions Adaptive control design in this course assumes
• the known system order n,
• the known relative degree ρ
• the minimum phase of the plant,
• the known sign of the high frequency gain sign[kp].

91
 We choose the reference model as SPR. If the parameters are unknown,
the control law based on the certainty equivalence principle is given as
u = θ̂T ω (174)
where θ̂ is the estimate of θ. The adaptive law is designed as
˙
θ̂ = −sign[kp]Γe1ω (175)
where Γ ∈ R2n, the adaptive gain, is a positive definite matrix, and ω is
generated in the same way as in the MRC case.

92
Stability Analysis In this can we can write
Zm T
e1 = kp (θ̂ ω − θT ω)
Rm
Zm k p T
= km [− θ̃ ω] (176)
Rm km
where θ̃ = θ − θ̂. Furthermore, we can put in the state space form as
kp T
ė = Ame + bm[− θ̃ ω]
km
e1 = cTme (177)
Zm(s)
where {Am, bm, cm} is a minimum state space realization of km R , ie,
m(s)

T −1 Zm(s)
cm(sI − Am) bm = km (178)
Rm(s)

93
Since {Am, bm, cm} is SPR, there exist positive definite matrices P and Q
such that
ATmP + P Am = −Q (179)
P bm = cm (180)
Define a Lyapunov function candidate as
1 T 1 kp T −1
V = e P e + | |θ̃ Γ θ̃ (181)
2 2 km
Its derivative is given by
1 T T kp T kp T −1 ˙
V̇ = − e Qe + e P bm[− θ̃ ω] + | |θ̃ Γ θ̃
2 km km
1 T kp T kp T −1 ˙
= − e Qe + e1[− θ̃ ω] + | |θ̃ Γ θ̃
2 km km
1 T
= − e Qe (182)
2
94
We can now conclude the boundedness of e and θ̂. Furthermore it can be
shown that e ∈ L2 and ė1 ∈ L∞. Therefore from Barbalat’s lemma we have
limt→∞ e1(t) = 0. The boundedness of other system state variables can be
established from the minimum phase property of the system.

95
9.3 Adaptive Control with ρ > 1
For systems with higher relative degrees, the certainty equivalence principle
can still be use to design the control, ie,
u = θ̂T ω (183)
with ω be the same as in (171). However the adaptive law is more involved
as the reference model is no longer SPR. If L(s) is a polynomial of order
ρ − 1 which makes km ZRmmL SPR, then the error model is given by
Zm L
e1 = km [k(uf − θT φ)] (184)
Rm
k 1 u and φ = 1 ω. An auxiliary error is constructed
where k = kmp , uf = L(s) L(s)
as
Zm L T Zm L 2
 = e1 − km [k̂(uf − θ̂ φ)] − km [ns] (185)
Rm Rm
96
where k̂ is an estimate of k, n2s = φT φ + u2f . The adaptive laws are designed
as
˙
θ̂ = −sign[bp]Γφ (186)
˙
k̂ = γ(uf − θ̂T φ) (187)

97
10. Robust Issues in Adaptive Control

10.1 Introduction
Adaptive control and stability analysis have been carried under the condition
that there are only parameter uncertainty in the system. However, many
types of non-parametric uncertainties do exist in practice. These include
• high-frequency unmodelled dynamics, such as actuator dynamics or struc-
tural vibrations
• low-frequency unmodelled dynamics, such as Coulomb frictions.
• measurement noise
• computation roundoff error and sampling delay
Such non-parametric uncertainties will affect the performance of adaptive
control systems when they are applied to control practical systems. They
may cause instability. Let us consider a simple example.
98
 Consider the system output is described by
y = θω (188)
The adaptive law
˙
θ̂ = γω (189)
 = y − θ̂ω (190)
will render the convergence of the estimate θ̂ by taking
1 2
V = θ̃ (191)
2γ
as a Lyapunov function candidate and the analysis
V̇ = −θ̃(y − θ̂ω)ω
= −θ̃2ω 2 (192)
Now, if the signal is corrupted by some unknown bounded disturbance d(t),
ie.,
y = θω + d(t) (193)
99
The same adaptive will have a problem. In this case,
V̇ = −θ̃(y − θ̂ω)ω
= −θ̃2ω 2 − θ̃dω
θ̃2ω 2 1 d 2
= − − (θ̃ω + d)2 + (194)
2 2 2
From the above analysis, we cannot conclude the boundedness of θ̃ even we
have ω bounded. In fact, if we take θ = 2, γ = 1 and ω = (1 + t)1/2 ∈ L∞
and let
−1/4 5
d(t) = (1 + t) ( − 2(1 + t)−1/4) (195)
4
It can then be obtained that
5
y(t) = (1 + t)−1/4, → 0 as t → ∞ (196)
4
˙ 5
θ̂ = (1 + t)−3/4 − θ̂(1 + t)−1 (197)
4
100
which has a solution
θ̂ = (1 + t)1/4 → ∞ as t → ∞ (198)
In this example, we have observed that adaptive law designed for disturbance-
free system fails to remain bounded even the disturbance is bounded and
converges to zero as t tends to infinity.
In this session, we conside the simple model
y = θω + d(t) (199)
with d as bounded disturbance. A number of robust adaptive laws will be
introduced for this simple model. In the following we keep use  = y − θ̂ω
θ̃ 2
and V = 2γ . Once the students understand the basic ideas, they can extend
the robust adaptive laws to adaptive control.

101
10.2 Dead-zone Modification
The adaptive law is modefied as
˙  γω || > g


θ̂ =  (200)
 0 || ≤ g
where g is a constant satisfying g > |d(t)| for all t. For  > g, we have
V̇ = −θ̃ω
= −(θω − θ̂ω)
= −(y − d(t) − θ̂ω)
= −( − d(t))
< 0 (201)
Therefore we have 

 < 0 || > g





V̇ =  (202)
 0 || ≤ g
and we can conclude that V is bounded.
102
10.3 σ-Modification
The adaptive law is modefied as
˙
θ̂ = γω − γσ θ̂ (203)
where σ is a positive real constant. In this case, we have
V̇ = −( − d(t)) + σ θ̃θ̂
= −2 + d(t) − σ θ̃2 + σ θ̃θ
2 d20 θ̃2 θ2
≤ − + −σ +σ
2 2 2 2
d20 θ2
≤ −σγV + + σ (204)
2 2
where d0 ≥ |d(t)|, ∀t ≥ 0. To establish the boundedness of V , we need a
standard result, which is a special case of the comparison lemma.

103
Comparison Lemma Let f, V : [0, ∞) 7→ R. Then
V̇ ≤ −αV + f, ∀t ≥ t0 ≥ 0 (205)
implies that
V (t) ≤ e−α(t−t0 ) V (t0) + tt0 e−α(t−τ )f (τ )dτ,
Z
∀t ≥ t0 ≥ 0 (206)
for any finite constant α.

104
 Applying the comparison lemma to (204), we have
−σγt t −σγ(t−τ ) d20 θ2
V (t) ≤ e
Z
+ V (0) 0 e [ + σ ]dτ, (207)
2 2
1 d20 θ2
V (∞) ≤ [ +σ ] (208)
σγ 2 2
Therefore we can conclude that V ∈ L∞.

105
10.4 Robust Adaptive Control
The robust adaptive laws introduced can be applied to various adaptive con-
trol schemes. We demonstrate the application a robust adaptive to MRAC
with ρ = 1. We start directly from the error model
ė = Ame + bm[−k θ̃T ω + d(t)]
e1 = cTme (209)
where k = kp/km and d(t) is a bounded disturbance with bound d0, which
represents the non-parametric uncertainty in the system. As discussed earlier,
we need a robust adaptive law to deal with the bounded disturbances. If we
take σ-modification, then the adaptive law is
˙
θ̂ = −sign[kp]Γe1ω − σΓθ̂ (210)
This adaptive law will ensure the boundedness of the variables.

106
Stability Analysis Let
1 T 1
V = e P e + |k|θ̃T Γ−1θ̃ (211)
2 2
Its derivative is given by
1 T
V̇ = − e Qe + e1[−k θ̃T ω + d] + |k|θ̃T Γ−1θ̃˙
2
1
= − λmin(Q)kek2 + e1d + |k|σ θ̃T θ̂
2
1
= − λmin(Q)kek2 + |e1d| − |k|σkθ̃k2 + |k|σ θ̃T θ (212)
2
Note
1 d2
0
|e1d| ≤ λmin(Q)kek2 +
4 λmin(Q)
1 1
|θ̃ θ| ≤ kθ̃k + kθk2
T 2 (213)
2 2
107
Hence we have
1 |k|σ d 2 |k|σ
2 2 0
V̇ ≤ − λmin(Q)kek − kθ̃k + + kθk2
4 2 λmin(Q) 2
d20 |k|σ
≤ −αV + + kθk2 (214)
λmin(Q) 2
where α is a positive real and
min{ 12 λmin(Q), |k|σ}
α= (215)
max{λmaxP, λ |k|(Γ) }
min
Therefore, we can conclude the boundedness of V from the comparison
lemma, which further implies the boundedness of the tracking error e1 and
the estimate θ̂.

108
11. Adaptive Control of Nonlinear Systems

11.1 First Order Nonlinear Systems

Consider a first order nonlinear system described by
ẏ = u + φT (y)θ (216)
where φ : R 7→ Rp is a smooth nonlinear function, and θ ∈ Rp is a unknow
vector of constant parameters. For this system, adaptive control law can be
designed as
u = −cy − φT (y)θ̂ (217)
˙
θ̂ = Γyφ(y) (218)
where c is a positive real constant, and Γ is a positive definite gain matrix.

109
 The closed-loop dynamics is given by
ẏ = −cy + φT (y)θ̃ (219)
with the usual notation θ̃ = θ − θ̂.
Stability Analysis Let
1 2 1 T −1
V = y + θ̃ Γ θ̃ (220)
2 2
and its derivative is obtained as
V̇ = −cy 2 (221)
which ensures the boundedness of y and θ̂. We can show limt→∞ y(t) = 0
in the usual way.

110
11.2 Adaptive Backstepping
Consider a second order nonlinear system
ẋ1 = x2 + φT1 (x1)θ
ẋ2 = u + φT2 (x1, x2)θ (222)
where φ1 : R 7→ R, φ2 : R2 7→ Rp are smooth nonlinear functions, and
θ ∈ Rp is an unknow constant parameters. From the previous section, we
know that if x2 = −c1x1 − φT1 (x1)θ̂, then the first part of the system is
stable. Hence if we set
α = −c1x1 − φT1 (x1)θ̂ (223)
then backstepping control design can be used to designed the control input
u in a similar way as for the systems without unknown parameters.

111
 For the adaptive backstepping design, we define
z1 = x1 (224)
z2 = x2 − α (225)
We have the dynamics for z1 as
ż1 = z2 + α + φT1 (x1)θ
= −c1z1 + z2 + φT1 (x1)θ̃ (226)
Consider the dynamics for z2
ż2 = ẋ2 − α̇(x1, θ̂)
T ∂α ∂α ˙
= u + φ2 (x1, x2)θ − ẋ1 − θ̂
∂x1 ∂ θ̂
∂α T ∂α ∂α ˙
= u + [φ2(x1, x2) − φ1(x1)] θ − x2 − θ̂ (227)
∂x1 ∂x1 ∂ θ̂
Notice that the adaptive law designed based on the dynamics of z1 will not
work because the dynamics of z2 involves the unknown parameters θ as well.
112
We leave the adaptive law to the Lyapunov function based analysis later.
From the dynamics of z2, we design the control input as
∂α T ∂α ∂α ˙
u = −z1 − c2z2 − [φ2(x1, x2) − φ1(x1)] θ̂ + x2 + θ̂ (228)
∂x1 ∂x1 ∂ θ̂
The resultant dynamics of z2 is given by
∂α
ż2 = −z1 − c2z2 + [φ2(x1, x2) − φ1(x1)]T θ̃ (229)
∂x1
Consider a Lyapunov function candidate
1 2
V = [z1 + z22 + θ̃T Γ−1θ̃] (230)
2

113
 We then have
V̇ = z1[−c1z1 + z2 + φT1 θ̃]
∂α T
+z2[−z1 − c2z2 + [φ2 − φ1] θ̃]
∂x1
˙ T −1
−θ̂ Γ θ̃
∂α ˙
2 2
= −c1z1 − c2z2 + [z1φ1 + z2(φ2 − φ1) − Γ−1θ̂]T θ̃ (231)
∂x1
From the above analysis, we decide the adaptive law as
˙ ∂α
θ̂ = Γ[z1φ1 + z2(φ2 − φ1)] (232)
∂x1
and we have
V̇ = −c1z12 − c2z22 (233)
Therefore we have the boundedness of z1, z1 and θ̂ and limt→∞ zi = 0 for
i = 1, 2.
114
11.3 Adaptive Control of Strict Feedback Systems
Consider a nonlinear system in the strict feedback form
ẋi = xi+1 + φTi (x1, x2, . . . , xi)θ for i = 1, . . . , n − 1
ẋn = u + φTn (x1, x2, . . . , xn)θ (234)
Adaptive control design can be carried using backstepping iteratively in n
steps. The main difficulty is in design of adaptive law, as the seam unknown
parameter appears in every step. The initial method used multiple estiamtes
for θ. Later the tuning function method was proposed to solve the multiple
estimate problem. For more details please refer to (Krstic et al, 1995)

115
Example 11.1 Design an adaptive control law for
ẋ1 = x2 + (ex1 − 1)θ
ẋ2 = u (235)
Example 11.2 Design an adaptive control law for
ẋ1 = x2 + x31θ + x21
ẋ2 = (1 + x21)u + x21θ (236)

116