EE5104 Adaptive Control Systems Part I

EE5104 Adaptive Control Systems
Part I
Dr Shuzhi Sam GE
Department of Electrical Engineering
National University of Singapore
Singapore 119260
Tel: 874 6821

Fax: 779 1103
E-mail: elegesz@nus.edu.sg
Office: E4-05-26F
http://vlab.ee.nus.edu.sg/~sge
References
1. K.J. Astrom and B. Wittenmark, Adaptive Control, Addison
Wesley, 1989 (1995).
2. K.S. Narendra and A.M. Annaswamy, Stable Adaptive Systems,
Prentice-Hall, 1989.
3. G.C. Goodwin and K.S. Sin, Adaptive Filtering, Prediction and
Control, Prentice, 1984.
Adaptive Control Systems --2-- S. Ge
Contents No. of
Hours
1 Introduction 3
Adaptive schemes. Adaptive control theory. Applications.
Situations when constant Gain feedback is insufficient.
Robust control. The adaptive control problem.
2 Model Reference Adaptive Systems 8

The model following problem. MRAS based on stability theory.
Model following when the full state is measurable. Direct MRAS
for general linear systems. Prior knowledge in MRAS. MRAS for
partially known systems. Use of robust estimation methods in
MRAS.
3 Self-Tuning Regulators 8
The basic idea. Indirect self-tuning regulators. Direct Self-tuning
regulators. Linear Quadratic STR. Adaptive Predictive control.
Prior knowledge in STR.
4 Real-Time Parameter Estimation 8
Linear-in-the-parameters model. Least squares estimation.
Experimental conditions. Recursive estimators. Extended least
squares. Robust estimation methods (dead zone, projection).
Implementation issues.
5 Auto-Tuning 3
PID Control. Transient response methods. Methods based on
relay feedback. Relay oscillations.
6 Gain Scheduling 3
The Principle. Design of Gain-scheduling Regulators. Nonlinear
Transformations. Applications of Gain-scheduling.
7 Alternatives to Adaptive Control 3

Situations where adaptive control should not be used. Robust
high gain feedback control. Self-oscillating adaptive systems.
Variable structure systems.
Pre-requisites for this course

　
•　　Linear Systems Theory
1. Controllability, Observability
2. General Observer Theory
• Digital Techniques
1. Digital filter design (e.g. Butterworth) and implementation

considerations)
2. Sampling considerations
3. Implementation of compensators, issues of finite word length
coefficient representation, and rounding errors
The course will assume that you are already very familiar with these
topics.
A simple adaptive control problem
(Narendra & Annaswamy, pp.111-114)

(Astrom & Wittenmark, pp.126-127)
Consider a system to be controlled
y = G yu u
with transfer function
kp
G yu ( s ) =
s −ap
In time domain
( p − a p ) y (t ) = k p u (t )
where we use the notation
d
p≡
dt
for the differential operator.
Thus
y& (t ) = a p y (t ) + k p u (t ) (1)
We usually refer to the system to be controlled as the “plant”.

Control Objective:
• Drive y(t) to follow some reference command trajectory

y m (t )
For example, y m (t ) may be generated by a reference system
y& m (t ) = am y m (t ) + k m r (t )
a m < 0 , i.e. stable reference model.

Consider a control input of the form
u (t ) = θ ∗ y (t ) + k ∗ r (t )
Then, the plant together with control (the closed-loop system) is
y& (t ) = a p y (t ) + k p {θ ∗ y (t ) + k ∗r (t )}
= ( a p + k pθ ∗ ) y (t ) + k p k ∗r (t )
If
am − a p km
θ∗ = and k ∗ =
kp kp
then
y& = a m y + k m r
and the plant together with the control will match the reference model
exactly because
e = y m − y , e& = ame
PROBLEM: We do not know a p and k p !

Adaptive Control: Find a way to achieve this automatically.
• Estimate a p and k p on-line, and try to use these estimates;

(Will work most of the time); and
• Use a rigorous procedure to “evolve” to the right values (Very

powerfull but applies only to restricted classes).
In the course we will look at both types of approaches.
For now, let us consider the rigorous approach.

Since θ ∗ , k ∗ not known, consider instead
u(t ) = θ (t ) y (t ) + k (t ) r(t ) , Time-varying controller gains!
How to choose {θ (t ), k (t ), t ∈[ 0, ∞)}?
Adaptive law
θ&(t ) = − sgn( k p )γ 1e(t ) y (t ); γ1 > 0
k& (t ) = − sgn( k p )γ 2 e (t ) r (t ); γ2 > 0
(Nonlinear updating mechanisms!)
Thus, adaptive controllers contain both time-varying and nonlinear

dynamics!!
Because of the NLTV characteristics, analysis and design usually

difficult.
For this system, consider the quadratic form (recall your 3rd year
mathematics!)
V (e, φ ,ψ ) = [
1 2
2
(
e + k p γ 1−1φ 2 + γ 2−1ψ 2 )]
with
e = y − ym , φ = θ −θ ∗, ψ = k − k∗
Re-membering that θ ∗ and k ∗ are constants
∗
φ& = θ& − θ& = − sgn( k p )γ 1ey
ψ& = k& − k& ∗ = − sgn( k p )γ 2 er
Note that
Reference Model:
y& m = am y m + k mr
Plant:
y& = a p y + k p u = a p y + k pθy + k p kr
= a p y + k p (φ + θ * ) y + k p (ψ + k * ) r
= (a p + k pθ * ) y + k pφy + k pψr + k p k *r
we have
e& = y& − y& m = ame + k pφy + k pψr
Consider derivative of quadratic form V

[
V& = ee& + | k p | γ 1−1φφ& + γ 2−1ψψ& ]
[
= am e 2 + k pφey + k pψer − | k p | sgn( k p )φey + sgn( k p )ψer ]
= am e 2 ≤ 0
Collecting together the intermediate results:
• Quadratic form V (e, φ ,ψ ) =

1 2
2
[ (
e + | k p | γ 1−1φ 2 + γ 2−1ψ 2 )]
results in V& = a m e 2 ≤ 0
• This means that

e 2 is bounded, φ 2 is bounded, and ψ 2 is bounded
∆
• φ (t ) = θ (t ) − θ ∗
∆
ψ (t ) = k (t ) − k ∗
∴ θ (t) and k(t) are bounded
∆
• e (t ) = y (t ) − y m (t )
y m (t ) is a reference signal, and obviously bounded
∴ y(t) is bounded
• V(e, φ,ψ) positive definite
V& = a m e 2 ≤ 0
∴ V is bounded over all time
t t
∫ V& (e (t ), φ (t ),ψ (t ) )dτ = ∫ ame (τ )dτ
2
0 0
t
⇒ −am ∫ e 2 (τ )dτ = V (e (0), φ (0),ψ ( 0) ) − V (e (t ), φ (t ),ψ (t ) )
0
Since V (e (t ), φ (t ),ψ (t ) ) is always bounded,
t
lim ∫ e 2 (τ ) dτ ≤ c1 a constant
t→∞ 0
n e& = a m e + k p φy + k pψr
∴ e& (t ) is bounded ∀t
e& bounded 
∞ 
n  ⇒ lim e (t ) = 0
∫ e (τ )dτ ≤ c1  t → ∞
2
0 
i.e. lim {y (t ) − y m (t )} = 0
t →∞
The adaptive controller achieves the designed objective!
Rigorous analysis of adaptive control is typically as difficult.

Summary
Plant: y& (t ) = a p y (t ) + k p u (t )
Reference Model: y& m (t ) = am y m (t ) + k m r (t ); am < 0
Control Law: u(t ) = θ (t ) y(t ) + k (t ) r (t ) , Time varying
Adaptive Law: θ&(t ) = − sgn( k p )γ 1e(t ) y (t )
k& (t ) = − sgn( k p )γ 2 e(t ) r(t )

e (t ) = y (t ) − y m (t )
θ ( 0) = k ( 0) = 0 , Arbitrary starting gains
Result:
If the adaptive controller is applied to the plant, then all signals
{y, u, θ, k} are bounded and lim [ y (t ) − y m (t ) ] = 0
t →∞
Adaptive Control - 14- S.Ge
A user-friendly picture
Gm
Model
Regulator parameter
Adjustment
mechanism
r u y
Regulator Plant
Figure 4.1 Block diagram of a model-reference adaptive system

(MRAS)
Lyapunov’s Direct Method
l Enables one to determine whether or not the equilibrium state

of system
x& = f ( x, t ) (2.1)
is stable without actually determining the solution x (t ; x0 , t0 ) .
l Involves finding a suitable scalar function V(x,t) and examining

its time derivative V& ( x , t ) along the trajectory of the system.
We will only need one result from this method, which we will state
without proof.
Theorem (Uniform Stability)
Assume that a scalar function V(x,t), with continuous first partial

derivatives w.r.t x and t, exists and that V(x,t) satisfies the
following conditions:
(i) V(x,t) is positive-definite, i.e.,∃ a continuous non-

decreasing scalar function α such that α (0) = 0 and
V ( x, t ) ≥ α ( x ) > 0 ∀t and all x ≠ 0 ;
(ii) V(x,t) is decrescent, i.e., ∃ a continuous non-decreasing

scalar function such that β (0) = 0 and β ( x ) ≥ V ( x, t ) for
all t;
(iii) V& ( x , t ) is negative semi-definite, i.e.,

∂V
V& ( x , t ) = + (∇ V ) T f ( x , t ) ≤ 0
∂t
(iv) V(x,t) is radial unbounded, i.e.,

α ( x ) → ∞ as x → ∞
then
(a) (i) and (iii) imply that the origin of differential equation (2-1)
is stable;
(b) (i), (ii) and (iii) imply the origin of d.e.(2-1) is uniformly
stable;
(c) (i) - (iv) imply that the origin of d.e. (2-1) is uniformly stable
in the large.
LTI System and Lyapunov Stability
Consider
x& = Ax ; x (t 0 ) = x 0
Theorem 2-2 (Narendra ….)
All solutions of the above equation tend to zero as t → ∞ if and

only if all the eigenvalues of A have negative real parts,
Such a matrix A is referred to as an asympototically stable matrix.
Theorem 2-10 (Narendra, …)
The equilibrium state x=0 of the LTI system

x& = Ax
is asymptotically stable if and only if, given any symmetric positive

definite matrix Q , there exists a symmetric positive definite matrix
n(n + 1)
P which is the unique solution of the linear equations
2
AT P + PA = −Q
Proof: See Narendra & Annaswamy pp.60.

Stability theory, in general, is a very wide field.
• See Narendra &Annaswamy, Chap.2 for a more comprehensive

overview.
• We have only summarized the concepts and results that we need

for adaptive control.
Adaptive control of plants with all state-variables measurable
(A similar treatment can be found in Narendra and Annaswamy

pp.128 ff. However, note that some aspects are different.)
Assume that the plant is described by
x& p = Ap x p + gbu
x p ∈ ℜ n , u ∈ ℜ1
A p is ( n× n) , b is ( n × 1)
b is known
This is possible in many situations.

Example 2.1: b is known
y = G yu u
g
G yu ( s ) =
s 2 + a1s + a2
Assume that position and velocity measurements are available.
Then, a suitable state-variable description is
x1 = y
x&1 = x2
x& 2 = &y& = − a1x 2 − a2 x1 + gu
or
 x&1   0 1   x1  0
 x&  =  − a − a1   x 2 
+ g 1 u
 2  2  
0
Thus b=  is known exactly.
1
Consider plant:
x& p = A p x p + gbu (1)
Non-adaptive case:
T
u = θ x∗ x p + θ r∗r (2)
[ ]
∆ T
θ ∗x = θ1∗ θ 2∗ L θ n∗
Then

T

( )
x& p =  Ap + gbθ x∗  x p + b gθ r∗ r
Assume that for these ＊ values, the feedback control (2) achieves
model matching in the sense that
(n × n) 
T
A p + gbθ x∗ ≡ Am
 Matching conditions
gθ r∗ ≡ gm scalar 
Thus, control gains θ x∗ ∈ ℜn , and θ r∗ ∈ ℜ exist to guarantee that

the closed-loop system match the reference model
x& m = Am xm + g mbr
Example 2-1 (continued)
Choice of reference model
For example, the following might be desirable
y m = G ym r r
ω n2
Gymr (s) =
s 2 + 2ξω n s + ω n2
• unity steady-state gain

• speed of response defined by natural frequency ω n
• damping specified by damping coefficient ξ
State-variable form
 x&1m   0 1   x1m  2 0 
 x&  = − ω 2 − 2ξω n   x2 m 
+ ω n  r
 2m   n 1
Step Response
Title:
a:step.eps
Creator:
MATLAB, The Mathworks, Inc.
Preview:
This EPS picture was not saved
with a preview included in it.
Comment:
This EPS picture will print to a
PostScript printer, but not to
other types of printers.
Impulse Response
Title:
a:impulse.eps
Creator:
MATLAB, The Mathworks, Inc.
Preview:
This EPS picture was not saved
with a preview included in it.
Comment:
This EPS picture will print to a
PostScript printer, but not to
other types of printers.
Adaptive Case:
To adaptive match the reference model, consider the control law
u(t ) = θ xT (t ) x p (t ) + θ r (t )r (t )
Define parameter errors
∆
φ x (t ) = θ x (t ) − θ ∗x
∆
φ r (t ) = θ r (t ) − θ r∗
Then, the control law applied to the plant results in
{
x& p = Ap x p + gb θ xT x p + θ r r }
=  A p + gbθ x∗  x p + gbφ xT x p + gbθ r r
T
 
= Am x p + gbφ xT x p + gbθ r r
Compared with
x& m = Am xm + g mbr
= Am x m + gbθ r∗ r
where g m = gθ r∗ .
Consider state error
∆
e = x p − xm
Then
e& = Am e + gbφ Tx x p + gbφ r r

= Am e + gbφ T x
where
∆ φ  ∆ x 
φ = x ; x = p 
φ r  r 
For a reference model, Am must be chosen to be a stable matrix.

Thus, it satisfies the Lyapunov equation.
AmT P + PAm = −Q
i.e., for any symmetric positive definite matrix Q, there exists a

symmetric positive definite P satisfying the above equation.
(From Linear Systems theory, or see Narendra & Annaswamy

pp.60)
Consider a Lyapunov function candidate
V (e (t ), φ (t ) ) = e (t )T Pe (t ) + | g | φ (t ) T Γ −1φ (t )
where
Γ is a symmetric positive definite (s.p.d.) matrix
We will sometime write V (e(t ),φ (t )) as V(t) for short.
Evaluate V& along the trajectory of the system
V& = 2eT Pe& + 2 g φ T Γ −1φ&

{ }
= 2e T P Am e + gbφ T x + 2 g φ T Γ −1φ&
= 2e T PAme + 2 geT Pbφ T x + 2 g φ T Γ −1φ&
(
= e T Am )
P + PAm e + 2 geT Pbφ T x + 2 g φ T Γ −1φ&
T T
= −eT Qe + 2 geT Pbφ T x + 2 g φ T Γ −1φ& (3)

l − eT Qe term always ≤ 0
l Ｕse the design
& θ&x 
θ =   = − sgn( g )Γxe T Pb
&
θ r 
and note that
φ = θ −θ ∗ , ⇒ φ& = θ&
Equation (3) becomes
V& = −eT Qe + 2 geT Pbφ T x − 2 g sgn( g )φ T xeT Pb

= −eT Qe ≤ 0
We can go through a similar analysis to show that

• V (t ) is positive definite and V& (t ) ≤ 0 ⇒ V (t ) is bounded
• e,φ (hence θ ) are bounded
• e& is bounded
∞
• ∫ e T Qed τ ≤ c1
0
• lim e = 0 (or lim x p − x m = 0 )

t →∞ t →∞
Summary
Plant
x& p = A p x p + gbu
with x p ∈ ℜ n measurable, and b known
Matching Conditions
T
A p + gbθ ∗x = Am
gθ r∗ = g m
Reference Model
x& m = Am x m + g mbr
Control Law
u (t ) = θ x T (t ) x p (t ) + θ r (t ) r (t )
Adaptive Law
e = x p − xm
ATm P + PAm = −Q Choose Q s.p.d., Calculate P s.p.d.
θ&x  x p  T
&  = − sgn( g ) Γ  r  e Pb
θ
 r  
{ }
Result: All signals x p , θ x , θ r are bounded, and lim x p − x m = 0
t →∞
Exercise 1
Consider the plant y = G yu u

where
g
G yu ( s ) =
s 2 + a1s + a2
with a1 = 0, a2 = 1, y and y& measurable.
Design an adaptive controller to match the reference model

y m = G ym r r
ω n2
Gymr (s) =
s 2 + 2ξω n s + ω n2
with ω n = 2 rad/s and ξ = 0.9 .
Simulate and verify your design using MATLAB.

Incorporation of integral control with all state variables

measurable
Plant:
x& p = Ap x p + gbu
y = x1 = [1 0 L 0]x p
Matching condition:
T
Ap + gbθ ∗x = Am
Integral control can be incorporated easily. Consider an additional

state
x& I = y − r
t
r(t) is the reference signal and x I = ∫ ( y − r ) dτ
0
Then, augmented plant
 x& p   A p 0  x p  b 0

=
 x&  1 0 ... 0 0  x  + g 0 u +  − 1 r
 I   I     
Consider
T
u = θ x∗ x p + θ I∗ x I
The closed-loop system is
 x& p   A p + gbθ x∗ gbθ I∗   x p  +  0  r

T
 x&  =     
 I   1 0 ... 0 0   x I   − 1
Thus, an additional matching condition is

 A + gbθ ∗T gbθ I∗  = A
 p x  m
 1 0 ... 0 0 
and
0
x& m = Am x m +   r
 − 1
For the corresponding adaptive controller, consider

u (t ) = θ x (t ) T x p (t ) + θ I (t ) x I (t )
and a reference model
0
x& m = Am x m +   r ; xm ∈ ℜ n +1
 − 1
Define
∆
x p 
x p = K 
 
 x I 
For e = x p − x m , we have
b 
(
e& = Ame + g K φ Tx x p + φ I x I
 
)
 0 
= Ame + gb φ T x
Clearly, it is now a well-known error equation, and

b
b = K is known
 
 0 
as b is known.
The adaptive law should be
AmT P + P Am = −Q
1st choose Q , 2nd calculate P
θ&x  x p  T
&  = − sgn( g ) Γ  x e Pb
θ
 I  I
Adaptive control with all state-variables measurable:

Incorporation of integral control
Example
Consider a plant
y& p = a p y p + gu
To incorporate integral control, we know we must augment the
state with
x& I = y p − r
The state of the new “plant” (including the augmented variable)

is
 x&1   a p 0  x1  1 0

 x&  =  1 0  x 2 
+ g 0 u + − 1 r
 2     
with x1 ≡ y p and x 2 = x I
The feedback control is of the form
u(t ) = θ 1 (t ) x1 (t ) + θ 2 (t ) x 2 (t )
(Refer to earlier notes.)
Consider constant gain feedback first. The plant with feedback is
 x&1  a p + gθ 1 gθ 2   x1   0 
 x&  =     +  r
0   x 2   − 1
 2  1
This means that the reference model used must have the form
 x&1m  a1m a2m   x1m   0 

 x&  =  1 +
0   x 2m  − 1
r
 2m  
• This is the only class of reference model that can be matched

-------“matching conditions”.
• The analysis then proceeds in the normal way.

Exercise 2
Consider the plant y = G yu u

where
g
G yu ( s ) =
s 2 + a1s + a2
with a1 = 0, a2 = 1, y and y& measurable.
Design an adaptive controller incorporating integral action.
Explain carefully how you choose the reference model to track.
In the limit as t → ∞ , what is the equivalent relationship between

y(t) and r(t)?
(Express this in transfer function form.)

Exercise 3: Proportional plus integral adaptation
Consider the plant
y& = a p y + u
and the reference model
y& m = a m y m + r
Then, the adaptive controller
e = y − ym
u = θy + r
θ& = −γye γ >0
will result in asymptotic tracking.
Astrom (pp.136,137) has suggested an adaptive law of the form
t
θ (t ) = −γ 1 ye − γ 2 ∫ y (τ )e (τ )dτ
0
Using the methods we have presented, analysis this adaptive law.

Positive Real Transfer Function
Definition: A rational transfer function H with real coefficients is

positive real (PR) if Re H(s)>0 for Re s>0.
A transfer function H is strictly positive real (SPR) if H(s-ε) is

positive real for some real ε>0.
Astrom…p129
s s1 s
Narendra…p65
PR SPR s = s1 + ε
Kalman Yakuborich Lemma, Lemma 2.3 (Narendra … p66)

Given a scalar γ ≥ 0 , a vector h, an
asymptotically stable matrix A, a vector
b such that (A, b) is controllable, and a
x& = Ax + bu
positive definite matrix L, there exist a
scalar ε>0, a vector q and a symmetric 1
y = hx + γu
positive-definite matrix P satisfying 2
AT P + PA = −qqT − εL
Pb − h = γ q
1
∆ ( A, b, c , d ) = ( A, b, h, γ )
1
if and only if H ( s ) = γ + hT ( sI − A) −1 b
2
2
is SPR.
Lemma 4-2 (Astrom p129) is a special case of preceeding with
γ =0
Positive real functions have been used extensively in network

analysis.
The use of the concept in adaptive control is relatively recent.
A passive network (consisting only of inductance, resistance and

capacitance) is positive real. If in addition, the network is
dissipative due to the presence of resistors, then the impedance
function is strictly positive real.
Viewed from realization, any PR (SPR) impedance function can be

realized by a passive (dissipative) network.
(Narendra pp.63)
H : is said to be passive if ∃ β ∈ R such that

t
< Hx , x >t = ∫ ( Hx )T xdτ ≥ β , ∀t ∈ R +
0
t
( < y , x > t = ∫ y T xdτ ≥ β , ∀t ∈ R + )
0
When H is a linear invariant operator, if H is passive, then H is
PR.
(Narendra pp.71)
Continuous-time adaptive control using only input-output

measurements
• Theory of SPR function will be used to prove stability.
References: Narendra……Chapter 5
Astrom………Section 4.5
Consider a system described by
R p ( p ) y (t ) = k p Z p ( p )u (t ) (1)
where
d
p≡
dt
R p ( p) = p n + a1 p n−1 + L + a n
Z p ( p ) = p m + b1 p m−1 + L + bm
R p is monic (i.e. leading coeff=1)

deg( R p ) = n, deg( Z p ) = m
Define “relative degree” n ∗

∆
∗
n =n−m (i.e. excess poles over zeros.)
Consider “Diophantine” identity
T ( p) R m ( p) = R p ( p ) E ( p ) + F ( p)
deg( T ) = n  Design polynomials, chosen to


deg( Rm ) = n∗  be stable, and monic
For given plant, i.e. R p , and specified T and Rm

E and F are unique, and
deg( E) = n∗ , E monic , deg( F ) = n − 1
T ( p) = p n + t1 p n −1 + ...+ tn −1 p + t n
Plant: R p ( p ) y = k p Z p ( p )u
ER p y = k p EZ p u
i.e.
TRm y = Fy + k p EZ p u = k p ( F y + G u)
F
( F= , G = EZ p )
kp
F G
Rm y = k p ( y + u) (5)
T T
Note that (a) Z p monic, and E monic, ∴ G monic;
(b) deg( G ) = deg( Z p ) + deg( E ) = n
Accordingly, we have
G G
= 1+ 1
T T
where
G1 ( p) = g1 p n−1 + g 2 p n −2 + L + g n
Similiarly, we can write
F ( p) = f 1 p n −1 + f 2 p n −2 + L + f n
Equation (5) then becomes
F G
Rm y = k p ( y + 1 u + u) (7)
T T
1 1
Define y f1 = y, u f1 = u
T T
[
θ ∗ = − f1 ,− f 2 , L − f n , − g1 , − g 2 , L − g n , k ∗ ]
T
[
ω = p n −1 y f1 , p n −2 y f1 , L , y f1 , p n −1u f1 , p n− 2 u f1 , L , u f1 , r ]
T
and r(t) is the set-point signal.
Consider
F G
u (t ) = − y − 1 u + k ∗ r = − F y f1 − G1u f1 + k ∗ r
T T (8)
∗T ∗T ∗T
=θ ω = θ u ω u + θ y ω y + k ∗r
Equation (7) becomes
Rm y = k p k ∗ r
= kmr for k pk ∗ = km
The choice of the input (8) ensure that the plant with feedback
becomes
R m ( p ) y (t ) = k m r (t )
In other words, the plant output will track a reference output

(reference model) given by
R m ( p ) y m (t ) = k m r (t ) ⇒ Ym k
= m
R Rm (s )
Remember,
deg( Rm ) = n∗
Rm is a design polynomial you choose.
Example:
n∗ Choice
n ∗ = 1, Rm ( s ) = ( s + a m )
k m = am ;
n∗ = 2 Rm ( s) = s 2 + 2ξω n s + ω n2
k m = ω n2
However,
while we can construct the signals in ω ,

we do not know the desired controller parameters θ ∗ .
Try using
u (t ) = θ (t ) T ω (t )
and adaptively determine θ (t ) in some way.
This is possible. Two separate cases:
(i) n ∗ = 1, convergence proof easy (relatively)

(ii) n ∗ > 1, convergence proof possible but very difficult
Recall equation (7)
F G 
Rm y = k p  y + 1 u + u 
T T 
F G  
= k p   y + 1 u − k ∗r  + k ∗r + u 
 T T  
= k p  − θ ∗ ω + k *r + u 
T
 
Control law:
u (t ) = θ (t )T ω (t )
Define parameter error
∆
φ (t ) = θ (t ) − θ ∗
Then above becomes
(
Rm y = k p φ (t )T ω (t ) + k *r (t ) )
Comparing with reference model
Rm y m = k mr = k p k *r
gives error equation
Rme1 = k pφ (t )T ω (t )
∆
where e1 = y − y m .
At this point, remembering n* = 1 and deg( Rm ) = n*
we have Rm (s ) = s + am (stable reference model)
It is straightforward to check that
ym k k
= Wm = m = m ; km > 0
r Rm ( s ) s + am
is SPR.
We will use this property together with the Ralman-Yakubovich

Lemma to prove stability and convergence of e1 (t ) to zero.
However, we need first a state realization which includes
∆
[
ω (t ) = p n −1 y f1 , p n − 2 y f1 ,..., y f1 , p n −1u f1 , p n − 2 u f1 ,..., u f1 ]
T
as states (in order to prove boundedness of all state variables).

Proposition 1: The vector ω (t ) is the state of a realization

(nonminimal) of the plant
R p y (t ) = k p Z pu (t )
Proof:
Recall that, T = p n + t1 p n −1 + ... + t n −1 p + tn

Tu f1 = u (1)
and clearly
R p y f1 = k p Z p u f1 (2)
Further, since deg( T ) = deg( R p ) = n , clearly
y = Ty f1 = Ty f1 + 0
( )
(3)
= T − Rp y f1 + k pZ p u f1
Equations (1), (2) and (3) form the basis for a 2n-dimensional
state representation of the plant.
From (1), we have
 u f1   u f1 
   f1 
d  pu f1 
= AT   +b u
pu
dt  M   M  n
 n −1 f   n −1 f 
 p u 1   p u 1 
where
 0 I ( n −1)× ( n −1) 0
AT =  and b =
L − t1  1
n
− tn − t n −1  
This can be combined with (2) and (3) to give
d A Ak pZ p  0
ω (t ) =  R p ω ( t ) + b  u(t )
dt 0 n× n AT   n
[ ]
y (t ) = t n − an , t n −1 − an −1 , L , t1 − a1 ,0, L ,0, k pbm , L , k p b1 ω (t )
The forms of the n × n matrices A R p and A k p Z p should be

obvious. The above is a state realization of the plant with ω (t ) as
the state.
QED
Proposition 2: For the state realization given previously, there

exist controller gains θ ∗ such that the closed-loop poles are at the
roots of the 2n degree polynomial TRm Z p .
Proof:
Most of the materials have been introduced already. It only
remains to put it into a formal basis.
Recall the Diophantine identity
TRm = R p E + F
and the plant equation
R p y = k p Z pu
From earlier, recall that our control law was

T
u = θ ∗ ω + k *r (1)
and it is equivalent to
F y f1 + G u f1 = k * r (2)
wih k pF = F .
[Why?
T F G
p.44, u(t ) = θ ∗ ω + k ∗r = − y − 1 u + k ∗r ⇒
T T
F G F G
y + (1 + 1 )u = k ∗ r ⇒ y + u = k ∗r
T T T T
]
Observe that
TRm Z p = R p EZ p + FZ p
= RpG + Z p F (3)
= RpG + k p Z p F
If the control law (1) or equivalent (2) is applied to the plant, we

have:
Plant: R p y = k p Z pu
R p y f1 = k p Z p u f1
From (2) R pG y f1 = k p Z p G u f1
{
= k p Z p k *r − F y f1 }
or
(R pG + k p Z p F )y f 1
= k pk ∗Z p r
From (3) TRm Z p y f1 = k m Z p r
i.e. the closed loop poles are at TRm Z p .
In addition, if T and Z p are stable, then the closed-loop transfer

function is
Rm y = k m r
QED
Corollary:
The reference model
R m ( p ) y m (t ) = k m r (t ) i.e., Ym k
= m
R Rm (s )
admits a 2n dimensional state-representation
ω& m = Amω m + bm r
y m = c Tmω m
Proof:
Straightforward by combining previous two propositions.
----------------------------------------------------------------
Note that:
Rm y m = k m r ⇒ ym km
Wm = =
r Rm
ω& m = Amω m + bm r ⇒ y
Wm = m = k m
T
( sI − Am ) −1bm
y m = c Tmω m r
⇒
km
Wm = k m
T
( sI − Am ) −1 bm =
Rm
Stability Analysis
Use the 2n-dimensional state-representation
ω& = A pω + b p u
y = cTp ω
Consider
u = θ (t ) T ω = θ (t )T ω + k (t ) r
{ }T {
= θ ∗ + φ (t ) ω + k * + φ k r }
Then, C.L.

T

( )
ω& =  A p + b pθ * ω + bp φ T ω + φ k r + k *b p r
y = c pω
T
For the exact values θ * , k * , the plant with feedback matches the
reference model,
T
Ap + b pθ * = Am ; k *b p = bm
and c p = c m
Consider now the state error

∆
e =ω − ωm, e1 = y − y m
Then
1
e& = Ame + bmφ T
ω
e1 = c Tm [sI − Am ]−1
* bm
k
*
φ T
ω
e1 = c Tme k
Noting that ( Am , bm , c m ) is a state representation of the reference

model, we have
k
c Tm (sI − Am ) bm = m
−1
Rm
1 km
This is strictly positive real, as is c Tm ( sI − Am )−1 bm = .
* *
k k Rm
Consider Lyapunov function condidate
V (e, φ ) = e T Pe + φ T Γ −1φ
where
ATm P + PAm = −Q
Then
1
V& = −eT Qe + 2e T P ∗ bmφ T ω + 2φ T Γ −1φ&
k
 
 1 
Since  Am , * bm , cm  is SPR, KY-Lemma ensures that
 k 
 
1
P bm = cm
k∗
which means that

V& = −eT Qe + 2eT c m sgn( k * )φ T ω + 2φ T Γ −1φ&
= −e T Qe + 2e T sgn( k * )φ T ω + 2φ T Γ −1φ&
1
If
θ& = φ& = − sgn( k * )Γω e1 = − sgn( k p )Γω e1 as k m > 0
we have
V& = −eT Qe ≤ 0
Thus
(a) e, φ (hence ω , θ ) are bounded;

∞
(b) e& bounded, ∫ e (τ )T Qe (τ ) dτ ≤ c1
0
∴ lim e (t ) = 0
t→∞
⇒ lim e1 (t ) = 0
t →∞
Case n* = 1 , Summary
Plant:
R p y = k p Z pu
Control:
1 1
y f1 = y, u f1 = u
T T
∆
[
ω (t ) = p n −1 y f1 , p n − 2 y f1 , L , y f1 , p n −1u f1 , p n − 2u f1 , L , u f1 , r ]
T
u (t ) = θ (t )T ω (t )
Adaptive law
θ& = − sgn( k p ) Γωe1

e1 = y − y m
Result: If
(a) order of R p , n, is known;
(b) Z p is stable polynomial;
(c) sgn( k p ) is known;
then the above system leads to y(t), u(t), θ (t ) bounded, and

lim ( y (t ) − y m (t ) ) = 0
t→∞

EE5104 Adaptive Control Systems Part I

Загружено:

Сведения о документе

Исходное описание:

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

EE5104 Adaptive Control Systems Part I

Загружено:

Авторское право:

Доступные форматы

EE5104 Adaptive Control Systems

Tel: 874 6821

2 Model Reference Adaptive Systems 8

7 Alternatives to Adaptive Control 3

Pre-requisites for this course

• Linear Systems Theory

1. Digital filter design (e.g. Butterworth) and implementation

A simple adaptive control problem

(Narendra & Annaswamy, pp.111-114)

Consider a system to be controlled

with transfer function

We usually refer to the system to be controlled as the “plant”.

• Drive y(t) to follow some reference command trajectory

For example, y m (t ) may be generated by a reference system

a m < 0 , i.e. stable reference model.

Consider a control input of the form

Then, the plant together with control (the closed-loop system) is

PROBLEM: We do not know a p and k p !

Adaptive Control: Find a way to achieve this automatically.

• Estimate a p and k p on-line, and try to use these estimates;

• Use a rigorous procedure to “evolve” to the right values (Very

In the course we will look at both types of approaches.

For now, let us consider the rigorous approach.

u(t ) = θ (t ) y (t ) + k (t ) r(t ) , Time-varying controller gains!

How to choose {θ (t ), k (t ), t ∈[ 0, ∞)}?

θ&(t ) = − sgn( k p )γ 1e(t ) y (t ); γ1 > 0

k& (t ) = − sgn( k p )γ 2 e (t ) r (t ); γ2 > 0

(Nonlinear updating mechanisms!)

Thus, adaptive controllers contain both time-varying and nonlinear

Because of the NLTV characteristics, analysis and design usually

Re-membering that θ ∗ and k ∗ are constants

Consider derivative of quadratic form V

Collecting together the intermediate results:

• Quadratic form V (e, φ ,ψ ) =

• This means that

∴ θ (t) and k(t) are bounded

• V(e, φ,ψ) positive definite

Since V (e (t ), φ (t ),ψ (t ) ) is always bounded,

The adaptive controller achieves the designed objective!

Rigorous analysis of adaptive control is typically as difficult.

Reference Model: y& m (t ) = am y m (t ) + k m r (t ); am < 0

Control Law: u(t ) = θ (t ) y(t ) + k (t ) r (t ) , Time varying

Adaptive Law: θ&(t ) = − sgn( k p )γ 1e(t ) y (t )

k& (t ) = − sgn( k p )γ 2 e(t ) r(t )

Figure 4.1 Block diagram of a model-reference adaptive system

Lyapunov’s Direct Method

l Enables one to determine whether or not the equilibrium state

l Involves finding a suitable scalar function V(x,t) and examining

Theorem (Uniform Stability)

Assume that a scalar function V(x,t), with continuous first partial

(i) V(x,t) is positive-definite, i.e.,∃ a continuous non-

(ii) V(x,t) is decrescent, i.e., ∃ a continuous non-decreasing

(iii) V& ( x , t ) is negative semi-definite, i.e.,

(iv) V(x,t) is radial unbounded, i.e.,

LTI System and Lyapunov Stability

Theorem 2-2 (Narendra ….)

All solutions of the above equation tend to zero as t → ∞ if and

Such a matrix A is referred to as an asympototically stable matrix.

Theorem 2-10 (Narendra, …)

The equilibrium state x=0 of the LTI system

is asymptotically stable if and only if, given any symmetric positive

Proof: See Narendra & Annaswamy pp.60.

Stability theory, in general, is a very wide field.

•　　Linear Systems Theory