Applied Nonlinear Control Nguyen Tan Tien - 2002.

6
_________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

8. Adaptive Control

In this chapter: tracking capacity in order to allow the possibility f tracking

- The nonlinear system with structured or unstructured convergence. Existing adaptive control designs normally
uncertainties (model imprecision) is considered. required linear parametrization of the controller in order to
- A so-called sliding control methodology is introduced. obtain adaptation mechanisms with guaranteed stability and
tracking convergence.
8.1 Basic Concepts in Adaptive Control
The adaptation mechanism is used to adjust the parameters in
- Why we need adaptive control ? the control law. In MRAC systems, the adaptation law
- What are the basic structures of adaptive control systems ? searches for parameters such that the response of the plant
- How to go about designing adaptive control system ? under adaptive control becomes the same as that of the
reference model. The main difference from conventional
8.1.1 Why Adaptive Control ? control lies in the existence of this mechanism.

8.1.2 What is Adaptive Control ? Example 8.1 MRAC control of unknown mass____________
An adaptive controller differs from an ordinary controller in
Consider the control of a mass on a frictionless surface by a
that the controller parameters are variable, and there is a
mechanism for adjusting these parameters on-line based on motor force u , with the plant dynamics being
signals in the system. There are two main approaches for
m &x& = u (8.1)
constructing adaptive controllers: so-called model-reference
adaptive control method and so-called self-tuning method.
Choose the following model reference
Model-Reference Adaptive Control (MRAC) &x&m + λ1 x& m + λ2 x m = λ2 r (t ) (8.2)
ym
reference model where,
λ1 , λ1 : positive constants chosen to reflect the performance
r
u y e specifications
controller plant xm : the reference model output (ideal out put of the
controlled system)
r (t ) : reference position
â
adaptation law
* m is known exactly, we can choose the following control
Fig. 8.3 A model-reference adaptive control system law to achieve perfect tracking &~
x& + 2λ ~
x& + λ2 ~
x = 0 , with
~
x = x − xm representing the tracking error and λ is a strictly
A MRAC can be schematically represented by Fig. 8.3. It is
composed of four parts: a plant containing unknown positive number. This control law leads to the exponentially
parameters, a reference model for compactly specifying the convergent tracking error dynamics: u = mˆ ( &x&m − 2λ ~
x& − λ2 ~
x).
desired output of the control system, a feedback control law
containing adjustable parameters, and an adaptation * m is not known exactly, we may use the control law
mechanism for updating the adjustable parameters.
u = mˆ ( &x&m − 2λ ~
x& − λ2 ~
x) (8.3)
The plant is assumed to have a known structure, although the
parameters are unknown. which contains the adjustable parameter m̂ . Substitution this
- For linear plants, the numbers of poles and zeros are control law into the plant dynamics, yields
assumed to be known, but their locations are not.
- For nonlinear plants, this implies that the structure of the m &x& = mˆ ( &x&m − 2λ ~x& − λ2 ~
x)
dynamic equations is known, but that some parameters are = (m + m ) ( &x&m − 2λ x& − λ2 ~
~ ~ x ), with m~ ≡ mˆ − m
not.
⇒ m &~ x& + 2m λ ~ x& + m λ2 ~x =m~ ( &x& − 2λ ~
x& − λ2 ~
x) m
A reference model is used to specify the ideal response of the
adaptive control system to external command. The choice of ⇒ m( ~
&x& + λ ~
x& ) + λ m ( ~
x& + λ ~ ~ ( &x& − 2λ ~
x)=m m x& − λ2 ~
x)
the reference model has to satisfy two requirements:
- It should reflect the performance specification in the control Let the combined tracking error measure be
tasks such as rise time, settling time, overshoot or frequency
s =~
x& + λ ~
x (8.5)
domain characteristics.
- This ideal behavior should be achievable for the adaptive
control system, i.e., there are some inherence constrains on and the signal quantity v is defined as v = &x&m − 2λ ~
x& − λ2 ~
x .
the structure of reference model given the assumed structure The closed-loop zero dynamics
of the plant model.
~v
m s& + λ m s = m (8.4)
The controller is usually parameterized by a number of
adjustable parameters. The controller should have perfect Consider Lyapunov function
___________________________________________________________________________________________________________
Chapter 8 Adaptive Control 37
 Applied Nonlinear Control Nguyen Tan Tien - 2002.6
_________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

1 1 1 ~2 For simplicity, assume that the acceleration can be measured

V= m s2 + m ≥0 (8.7) by an accelerometer. From (8.1), the simplest way of
2 2γ
estimating m is
Its derivative yields
~m
V& = s m s& + γ −1m ~& u (t )
mˆ (t ) = (8.9)
~ v ) − γ −1m~ m&ˆ &x&(t )
= s ( −λ m s − m
= −λ m s 2 − γ −1m ~ m&ˆ + γ s v ( ) However this is not good method because there may be
considerable noise in the measurement &x& , and, furthermore,
If the update law is chosen to satisfy the acceleration may be close to zero. A better approach is to
estimate the parameter using a least-squares approach, i.e.,
m&ˆ = −γ s v (8.6)
choosing the estimate in such a way that the total prediction
The derivative of Lyapunov function becomes error
t
V& = −λ m s 2 ≤ 0 (8.8)
∫ e (r ) dr
2
(8.10)
0
Using Barbalat’s lemma, it is easily to show that s converges
to zero. The convergence of s to zero implies that of the is minimal, with the prediction error e defined as
x and the velocity tracking error ~
position tracking error ~ x& . For e(t ) ≡ mˆ (t ) &x&(t ) − u (t ) . The prediction error is simply the error
illustration, the results of simulation for this example are given in fitting the known input u using the estimated parameter m̂ .
in Fig. 8.4 and 8.5. The numerical values are chosen as m = 2 , This total error minimization can potentially average out the
mˆ (0) = 0, γ = 0.5, λ1 = 10, λ2 = 25, λ = 6, x& (0) = x& m (0) = 0 . effects of measurement noise. The resulting estimating is
In Fig. 8.4, x(0) = xm (0) = 0 and r (t ) = 0 . In Fig. 8.5,
t
x(0) = x m (0) = 0.5 and r (t ) = sin( 4t ) .
mˆ =
∫ w u dr 0
(8.11)
t
∫ w dr
0.6 2.5

0.5
2
2.0
Tracking Performance

Parameter Estimation

0.4 0
1.5
0.3

0.2
1.0 with w = &x& . If actually, the unknown parameter m is slowly
time-varying, the above estimate has to be recalculated at
0.1
0.5
0.0

-0.1
0.0 0.5 1.0 1.5 2.0 2.5 3.0
0.0
1.0 1.5 2.0 3.0
every new time instant. To increase computational efficiency,
0.0 0.5 2.5
Time (s) Time (s) it is desirable to adopt a recursive formulation instead of
Fig. 8.4 Tracking performance and parameter estimation for repeatedly using (8.11). To do this, we define
an unknown mass with reference path r (t ) = 0
0.8 2.5 1
0.6 P (t ) ≡ t
(8.12)

∫w
2.0
Tracking Performance

Parameter Estimation

0.4
2
0.2 1.5
dr
0.0 0
-0.2 1.0

The function P (t ) is called the estimation gain, its update can

-0.4
0.5
-0.6

-0.8
0.0 0.5 1.0 1.5 2.0 2.5 3.0
0.0
0.0 0.5 1.0 1.5 2.0 2.5 3.0 be directly obtained by using
Time (s) Time (s)

( )
Fig. 8.5 Tracking performance and parameter estimation for d −1
an unknown mass with reference path r (t ) = sin( 4t ) P = w2 (8.13)
__________________________________________________________________________________________
dt

Self-Tuning Controller (STC) Then differentiation of Eq. (8.11)(which can be written

t
r
controller
u
plant
y ∫
P −1mˆ = w u dr ) leads to
0

mˆ& = − P (t ) w e (8.14)

â In implementation, the parameter estimate m̂ is obtained by

estimator
numerically integrating Eqs. (8.13) and (8.14).
__________________________________________________________________________________________
Fig. 8.5 A self-tuning controller
Relations between MRAC and ST methods
A self-tuning controller is a controller which performs
simultaneous identification of the unknown plant.
8.1.3 How to Design Adaptive Controllers ?
Example 8.2 Self-tuning control of unknown mass________ The design of an adaptive controller usually involves the
Consider the control of a mass of Example 8.1. Let us still use following three steps:
the pole-placement (placing the poles of the tracking error - choose a control law containing variable parameters
dynamics) control laws (8.3) for generating the control input, - Choose an adaptation law for adjusting those parameters
but let us now generate the estimated mass parameter using a - analyze the convergence properties of the resulting control
estimation law. system.
___________________________________________________________________________________________________________
Chapter 8 Adaptive Control 38
 Applied Nonlinear Control Nguyen Tan Tien - 2002.6
_________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

Lemma 8.1: Consider two signals e and φ related by the Now we choose the adaptation laws for â r and â y . Let the
following dynamic equation tracking error be e = y − y m and the error of parameter
estimation be
e(t ) = H ( p ) [ k φT (t ) v (t )] (8.15)
a~r = aˆ r − a r a~ y = aˆ y − a y (8.25)
where e(t ) is a scalar output signal, H ( p) is strictly positive
real transfer function, k is an unknown constant with know The dynamics of tracking error can be found by subtracting
sign, φ(t ) is m × 1 vector function of time, and v (t ) is a (8.23) and (8.21)
measurable m × 1 vector. If the vector φ(t ) varies according to
e& = − a m ( y − y m ) + (a m − a p + b p aˆ y ) y + (b p aˆ r − bm )r
φ& (t ) = − sgn(k ) γ e v (t ) (8.16) = −a m e + b p (a~r r − a~ y y ) (8.26)

with γ being a positive constant, then e(t ) and φ(t ) are The Lemma 8.1 suggests the following adaptation laws
globally bounded. Furthermore, if v (t ) is bounded, then
aˆ& r = − sgn(b p ) γ e r (8.27)
e(t ) → 0 as t → ∞ .
a&ˆ y = − sgn(b p ) γ e y (8.28)
8.2 Adaptive Control of First-Order Systems
Let us discuss the adaptive control of first-order plants using with γ being a positive constant representing the adaptation
MRAC method. Consider the first-order differential equation gain. The sgn(b p ) in (8.27-28) determines the direction of the
y& = − a p y + b p u (8.20) search for the proper controller parameters.

Tracking convergence analysis

where, y is the plant output, a p and b p are constant unknown We analyze the system’s stability and convergence behavior
plant parameters. using Lyapunov theory. Choose the Lyapunov function
candidate
Choice of reference model
Let the desired performance of the adaptive control system be 1 2 1
V= e + b p ( a~r2 + a~ y2 ) ≥ 0 (8.29)
specified by a first-order reference model 2 2γ

y& m = − a m y m + bm r (t ) (8.21) Its derivative yields

1
where a m ≥ 0, bm > 0 are constant parameters, and r (t ) is V& = e [− a m e + b p (a~r r − a~ y y )] + b p ( a~r a~& r + a~ y a~& y )
γ
bounded external reference signal. 1
= −a m e + e b p sgn(b p ) (a~r r − a~ y y ) + b p ( a~r aˆ& r + a~y aˆ& y )
2

Choice of control law γ

As first step in adaptive controller design, let us choose the
control law to be
= −a m e + 2
γ
1
(
b p a~r aˆ& r + sgn(b p ) γ e r )
u = aˆ r (t )r (t ) + aˆ y (t ) y (t ) (8.22) +
1
γ
(
b p a~ y aˆ& y + sgn(b p ) γ e y )
where aˆ r , aˆ y are variable feedback gains. The reason for the With adaptation laws (8.27) and (8.28), the derivative of
choice of control law (8.21) is clear: it allows the possibility of Lyapunov function becomes V& = − a m e 2 ≤ 0 . Thus, the
perfect model matching. With this control law, the closed-loop adaptive control system is globally stable, i.e., the signals e ,
dynamics is
a~r and a~ y are bounded. Furthermore, the global asymptotic
y& = −(a p − aˆ y b p ) y + aˆ r b p r (8.23) convergence of the tracking error e(t ) is guaranteed by
Barbalat’s lemma, because the boundedness of e , a~r and a~ y
If the plant parameters were known, such as aˆ r = a r , aˆ y = a y , imply the boundedness of e& and therefore the uniform
comparing (8.21) and (8.23), we get continuity of V& .

bm a p − am Example 8.3 A first-order plant________________________

ar = ay = (8.24)
bp bp Consider the control of the unstable plant y& = y + 3 u using
the previous designed adaptive controller. The numerical
which lead to the closed-loop dynamics y& = − a m y + bm r
parameters are: a p = −1 , b p = 3 , a m = 4 , bm = 4 , γ = 2 , and
which is identical to the reference model dynamics, and yields
zero tracking error. y (0) = y m (0) = 0 . Two reference signals are used:
* r (t ) = 4 ⇒ simulation results in Fig. 8.9
Choice of adaptation law * r (t ) = 4 sin(3t ) ⇒ simulation results in Fig. 8.10

___________________________________________________________________________________________________________
Chapter 8 Adaptive Control 39
 Applied Nonlinear Control Nguyen Tan Tien - 2002.6
_________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

6 2.0 5.0 2.0

1.5 4.5 1.5 âr

5
4.0
Tracking Performance

Tracking Performance
Parameter Estimation

Parameter Estimation
1.0 1.0
3.5
4
0.5 âr 3.0 0.5 â y
3 0.0 2.5 0.0

-0.5 2.0 -0.5

2 â y 1.5
-1.0 -1.0
1.0 â f
1
-1.5 -1.5
0.5
0 -2.0 0.0 -2.0
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 0 0.5 1.0 1.5 2.0 2.5 3 0.0 0.5 1.0 1.5 2.0 2.5 3.0
Time (s) Time (s) Time (s) Time (s)

Fig. 8.9 Tracking performance and parameter estimation with Fig. 8.11 Tracking performance and parameter estimation
reference path r (t ) = 4 with reference path r (t ) = 4
5 2.0 5 2.0
4 1.5 4 1.5 âr
3 3
Tracking Performance

Tracking Performance
Parameter Estimation

Parameter Estimation
1.0 1.0
2 2
0.5 âr 0.5
1 1
â f
0 0.0 0 0.0
-1 -0.5 -1 -0.5
-2
-1.0
â y -2
-1.0
-3 -3 â y
-1.5 -1.5
-4 -4
-5 -2.0 -5 -2.0
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
Time (s) Time (s) Time (s) Time (s)

Fig. 8.10 Tracking performance and parameter estimation Fig. 8.12 Tracking performance and parameter estimation
with reference path r (t ) = 4 sin(3t ) with reference path r (t ) = 4 sin(3t )
__________________________________________________________________________________________ __________________________________________________________________________________________

Parameter convergence analysis ⇒ refer text book 8.3 Adaptive Control of Linear Systems with Full States
Feedback
Extension to nonlinear plant
The same method of adaptive control design can be used for Consider the nth-order linear system in the canonical form
the non-linear first-order plant describe by the differential
equation a n y ( n ) + a n−1 y ( n−1) + K + a0 y = u (8.36)

y& = − a p y − c p f ( y ) + b p u (8.32) where the state components y, y& ,K, y ( n−1) are measurable,
coefficient vector a = [a n L a1 a0 ]T is unknown, but
where f is any known nonlinear function. The nonlinear in
their signs are known. The objective of the control system is to
these dynamics is characterized by its linear parametrization in make y closely track the response of a stable reference model
terms of the unknown constant c . Instead of using (8.22), now
we use the control law α n y m ( n) + α n−1 y m ( n−1) + K + α 0 y m = r (t ) (8.37)

u = aˆ y y + aˆ f f ( y ) + aˆ r r (8.33) with r (t ) being a bounded reference signal.

where the second term in (8.33) is introduced with the Choice of control law
intention of adaptively canceling the nonlinear term. Using the Define a signal z (t ) as follows
same procedure for the linear plant,
(n)
z (t ) = y m − β n−1e ( n−1) − K − β 0 e (8.38)
y& = − a p y − c p f ( y ) + b p [aˆ y y + aˆ f f ( y ) + aˆ r r ]
= −(a p − b p aˆ y ) y − (c p − b p aˆ f ) f ( y ) + b p aˆ r r with β1 ,K, β n being positive constants chosen such that
p n + β n−1 p n−1 + K + β 0 is a stable (Hurwitz) polynomial.
Comparing to (8.21) and define a f ≡ c p / bp and Adding both side of (8.36) and rearranging, we can rewrite the
a~ f ≡ aˆ f − a f . The adaptation laws are plant dynamics as

a n [ y ( n) − z ] = u − a n z − a n−1 y ( n−1) − K − a0 y
a&ˆ y = − sgn(b p ) γ e y (8.34a)
Let us choose the control law to be
a&ˆ f = − sgn(b p ) γ e f (8.34b)

aˆ& r = − sgn(b p ) γ e r (8.34c) u = aˆ n z + aˆ n−1 y ( n−1) + K + aˆ 0 y = v T (t )aˆ (t ) (3.39)

with v (t ) = [z (t ) y n−1 L y& y ]T

Example 8.4 A first-order non-linear plant_______________
aˆ (t ) = [aˆ n aˆ n−1 L aˆ1 aˆ 0 ]T
Consider the control of the unstable plant y& = y + y 2 + 3 u
using the previous designed nonlinear adaptive controller. The denoting the estimated parameter vector. This represents a
numerical parameters are: a p = −1 , b p = 3 , a m = 4 , bm = 4 , pole-placement controller which places the poles at positions
specified by the coefficients β i . The tracking error
γ = 2 , and y (0) = y m (0) = 0 .
e = y − y m then satisfies the closed-loop dynamics
Two reference signals are used:
a n [e ( n) + β n−1e ( n−1) + K + β 0 e = v T (t )~
a (t ) (3.40)
* r (t ) = 4 ⇒ simulation results in Fig. 8.11
* r (t ) = 4 sin(3t ) ⇒ simulation results in Fig. 8.12 where ~ a = aˆ − a
___________________________________________________________________________________________________________
Chapter 8 Adaptive Control 40
 Applied Nonlinear Control Nguyen Tan Tien - 2002.6
_________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

Choice of adaptation law theory that the relative degree of the reference model has to be
Rewrite the closed-loop system (3.40) in state space form larger or equal to that the plant in order to allow the possibility
of perfect tracking. Therefore, in our treatment, we will
x& = A x + b [(1 / a n ) v T ~
a] (3.41a) assume that nm − mm ≥ n − m .
e = cx (3.41b)
where The objective of the design is to determine a control law, and
 0 1 0 L 0  0  1 an associated adaptation law, so that the plant
      output y asymptotically approaches y m . We assume as follows
 0 0 1 L 0  0  0  - the plant order n is known
A= M M M O M  , b =  M  , cT =  M 
      - the relative degree n − m is known
 0 0 0 L 1  0  0  - the sign of k p is known
− β − β n−1  1 0 
 0 − β1 − β 2 L     - the plant is minimum phase

Consider Lyapunov function candidate 8.4.1 Linear systems with relative degree one
Choice of the control law
V ( x, ~
a ) = xT P x + ~
a T Γ −1~
a To determine the appropriate control law for the adaptive
controller, we must first know what control law can achieve
perfect tracking when the plant parameters are perfect known.
where both Γ and P are symmetric positive constant matrix, Many controller structures can be used for this purpose. The
and P satisfies following one is particularly convenient for later adaptation
design.
PA + A T P = −Q Q = QT > 0
Example 8.5 A controller for perfect tracking_____________
for a chosen Q . The derivative V& can be computed easily as Consider the plant described by

V& = − xT Q x + 2~
a T v bT P x + 2~
a Γ −1~
a& k p ( p + bp )
y= 2
u (8.45)
p + a p1 p + a p2
Therefore, the adaptation law
and the reference model
aˆ& = − Γ v bT P x (8.42)
k m ( p + bm )
ym = r (8.46)
leads to V& = − x T Q x ≤ 0 . p 2 + a m1 p + a m2
ym (t )
8.4 Adaptive Control of Linear Systems with Output r (t ) Wm ( p )
Feedback u0 u1 e
u
Consider the linear time-invariant system presented bu the k W p ( p)
transfer function α1
p + bm
Z p ( p) b0 + b1 p + K + bm−1 p m−1 + p m
W ( p) = k p = kp (8.43) β1 p + β 2
R p ( p) a0 + a1 p + K + a n−1 p n−1 + p n p + bm

where k p is called the high-frequency gain. The reason for Fig. 8.13 Model-reference control system for relative degree 1
this term is that the plant frequency response at high frequency
Let the controller be chosen as shown in Fig. 8.13, with the
kp
verifies W ( jω ) = n−m , i.e., the high frequency response is control law being
ω
essentially determined by k p . The relative degree r of this β1 p + β 2
u = α1 z + y+kr (8.47)
system is r = n − m . In our adaptive control problem, the p + bm
coefficients ai , b j (i = 0,1,K, n − 1; j = 0,1,K, m − 1) and the
high frequency gain k p are all assumed to be unknown. where z = u /( p + bm ) , i.e., z is the output of a first-order
filter with input u , and α1 , β1 , β 2 , k are controller parameters.
The desired performance is assumed to be described by a If we take these parameters to be α1 = b p − bm ,
reference model with transfer function
a m1 − a p1 a m2 − a p2 km
β1 = , β2 = , k= , the transfer
Zm kp kp kp
Wm ( p ) = k m (8.44)
Rm function from the reference input r to the plant output is

where Z m and Rm are monic Hurwitz polynomials of degrees k m ( p + bm )

Wry = = Wm ( p )
nm and mm , and k m is positive. It is well known from linear p 2 + a m1 p + a m2
___________________________________________________________________________________________________________
Chapter 8 Adaptive Control 41
 Applied Nonlinear Control Nguyen Tan Tien - 2002.6
_________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

Therefore, perfect tracking is achieved with this control law, - The vector θ1* contains (n − 1) parameters which intend to
i.e., y (t ) = y m (t ), ∀t ≥ 0 .
cancel the zeros of plant.
Why the closed-loop transfer function can become exactly the - The vector θ 2* contains (n − 1) parameters which, together
same as that of the reference model ? To know this, note that with the scalar gain θ 0* can move the poles of the closed-loop
the control input in (8.47) is composed of three parts:
- The first part in effect replaces the plant zero by the control system to the locations of the reference model poles.
reference model zero, since the transfer function from u1 to
As before, the control input in this system is a linear
y is combination of:
- the reference signal r (t )
p + bm k p ( p + b p ) k p ( p + bm )
Wu1, y = = 2 - the vector signal ω1 obtained by filtering the control input u
p + bp p2 + a p p + a p p + a p1 p + a p2
1 2 - the vector signal ω 2 obtained by filtering the plant output y
- The second part places the closed-loop poles at locations of and the output itself.
those of reference model. This is seen by noting that the The control input can be rewritten in terms of the adjustable
transfer function from u 0 to y is parameters and the various signals, as

Wu1, y k p ( p + bm ) u * (t ) = k *r + θ1*ω1 + θ*2ω 2 + θ*0 y (8.49)

Wu0 , y = =
1 + W f Wu1, y 2
p + (a p1 + β1k p ) p + (a p2 + β 2 k p )
Corresponding to this control law and any reference input r (t ) ,
- The third part of the control law (k m / k p ) r obviously the output of the plant is
replaces k p , the high frequency gain of the plant, by k m . As
B( p) *
a result of the above three parts, the closed-loop system has y (t ) = u (t ) = Wm r (t ) (8.50)
the desired transfer function. A( p )
__________________________________________________________________________________________

since these parameters result in perfect tracking. At this point,

The above controller in Fig. 8.13 can be extended to any plant we can see the reason for assuming the plant to be minimum-
with relative degree one. The resulting structure of the control phase: this allows the plant zeros to be cancelled by the
system is shown in the Fig. 8.14, where k * , θ1* , θ*2 and θ*0 controller poles.
represents controller parameters which lead to perfect tracking
In adaptive control problem, the plant parameters are unknown,
when the plant parameters are known.
and the ideal control parameters described above are also
ym (t ) unknown. Instead (8.49), the control law is chosen to be
r (t ) Wm ( p )
u0 u1 e u (t ) = k r + θ1ω1 + θ 2 ω 2 + θ 0 y (8.49)
u
k* W p ( p)
ω1 where, k , θ1 , θ 2 and θ 0 are controller parameters to be
θ1* Λ, h
provided by the adaptation law.
ω2
θ 2* Λ, h
Choice of adaptation law
For the sake of simplicity, define as follows
θ 0*
Fig. 8.14 A control system with perfect tracking θ(t ) = [k (t ) θ1 (t ) θ 2 (t ) θ 3 (t )]T
The structure of this control system can be described as ω(t ) = [r (t ) ω1 (t ) ω 2 (t ) ω 3 (t )]T
follows:
- The block for generating the filter signal ω1 represent an Then the control law (8.51) becomes
(n − 1) th order dynamics, which can be described by
ω& 1 = Λ ω1 + hu , where ω1 is an (n − 1) × 1 vector, Λ is an u (t ) = θ T (t ) ω (t ) (8.52)
(n − 1) × (n − 1) matrix, and h is constant vector such that
( Λ, h) is controllable. The poles of the matrix Λ are chosen to Let the ideal value of θ be θ* and the error φ(t ) = θ(t ) − θ* ,
be the same as the roots of polynomial Z m ( p ) , i.e., then θ(t ) = θ* + φ(t ) . Therefore, the control law (8.52) can
T
also be written as u (t ) = θ* ω + φT (t )ω . With the control law
det[ pI − Λ ] = Z m ( p ) (8.48)
(8.52), the control system with variable gains can be
equivalently represented as shown in Fig. 8.15,
- The block for generating the (n − 1) × 1 vector ω 2 has the
with φT (t ) ω / k * regarded as an external signal. The output
& 2 = Λ ω 2 + hy
same dynamics but with y as input, i.e., ω
here must be
- The scalar gain k * is defined to be k * = k m / k p and is
intended to modulate the high frequency gain of the control y (t ) = Wm ( p )r + Wm ( p )[φT ω / k * ] (8.53)
system.
___________________________________________________________________________________________________________
Chapter 8 Adaptive Control 42
 Applied Nonlinear Control Nguyen Tan Tien - 2002.6
_________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

u0 u1
r (t ) u y Let the controller be chosen as shown in Fig. 8.16. Noting that
* W p ( p)
k
bm in the filter in Fig. 8.13 has been replaced by a positive
ω1
θ1* Λ, h number λ . The closed-loop transfer function from the
T
φ ω ω2 reference signal r to the plant output y is
k* θ 2* Λ, h
p + λ0 kp
θ 0* p + λ0 + α1 p 2 + a p p + a p
Fig. 8.15 An equivalent control system for time-varying gains W ry = k 1 2

p + λ0 β1 p + β 2 kp
1+
Since y m (t ) = Wm ( p ) r , the tracking error is seen to be related p + λ0 + α 1 p + λ0 p 2 + a p p + a p
1 2

to the parameter error by the simple equation k k p ( p + λ0 )

=
2
( p + λ 0 + α 1 )( p + a p1 p + a p2 ) + k p ( β 1 p + β 2 )
e(t ) = Wm ( p ) [φT (t ) ω(t ) / k * ] (8.54)
Therefore, if the controller parameters α 1 , β 1 , β 2 , and k are
Since this is the familiar equation seen in Lemma 8.1, the chosen such that
following adaptation law is chosen
( p + λ 0 + α 1 )( p 2 + a p1 p + a p2 ) + k p ( β 1 p + β 2 ) =
θ& = − sgn(k p ) γ e(t ) ω(t ) (8.55)
( p + λ 0 )( p 2 + a m1 p + a m2 )
where γ is positive number representing the adaptation gain
and we have used the fact that the sign of k * is the same as that and k = k m / k p , then the closed-loop transfer function
of k p , due to the assumed positiveness of k m . W ry becomes identically the same as that of the reference
model. Clearly, such choice of parameters exists and is unique.
__________________________________________________________________________________________
Based on Lemma 8.1 and through a straightforward procedure
for establishing signal boundedness, we can show that the
For a general plants of relative degree larger than 1, the same
tracking error in the above adaptive control system converges
control structure as given in Fig. 8.14 is chosen. Note that the
to zero asymptotically.
order of the filters in the control law is still (n − 1) . However,
8.4.2. Linear system with higher relative degree since the model numerator polynomial Z m ( p ) is of degree
The design of adaptive controller for plants with relative smaller than (n − 1) , it is no longer possible to choose the
degree larger than 1 is both similar to, and different from, that poles of the filters in the controller so
for plants with relative degree 1. Specifically, the choice of
that det[ pI − Λ] = Z m ( p ) as in (8.48). Instead, we now choose
control law is quite similar but the choice of adaptation law is
very different.
λ ( p ) = Z m ( p ) λ1 ( p ) (8.57)
Choice of control law
Let us start from a simple example. where λ ( p ) = det[ pI − Λ ] and λ1 ( p) is a Hurwitz polynomial
of degree (n − 1 − m) . With this choice, the desired zeros of
Example 8.6 _______________________________________ the reference model can be imposed.
Consider the second-order plant described by the transfer
function Let us define the transfer function of the feed-forward
part u / u1 of the controller by λ ( p ) /(λ ( p ) + C ( p )) , and that of
kp
y= u the feedback part by D( p ) / λ ( p ) , where the polynomial C ( p)
2
p + a p1 p + a p2 contains the parameter in the vector θ1 , and the
polynomial D( p ) contains the parameter in the vector θ 2 .
and the reference model Then the closed-loop transfer function is easily found to be

km k k p Z p λ1 ( p ) Z m ( p )
ym = r Wry = (8.58)
2
p + a m1 p + a m2 R p ( p )[λ ( p ) + C ( p )] + k p Z p D( p )
ym (t )
Wm ( p ) The question now is whether in this general case, there exists
r (t )
u0 u1 choice of values for k , θ1 , θ 2 and θ 0 such that the above
e
u transfer function becomes exactly the same as Wm ( p ) , or
k W p ( p)
equivalently
α1
p + λ0 R p (λ ( p ) + C ( p )) + k p Z p D( p ) = λ1Z p Rm ( p ) (8.59)
β1 p + β 2
p +λ 0 The answer to this question can be obtained from the
Fig. 8.16 Model-reference control system for relative degree 2 following lemma
___________________________________________________________________________________________________________
Chapter 8 Adaptive Control 43
 Applied Nonlinear Control Nguyen Tan Tien - 2002.6
_________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

Lemma 8.2: Let A( p ) and B( p ) be polynomials of 1

ε (t ) = φT (t ) ω + φα η (t ) (8.65)
degree n1 and n2 , respectively. If A( p ) and B( p ) are relative k*
prime, then there exist polynomials M ( p) and N ( p) such that where

A( p ) M ( p ) + B( p ) N ( p ) = A* ( p ) (5.60) ω(t ) = Wm ( p )[ω] (8.66)

This implies that the augmented error can be linearly

where A* ( p ) is an arbitrary polynomial.
parameterized by the parameter error φ(t ) and φα . Using the
This lemma can be used straight forward to answer our gradient method with normilazation, the controller parameters
question regarding to (8.59). θ(t ) and the parameter α (t ) for forming the augmented error
are updated by
Choice of adaptation law
sgn(k p ) γ ε ω
When the plant parameters are unknown, the controller (8.52) θ& = − (8.67a)
is used again 1 + ωT ω
γ εη
u (t ) = θT (t ) ω(t ) (8.61) α& = − (8.67b)
1 + ωT ω
and the tracking error from (8.54)
With the control law (8.61) and adaptation law (8.67), global
T * convergence of the tracking error can be shown.
e(t ) = Wm ( p ) [φ (t ) ω(t ) / k ] (8.62)
8.5 Adaptive Control of Nonlinear System
However, the adaptation law (8.55) cannot be used. A famous
technique called error augmentation can be used to avoid the Consider a class of nonlinear system satisfying the following
difficulty in finding an adaptation law for (8.62). The basic conditions:
idea of the technique is to consider a so-called augmented 1. the nonlinear plant dynamics can be linearly
error ε (t ) which correlates to the parameter error φ in a more parameterized.
desirable way than the tracking error e(t ). 2. the full state is measurable
3. nonlinearities can be cancelled stably (i.e., without
ω (t ) 1 e(t ) ε (t ) unstable hidden modes or dynamics) by the control input
φT W p ( p) if the parameters are known.
k*
In this section, we consider the case of SISO system.
θT W p ( p)
η (t ) Problem statement
α (t ) Consider nth-order nonlinear systems in companion form
W p ( p) θT
n
Fig. 8.17 The augmented error y ( n) + ∑ α f ( x, t ) = b u
i =1
i i (8.68)

Let define an auxiliary error η (t ) by where

η (t ) = θT (t )Wm ( p )[ω(t )] − Wm ( p )[θT (t ) ω(t )] (8.63) x= y [ y& L y ( n−1) ] T

: the state vector
f i ( x, t ) : known nonlinear functions
as shown in Fig. 8.17. It is useful to note two features about αi ,b : unknown constant
this error
- Firstly, η (t ) can be computed on-line, since the estimated
The control objective is track a desired output y d (t ) despite
parameter vector θ(t ) and the signal vector ω(t ) are both
the parameter uncertainty. (8.68) can be rewritten in the form
available.
- Secondly, η (t ) is caused by time-varying nature of the n
estimated parameters θ(t ) , in the sense that when θ(t ) is h y ( n) + ∑ a f ( x, t ) = u
i =1
i i (8.70)
replaced by the true (constant) parameter vector θ* , we
where, h = 1 / b and ai = α i / b .
have θ*Wm ( p )[ω(t )] − Wm ( p )[θ* (t ) ω(t )] = 0. This also
implies that η can be written: η (t ) = φT Wm (ω) − Wm (φT ω) Choice of control law
Similarly to the sliding control approach, define a combined
Define an augmented error ε (t ) error s = e ( n−1) + λn−2 e ( n−2) + K + λ0 e = ∆ ( p ) e , where the
output tracking error is e = y − y d and a stable (Hurwitz)
ε (t ) = e(t ) + α (t )η (t ) (8.64)
polynomial is ∆ ( p ) = p n−1 + λn−2 p n−2 + K + λ0 . Note that s
where α (t ) is a time-varying parameter to be determined by
can be rewritten as s = y ( n−1) − y r( n−1) with y r( n−1) is defined as
adaptation. For convenience, let us write α (t ) in the form
y r( n−1) = y d( n−1) − λn−2 e ( n−2) − K − λ0 e .
α (t ) = 1 / k * + φα (t ) . From (8.62)-(8.64) we obtain
___________________________________________________________________________________________________________
Chapter 8 Adaptive Control 44
 Applied Nonlinear Control Nguyen Tan Tien - 2002.6
_________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

Consider the control law s ≡ e& + λ0 e = x& − ( x& d − λ0 e) ≡ x& − x& r

n
u = h y r( n) − k s + ∑ a f (x, t )
i =1
i i (8.71) Chose the control law

u = m &x&r − α s + c f1 + k f 2 ( → 8.71)
where k is constant of the same sign as h , and y r(n) is the
which yields m &x& + c f1 + k f 2 = m &x&r − α s + c f1 + k f 2 or
derivative of y r( n−1) , i.e., y r( n) = y d( n) − λn−2 e ( n−1) − K − λ0 e& .
m ( &x& − &x&r ) + α s = 0 . Because the unknown parameters, the
Noting that y r(n) , the so-called “reference” value of y (n) , is controller is
obtained by modifying y d(n ) according to the tracking errors. If
u → uˆ = mˆ &x&r − α s + cˆ f1 + kˆ f 2 ( → 8.72)
the parameters are all known, this choice leads to the tracking
error dynamics h s& + k s = 0 and therefore gives exponential which leads to the tracking error
convergence of s , which in turn, guarantees the convergence
of e . m &x& + c f1 + k f 2 = mˆ &x&r − α s + cˆ f1 + kˆ f 2
Choice of adaptation law (m &x& − m &x&r ) − (mˆ &x&r − m &x&r ) − (cˆ − c) f1 − (kˆ − k ) f 2 + α s = 0
For our adaptive control, the control law (8.71) is replaced by ~ &x& − c~ f − k~ f + α s = 0
m ( &x& − &x& ) − mr r 1 2
n
 &x&r 
u = hˆ y r( n) − k s + ∑ aˆ f (x, t )
i =1
i i (8.72) ~ c~ k~  f 
m s& + α s = m  1 [ ] ( → 8.74)
where h, ai have been replaced by their estimated values. The  f 2 
tracking error yields &
Adaptation laws: m&ˆ = − γ s &x&r , c&ˆ = − γ s f1 , kˆ = − γ s f 2
n

∑ a~ f (x, t )
~
h s& + k s = h y r( n) + i i (8.73) ~ 2 + c~ 2 + k~ 2 ) and its
Lyapunov function: V = m s 2 + γ −1 ( m
i =1
~ derivative with the above adaptation laws yields
where, hi = hˆi − hi , a~i = aˆ i − ai . (8.73) can be written in the
form
V& = 2m s s& + 2γ −1 ( m~ m&ˆ + c~ c&ˆ + k~ k&ˆ)

~ &x& + c~ f + k~ f − α s) + 2γ −1 ( m ~ m&ˆ + c~ c&ˆ + k~ k&ˆ)

1 / h  ~ ( n) 
n
= 2 s (m
s=
p + ( k / h) 
h yr + a~i f i (x, t ) 
 ∑ r 1

= −2α s 2 + 2γ −1  m
2
~ (mˆ& + γ s &x& ) + c~ (cˆ& + γ s f ) + k~ (kˆ& + γ s f ) 
 i =1  r 1 2 
 
  y r( n)   (8.74)
   = −2α s 2 ≤ 0
=
1/ h  ~ ~
 h a1 L a n 
p + ( k / h) 
[
~  f1 (x, t )  
M 
 ] __________________________________________________________________________________________

  8.6 Robustness of Adaptive Control System

  f n (x, t ) 

The above tracking and parameter convergence analysis has
Lemma 8.1 suggests the following adaptation law provided us with considerable insight into the behavior of the
& adaptive control system. The analysis has been carried out
hˆ = − γ sgn(h) s y r( n) assuming that no other uncertainties exist in the control system
besides parametric uncertainties. However, in practice, many
a&ˆ = − γ sgn(h) s f
i i types of non-parametric uncertainties can be present. These
include
Specially, using the Lyapunov function candidate
- high-frequency un-modeled dynamics, such as actuator
 n 
∑
−1  2 dynamics or structural vibrations
V = h s +γ 2
h + ai2  ≥ 0
  - low-frequency un-modeled dynamics, such as Coulomb
 i =1  friction and stiction
- measurement- noise
it is straight forward to verify that V& = −2 k s 2 ≤ 0 and - computation round-off error and sampling delay
therefore the global tracking convergence of the adaptive Since adaptive controllers are designed to control real physical
control system can be easily shown. systems and such non-parametric uncertainties are
unavoidable, it is important to ask the following questions
Example___________________________________________ concerning the non-parametric uncertainties:
- what effects can they have on adaptive control systems ?
Consider a mass-spring-damper system with nonlinear friction - when are adaptive control systems sensitive to them ?
and nonlinear damping described by the equation - how can adaptive control systems be made insensitive to
them ?
m &x& + c f1 ( x& ) + k f 2 ( x) = u (8.69)
While precise answers to such questions are difficult to obtain,
Apply the above analysis procedure. Define the error because adaptive control systems are nonlinear systems, some
___________________________________________________________________________________________________________
Chapter 8 Adaptive Control 45
 Applied Nonlinear Control Nguyen Tan Tien - 2002.6
_________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

qualitative answers can improve our understanding of adaptive of non-parametric uncertainties present in the above example,
control system behavior in practical applications. the observed instability can seem quite surprising.

Parameter drift Dead-zone

When the signal v is persistently exciting, both simulations
and analysis indicate that the adaptive control systems have 8.7 On-line Parameter Estimation
some robustness with respect to non-parametric uncertainties.
However, when the signals are not persistently exciting, even Few basic methods of on-line estimation are studied.
small uncertainties may lead to severe problems for adaptive Continuous-time formulation is used.
controllers. The following example illustrates this situation.
8.7.1 Linear parameterization model
Example 8.7 Rohrs’s example_________________________ The essence of parameter estimation is to extract parameter
information from available data concerning the system. The
Consider the plant described by the following nominal model quite general model for parameter estimation applications is in
the linear parameterization form
kp
H 0 ( p) =
p+ap y (t ) = W (t ) a (8.77)

The reference model has the following SPR function where

y ∈ Rn : known “output” of the system
km 3 m
M ( p) = = a∈R : unknown parameters to be estimated
p + am p+3
W (t ) ∈ R n×m : known signal matrix
The real plant, however, is assumed to have the transfer
function relation (8.77) is simply a linear equation in terms of the unknown a.
Model (8.77), although simple, is actually quite general. Any
linear system can be rewritten in this form after filtering both
2 229
y= u side of the system dynamics equation through an
p + 1 p 2 + 30 p + 229 exponentially stable filter of proper order, as seen in the
following example.
This means that the real plant is of third order while the
nominal plant is of only first order. The un-modeled dynamics Example 8.9 Filtering linear dynamics__________________
are thus to seen to be 229 /( p 2 + 30 p + 229) , which are high- Consider the first-order dynamics
frequency but lightly-damped poles at (−15 ± j ) . Beside the
un-modeled dynamics, it is assumed that there is some y& = − a1 y + b1u (8.78)
measurement noise m(t ) in the adaptive system. The whole
adaptive control system is shown in Fig. 8.18. The Assume that a1 ,b1 in model are unknown, and that the output
measurement noise is assumed to be n(t ) = 0.5 sin(16.1t ) . y and the input u are available. The above model cannot be
reference model directly used for estimation, because the derivative of
3 ym (t ) y appears in the above equation (noting that numerically
r (t ) p+3 n(t ) e(t ) differentiating y is usually undesirable because of noise
u consideration). To eliminate y& in the above equation, let us
2 229
âr p + 1 p 2 + 30 p + 229 filter (multiply) both side of the equation by 1 /( p + λ f )
y1 (where p is the Laplace operator and λ f is a known positive
nominal un-modeled
constant). Rearranging, this leads to the form
â y
y (t ) = y f (λ f − a1 ) + u f b1 (8.78)
Fig. 8.18 Adaptive control with un-modeled dynamics and
measurement noise where y f = y /( p + λ f ) and u f = u /( p + λ f ) with subscript
Corresponding to the reference input r (t ) = 2 , the results of f denoting filtered quantities. Note that, as a result of the
adaptive control system are shown in Fig. 8.19. It is seen that filtering operation, the only unknown quantities in (8.79) are
the output y (t ) initially converges to the vicinity of y = 2 , the parameters (λ f − a1 ) and b1 .
then operates with a small oscillatory error related to the
measurement noise, and finally diverges to infinity. The above filtering introduces a d.c. gain of 1 / λ f , i.e., the
magnitudes of y f and u f are smaller than those of y and
Fig. 8.19 Instability and parameter drift
u by a factor of λ f at low frequencies. Since smaller signals

_________________________________________________________________________________________
may lead to slower estimation, we may multiply both side of
(8.79) by a constant number, i.e., λ f .
In view of the global tracking convergence proven in the
absence of non-parametric uncertainties and the small amount
___________________________________________________________________________________________________________
Chapter 8 Adaptive Control 46
 Applied Nonlinear Control Nguyen Tan Tien - 2002.6
_________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

Generally, for a linear SISO system, its dynamics can be H 12 = H 21 = m 2 l1c 2 cos q 2 + m 2 l c22 + I 2
described by
H 22 = m 2 l c22 + I 2
A( p ) y = B ( p ) u (8.80) h = m 2 l1l c 2 sin q 2
where g1 = m1l c1 g cos q1 + m 2 g[l c 2 cos(q1 + q 2 ) + l1 cos q1 ]
A( p ) = a0 + a1 p + K + a n−1 p n−1 + p n g 2 = m 2 l c 2 g cos(q1 + q 2 )
B ( p ) = b0 + b1 p + K + a n−1 p n−1
Let us define a1 = m2 , a 2 = m2lc 2 , a3 = I1 + m1lc21 , and
Divide both sides of (8.80) by a known monic polynomial of
order n, leading to a4 = I 2 + m2 lc22 .
Then each term on the left-hand side of (6.9)
is linear terms of the equivalent inertia parameters
y=
A0 ( p ) − A( p) B( p )
u (8.81) a = [a1 a 2 a3 a 4 ]T . Specially,
A0 ( p ) A0 ( p )
where H11 = a3 + a 4 + a1l12 + 2a 2 l1 cos q 2
n −1 n
A0 = α 0 + α1 p + K + α n−1 p +p H 22 = a 4
H12 = H 21 = a 2 l1 cos q 2 + a 4
has known coefficients. In view of the fact that
Thus we can write
A0 ( p ) − A( p ) = (α 0 − a0 ) + (α1 − a1 ) p + K + (α n−1 − a n−1 ) p n−1
τ = Y1 (q, q& , q
&&) a (8.83)
(8.81) can be rewritten in the form
This linear parametrization property actually applies to any
y = θT w (t ) (8.82) mechanical system, including multiple-link robots.

where Relation (8.83) cannot be directly used for parameter

estimation, because of the present of the un-measurable joint
θ = [α 0 − a0 α1 − a1 L α n−1 − a n−1 b0 L bn−1 ] T
&& . To avoid this, we can use the above filtering
acceleration q
T
 y py p n−1 y u p n−1u  technique. Specially, let w(t ) be the impulse response of a
w= L L 
A A A A A0  stable, proper filter. Then, convolving both sides of (6.9),
 0 0 0 0
yields
Note that w can be computed on-line based on the available
values of y and u . t t
_________________________________________________________________________________________
∫ w(t − r )τ(r )dr = ∫ w(t − r )[Hq&& + Cq& + G]dr
0 0
(8.84)
Example 8.9 Linear parametrization of robot dynamics_____
Using partial integration, the first term on the right hand side
of (8.84) can be rewritten as

t t d
∫ ∫ dr [wH]q& dr
I2, m2 t
w(t − r )Hq
&& dr = w(t − r )Hq& −
lc2 0 0 0
l2
l1 q 2, τ 2 = w(0)H (q)q& − w(0)H[q (0)] q& (0) −
t

lc1 ∫ [w(t − r )H& q& - w& (t − r )Hq& ] dr

0
I 1, m 1 This means that the equation (8.48) can be rewritten as
q1,τ1
y (t ) = W(q, q& )a (8.85)

Fig. 6.2 A two-link robot where y is the filtered torque and W is the filtered version of
Y1 . Thus the matrix W can be computed from available
Consider the nonlinear dynamics of a two-link robot measurements of q and q& . The filtered torque y can also be
computed because the torque signals issued by the computer
 H 11 H 12   q&&1  − h q& 2 − h q&1 − h q& 2   q&1   g1  τ 1  are known.
   =   &  +   =   _________________________________________________________________________________________

 H 21 H 22  q&&2   h q&1 0  q 2   g 2  τ 2 
(6.9) 8.7.2 Prediction-error-based estimation model
where,
8.7.3 The gradient estimator
q = [q1 q 2 ]T : joint angles
τ = [τ 1 τ 2 ]T : joint inputs (torques) 8.7.4 The standard least-squares estimator

H 11 = m1l c21 + I 1 + m 2 (l12 + l c22 + 2l1l c 2 cos q 2 ) + I 2

8.7.5 Least-squares with exponential forgetting
___________________________________________________________________________________________________________
Chapter 8 Adaptive Control 47
 Applied Nonlinear Control Nguyen Tan Tien - 2002.6
_________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

8.7.6 Bounded gain forgetting

8.7.7 Concluding remarks and implementation issues

8.8 Composite Adaptation

___________________________________________________________________________________________________________
Chapter 8 Adaptive Control 48