Robust Adaptive Control by Petros A. Ioannou, Jing Sun - HTML preview

PLEASE NOTE: This is an HTML preview only and some elements such as links or page numbers may be incorrect.
Download the book in PDF, ePub, Kindle for a complete version.

p( s)

Reference signal

Qm( s) ym = 0

Assumptions

Same as in Table 9.5

Robust

Same as in Table 9.5 to generate ˆ

Zp( s, t) , ˆ

Rp( s, t)

adaptive law

and ˆ

A, ˆ

B

State observer

Same as in Table 9.6

Calculation

ˆ

Kc = λ− 1 ˆ

B P

of controller

ˆ

A P + P ˆ

A − 1 P ˆ

B ˆ

B P + CC = 0

λ

parameters

C = [1 , 0 , . . . , 0] , P = P

> 0

Control law

¯

up = ˆ

Kc( te, up = Q 1 ¯ u

Q

p

m

Design variables

λ > 0 and Q 1 , Qm as in Table 9.6

Example 9.5.3 Consider the same plant as in Example 9.5.2, i.e.,

b

yp =

(1 + ∆

s + a

m( s)) up

and the same control objective that requires the output yp to track the constant

reference signal ym = 1. A robust ALQC scheme can be constructed using Table 9.7

as follows:

Adaptive Law

˙ θp = Pr[Γ( φ − σsθp)]

732

CHAPTER 9. ROBUST ADAPTIVE CONTROL SCHEMES

z − θ

=

p φ , m 2 = 1 + n 2

m 2

s,

n 2 s = ms

˙

ms = −δ 0 ms + u 2 p + y 2 p, ms(0) = 0

s

1

z =

y

[ u

s + λ p, φ = s + λ p, −yp]

θp = [ˆ b, ˆ a]

The projection operator Pr[ ·] constraints ˆ b to satisfy |ˆ b( t) | ≥ b 0 ∀t ≥ 0 where b 0 is a known lower bound for |b| and σs is the switching σ-modification.

State Observer

˙ˆ e = ˆ

Aˆ

e + ˆ

B ¯

up − ˆ

Ko([1 0]ˆ e − e 1)

where

ˆ

ˆ a 1

1

10 ˆ a

A =

,

ˆ

B = ˆ b

,

ˆ

K

0

0

1

o =

25

Controller Parameters ˆ

Kc = λ− 1 ˆ

B P where P = P

> 0 is solved pointwise

in time using

ˆ

A P + P ˆ

A − P ˆ

Bλ− 1 ˆ

B P + CC = O, C = [1 0]

s + 1

Control Law

¯

up = ˆ

Kc( te, up =

¯

u

s

p

9.6

Adaptive Control of LTV Plants

One of the main reasons for considering adaptive control in applications is to

compensate for large variations in the plant parameters. One can argue that

if the plant model is LTI with unknown parameters a sufficient number of

tests can be performed off-line to calculate these parameters with sufficient

accuracy and therefore, there is no need for adaptive control when the plant

model is LTI. One can also go further and argue that if some nominal values

of the plant model parameters are known, robust control may be adequate

as long as perturbations around these nominal values remain within certain

bounds. And again in this case there is no need for adaptive control. In

many applications, however, such as aerospace, process control, etc., LTI

plant models may not be good approximations of the plant due to drastic

changes in parameter values that may arise due to changes in operating

9.6. ADAPTIVE CONTROL OF LTV PLANTS

733

points, partial failure of certain components, wear and tear effects, etc. In

such applications, linear time varying (LTV) plant models of the form

˙ x = A( t) x + B( t) u

y = C ( t) x + D( t) u

(9.6.1)

where A, B, C, and D consist of unknown time varying elements, may be

necessary. Even though adaptive control was motivated for plants modeled

by (9.6.1), most of the work on adaptive control until the mid 80’s dealt

with LTI plants. For some time, adaptive control for LTV plants whose

parameters vary slowly with time was considered to be a trivial extension of

that for LTI plants. This consideration was based on the intuitive argument

that an adaptive system designed for an LTI plant should also work for a

linear plant whose parameters vary slowly with time. This argument was

later on proved to be invalid. In fact, attempts to apply adaptive control to

simple LTV plants led to similar unsolved stability and robustness problems

as in the case of LTI plants with modeling errors. No significant progress was

made towards the design and analysis of adaptive controllers for LTV plants

until the mid-1980s when some of the fundamental robustness problems of

adaptive control for LTI plants were resolved.

In the early attempts [4, 74], the notion of the PE property of certain

signals in the adaptive control loop was employed to guarantee the exponen-

tial stability of the unperturbed error system, which eventually led to the

local stability of the closed-loop time-varying (TV) plant. Elsewhere, the

restriction of the type of time variations of the plant parameters also led to

the conclusion that an adaptive controller could be used in the respective

environment. More specifically, in [31] the parameter variations were as-

sumed to be perturbations of some nominal constant parameters, which are

small in the norm and modeled as a martingale process with bounded covari-

ance. A treatment of the parameter variations as small TV perturbations,

in an L 2-gain sense, was also considered in [69] for a restrictive class of LTI-

nominal plants. Special models of parameter variations, such as exponential

or 1 /t-decaying or finite number of jump discontinuities, were considered in

[70, 138, 181]. The main characteristic of these early studies was that no

modifications to the adaptive laws were necessary due to either the use of

PE or the restriction of the parameter variations to a class that introduces

no persistent modeling error effects in the adaptive control scheme.

734

CHAPTER 9. ROBUST ADAPTIVE CONTROL SCHEMES

The adaptive control problem for general LTV plants was initially treated

as a robustness problem where the effects of slow parameter variations were

treated as unmodeled perturbations [9, 66, 145, 222]. The same robust adap-

tive control techniques used for LTI plants in the presence of bounded distur-

bances and unmodeled dynamics were shown to work for LTV plants when

their parameters are smooth and vary slowly with time or when they vary

discontinuously, i.e., experience large jumps in their values but the disconti-

nuities occur over large intervals of time [225].

The robust MRAC and APPC schemes presented in the previous sections

can be shown to work well with LTV plants whose parameters belong to the

class of smooth and slowly varying with respect to time or the class with

infrequent jumps in their values. The difficulty in analyzing these schemes

with LTV plants has to do with the representations of the plant model. In

the LTI case the transfer function and related input/output results are used

to design and analyze adaptive controllers. For an LTV plant, the notion

of a transfer function and of poles and zeros is no longer applicable which

makes it difficult to extend the results of the LTI case to the LTV one. This

difficulty was circumvented in [223, 224, 225, 226] by using the notion of

the polynomial differential operator and the polynomial integral operator to

describe an LTV plant such as (9.6.1) in an input-output form that resembles

a transfer function description.

The details of these mathematical preliminaries as well as the design

and analysis of adaptive controllers for LTV plants of the form (9.6.1) are

presented in a monograph [226] and in a series of papers [223, 224, 225].

Interested readers are referred to these papers for further information.

9.7

Adaptive Control for Multivariable Plants

The design of adaptive controllers for MIMO plant models is more complex

than in the SISO case. In the MIMO case we are no longer dealing with a

single transfer function but with a transfer matrix whose elements are trans-

fer functions describing the coupling between inputs and outputs. As in the

SISO case, the design of an adaptive controller for a MIMO plant can be

accomplished by combining a control law, that meets the control objective

when the plant parameters are known, with an adaptive law that generates

estimates for the unknown parameters. The design of the control and adap-

9.7. ADAPTIVE CONTROL OF MIMO PLANTS

735

tive law, however, requires the development of certain parameterizations of

the plant model that are more complex than those in the SISO case.

In this section we briefly describe several approaches that can be used to

design adaptive controllers for MIMO plants.

9.7.1

Decentralized Adaptive Control

Let us consider the MIMO plant model

y = H( s) u

(9.7.1)

where y ∈ RN , u ∈ RN and H( s) ∈ CN×N is the plant transfer matrix that

is assumed to be proper. Equation (9.7.1) may be also expressed as

yi = hii( s) ui +

hij( s) uj,

i = 1 , 2 , . . . , N

(9.7.2)

1 ≤j≤N

j= i

where hij( s), the elements of H( s), are transfer functions.

Another less obvious but more general decomposition of (9.7.1) is

yi = hii( s) ui +

( hij( s) uj + qij( s) yj) , i = 1 , 2 , . . . , N

(9.7.3)

1 ≤j≤N

j= i

for some different transfer functions hij( s) , qij( s) .

If the MIMO plant model (9.7.3) is such that the interconnecting or

coupling transfer functions hij( s) , qij( s) ( i = j) are stable and small in some sense, then they can be treated as modeling error terms in the control design.

This means that instead of designing an adaptive controller for the MIMO

plant (9.7.3), we can design N adaptive controllers for N SISO plant models

of the form

yi = hii( s) ui, i = 1 , 2 , . . . , N

(9.7.4)

If these adaptive controllers are designed based on robustness considerations,

then the effect of the small unmodeled interconnections present in the MIMO

plant will not destroy stability. This approach, known as decentralized adap-

tive control, has been pursued in [38, 59, 82, 185, 195].

The analysis of decentralized adaptive control designed for the plant

model (9.7.4) but applied to (9.7.3) follows directly from that of robust adap-

tive control for plants with unmodeled dynamics considered in the previous

sections and is left as an exercise for the reader.

736

CHAPTER 9. ROBUST ADAPTIVE CONTROL SCHEMES

9.7.2

The Command Generator Tracker Approach

The command generator tracker (CGT) theory was first proposed in [26] for

the model following problem with known parameters. An adaptive control

algorithm based on the CGT method was subsequently developed in [207,

208]. Extensions and further improvements of the adaptive CGT algorithms

for finite [18, 100] as well as for infinite [233, 234] dimensional plant models

followed the work of [207, 208].

The adaptive CGT algorithms are based on a plant/reference model

structural matching and on an SPR condition. They are developed using

the SPR-Lyapunov design approach.

The plant model under consideration in the CGT approach is in the state

space form

˙ x = Ax + Bu ,

x(0) = x 0

y = C x

(9.7.5)

where x ∈ Rn; y, u ∈ Rm; and A, B, and C are constant matrices of appro-

priate dimensions. The control objective is to choose u so that all signals in

the closed-loop plant are bounded and the plant output y tracks the output

ym of the reference model described by

˙ xm = Amxm + Bmr

xm(0) = xm 0

ym = Cmxm

(9.7.6)

where xm ∈ Rnm, r ∈ Rpm and ym ∈ Rm.

The only requirement on the reference model at this point is that its

output ym has the same dimension as the plant output. The dimension of

xm can be much lower than that of x.

Let us first consider the case where the plant parameters A, B, and C

are known and use the following assumption.

Assumption 1 (CGT matching condition) There exist matrices S∗ 11,

S∗ 12, S∗ 21, S∗ 22 such that the desired plant input u∗ that meets the control

objective satisfies

x∗

S∗ 11 S∗ 12

xm

(9.7.7)

u∗

=

S∗ 21 S∗ 22

r

9.7. ADAPTIVE CONTROL OF MIMO PLANTS

737

where x∗ is equal to x when u = u∗, i.e., x∗ satisfies

˙ x∗ = Ax∗ + Bu∗

y∗ = C x∗ = ym

Assumption 2 (Output stabilizability condition) There exists a ma-

trix G∗ such that A + BG∗C

is a stable matrix.

If assumptions 1, 2 are satisfied we can use the control law

u = G∗( y − ym) + S∗ 21 xm + S∗ 22 r

(9.7.8)

that yields the closed-loop plant

˙ x = Ax + BG∗e 1 + BS∗ 21 xm + BS∗ 22 r

where e 1 = y − ym. Let e = x − x∗ be the state error between the actual and

desired plant state. Note the e is not available for measurement and is used

only for analysis. The error e satisfies the equation

˙ e = ( A + BG∗C ) e

which implies that e and therefore e 1 are bounded and converge to zero

exponentially fast. If xm, r are also bounded then we can conclude that all

signals are bounded and the control law (9.7.8) meets the control objective

exactly.

If A, B, and C are unknown, the matrices G∗, S∗ 21, and S∗ 22 cannot be

calculated, and, therefore, (9.7.8) cannot be implemented. In this case we

use the control law

u = G( t)( y − ym) + S 21( t) xm + S 22( t) r

(9.7.9)

where G( t) , S 21( t) , S 22( t) are the estimates of G∗, S∗ 21 , S∗ 22 to be generated by an adaptive law. The adaptive law is developed using the SPR-Lyapunov

design approach and employs the following assumption.

Assumption 3 (SPR condition) There exists a matrix G∗ such that

C ( sI − A − BG∗C ) 1 B is an SPR transfer matrix.

738

CHAPTER 9. ROBUST ADAPTIVE CONTROL SCHEMES

Assumption 3 puts a stronger condition on the output feedback gain

G∗, i.e., in addition to making A + BG∗C stable it should also make the

closed-loop plant transfer matrix SPR.

For the closed-loop plant (9.7.5), (9.7.9), the equation for e = x − x∗

becomes

˙ e = ( A + BG∗C ) e + B ˜

Ge 1 + B ˜

e 1 = C e

(9.7.10)

where

˜

G( t) = G( t) − G∗,

˜

S( t) = S( t) − S∗

S( t) = [ S 21( t) , S 22( t)] ,

S∗ = [ S∗ 21 , S∗ 22]

ω = [ xm, r ]

We propose the Lyapunov-like function

V = e P

˜

˜

ce + tr[ ˜

G Γ 1

1 G] + tr[ ˜

S Γ 1

2 S]

where Pc = Pc > 0 satisfies the equations of the matrix form of the LKY

Lemma (see Lemma 3.5.4) because of assumption 3, and Γ1 , Γ2 are symmet-

ric positive definite matrices.

Choosing

˙˜

G = ˙

G = Γ1 e 1 e 1

˙˜

S = ˙

S = Γ2 e 1 ω

(9.7.11)

it follows as in the SISO case that

˙

V = −e QQ e − νce Lce

where Lc = Lc > 0 and νc > 0 is a small constant and Q is a constant matrix.

Using similar arguments as in the SISO case we can establish that e, G, S

are bounded and e ∈ L 2. If in addition xm, r are bounded we can establish

that all signals are bounded and e, e 1 converge to zero as t → ∞. Therefore

the adaptive control law (9.7.9), (9.7.11) meets the control objective.

Due to the restrictive nature of Assumptions 1 to 3, the CGT approach

did not receive as much attention as other adaptive control methods. As

9.7. ADAPTIV