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.

( τ ) φ( τ ) dτ, k = 0 , 1 , . . .

k

α

( s)

α

( s)

φ =

n− 1

u

n− 1

y

y

Λ

p, −

p

,

z = sn

p

p( s)

Λ p( s)

Λ p( s)

z−θ

φ

=

pk

, m 2 = 1 + φ φ, ∀t ∈ [ t

Adaptive law

m 2

k, tk+1)

θpk = [ θbk, θak]

ˆ

Rp( s, tk) = sn + θ

α

a( k− 1) n− 1( s)

ˆ

Zp( s, tk) = θ

α

b( k− 1) n− 1( s)

T

Design variables

s = tk+1 − tk > Tm; 2 − Tsλ max(Γ) > γ, for some

γ > 0; Λ p( s) is monic and Hurwitz with degree n

with respect to such computational real time delays can be considerably

improved by using a hybrid adaptive law for parameter estimation. The

sampling rate of the hybrid adaptive law may be chosen appropriately to

allow for the computations of the control law to be completed within the

sampling interval.

Let Tm be the maximum time for performing the computations required

to calculate the control law. Then the sampling period Ts of the hybrid

adaptive law may be chosen as Ts = tk+1 − tk > Tm where {tk : k =

1 , 2 , . . .} is a time sequence. Table 7.7 presents a hybrid adaptive law based

on parametric model (7.4.2). It can be used to replace the continuous-time

496

CHAPTER 7. ADAPTIVE POLE PLACEMENT CONTROL

Table 7.8 Hybrid APPC scheme: polynomial approach

Z

Plant

y

p( s)

p =

u

R

p

p( s)

Reference signal

Qm( s) ym = 0

Adaptive law

Hybrid adaptive law of Table 7.7

Algebraic

Solve for ˆ

L( s, tk)= sn− 1+ l ( tk) αn− 2( s)

êquation

P ( s, tk) = p ( tk) αn+ q− 1( s) from equation

ˆ

L( s, tk) Qm( s) ˆ

Rp( s, tk) + ˆ

P ( s, tk) ˆ

Zp( s, tk) = A∗( s)

Control law

Λ( s) ˆ

L( s, t

ˆ

P ( s, t

u

k) Qm( s)

k)

p =

u

( y

Λ( s)

p −

Λ( s)

p − ym)

A∗ monic Hurwitz of degree 2 n + q − 1; Qm( s)

Design variables

monic of degree q with nonrepeated roots on the

axis; Λ( s) = Λ p( sq( s); Λ p( s) , Λ q( s) monic and Hurwitz with degree n, q − 1, respectively

adaptive laws of the APPC schemes discussed in Section 7.4 as shown in

Tables 7.8 to 7.10.

The controller parameters in the hybrid adaptive control schemes of Ta-

bles 7.8 to 7.10 are updated at discrete times by solving certain algebraic

equations. As in the continuous-time case, the solution of these equations ex-

ist provided the estimated polynomials ˆ

Rp( s, tk) Qm( s) , ˆ

Zp( s, tk) are strongly

coprime at each time tk.

The following theorem summarizes the stability properties of the hybrid

APPC schemes presented in Tables 7.8 to 7.10.

Theorem 7.5.1 Assume that the polynomials ˆ

Rp( s, tk) Qm( s) , ˆ

Zp( s, tk) are

strongly coprime at each time t = tk. Then the hybrid APPC schemes given

7.5. HYBRID APPC SCHEMES

497

Table 7.9 Hybrid APPC scheme: state variable approach

Plant

Z

y

p( s)

p =

u

R

p

p( s)

Reference signal

Qm( s) ym = 0

Adaptive law

Hybrid adaptive law of Table 7.7.

˙ˆ e = ˆ

Ak−e + ˆ

Bk−up − ˆ

Ko( k− 1)[ C ˆ e − e 1] ˆ

Ak− 1 =

In+ q− 1

−θ 1( k− 1) − − −−

ˆ

Bk− 1 = θ 2( k− 1) , C

=

State observer

0

[1 , 0 , . . . , 0] ˆ

Ko( k− 1) = a∗− θ 1( k− 1) θ 1( k− 1), θ 2( k− 1)

are the coefficient vectors of ˆ

Rp( s, tk) Qm( s) −sn+ q,

ˆ

Zp( s, tk) Q 1( s), respectively, a∗ is the coefficient

vector of A∗o( s) − sn+ q

Algebraic

Solve for ˆ

Kc( k− 1) the equation det[ sI − ˆ

Ak− 1 +

equation

ˆ

B

ˆ

k− 1 Kc( k− 1)] = A∗c( s)

Control law

Q

¯

u

1( s)

p = ˆ

Kc( k− 1)ˆ e, up =

¯

u

Q

p

m( s)

Design variables

Choose Qm, Q 1 , A∗o, A∗c, Q 1 as in Table 7.5

in Tables 7.8 to 7.10 guarantee signal boundedness and convergence of the

tracking error to zero asymptotically with time.

The proof of Theorem 7.5.1 is similar to that of the theorems in Sec-

tion 7.4, with minor modifications that take into account the discontinuities

in the parameters and is given in Section 7.7.

Table 7.10 Hybrid adaptive LQ control scheme

498

CHAPTER 7. ADAPTIVE POLE PLACEMENT CONTROL

Z

Plant

y

p( s)

p =

u

R

p

p( s)

Reference signal

Qm( s) ym = 0

Adaptive law

Hybrid adaptive law of Table 7.7.

˙ˆ e = ˆ

A

State observer

k− 1 ˆ

e + ˆ

Bk−up − ˆ

Ko( k− 1)[ C ˆ e − e 1]

ˆ

Ko( k− 1) , ˆ

Ak− 1 , ˆ

Bk− 1 , C as in Table 7.9

Solve for P

Riccati equation

k− 1 = Pk− 1 > 0 the equation

ˆ

A

ˆ

ˆ

ˆ

k− 1 Pk− 1+ Pk− 1 Ak− 1 1 P

B

B

λ k− 1 k− 1 k− 1 Pk− 1+ CC = 0

1

Q

Control law

¯

u

ˆ

1( s)

p =

B

¯

u

λ k− 1 Pk−

e, up = Q

p

m( s)

Design variables

Choose λ, Q 1( s) , Qm( s) as in Table 7.6

The major advantage of the hybrid adaptive control schemes described in

Tables 7.7 to 7.10 over their continuous counterparts is the smaller computa-

tional effort required during implementation. Another possible advantage is

better robustness properties in the presence of measurement noise, since the

hybrid scheme does not respond instantaneously to changes in the system,

which may be caused by measurement noise.

7.6

Stabilizability Issues and Modified APPC

The main drawbacks of the APPC schemes of Sections 7.4 and 7.5 is that

the adaptive law cannot guarantee that the estimated plant parameters or

polynomials satisfy the appropriate controllability or stabilizability condi-

tion at each time t, which is required to calculate the controller parameter

vector θc. Loss of stabilizability or controllability may lead to computational

problems and instability.

7.6. STABILIZABILITY ISSUES AND MODIFIED APPC

499

In this section we concentrate on this problem of the APPC schemes and

propose ways to avoid it. We call the estimated plant parameter vector θp at

time t stabilizable if the corresponding algebraic equation is solvable for the

controller parameters. Because we are dealing with time-varying estimates,

uniformity with respect to time is guaranteed by requiring the level of sta-

bilizability to be greater than some constant ε∗ > 0 . For example, the level

of stabilizability can be defined as the absolute value of the determinant of

the Sylvester matrix of the estimated plant polynomials.

We start with a simple example that demonstrates the loss of stabiliz-

ability that leads to instability.

7.6.1

Loss of Stabilizability: A Simple Example

Let us consider the first order plant

˙ y = y + bu

(7.6.1)

where b = 0 is an unknown constant. The control objective is to choose u

such that y, u ∈ L∞, and y( t) 0 as t → ∞.

If b were known then the control law

2

u = − y

(7.6.2)

b

would meet the control objective exactly. When b is unknown, a natural

approach to follow is to use the certainty equivalence control (CEC) law

2

uc = − y

(7.6.3)

ˆ b

where ˆ b( t) is the estimate of b at time t, generated on-line by an appropriate

adaptive law.

Let us consider the following two adaptive laws:

(i) Gradient

˙ˆ b = γφε , ˆ b(0) = ˆ b 0 = 0

(7.6.4)

where γ > 0 is the constant adaptive gain.

500

CHAPTER 7. ADAPTIVE POLE PLACEMENT CONTROL

(ii) Pure Least-Squares

˙ˆ b = Pφε , ˆ b(0) = ˆ b 0 = 0

˙

P

= −P 2

φ 2

, P (0) = p

1 + β

0 > 0

(7.6.5)

0 φ 2

where P, φ ∈ R 1,

z − ˆ

1

1

ε =

, y

y , φ =

u

1 + β

f =

0 φ 2

s + 1

s + 1

z =

˙ yf − yf

(7.6.6)

It can be shown that for β 0 > 0, the control law (7.6.3) with ˆ b generated

by (7.6.4) or (7.6.5) meets the control objective provided that ˆ b( t) = 0 ∀t ≥ 0 .

Let us now examine whether (7.6.4) or (7.6.5) can satisfy the condition

ˆ b( t) = 0 , ∀t ≥ 0 .

From (7.6.1) and (7.6.6), we obtain

˜

ε =

(7.6.7)

1 + β 0 φ 2

where ˜ b = ˆ b − b is the parameter error. Using (7.6.7) in (7.6.4), we have

˙ˆ

φ 2

b = −γ

b − b) , ˆ b(0) = ˆ b

1 + β

0

(7.6.8)

0 φ 2

Similarly, (7.6.5) can be rewritten as

˙ˆ

φ 2

b = −P

b − b) , ˆ b(0) = ˆ b

1 + β

0

0 φ 2 p

P ( t) =

0

, p 0 > 0

(7.6.9)

1 + p

t

φ 2

0

0 1+ β 0 φ 2

˙

It is clear from (7.6.8) and (7.6.9) that for ˆ b(0) = b, ˆ b( t) = 0 and ˆ b( t) =

b, ∀t ≥ 0; therefore, the control objective can be met exactly with such an

initial condition for ˆ b.

˙

If φ( t) = 0 over a nonzero finite time interval, we will have ˆ b = 0 , u =

y = 0, which is an equilibrium state (not necessarily stable though) and the

control objective is again met.

7.6. STABILIZABILITY ISSUES AND MODIFIED APPC

501

10

9

8

7

6

5

Output y(t)

4

3

2

1

0

-1

-0.5

0

0.5

1

1.5

2

2.5

Estimate b(t)

^

Figure 7.5 Output y( t) versus estimate ˆ b( t) for different initial conditions

y(0) and ˆ b(0) using the CEC uc = 2 y/ˆ b.

For analysis purposes, let us assume that b > 0 (unknown to the designer).

For φ = 0, both (7.6.8), (7.6.9) imply that

˙

sgn(ˆ b) = sgn(ˆ b( t) − b)

and, therefore, for b > 0 we have

˙ˆ

˙

b( t) > 0 if ˆ b(0) < b and ˆ b( t) < 0 if ˆ b(0) > b Hence, for ˆ b(0) < 0 < b, ˆ b( t) is monotonically increasing and crosses zero

leading to an unbounded control uc.

Figure 7.5 shows the plots of y( t) vs ˆ b( t) for different initial conditions

ˆ b(0), y(0), demonstrating that for ˆ b(0) < 0 < b, ˆ b( t) crosses zero leading to unbounded closed-loop signals. The value of b = 1 is used for this simulation.

The above example demonstrates that the CEC law (7.6.3) with (7.6.4)

or (7.6.5) as adaptive laws for generating ˆ b is not guaranteed to meet the

control objective. If the sign of b and a lower bound for |b| are known,

then the adaptive laws (7.6.4), (7.6.5) can be modified using projection to

502

CHAPTER 7. ADAPTIVE POLE PLACEMENT CONTROL

constrain ˆ b( t) from changing sign. This projection approach works for this

simple example but its extension to the higher order case is awkward due to

the lack of any procedure for constructing the appropriate convex parameter

sets for projecting the estimated parameters.

7.6.2

Modified APPC Schemes

The stabilizability problem has attracted considerable interest in the adap-

tive control community and several solutions have been proposed. We list

the most important ones below with a brief explanation regarding their ad-

vantages and drawbacks.

(a) Stabilizability is assumed. In this case, no modifications are introduced

and stabilizability is assumed to hold for all t ≥ 0 . Even though there is no

theoretical justification for such an assumption to hold, it has been often

argued that in most simulation studies, no stabilizability problems usually

arise. The example presented above illustrates that no stabilizability prob-