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.

whose understanding helps find ways to counteract them by redesigning the

adaptive schemes of the previous chapters.

8.3.1

Parameter Drift

Let us consider the same plant as in the example of Section 4.2.1, where the

plant output is now corrupted by some unknown bounded disturbance d( t),

i.e.,

y = θ∗u + d

The adaptive law for estimating θ∗ derived in Section 4.2.1 for d( t) = 0

∀t ≥ 0 is given by

˙ θ = γ 1 u,

1 = y − θu

(8.3.1)

where γ > 0 and θ( t) is the on-line estimate of θ∗. We have shown that for

d( t) = 0 and u, ˙ u ∈ L∞, (8.3.1) guarantees that θ, 1 ∈ L∞ and 1( t) 0 as t → ∞. Let us now analyze (8.3.1) when d( t) = 0. Defining ˜

θ = θ − θ∗, we

have

1 = ˜

θu + d

and

˙˜ θ = −γuθ+ γdu

(8.3.2)

546

CHAPTER 8. ROBUST ADAPTIVE LAWS

We analyze (8.3.2) by considering the function

˜

θ 2

V

θ) =

(8.3.3)

2 γ

Along the trajectory of (8.3.2), we have

˜

˙

θ 2 u 2

1

d 2

V = ˜

θ 2 u 2 + d˜

θu =

θu − d)2 +

(8.3.4)

2

2

2

For the class of inputs considered, i.e., u ∈ L∞ we cannot conclude that

˜

θ is bounded from considerations of (8.3.3), (8.3.4), i.e., we cannot find a

constant V 0 > 0 such that for V > V 0, ˙ V ≤ 0. In fact, for θ∗ = 2 , γ = 1, u = (1 + t) 12 ∈ L∞ and

5

d( t) = (1 + t) 14

2(1 + t) 14

0 as t → ∞

4

we have

5

y( t) = (1 + t) 14 0 as t → ∞

4

1

1( t) =

(1 + t) 14 0 as t → ∞

4

and

1

θ( t) = (1 + t) 4 → ∞ as t → ∞

i.e., the estimated parameter θ( t) drifts to infinity with time. We refer to

this instability phenomenon as parameter drift.

It can be easily verified that for u( t) = (1 + t) 12 the equilibrium ˜

θe = 0

of the homogeneous part of (8.3.2) is u.s. and a.s., but not u.a.s. Therefore,

it is not surprising that we are able to find a bounded input γdu that leads

to an unbounded solution ˜

θ( t) = θ( t) 2. If, however, we restrict u( t) to be

PE with level α 0 > 0, say, by choosing u 2 = α 0, then it can be shown (see

Problem 8.4) that the equilibrium ˜

θe = 0 of the homogeneous part of (8.3.2)

is u.a.s. (also e.s.) and that

d

lim sup |˜

θ( τ ) | ≤ 0 ,

d 0 = sup |d( t) u( t) |

t→∞ τ≥t

α 0

t

i.e., ˜

θ( t) converges exponentially to the residual set

d

D

0

θ =

˜

θ |˜

θ| ≤ α 0

8.3. INSTABILITY PHENOMENA IN ADAPTIVE SYSTEMS

547

indicating that the parameter error at steady state is of the order of the

disturbance level.

Unfortunately, we cannot always choose u to be PE especially in the case

of adaptive control where u is no longer an external signal but is generated

from feedback.

Another case of parameter drift can be demonstrated by applying the

adaptive control scheme

u = −kx,

˙ k = γx 2

(8.3.5)

developed in Section 6.2.1 for the ideal plant ˙ x = ax + u to the plant

˙ x = ax + u + d

(8.3.6)

where d( t) is an unknown bounded input disturbance. It can be verified that

for k(0) = 5 , x(0) = 1 , a = 1 , γ = 1 and

d( t) = (1 + t) 15 5 (1 + t) 15 0 . 4(1 + t) 65 0 as t → ∞

the solution of (8.3.5), (8.3.6) is given by

x( t) = (1 + t) 25 0 as t → ∞

and

1

k( t) = 5(1 + t) 5 → ∞ as t → ∞

We should note that since k(0) = 5 > 1 and k( t) ≥ k(0) ∀t ≥ 0 parameter

drift can be stopped at any time t by switching off adaptation, i.e., setting

γ = 0 in (8.3.5), and still have x, u ∈ L∞ and x( t) 0 as t → ∞. For

example, if γ = 0 for t ≥ t 1 > 0, then k( t) = ¯ k ∀t ≥ t 1, where ¯ k ≥ 5 is a stabilizing gain, which guarantees that u, x ∈ L∞ and x( t) 0 as t → ∞.

One explanation of the parameter drift phenomenon may be obtained by

solving for the “quasi” steady state response of (8.3.6) with u = −kx, i.e.,

d

xs ≈ k − a

Clearly, for a given a and d, the only way for xs to go to zero is for k → ∞.

That is, in an effort to eliminate the effect of the input disturbance d and send

x to zero, the adaptive control scheme creates a high gain feedback. This

high gain feedback may lead to unbounded plant states when in addition to

bounded disturbances, there are dynamic plant uncertainties, as explained

next.

548

CHAPTER 8. ROBUST ADAPTIVE LAWS

8.3.2

High-Gain Instability

Consider the following plant:

y

1 − µs

1

2 µs

=

=

1

(8.3.7)

u

( s − a)(1 + µs)

s − a

1 + µs

where µ is a small positive number which may be due to a small time constant

in the plant.

A reasonable approximation of the second-order plant may be obtained

by approximating µ with zero leading to the first-order plant model

¯

y

1

=

(8.3.8)

u

s − a

where ¯

y denotes the output y when µ = 0. The plant equation (8.3.7) is in

the form of y = G 0(1 + ∆ m) u where G 0( s) = 1 and ∆

. It

s−a

m( s) = 2 µs

1+ µs

may be also expressed in the singular perturbation state-space form discussed

in Section 8.2, i.e.,

˙ x = ax + z − u

µ ˙ z = −z + 2 u

(8.3.9)

y = x

Setting µ = 0 in (8.3.9) we obtain the state-space representation of (8.3.8),

i.e.,

˙¯ x = a¯ x + u

(8.3.10)

¯

y = ¯

x

where ¯

x denotes the state x when µ = 0. Let us now design an adaptive

controller for the simplified plant (8.3.10) and use it to control the actual

second order plant where µ > 0. As we have shown in Section 6.2.1, the

adaptive law

u = −k¯

x,

˙ k = γ x,

1 = ¯

x

can stabilize (8.3.10) and regulate ¯

y = ¯

x to zero. Replacing ¯

y with the actual

output of the plant y = x, we have

u = −kx,

˙ k = γ 1 x,

1 = x

(8.3.11)

which, when applied to (8.3.9), gives us the closed-loop plant

˙ x = ( a + k) x + z

(8.3.12)

µ ˙ z = −z − 2 kx

8.3. INSTABILITY PHENOMENA IN ADAPTIVE SYSTEMS

549

whose equilibrium xe = 0 , ze = 0 with k = k∗ is a.s. if and only if the

eigenvalues of (8.3.12) with k = k∗ are in the left-half s-plane, which implies

that

1 − a > k∗ > a

µ

Because ˙ k ≥ 0, it is clear that

1

1

k(0) >

− a ⇒ k( t) >

− a, ∀t ≥ 0

µ

µ

that is, from such a k(0) the equilibrium of (8.3.12) cannot be reached.

Moreover, with k > 1 − a the linear feedback loop is unstable even when the

µ

adaptive loop is disconnected, i.e., γ = 0.

We refer to this form of instability as high-gain instability. The adaptive

control law (8.3.11) can generate a high gain feedback which excites the

unmodeled dynamics and leads to instability and unbounded solutions. This

type of instability is well known in robust control with no adaptation [106]

and can be avoided by keeping the controller gains small (leading to a small

loop gain) so that the closed-loop bandwidth is away from the frequency

range where the unmodeled dynamics are dominant.

8.3.3

Instability Resulting from Fast Adaptation

Let us consider the second-order plant

˙ x = −x + bz − u

µ ˙ z = −z + 2 u

(8.3.13)

y = x

where b > 1 / 2 is an unknown constant and µ > 0 is a small number. For

u = 0, the equilibrium xe = 0 , ze = 0 is e.s. for all µ ≥ 0. The objective

here, however, is not regulation but tracking. That is, the output y = x is

required to track the output xm of the reference model

˙ xm = −xm + r

(8.3.14)

where r is a bounded piecewise continuous reference input signal. For µ = 0,

the reduced-order plant is

˙¯ x = ¯ x + (2 b − 1) u

(8.3.15)

¯

y = ¯

x

550

CHAPTER 8. ROBUST ADAPTIVE LAWS

which has a pole at s = 1, identical to that of the reference model, and an

unknown gain 2 b − 1 > 0. The adaptive control law

u = lr,

˙ l = −γ 1 r,

1 = ¯

x − xm

(8.3.16)

guarantees that ¯

x( t) , l( t) are bounded and |¯

x( t) − xm( t) | → 0 as t → ∞ for

any bounded reference input r. If we now apply (8.3.16) with ¯

x replaced by

x to the actual plant (8.3.13) with µ > 0, the closed-loop plant is

˙ x = −x + bz − lr

µ ˙ z = −z + 2 lr

(8.3.17)

˙ l = −γr( x − xm)

When γ = 0 that is when l is fixed and finite, the signals x( t) and z( t) are

bounded. Thus, no instability can occur in (8.3.17) for γ = 0.

Let us assume that r = constant. Then, (8.3.17) is an LTI system of the

form

˙

Y = AY + Bγrxm

(8.3.18)

where Y = [ x, z, l] , B = [0 , 0 , 1] ,

1

b

−r

A = 

0

1 2 r/µ

−γr

0

0

and γrxm is treated as a bounded input. The characteristic equation of

(8.3.18) is then given by

1

1

γ

det( sI − A) = s 3 + (1 + ) s 2 + ( − γr 2) s + r 2(2 b − 1)

(8.3.19)

µ

µ

µ

Using the Routh-Hurwitz criterion, we have that for

1 (1 + µ)

γr 2 >

(8.3.20)

µ (2 b + µ)

two of the roots of (8.3.19) are in the open right-half s-plane. Hence given any

µ > 0, if γ, r satisfy (8.3.20), the solutions of (8.3.18) are unbounded in the

sense that |Y ( t) | → ∞ as t → ∞ for almost all initial conditions. Large γr 2

increases the speed of adaptation, i.e., ˙ l, which in turn excites the unmodeled

dynamics and leads to instability. The effect of ˙ l on the unmodeled dynamics

8.3. INSTABILITY PHENOMENA IN ADAPTIVE SYSTEMS

551

can be seen more clearly by defining a new state variable η = z − 2 lr, called

the parasitic state [85, 106], to rewrite (8.3.17) as

˙ x = −x + (2 b − 1) lr +

µ ˙ η = −η + 2 µγr 2( x − xm) 2 µl ˙ r

(8.3.21)

˙ l = −γr( x − xm)

where for r = constant, ˙ r ≡ 0. Clearly, for a given µ, large γr 2 may lead to

a fast adaptation and large parasitic state η which acts as a disturbance in

the dominant part of the plant leading to false adaptation and unbounded

solutions. For stability and bounded solutions, γr 2 should be kept small, i.e.,

the speed of adaptation should be slow relative to the speed of the parasitics

characterized by 1 .

µ

8.3.4

High-Frequency Instability

Let us now assume that in the adaptive control scheme described by (8.3.21),

both γ and |r| are kept small, i.e., adaptation is slow. We consider the case

where r = r 0 sin ωt and r 0 is small but the frequency ω can be large. At

lower to middle frequencies and for small γr 2 , µ, we can approximate µ ˙ η ≈ 0

and solve (8.3.21) for η, i.e., η ≈ 2 µγr 2( x − xm) 2 µl ˙ r, which we substitute into the first equation in (8.3.21) to obtain

˙ x = (1 2 µbγr 2) x + ( gr − 2 ˙ r) l − 2 µbγr 2 xm

˙

(8.3.22)

l = −γr( x − xm)

where g = 2 b − 1. Because γr is small we can approximate l( t) with a

constant and solve for the sinusoidal steady-state response of x from the

first equation of (8.3.22) where the small µγr 2 terms are neglected, i.e.,

r

x

0

ss =

[( g − 2 bµω 2) sin ωt − ω( g + 2 )cos ωt] l

1 + ω 2

Now substituting for x = xss in the second equation of (8.3.22) we obtain

˙

γr 2

l

0 sin ωt

s =

( g − 2 bµω 2) sin ωt − ω( g + 2 )cos ωt l

1 + ω 2

s + γxmr 0 sin ωt

i.e.,

˙ ls = α( t) ls + γxmr 0 sin ωt

(8.3.23)

552

CHAPTER 8. ROBUST ADAPTIVE LAWS

γr 2