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.

where A =

|b|

, ϕ = (sgn( b) 1)90 ◦ − tan 1 1 , which can be shown to

1+ a 2

a

satisfy (5.2.6). Therefore, for u = sin t the signal vector φ carries sufficient

information about a and b, φ is P E and θ( t) → θ∗ exponentially fast.

We say that u = sin t is sufficiently rich for identifying the plant (5.2.3),

i.e., it contains a sufficient number of frequencies to excite all the modes of

the plant. Because u = c 0 = 0 can excite only the zero frequency gain of the

plant, it is not sufficiently rich for the plant (5.2.3).

5.2. PARAMETER IDENTIFIERS

255

Let us consider the second order plant

b

y =

1 s + b 0

u = G( s) u

(5.2.7)

s 2 + a 1 s + a 0

where a 1 , a 0 > 0 and G( s) has no zero-pole cancellations. We can show

that for u = sin ω 0 t, y( t) at steady state does not carry sufficient infor-

mation to be able to uniquely determine a 1 , a 0 , b 1 , b 0. On the other hand,

u( t) = sin ω 0 t + sin ω 1 t where ω 0 = ω 1 leads to the steady-state response y( t) = A 0 sin( ω 0 t + ϕ 0) + A 1 sin( ω 1 t + ϕ 1) where A 0 = | G( 0) |, ϕ 0 = G( 0) , A 1 = | G( 1) |, ϕ 1 = G( 1). By measuring A 0 , A 1 , ϕ 0 , ϕ 1 we can determine uniquely a 1 , a 0 , b 1 , b 0 by solving four algebraic equations.

Because each frequency in u contributes two equations, we can argue that

the number of frequencies that u should contain, in general, is proportional

to the number of unknown plant parameters to be estimated.

We are now in a position to give the following definition of sufficiently

rich signals.

Definition 5.2.1 A signal u : R+ → R is called sufficiently rich of order

n if it consists of at least n distinct frequencies.

2

For example, the input

m

u =

Ai sin ωit

(5.2.8)

i=1

where m ≥ n/ 2, Ai = 0 are constants and ωi = ωk for i = k is sufficiently

rich of order n.

A more general definition of sufficient richness that includes signals that

are not necessarily equal to a sum of sinusoids is presented in [201] and is

given below.

Definition 5.2.2 A signal u : R+ → Rn is said to be stationary if the

following limit exists uniformly in t 0

1

t 0+ T

Ru( t) = lim

u( τ ) u ( t + τ )

T →∞ T

t 0

256

CHAPTER 5. IDENTIFIERS AND ADAPTIVE OBSERVERS

The matrix Ru( t) ∈ Rn×n is called the autocovariance of u. Ru( t) is a positive semidefinite matrix and its Fourier transform given by

Su( ω) =

e−jωτ Ru( τ)

−∞

is referred to as the spectral measure of u. If u has a sinusoidal component at

frequency ω 0 then u is said to have a spectral line at frequency ω 0 and Su( ω) has a point mass (a delta function) at ω 0 and −ω 0. Given Su( ω), Ru( t) can be calculated using the inverse Fourier transform, i.e.,

1

Ru( t) =

ejωtS

2 π

u( ω)

−∞

Furthermore, we have

∞ Su( ω) = 2 πRu(0)

−∞

For further details about the properties of Ru( t) , Su( ω), the reader is referred

to [186, 201].

Definition 5.2.3 A stationary signal u : R+ → R is called sufficiently

rich of order n, if the support of the spectral measure Su( ω) of u contains

at least n points.

Definition 5.2.3 covers a wider class of signals that includes those specified

by Definition 5.2.1. For example, the input (5.2.8) has a spectral measure

with 2 m points of support, i.e., at ωi, −ωi for i = 1 , 2 , . . . m, where m ≥ n/ 2, and is, therefore, sufficiently rich of order n.

Let us now consider the equation

φ = H( s) u

(5.2.9)

where H( s) is a proper transfer matrix with stable poles and φ ∈ Rn. The

P E property of φ is related to the sufficient richness of u by the following

theorem given in [201].

Theorem 5.2.1 Let u : R+ → R be stationary and assume that H( 1) ,

. . ., H( jωn) are linearly independent on Cn for all ω 1 , ω 2 , . . . , ωn ∈ R, where

ωi = ωk for i = k. Then φ is P E if, and only if, u is sufficiently rich of

order n.

5.2. PARAMETER IDENTIFIERS

257

The proof of Theorem 5.2.1 is given in Section 5.6.

The notion of persistence of excitation of the vector φ and richness of the

input u attracted the interest of several researchers in the 1960s and 1970s

who gave various interpretations to the properties of P E and sufficiently

rich signals. The reader is referred to [1, 171, 201, 209, 242] for further

information on the subject.

Roughly speaking, if u has at least one distinct frequency component for

each two unknown parameters, then it is sufficiently rich. For example, if

the number of unknown parameters is n, then m ≥ n distinct frequencies

2

in u are sufficient for u to qualify as being sufficiently rich of order n. Of

course, these statements are valid provided H( 1) , . . . , H( jωn) with ωi = ωk

are linearly independent on Cn for all ωi ∈ R, i = 1 , 2 , . . . , n. The vectors

H( jωi) , i = 1 , 2 , . . . , n may become linearly dependent at some frequencies

in R under certain conditions such as the one illustrated by the following

example where zeros of the plant are part of the internal model of u.

Example 5.2.1 Let us consider the following plant:

b

y = 0( s 2 + 1) u = G( s) u

( s + 2)3

where b 0 is the only unknown parameter. Following the procedure of Chapter 2, we

rewrite the plant in the form of the linear parametric model

y = θ∗φ

where θ∗ = b 0 is the unknown parameter and

s 2 + 1

φ = H( s) u,

H( s) = ( s + 2)3

According to Theorem 5.2.1, we first need to check the linear independence of

H( 1) , . . . , H( jωn). For n = 1 this condition becomes H( ) = 0 , ∀ω ∈ R. It is clear that for ω = 1 , H( j) = 0, and, therefore, φ may not be PE if we simply choose

u to be sufficiently rich of order 1. That is, for u = sin t, the steady-state values of

φ, y are equal to zero and, therefore, carry no information about the unknown b 0.

We should note, however, that for u = sin ωt and any ω = 1 , 0, φ is PE.

Remark 5.2.1 The above example demonstrates that the condition for the

linear independence of the vectors H( jωi) , i = 1 , 2 , · · · , n on Cn is suf-

ficient to guarantee that φ is P E when u is sufficiently rich of order n.

258

CHAPTER 5. IDENTIFIERS AND ADAPTIVE OBSERVERS

It also demonstrates that when the plant is partially known, the input

u does not have to be sufficiently rich of order n where n is the order of

the plant. In this case, the condition on u can be relaxed, depending

on the number of the unknown parameters. For further details on the

problem of prior information and persistent of excitation, the reader

is referred to [35].

In the following sections we use Theorem 5.2.1 to design the input signal

u for a wide class of parameter estimators developed in Chapter 4.

5.2.2

Parameter Identifiers with Full-State Measurements

Let us consider the plant

˙ x = Ax + Bu, x(0) = x 0

(5.2.10)

where C = I , i.e., the state x ∈ Rn is available for measurement and A, B

are constant matrices with unknown elements that we like to identify. We

assume that A is stable and u ∈ L∞.

As shown in Chapter 4, the following two types of parameter estimators

may be used to estimate A, B from the measurements of x, u.

Series-Parallel

˙ˆ x = Amˆ x + ( ˆ

A − Am) x + ˆ

Bu

(5.2.11)

˙ˆ

A = γ 1 1 x ,

˙ˆ

B = γ 2 1 u

where Am is a stable matrix chosen by the designer, 1 = x − ˆ x, γ 1 , γ 2 > 0

are the scalar adaptive gains.

Parallel

˙ˆ x = ˆ

Aˆ

x + ˆ

Bu

(5.2.12)

˙ˆ

A = γ 1 1ˆ x ,

˙ˆ

B = γ 2 1 u

where 1 = x − ˆ x and γ 1 , γ 2 > 0 are the scalar adaptive gains.

As shown in Chapter 4, if A is a stable matrix and u ∈ L∞ then

ˆ

x, ˆ

A, ˆ

B ∈ L∞;

˙ˆ

A( t) , ˙ˆ

B( t) , 1 ∈ L 2 ∩ L∞ and 1( t) and the elements

of ˙ˆ

A( t) , ˙ˆ

B( t) converge to zero as t → ∞.

5.2. PARAMETER IDENTIFIERS

259

For the estimators (5.2.11) and (5.2.12) to become parameter identifiers,

the input signal u has to be chosen so that ˆ

A( t) , ˆ

B( t) converge to the un-

known plant parameters A, B, respectively, as t → ∞.

For simplicity let us first consider the case where u is a scalar input, i.e.,

B ∈ Rn× 1.

Theorem 5.2.2 Let ( A, B) be a controllable pair. If the input u ∈ R 1 is suf-

ficiently rich of order n + 1 , then the estimates ˆ

A, ˆ

B generated by (5.2.11) or

(5.2.12) converge exponentially fast to the unknown plant parameters A, B,

respectively.

The proof of Theorem 5.2.2 is quite long and is presented in Section 5.6.

An example of a sufficiently rich input u for the estimators (5.2.11),

(5.2.12) is the input

m

u =

Ai sin ωit

i=1

for some constants Ai = 0 and ωi = ωk for i = k and for some integer

m ≥ n+1 .

2

Example 5.2.2 Consider the second-order plant

˙ x = Ax + Bu

where x = [ x 1 , x 2] , and the matrices A, B are unknown, and A is a stable matrix.

1

0

Using (5.2.11) with Am =

, the series-parallel parameter identifier is

0

1

given by

ˆ

˙

1

0

ˆ a

b

ˆ

x =

x − x) +

11( t)

ˆ a 12( t)

x +

1( t)

u

0

1

ˆ a

ˆ

21( t)

ˆ a 22( t)

b 2( t)

where ˆ

x = [ˆ

x 1 , ˆ x 2] and

˙ˆ aik = ( xi − ˆ xi) xk,

i = 1 , 2; k = 1 , 2

˙ˆ bi = ( xi − ˆ xi) u, i = 1 , 2

The input u is selected as

u = 5 sin 2 . 5 t + 6 sin 6 . 1 t

260

CHAPTER 5. IDENTIFIERS AND ADAPTIVE OBSERVERS

which has more frequencies than needed since it is sufficiently rich of order 4. An

input with the least number of frequencies that is sufficiently rich for the plant

considered is

u = c 0 + sin ω 0 t

for some c 0 = 0 and ω 0 = 0.

If u is a vector, i.e., u ∈ Rq, q > 1 then the following theorem may be

used to select u.

Theorem 5.2.3 Let ( A, B) be a controllable pair. If each element ui, i =

1 , 2 , . . . q of u is sufficiently rich of order n + 1 and uncorrelated, i.e., each ui

contains different frequencies, then ˆ

A( t) , ˆ

B( t) converge to A, B, respectively,

exponentially fast.

The proof of Theorem 5.2.3 is similar to that for Theorem 5.2.2, and is

given in Section 5.6.

The controllability of the pair ( A, B) is critical for the results of The-

orems 5.2.2 and 5.2.3 to hold. If ( A, B) is not a controllable pair, then

the elements of ( A, B) that correspond to the uncontrollable part cannot be

learned from the output response because the uncontrollable parts decay to

zero exponentially fast and are not affected by the input u.

The complexity of the parameter identifier may be reduced if some of

the elements of the matrices ( A, B) are known. In this case, the order of

the adaptive law can be reduced to be equal to the number of the unknown

parameters. In addition, the input u may not have to be sufficiently rich of

order n + 1. The details of the design and analysis of such schemes are left

as exercises for the reader and are included in the problem section.

In the following section we extend the results of Theorem 5.2.2 to the case

where only the output of the plant, rather than the full state, is available

for measurement.

5.2.3

Parameter Identifiers with Partial-State Measurements

In this section we concentrate on the SISO plant

˙ x = Ax + Bu,

x(0) = x 0

(5.2.13)

y = C x

5.2. PARAMETER IDENTIFIERS

261

where A is a stable matrix, and y, u ∈ R 1 are the only signals available for

measurement.

Equation (5.2.13) may be also written as

y = C ( sI − A) 1 Bu + C ( sI − A) 1 x 0

(5.2.14)

where, because of the stability of A, t = L− 1 {C ( sI − A) 1 }x 0 is an exponentially decaying to zero term. We would like to design an on-line parameter

identifier to estimate the parameters A, B, C. The triple ( A, B, C) contains

n 2 + 2 n unknown parameters to be estimated using only input/output data.

The I/O properties of the plant (5.2.14) at steady state (where t = 0), how-

ever, are uniquely determined by at most 2 n parameters. These parameters

correspond to the coefficients of the transfer function

y( s)

b

= C ( sI − A) 1 B = msm + bm− 1 sm− 1 + . . . + b 0

(5.2.15)

u( s)

sn + an− 1 sn− 1 + . . . + a 0

where m ≤ n − 1.

Because there is an infinite number of triples ( A, B, C) that give the same

transfer function (5.2.15), the triple ( A, B, C) associated with the specific

or physical state space representation in (5.3.13) cannot be determined, in

general, from input-output