Scribd
Upload
91 views

523 M1380: Adaptive Control Systems Lecture 5: Sufficiently Rich Signals and Parameter Convergence

This document discusses the concept of sufficiently rich signals and how they relate to parameter convergence in adaptive control systems. It defines what constitutes a sufficiently rich signal and proves that if the signal vector φ is defined as φ = H(s)u, where u is a sufficiently rich signal and H(s) is a proper transfer matrix, then φ will have the persistence of excitation (PE) property. Specifically, it shows that if u is stationary and sufficiently rich of order n, and the vectors {H(jω1), ..., H(jωn)} are linearly independent, then φ will be PE. An example is also given to illustrate how the linear independence condition can sometimes be relaxed if the plant is partially known.

Uploaded by

Cuter Hsu
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
91 views

523 M1380: Adaptive Control Systems Lecture 5: Sufficiently Rich Signals and Parameter Convergence

This document discusses the concept of sufficiently rich signals and how they relate to parameter convergence in adaptive control systems. It defines what constitutes a sufficiently rich signal and proves that if the signal vector φ is defined as φ = H(s)u, where u is a sufficiently rich signal and H(s) is a proper transfer matrix, then φ will have the persistence of excitation (PE) property. Specifically, it shows that if u is stationary and sufficiently rich of order n, and the vectors {H(jω1), ..., H(jωn)} are linearly independent, then φ will be PE. An example is also given to illustrate how the linear independence condition can sometimes be relaxed if the plant is partially known.

Uploaded by

Cuter Hsu
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 6

523 M1380: Adaptive Control Systems

Lecture 5: Sufficiently Rich Signals and Parameter Convergence


Spring 2005
As shown in Lecture 4, parameter convergence requires additional conditions on
a signal vector that are independent of the type of the adaptive law employed. In
general the signal is related to the plant input or external input command u through
the equation

= H ( s )u

(5.0.1)

where H(s) is some proper transfer matrix with stable poles. In this appendix, we introduce some class of input signals u that guarantee the PE property of the signal vector in (5.0.1).

5.1 SUFFICIENTLY RICH SIGNALS

In this section, we introduce the signals of sufficient richness. For convenience,


we first define the following concepts:
+

Definition 5.1.1: A signal u : R R is said to be stationary if the following limit


n

exists uniformly in t0
1 t 0 +T
u ()u T (t + )d.
T T t0

Ru (t ) lim

Remark 5.1.1: The matrix Ru(t) R

nn

is called the autocovariance of u. Ru(t) is a

positive semidefinite matrix, and its Fourier transform given by

Su () = e j Ru ()d

is referred to as the spectral measure of u. If u has a sinusoidal component at 0


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

1 j t
e Su ()d.
2

Furthermore, we have

Su ()d = 2Ru (0).


1

Definition 5.1.2: A stationary signal u : R R is sufficiently rich of order n if the

support of the spectral measure Su() of u contains at least n point.


Example 5.1.1: The input
m

u = Ai sin i t ,
i =1

where m n /2, Ai 0 are constants and i k for i k is sufficiently rich of order n.


(Note that this input has a spectral measure with 2m points of support, i.e., i, i for
for i = 1, 2, , m, where m n /2, and is, therefore, sufficiently rich of order n.)

5.2 PE PROPERTY AND SUFFICIENT RICHNESS

Let us now consider the equation


= H ( s )u

(5.2.1)

where H(s) is a proper transfer matrix with stable poles and R . The PE property
n

of is related to the sufficiently richness of u. Before we guaranteed this property, we


need several useful lemmas.
+

Lemma 5.2.1: If the auoconvariance of a function x : R R defined by


n

1 t 0 +T
x() x T (t + )d.
T T t0

Rx (t ) lim

exists and is uniform with respect to t0, then x is PE if and only if Rx(0) is positive
definite.
Proof: () The definition of the autocovariance Rx(0) implies that there exists a T0 >

0 such that
1
1
Rx (0)
2
T0

t 0 + T0
t0

x() x T (t + )d

3
Rx (0), t 0.
2

If Rx(0) is positive definite, there exist 1, 2 > 0 such that 1 I Rx(0) 2 I. Therefore,
1
1
1I
2
T0

t 0 + T0
t0

3
x() x T ()d 2 I ,
2

for all t0 0 and thus x is PE.


() If x is PE, then there exist constants 0, T1 > 0 such that

t + T1
t

x() x T ()d 0T1I

for all t 0. For any T > T1, we can write

t 0 +T
t0

k 1

x() x T ()d =
i =0

t 0 + ( i +1)T1
t 0 + iT1

x() x T ()d +

t 0 + T1
t 0 + kT1

x() x T ()d,

where k is the largest integer that satisfies k T/ T1, i.e.,


kT1 T < (k + 1)T1.

Therefore, we have
1 t +T
kT
x() x T ()d 1 0 I .

t
T
T
For k 2, we have
kT1 (k + 1)T1 T1
T 1
=
1 1
T
T
T
T 2
such that

1 t 0 +T
x() x T ()d 0 I ,

T t0
2
and

1 t0 +T
x() x T ()d 0 I ,

T T t0
2

Rx (0) = lim

which implies that Rx(0) is positive definite.

Lemma 5.2.2: Consider the system

y = H ( s )u,
where H(s) is a strictly proper transfer function matrix of dimension m n with stable
poles and real impulse response h(t). If u is stationary with autocovariance Ru(t), then
y is stationary, with autocovariance
R y (t ) =

h(1 ) Ru (t + 1 2 )h T (2 )d1d2 ,

and spectral distribution


S () = H ( j) Su () H T ( j).

Theorem 5.2.1: Let u : R R be stationary and assume that {H( j1), , H( jn)}

are linearly independent on C for all 1, , n R, where i k for i k. Then


n

the signal vector is PE if and only if u is sufficiently rich of order n.


Proof: According to Lemma 5.2.1, Theorem 5.2.1 can be proved if we establish that

R(0) is positive definite if and only if u is sufficiently rich of order n.


() We will show the results by contradiction. Since u is stationary and R(t) is
3

uniform with respect to t, we obtain


1 T
1
u ()u T ()d =
S ()d,

T T 0
2

R (0) = lim

(5.2.2)

where S() is the spectral distribution of . From Lemma 5.2.2, we have


S () = H ( j) Su () H T ( j).

(5.2.3)

Using the condition that u is sufficiently rich of order n, i.e., u has spectral lines at n
points, we can express Su() as
n

Su () = f u (i )( i ),
i =1

f u (i ) > 0.

(5.2.4)

From (5.2.2) ~ (5.2.4), we have


1
S ()d
2
1
=
H ( j) Su () H T ( j)d

2
n

1
=

H
(
j
)
f u (i )( i ) H T ( j)d

2
i =1

R (0) =

1 n
fu (i ) H ( ji ) H T ( ji ).
2 i =1

Suppose that R(0) is not positive definite, then there exists x R with x 0 such
n

that
x T R (0) x =

1 n
fu (i ) xT H ( ji ) H T ( ji ) x = 0,
2 i =1

Since fu(i) > 0 and each term under the summation is nonnegative, the result above
can be true only if

x T H ( ji ) H T ( ji ) x = 0, i = 1, 2, K , n
or equivalently
x T H ( ji ) = 0, i = 1, 2, K , n
However, we obtain that {H( j1), , H( jn)} are linearly dependent, which contradicts with the condition that {H( j1), , H( jn)} are linearly independent. Thus
R(0) is positive definite.
() We also prove this by contradiction. Assume that R(0) is positive definite

but u is sufficiently rich of order r < n, then we can express R(0) as

R (0) =

1 r
f u (i ) H ( ji ) H T ( ji ),

2 i =1

where fu(i) > 0. Note that the right hand side is the sum of r-dyads, and the rank of
R(0) can be at most r < n, which contradicts with the assumption that R(0) is posi

tive definite.

Remark 5.2.1: If the number of unknown parameters is n, then m n /2 distinct fre-

quencies in u are sufficiently for u to quality as being sufficiently rich of order n.


Note that Theorem 5.2.1 is valid provided H( j1), , H( jn) are linearly independent on C for all 1, , n R, where i k for i k. The vectors may ben

come 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:
y=

b0 ( s 2 + 1)
u = G ( s )u ,
( s + 2) 3

where b0 is the only unknown parameter. Following the procedure of Lecture 2, i.e.,
the lecture on Models for Dynamic Systems, we rewrite the plant in the form of the
linear parametric model
y =

where = b0 is the unknown parameter and


= H ( s )u , H ( s ) =

s2 + 1
( s + 2) 3

According to Theorem 5.2.1, we first need to check the linear independence of H( j1),
, H( jn). For n = 1, this condition becomes H( j) 0, R. It is clear hat for
= 1, H( j) = 0, and, hence, 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 , y are equal to zero
and, hence, carrying no information about the unknown b0. We should note, however,
that u = sin t and any 1, is PE.

Remark 52.2: The above example demonstrates that the condition for the linear in-

dependent of the vectors H( ji), i = 1, , n, on C is sufficient to guarantee is PE


n

when u is sufficiently rich of order n. It also demonstrates that when the plant is par5

tially 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.

You might also like