Black-Box Identification of MIMO Transfer Functions: Asymptotic Properties of Prediction Error Models
Black-Box Identification of MIMO Transfer Functions: Asymptotic Properties of Prediction Error Models
Black-Box Identification of MIMO Transfer Functions: Asymptotic Properties of Prediction Error Models
by
ZHU Yu-Cai
Eindhoven
November 1987
!i 3-~U1~~~k~.t4-¥~j
-7~;J~Y~l~l ~ ~)f~~ ~t~
Abstract iv
1. Introduction
Kronecker products 3
Acknowledgement 22
References 22
Appendix 1 25
Appendix 2 26
Author's address:
ZHU Yu-cai
ZHU Yu-cai
1 INTRODUCTION
y( t) = L G • u(t-k) + vet)
k
k=l
where: yet) is a p-dimensional column output vector at time t; u(t) is an
m-dimensional column input vector at time t;
G is a sequence of p x m matrices; and
k
is assumed to be a stoch- {vet)}
astic stationary process with zero mean values.
When the delay operator q-l is introduced as
q-1u(t) = u(t-1)
where
G(q) = ( 1 .3)
For the disturbance, the most common approach is to assume that vet) is
the output vector of a stable filter driven by a white noise vector
where
where {e(t)} is white noise with covariance matrix R. Both H(q) and
2
H-1(q) are stable. Then the disturbance vet) will be a stationary pro-
cess with spectral density
( 1 .6)
iw
where H(e ) is the p x p transfer function matrix of H(q)
..
\'
L Hk_e
-iwk
(1.7a)
k=O
y(tI8) ( 1.10)
is minimized, where
Expression (1.11) can cover most of the time domain identification tech-
niques in practice. It can be shown that specific methods, e.g. the
least squares or maximum likelihood method or k-step ahead prediction
3
( 1.13)
In section 2 the Kronecker matrix product and some of its basic proper-
ties will be presented. This will prove useful in the derivation of the
result. In section 3 the Box-Jenkins model will be introduced and the
shift property of the polynomial-type models will be emphasized. The
main result is in section 4. In section 5 an application of the theory
is proposed. Section 6 gives conclusions.
2 KRONECKER PRODUCTS
The results here have been adapted from BREWER (1978) and Yuan and Ljung
(1984).
Let
am,B a B a B
m2 ron
matrices, then
(2.3)
l!, (2.5)
col B = (ron x 1)
B
n
where B. is the j-th column of B.
)
If A is a p x m matrix and B is an m x r matrix, we have 1:.he following
With the help of the Kronecker product, we can now presen"o a matrix cal-
culus and some of the properties.
Given A(m x n) and B(p x r), the matrix derivative is defined
5
aA aA
ab " ab'2
aA /!, aA
as =
ab ,
2
(2.7)
aA
aiJ
p'
Given A(dim m x n), F(dim s x t) and B(dim p x r), it can be shown that
(2.8)
In order to show the idea in a concrete way, we will take a special model
structure, the so-called Box-Jenkins model. But the results holds for
all the models which have the shift property.
The Box-Jenkins model is given as
where A(q,8), B(q,8), C(q,8) and D(q,8) are polynomial matrices with
dimension p x P, P x m, P x P and p x p respectively
A(q,8) = I
P
+ A,q-l + ... + A q
n
-n
,
}
-n
B(q,8) = B q-l + + Bnq
(3.2 )
-n
C (q, 8) I + C,q-l + + C q
P n
-n
D(q,8) I + D,q-l + + D q
P n
A = I,
o p
B
0
= 0, C
0
= Ip and 0
0
I (3.3)
P
Remark
When A
o
= I, then [A(q,S), B(q,S) 1 is called a monic ARMA model of
G(q,S). It can easily be shown that any ARMA model can be transferred
into the monic ARMA model provided A is invertible. B 0 means that
o 0
G(q,S) is strictly proper. This assumption is justified by the fact that
most input-output systems are strictly proper. co oo = I
p
means that
H = I as in ('.7b). (3.2) has the order n.
o p
ABC D
n n n n
1
= (d x ') (3.4)
S
n
where
Let
= :pm(q,S) (3.6)
h,,(q,S)
h (q, S)
pp
7
where g, ,(q,9) and h ,(q,9) are the entries of rational matrices G(q,9)
~) i)
and H(q,9) respectively.
It is easy to verify that
-k
q Z(q,9) (3.7)
II
where Z(q,9) = ae-
0 T
T (q,9).q
1
oTT
Here ae-- (q,9) is a s x p(p+m) matrix. ( 3.7) holds because g"
~)
and h, ,
1)
k
are rational functions of q-l and 9 is specially decomposed as in (3.4).
The reader can verify (3.7) by taking a 5150 ARMA example.
Equation (3.7) is the so-called "shift property" of model set (3.1) and
(3.2), which is one of the keys for deriving our result.
.p(t,9) (d x p) (3.8)
AT T T T T
y (tI9)H (q,9) (col G(q,e» (u(t)~ )-y (t)+(colH(q,9» (y(t)®1 )
p p
(3.11)
Using (2.8) we obtain the relation
8
d AT T AT dH T ( q e)
de y (tle)H (q,e) + (I!N (tie» de , =
d T d T
de (colG(q,e» (Uelp) + de (colH(q,e» (Y®[p) (3.12)
where
U(t) ]
and E(t,e) = y(t) - y(tle)
[ E(t,e)
(3.15)
(3.16)
In this section the main result of the paper will be developed. First
some formal assumptions will be given. Then several lemmas will be prov-
ed. Finally, we will end up with Theorem 4.1 which gives the expression
of the covariance matrix of the transfer function estimates.
problem. Since the system is viewed as a black box, the internal para-
9
n n(N) (4.1)
n(N) + ~ as N + ~ (4.2)
When the model order n increases, the model may lose "parameter identifi-
ability", but it will retain "system identifiability" under weak condi-
tions on the experiment design. See Gustavsson et al. (1977) for a dis-
cussion of this point. To deal with this problem, we introduce a regu-
larization procedure in the following way. Let
-AT A
6*(n) = arg min Ee (t,6)e(t,6) (4.3)
~D
n
where
-~T A 1 N AT •
Ee (t,6)e(t,6) = lim L E e (t,6)e(t,6)
N+~ N t=l
(If the minimum is not unique, let 6*(n) denote any of the parameter
vectors leading to such a minimum).
Here n emphasises that the minimum is carried out over n-th order mod-
els.
1 1 N
= 2 [N L
t=l
(4.5)
"
Here 0 is a regularization parameter, helping us to select a unique mini-
mizing element in (4.4) in cases where 6 = 0 leads to non-unique minima.
10
The procedure here is a technical way of dealing with the unique esti-
mate
• iw iw •
~(e ) = G(e ,eN) = (4.6)
Further assumptions
Assume that the true system can be described by
where {eCt)} is a white noise vector with covariance matrix R and bounded
fourth moments. Moreover, Go and Ho are stable filters. The output
noise spectrum is then
( 4.8)
Assume the predictor filters Hrl(q,e) and Hr1G(q,e) in (1.10) along with
their first-, second-, and third-order derivatives with rep sect to e are
uniformly stable filters in e E 0 for each given n. Let
n
• iw
T(e ,n,o)
(4.9)
Assume that
• T •
lim n 2 E[€(t,e*(n»-e(t)] [€(t,e*(n»-e(t)] o (4.10)
n-
iw iw
which implies that T*(e ) tends to T (e ) as n tends. to infinity, i.e.
n 0
the transfer functions estimates are consistent.
In the same way that Z(q,e) defined in (3.7), we denote Z (q) as
o
11
Z (q) .q (4.11)
o a6
1
and
iw
e (4-12)
a6
1
Assume that
(4.13)
N
r (T)
u
= -E T
u(t)u (t-T) = lim N I (4.14a)
N+" t=1
N
T
r
eu
(T) =E e(t) u (t-T) lim 1 I
E[e(t) U(t-T)l (4.14c)
N- N t=1
r
ue
(T) o for T < 0
(4.15)
r
eu
(T) = 0 for T > 0
" -iTw
~
u
(w) = I r (T) e
u
(4.16)
,[=-«1
lim -
tIN
1 N
I
t=1
d 'T
E[ d6 € (t,S(n»€ (t,S(n»
'
I
S=S*
1= 0 (n fixed) (4.17)
a;T (t! B)
X(t-l,B) = (4.18)
aB
1
where
T
Z(q,B) = aT (q,B) .q
(4.19)
aB
1
Then from (3.16) we have
1/I(t,e) =
X(t-l
~(t-2,e)
'B)J (4.20)
.
x(t-n,e)
Denote the d x d matrix
T 6
E 1/I(t,B) 1/1 (t,e) = M (e) (4.21)
n
1 T
~ (lIl,e) lim - W (w) M (e) Wn(-IIl)
X n n n
13
(4.24)
Now we have the following result:
Lemma 4.1
Assume that (4.14)-(4.17) hold. Suppose also that
n(N)
L nr(T)n+Q
u
as N .. m, n(N) + m (4.26)
T=-n( N)
(s x s) as n(N) + ~ (4.27)
and
1
lim
n(N)
n(N)-
(4.28)
Proof: The matrix M (e) + 6Id is the block Toeplitz covariance matrix of
----- n
the s x p dimensional process
x(t,e) + n; w(t)
T
RW = EW(t) w (t) = Is. The spectrum of this process is given by
(~ (w) + 51). The result follows from the corollary to Yuan and Ljung
X s
(1984) Lemma 4.3. (Take Wd(W) = Wn(w), Rd = Mn(6) + 5I ).
d
Similarly we have
Corollary 4. 1
Let us now consider the parameter estimate (4.4). First, from (4.3) and
(4.4) we have as in Ljung (1978)
(4.31)
n
where ~N belongs to a neighbourhood of 6*(n) and from (4.30)
Hence
t:,
E(t,e*(n}} = e(t} + r(t,e*(n}} (4.34)
where
1 T
lim W (w) Q(n} W (-w)
n(N} n n
= z(eiW,e*(n}}(~~(W} [(HT(eiW,6*}}-IR(H-l(e-iW,6*(n}}]zT(e-iW,6*(n})
(4.38)
The proof is given in Appendix 2.
Lemma 4.4
under the conditions of Lemma 4.3 we have
(4.39)
where
T
lim W (w}) R(n,5} W (-w)
n(N} n n
16
(4.41)
Applying Lemma 4.2, Lemma 4.1 and Corollary 4.1 successively, we obtain
then (4.40).
Now we consider
By Taylor's expansion
ioo ioo
A
TN(e ,n,o) - T*(e ,n) = T(e ioo ,6 N(n,5»
A
- T(e
ioo
,6*(n»
(4.43)
Thus (4.39) implies that the variable in (4.42) will have an asymptotic
normal distribution with covariance matrix
- d ioo ~ TT( -ioo 6)
P(oo,n,5) = ~ T(e ,6) R(n,o) d6 e , (4.44)
d6
Now consider the p(p+m) x d matrix
T(e ioo ,6) = [ 3 -or T(e ioo ,6) T(e ioo ,6) ]
36
1
Using the shift property (3.7) we have
= e -ikoo z T( e ioo , 6)
17
-d- T (iw
e , 6) (4.45)
T
d6
-
P(w,n,o) = ZT (e ioo ,6) Wn (-w) R(n,o) WnT (+w) Z(e -ioo ,6) (4.46)
f\
According to Lemma 4.4 and the fact that o(t,6*(n» + e(t)
(4.47)
where
(4.48)
to [ Z (e
-iw T iw
)S(-w)Z (e ) + 01
1-IZ (e
-iw
), suppressing arguments and
o 0 S 0
indices
18
(4.49)
• ioo
Theorem 4.1: Consider the estimate TN(e ,n,6) under the assumptions
Then
• iw iw
IN[TN(e ,n,6) - T~(e )
1e As N(0,p(w,n,6» (4.50)
as N + - for fixed n,o
where
lim lim
0=0 n+CD
~ (00)
v
~
ue
(w) = ~
eu
(w) = 0
and
... iw
cov[col GN(e ,n)] ~ (4.52)
;.. iw
cov[col ~(e ,n)] ~ (4.53)
In the development of Theorem 4.1, we have used the shift property of the
model structure, and the prediction error criterion. Therefore, it
should be clear that the result holds for all the polynomial-type models
which have the shift property. The Box-Jenkins model can cover many
special parametrizations of this class, but not all of them. (4.52) is
consistent with the result of Yuan and Ljung (1984), taking note of dif-
ferent definitions of r (T).
u
The author would like to point out that the right-hand side of (3.22) of
Ljung (1985) should be complex conjugated (or transposed).
corollary 4.2
Consider the same situation as in Theorem 4.1, but assume that H(q,6) is
fixed, and independent of 6.
20
Then
• ioo .
/N[col ~(e ,n,6) - col G~(e~oo)l ~ As N(O,P(oo,n,6» (4.55)
1
lim lim P(w,n,cS) = (4.56)
n
6+0 n+ClO
H(q,8) = I (4.57)
Because the expressions of the result are remarkably simple, they are
very useful in applications. Ljung and Yuan have used thE' (other version
of the) result for input design and order selection. HerE', another ap-
plication of the result will be proposed.
( 5. 1 )
iw
Then, from Theorem 4.1 and (4.52) we know that 6 g (e ) follows, asymp-
ij
totically, the normal distribution and
(5.2)
where [~-l(w)l .. is the (j,j) entry of the matrix ~-l(w), and ~ (w) is
U JJ U vi
the spectrum of Vi(t), and equals the (i,i) entry of the matrix ~v(w).
(5.3)
with
We can compute UB(w) by (5.3), using the estimates of ~ (w) and ~ (w) and
u v
this quantity can be used for robust controller design of the feedback
system.
Details on the estimation of the upper bound can be found in Zhu (1987a,
1987b).
22
The derivation of the upper bound of the identification errors from this
theory completes the contribution of identification to robust control.
ACKNOWLEDGEMENT
REFERENCES
~, T. (1987)
IDENTIFICATION OF AN INDUSTRIAL PROCESS: A MARKOV PARAMETER APPROACH
Ph.D.-thesis, Eindhoven University of Technology
23
Ljung, L. (1985)
ASYMPTOTIC VARIANCE EXPRESSIONS FOR IDENTIFIED BLACK-BOX TRANSFER
FUNCTION MODELS
IEEE 'rrans. Autom. Control, Vol. AC-30, p. 834~844.
Ljung, L. (1978)
CONVERGENCE ANALYSIS OF PARAMETRIC IDENTIFICATION METHODS
IEEE Trans. Autom. Control, vol. AC-23, pp. 770-738.
Vidyasagar, M. (1985)
CONTROL SYSTEM SYNTHESIS: A factorization approach
Cambridge, Mass.: MIT Press.,
MIT Press series in signal processing, optimization and control, Vol. 7.
~, Y.C. (1987a)
ON A BOUND OF THE MODELLING ERRORS OF BLACK-BOX TRANSFER FUNCTION
ESTIMATES
EUT Report 87-E-173, Faculty of Electrical Engineering, Eindhoven Univer-
sity of Technology, The Netherlands.
24
~, Y.C. (1987b)
ON THE BOUNDS OF THE MODELLING ERRORS OF BLACK-BOX MIMO TRANSFER FUNCTION
ESTIMATES
EUT Report 87-E-183, Faculty of Electrical Engineering, Eindhoven Univer-
sity of Technology, The Netherlands.
25
By standard arguments and a law of large numbers (see e.g. Ljung, 1978)
we have
Remark
The reason why ~(t,9) and ~'(t,9) are independent of e(t) is due to the
fact that the "prediction error" criterion is used: y(tle) is dependent
only on the previous y and u, i.e. y(tle) is only dependent on the
previous e, and e(t) is an independent variable, therefore y(tl9) and
e(t) are independent, so that ~(t,9), ~'(t,9) are also independent of
e( t) •
It remains to be shown that the operator norm of the last term of (Al.3)
"T
Iii Y 8 (t 8) .r(t,e*(n»
8
I (
k j
Hence
d
"T
L Iii Y6 6 (tI6) r(t,e*(n) I ( s.C.C n (A1.4)
k=1 k j
NOw, the operator norm of any symmetric matrix is bounded by its absolute
where
Q(n)
1 N N
= lim N L L E[ljI(t,6*(n»(e(t)+r(t,8*(n».
N- t=1 s=1
27
N
1
lim I E[~(t,6*(n) e(t).eT(t) ~T(t,6*(n»]
N
N+~ t=l
N N
2
+ lim I I E ~(t,6*(n»r(t,6*(n» eT(s)~T(s,6*(n»
N
N- t=l s=l
N N
T
+ lim I I E ~(t,6*(n».r(t,6*(n»rT(s,6.(n» ~ (s,6*(n»
N
N- t=l s=l
(A2.3)
The second and the third sums are obtained as filtered white noise and
filtered deterministic input. According to Ljung (1985) the values of
the entries of these limits are bounded by
T
G/[Er (t,6*(n» r(t,6*(n»T = C.C
n
In
(A2.4)
Hence
T
lim W (w) M (6*(n» W (-w) =
n(N) n e n
n-
z(eiW,6*(n»){~c(W) [HT(eiW,6*(n»)-IRH-l(e-iw,9*(n»)]lzT(e-iW,9*(n»)
(A2.6)
and the lemma is proved.
Eindhoven Universitv of Technolo Resenrch Re arts 1551'1 1J167-970J3
acultv of Electrical E:.nuineerinq Coden: TEUEDE
ti -'01 Yibin
i-I:.l!.l
DAS1~: A tool f;JC deccmposlt.im: a:1d analysls ot: ~equentiu.l l!'.i1chines.
EL'T R-=port 87-E-170. 19R7. ISBN 90-6144-170-6
(184) Kadete, H.
ENHANCEMENT OF HEAT TRANSFER BY CORONA WIND.
EUT R~port 87-E-184. 1987. ISBN 90-6144-184-6
(185) ~, P.A.M. and A.M,.]. ~, LV. ~, J. [Jijk
THE IMPACT OF TELECOMMU:-.I!CATION ON RURAL AREAS IN DEVELOPING COUNTRIES.
EUT Repurt: 8";-E-185. 19H7. 1:SBN YO-6144-185-4
\186) ~ Yanhong
THE INfLUEN('~ OF COt,TACT SURFACE ~ICROSTRlJCTURE UN VACUUM ARC STABILITY AND ARC VOLTAGE.
El'T Repol-t S'-j,;-lB6. 1987. ISBN ':JO-61<l4-1H6-}