Multivariable Process

Identification for Robust

Control
By
Y.C. Zhu
A.C.P.M. Backx
P. Eykhoff
EUT Report 91-E-249
ISBN 90-6144-249-4
January 1991
f

'[,
1
I
ISSN 0167 9708
Eindhoven University of Technology Research Reports
EINDHOVEN UNIVERSITY OF TECHNOLOGY
Faculty of Electrical Engineenng
Eindhoven The Netherlands
Coden: TEUEDE
MUL TI VARIABLE PROCESS IDENT! FICATION FOR ROBUST CONTROL
by
V.C. Zhu
A.C.P.M. Backx
P. Eykhoff
EUT Report 91-E-249
ISBN 90-6144-249-4
Eindhoven
January 1991
CIP-GEGEVENS KONINKLIJKE BIBLIOTHEEK, DEN HAAG
Zhu, V.C.
Multivariable process identification for robust control/by
V.C. Zhu, A.C.P.M. Backx and P. Eykhoff. - Eindhoven: Eindhoven
University of Technology, Faculty of Electrical Engineering. -
(EUT report, ISSN 0167-9708; 91-E-249)
Met lit. opg., reg.
ISBN 90-6144-249-4
SISO 656.2 UDC 519.71.001.3 NUGI 832
Trefw.: systeemidentificatie; procesregeling.
MULTIVARIABLE PROCESS IDENTIFICATION FOR ROBUST CONTROL
*
**
Contents
Abstract

y.c. ZHU , A.C.P.M. BACKX and P. EYKHOFF
IPooS Group, Datex Industry
Tinnegieterstraat 12, NL-5232 BM 's-Hertogenbosch
The Netherlands
tel. 31-73-412225, fax: 31-73-426055
Dept. EE, Eindhoven University of Technology
P.O. Box 513, NL-5600 MB Eindhoven
The Netherlands
Tel. 31-40-473300
1
1 Introduction 1
2 Black-Box Models Identified by Prediction Error Methods 2
3 Nominal Model and Error Bounds 7
3.1 An Algorithm for Estimating a Nominal Model with an Error Bound 8
3.2 Performing Input Design before any Identification and Control Test 11
3.3 On the Use of Upper Bound Matrix 13
4 Model Reduction of Identified High Order ARX Model 14
5 Model Structure Determination 20
6 A Simulation Study 22
7 Conclusions 24
References 24
Figures
28
MULTIVARIABLE PROCESS IDENTIFICATION FOR ROBUST CONTROL
Y.C. ZHU, A.C.P.M. BACKX and P. EYKHOFF
Abstract-In this work, multi-input multi-output (MIMO) process identification
is studied, where the purpose of identification is control system design. An
identification procedure is presented by which one can . estimate not only a
nominal parametric process model, but also an upper bound of the model
uncertainty (modelling errors) in the frequency domain. The basic steps of
this method consist of high order model estimation and subsequent model
reduction. One advantage is that, in this framework, fundamental problems such
as input design and model structure selection can easily be solved. Another
advantage of the method is numerical simplicity and reliability. The
identified nominal model and the error bound can readily be used for robust
control system analysis and design. A simulation example is given to
illustrate the method.
I Introduction
Because of our limited knowledge about reality, mathematical models can never
give an exact description of the system behavior under study; hence model
uncertainties or modelling errors always exist. In the previous decade, robust
control theory has been proposed and developed, cf. Zames (1981), Doyle
(1984), Vidyasagar (1985) and Morari and Zaftriou (1989). The advantage of
robust control is its capability to cope with modelling errors in the analysis
and design of control systems. In order to apply robust control theory, one
needs not only a nominal process model, but also a suitable description of the
modelling errors which are typically in the form of some bounds of either
parameter variations or transfer function variations.
Although the topic of linear system identification has been studied
extensively in the previous two decades, there are no straightforward methods
which can derive error bounds of identified models in the frequency domain.
Only very recently some researchers started to study this problem. Cloud and
Kouvaritakis (1986) have proposed an error bound for MIMO FIR models when
output disturbances are white noises; Kosut (1986) derived an error bound for
1
SISO prediction error models; Goodwin and Salgado (1989) proposed a method for
quantifying uncertainty in the estimation of simplified SISO models; Helmicki
et al. (1989) studied the problem in a deterministic setting for SISO
processes; Correa (1989) presented some discussion about the problem in the
framework of prediction error method; Van den Boom et al. (1991) proposed a
method of SISO process identification in the frequency domain, where the
H -norm is used as error criterion, in a deterministic 'setting. The work on
00
this topic is as yet so limited that this shGft list almost completes the
story. It seems that identification is not ready yet for its use in robust
control, at least r.;){ ior IYlIMO processes.
In this W\ir.: we will present an identification method for linear time-
invariant MIMO processes, in which a nominal parametrical model of the process
can be estimated together with an upper bound matrix of the modelling errors
in the frequency domain. The idea of this method was first proposed in Zhu
(1987a, 1989a) for SISO processes, based on a newly developed asymptotic
theory of black-box model identification in Ljung (1985) and Ljung and Yuan
(1985). Because the asymptotic theory can be extended to the MIMO case and the
result has a nice structure (Yuan and Ljung, 1984, and Zhu 1987b,1989b), the
method can be generalized for MIMO processes in a straightforward way (Zhu,
1987c, 1990). The method has been applied to analyze the robust stability of a
2-input 2-output glass tube production process (Zhu, 1990, Zhu and Bach,
1991).
First we introduce an asymptotic theory of the MIMO prediction error
method (Section 2); then, based on this theory, the identification procedure
is presented in Section 3, where identification input design is also
discussed. In Section 4 we propose a method for a frequency weighted
Frobenius-norm model reduction, which is needed in deriving an optimal reduced
order model. The problem of model order determination and structure selection
will be treated in Section 5. Section 6 presents a simulation study. Section 7
gives the conclusions of this work.
2 Black-Box Models Identified by h _ .'ktion Error Methods
In the previous decade, many sophisticated time domain parametric
identification methods have been developed. Most of these techniques can be
classified in the set of prediction error methods. Recenily, Ljung and Yuan
developed an asymptotic theory on the frequency domain properties of
prediction error models; d. Ljung and Yuan (1985), Ljung (1985), and Yuan and
Ljung (1984). The MIMO version of the theory (Zhu, 1989b) will briefly be
2
presented.
Consider a discrete time process with m inputs and p outputs. A general
linear time-invariant model for the relationship between inputs and outputs
can be written
00
y(1) = I. Gku(t-k) + v(t) (2.1)
k=1
where: y(t) is the p-dimensional column output vector at time 1; u(t) is the
m-dimensional column input vector at time 1; (G k
J
is the impulse response of
the process, which is a sequence of pxm matrices; and ( v(t) J is a
p-dimensional stochastic stationary process with zero mean values.
When the unit time delay operator q.1 is introduced:
q'IU(t) = u(t-I)
the model (2.1) can be written as
where
y(t) = G(q)u(t) + v(t)
00 k
G(q) = I. Gk'q'
k=1
is called the transfer operator of the process model.
The transfer function matrix of the model is defined as
= Gk'e,irok
k=1
-1t $ ro $ 1t
For the output disturbances, a common approach is to assume that they are
mutually independent and generated as filtered white noises:
(2.2)
(2.3)
(2.4)
v(t) = H(q)e(t) (2.5)
where
H k is a sequence of pxp matrices with
Ho = I p (pxp identity matrix)
and (e(t) J is a p-dimensional white noise vector, mutually independent, with
covariance matrix R = diag[R ... R ], where R. is the variance of (e.(t) J.
I P 1 1
Both H(q) and lfl(q) are stable. Then (v(t)} will be a stationary process with
spectral density
00
= I. [E{v(t)v
T
(t)J]e'IClY1: =
t=-oo
3
(2.6)
where E means expectation. T means transpose. H(iCf) is the transfer function
matrix of H(q):
H(iCf) = Hk"e-
irok
k=o
-It OJ It
which is diagonal.
(2_7)
The identification problem is to estimate an approximate model from
observed input-output data. Denote the data sequence by z!'I:
z!'I := y(I). u(I) ...... yeN). u(N) (2.8)
where N is the .. :If samples of the data sequence.
If we ha', "parametrized model:
y(t) = G(q.B)u(t) + H(q.B)e(t) (2.9)
where B is a d-dimensional parameter vector. we can determine the one step
ahead prediction:
;(tiB) = [Ip - l1\q.B)]y(t) + 11
1
(q.B)G(q.B)u(t) (2.10)
and compute the prediction error
e(t.B) = yet) - ;(tiB) = 11
1
(q.B)[y(t) - G(q.B)u(t)] (2.11 )
Then a common way to determine the parameters is to minimize the squared sum
of the prediction errors
(2_12)
where
N
= k L eT(t.B)e(t.B)
t=l
and D nC Rd is the parameter space. Here the subscript N is to emphasize that N
data samples are used for the estimation; the dimension of the parameter space
d depends on the numbers of process inputs and outputs. the model order and
the model parametrization. The model degree n will be defined in the
following.
The general MIMO black-box moo'" is defined as
A(q)y(t) = F-1(q)B(q)u(t) + D-1(q)C(q)e(t)
where A(q). B(q). C(q). D(q) and F(q) are polynomial matrices with dimension
pxp. pxm. PXP. pxp and pxp respectively:
4
(2.13)
A (q) = I + A q-I + ... + A q-n
pin
B(q) = Blq-I + ... + Bnq-n
C(q) = I + C q-I + ... + C q-n }
pin
D(q) = I + D q-I + ... + D q-n
pin
F(q) = 1 + F q-I + ... + F q-n
pin
The assumption that output disturbances are mutually independent implies that
A(q), C(q) and D(q) are diagonal matrices.
(2.14)
In order to have a unique representation of a given MIMO process, some
structural conditions should be imposed on the general model (2.14). A natural
way to do this is by letting F(q) be diagonal matrices. This arrangement
decomposes the MIMO process into p MISO (multi-input single--{)utput)
sub-processes where the i-th subprocess is given as
A.(q)y(t) = rl(q)B(q)u(t) + D:I(q)C(q)e(t)
11 II I 11 11 1
(2.15)
The parameter vector for model (2.14) is
9 = col [AI BI C
I
01 FI A2 B2 C
2
02 F
2
An Bn C
n
On Fn1 (2.16)
The degree of this model is defined as the degrees of the polynomials which
are equal to n. Note that, in general, this degree is not the same as the
McMillan degree which is defined as the dimension of the minimal state space
realization of the model. Generally, the McMillan degree of the model equals
pn.
Various model parametrizations used in practice are special cases of this
general model:
(I) Box-Jenkins model, A(q) = I;
(2) ARMAX model, F(q) = D(q) = I;
(3) OE (output error) model, A(q) = D(q) = C(q) = I;
(4) ARX (equation error) model, F(q) = D(q) = C(q) = I; and
(5) finite impulse response (FIR) or Markov parameter model,
A(q) = F(q) = D(q) = C(q) = I
Hence the problem statement (2.9) - (2.16) can cover most of the time domain
identification techniques in practice. Specific methods can be obtained by
taking a specific model parametrization.
After the parameter estimation, the transfer function estimates are
5
denoted as:
= }
=
Asymptotic Properties of the Transfer Function Estimates
When the identified model is used in controller design, we are more concerned
about the quality of the transfer function estimates than about the accuracy
of parameters. The eO';mation of a transfer function matrix is basically a
non-parametric r" . :_ul. Since the process is viewed as a black-box, the
internal parametrization via 6 is merely a vehicle to arrive at this estimate.
Then, it is natural to let the model order n depend on the number of observed
data samples, n = n(N). Typically, in order to have a model set that is large
enough to contain the true process dynamics, or to give a good approximation
of the true dynamics, we will allow the order n to increase when the number of
data samples N increases, but n should be small compared to N. This can be
formally expressed by the following assumptions
n(N) .. 00 (2.17)
and
as N .. 00 (2.18)
Assume that the true process is described by
y(t) = GO(q)u(t) + Jt'(q)e(t) (2.19)
where GO(q) and Jt'(q) are the true transfer operators of the process and the
disturbance. They are both assumed to be stable filters. Denote the true
transfer function matrices as and Then, for an open loop
experiment, under some suitable conditions of the inputs, the following
results hold:
w.p.! as N .. 00
6
(2.20)
(2.21)
(2.22)
where colO denotes the vector operator on a matrix, EO denotes mean value,
q, (0) and q, (0) are the spectrum matrices of the inputs and disturbances
u v
respectively, T means inverse and transpose, denotes the Kronecker product.
Result (2.21) and (2.22) are the MIMO extension of the result of Ljung (1985);
the proof can be found in Zhu (1989b). Similar results hold for closed loop
experiments, which are not shown here.
From (2.20)(2.22) we can say heuristically that the errors of the
transfer function estimates. have an asymptotically normal distribution; the
covariance matrix of G at a given frequency is proportional to the
(generalized) noise-to-signal-ratio at that frequency and inversely
proportional to the number of data N; the covariance increases with the model
degree n, not with the number of parameters d. We also find that different
model parametrizations will have the same asymptotic properties.
Because the expression of the covariance matrix is remarkably simple, this
result is very useful in applications, as will be shown in the following
sections.
3 Nominal Model and Error Bounds
In this section, we will propose an identification method which will deliver
both a nominal model and an upper bound matrix of the modelling errors, using
the asymptotic theory of Section 2. Also the input design problem will be
treated.
Because the theory in the previous section is asymptotic both in the
number of data samples N and the model order n, the results will be valid if
the number of data samples is large and the model order is high. But in
practice low order models are used for simulation, prediction and controller
design. So our approach is to start with a high order model estimation and
then to apply model reduction techniques to arrive at low order models. This
approach has its practical background. For industrial process identification
and control, we may have a large amount of data; and we have to use high order
models to initially fit the complex dynamics of the process, as we in general
do not have detailed knowledge of process dynamics. We can avoid numerical
problems by combining high order estimation and model reduction, cf. Backx
(1987) and Wahlberg (1989). By this approach one also can simplify the step of
model order determination and structure selection, which is a hard job for
MIMO processes. Input design for high order models is an easy task; cf. Yuan
and Ljung (1984).
7
3.1 An Algorithm for Estimating a Nominal Model with an Error Bound
The asymptotic theory implies that the model parametrization is not a crucial
issue for high order models. In this case the ARX model is the most simple
parametrization which can supply both the process model and disturbance model.
It is well known that the least squares estimation of ARX model parameters is
a linear regression problem, which can be solved reliably with simple
numerical techniques. Assume that an open loop experiment is performed; also
assume that the inputs are mutually independent. The following steps are
proposed:
Step 1 EstiL.iLe the parameters of the ARX model with a high order, e.g.
n = 20 - 30. The matrix A(q) has a diagonal form:
A(q) = diag [ All (q) ... A pp (q)].
Then we have
and the spectrum estimates of the disturbances are
where R is the estimated covariance of the equation error residuals.
(3.1)
(3.2)
Denote and as the (ij) element of and
1 J IJ N'
respectively. Then according to (2.20) - (2.22), and remembering that the
inputs are mutually independent, we have asymptotically
- E ASN(O,
J I
(3.3)
and we can define the 30' bound for - as
IJ IJ
- 3/
) I
w.p.99.99% (3.4)
where <I> (00) is the spectrum of u.(t), and <l>y(OO) is the spectrum of y.(t).
u. J. 1
J I
Step 2 , Perform a ,model re<!uction on GZ<q) to obtain a low process
model d(q), and on HZ<q) = Aiv'(q) to obtain a disturbance model H (q), where
I means the low order. Model reduction by truncated balanced realization and
by optimal Hankel norm approximation are two well known methods (Glover,
1984). When the high order model is obtained by identification, a model
8
reduction method which can take the properties of the high order model into
account will be preferred; cf., Wahlberg (1989). In the next section a method
of frequency weighted Frobenius norm MIMO model reduction will be proposed
which works on polynomial matrices. The order and structure selection for the
low order model will be treated in Section 5.
Define the modelling errors as
\fi,j
where ( ; l . ( i ~ is the (i,j) element of O! 1 ~ . Then
II N'
The first term in (3.6) we call the bias part of the modelling errors and the
second term the variance part of the modelling errors. Now
From (3.4) it follows that,
(3.5)
(3.6)
(3.7)
w.p. 2! 99.99% (3.8)
This is the basis for estimating the upper bound matrix.
Step 3 Define the upper bound matrix
il(ro) = {il(ro) J
II
(3.9)
such that
I . ( i ~ I ::; Ll(ro)
lJ IJ
\fi,j \fro (3.10)
Finally we estimate Ll(ro) by
(3.11)
where Cf) u(ro) is calculated as the sample spectrum:
I
9
and <l>v(ro) is given in (3.2).
i
N
~ u.(t)u.(t-t)
t = 1 ) )
ion
e

Remarks-If a good model structure is selected and a proper model reduction
technique is used in Step 2, it is possible to obtain a low order model which
is more accurate than the high order modeL In this case, the 3cr bounds will
be the bounds for the errors of the low order model as welL Then the formula
for the upper bol.>,,,1 hp.comes
(3.llb)
This bound tighter than that is given in (3.11).
Notice that the probability of 3cr bound of a normal complex variable IS
greater than that of a normal real variable.
In formula (3.11) we note that the bias part of the modelling errors is
caused by model reduction; the variance part can be affected by the input
spectrum, the model order and the number of data samples which are design
variables chosen by the user. If at some frequencies the modelling errors are
too large for a specific application of the model, we know from (3.11) that we
can reduce the modelling errors at those frequencies by: (I) performing
another model reduction, (2) modifying the input spectrum (input design), (3)
using more data (increase experiment time) and (4) reducing the model degree
n. The last choice should be made with caution, because too Iowan order will
introduce more bias.
This type of algorithm can be applied also to closed loop experiments. In
this case, use the expression of the asymptotic covariance matrix for closed
loop experiments; see Zhu (1989b).
This type of algorithm can also be used for a FIR model, see Zhu (1990).
Here we prefer an ARX model to FIR mo..:_. because: (I) for modelling a given
process the degree (see the definition in Section 2) and number of parameters
of an ARX model are usually less than those of a FIR model; (2) a FIR model
does not supply a disturbance modeL
10
3.2 Performing Input Design before any Identification and Control Test
Input design is important in detennining the quality of an identified model.
Conventional methods of optimal input design use the covariance of the model
parameters as the measure of model quality (see Mehra, 1974). This methodology
can not address directly the intended use of the model; and the calculation of
the optimal input needs the true model which is not known.
The problem of input design is very much related to the intended use of
the model. Now let us assume that the model is used for controller design in
an internal model control (see Morari and Zafrriou, 1989) which is a very
suitable scheme for controlling e.g. an indusrrial process with time delays.
In this control scheme, the identified process model is placed in parallel
with the process, and the differences between the process outputs and
simulated model outputs are fedback to the controller. Therefore a model which
can optimally simulate the underlying process for certain given inputs will be
most suitable for a internal model control scheme.
Denote u'(t) as the control input to the process (the simulation input to
the model) with spectrum (which, in general, is different from the
spectrum of the input used in the identification experiment). Yuan and Ljung
(1985) have shown that in the SISO case, the optimal input which minimizes the
mean square of the simulation error for the high order model is given by
<l>:P(ro) = (3.12)
where is a constant which is adjusted with respect to the constraint of
input amplitude or output variance. This result agrees with our engineering
intuition: it says that the process should be excited more at those
frequencies where the simulation input has more power and where disturbances
make more trouble. Notice that this spectrum is not related to the process
transfer function.
It can be shown (see Lenssen, 1988) that for MIMO processes, under the
assumption that the input signals are mutually independent, the optimal input
spectrum for the i-th input is given by
p
<l>v(ro)
i j = 1 j
(3.13)
Similarly, the optimal input spectrum for other uses, such as prediction, pole
placement control design, can be derived; see Yuan and Ljung (1985), Gevers
and Ljung (1986).
Initially one might think that fonnula (3.12) or (3.13) is useless
11
because one does not know (us(t) before identifying the process, designing
the controller and testing the closed loop control system. Examine a feedback
control system in which the main goal of the control is disturbance reduction.
Then the control input is the filtered disturbance
u'(t) = T(q)v(t) (3.14)
where T(q) is the unknown transfer operator from the disturbance to the input.
The optimal input is (for SISO process)
= (3.15)
Because T(q) is unknown, we can simply let T(q) = 1. Then we have a first
approximation to :he optimal input:
If T( is flat over the
very good approximation.
<f> u(O) = J.1<f> v(O)
bandwidth of the disturbance
(3.16)
{v(t), this will be a
Similarly, for a MIMO process, we can have an approximation to the i-th
input
<f> (0) = J.1 ! <f> (0) (3.17)
u. . V
I J = 1 J
If the number of inputs and the number of outputs of the process are equal and
the decoupling technique is used in controller design, will be nearly
diagonal. In this case, the approximation of the i-th optimal input signal is
<f> (0) = J.1
u .

(3.18)
Hence, the nearly optimal input spectrum can be derived without knowing the
process model and the controller; the output disturbance (v(t) can be
measured from the uncontrolled process when keeping the input constant.
If some readers are still not convinced yet, let us look at the
consequence of this input design method for model errors in the SISO case.
Substituting (3.16) into (3.4), we obtain a constant 30 bound of the errors of
the high order model as
I - I $ 3/ N J.1 w.p. 99.99%
This means that we will minimize, in a stochastic sense, the L -norm of the
00
errors of the high order model by using (3.16). This is not at all a bad
choice.
When taking into account other control requirements, such as setpoint
12
tracking and robust stability, the following modification of (3.16) is
recommended:
(3.19)
Here <1>(00) is a spectrum which adds some energy to the input signal at
frequency bands which are important for tracking and robust stability; ILl and
IL are adjusted to meet the constraint on input amplitude or output variance,
2
and to weight the relative importance of the two terms. The spectrum can be
realized by filtering a white noise signal or a PRBNS (pseudo random binary
noise sequence) signal, with a filter which has the property that the square
of the filter transfer function approximates the desired input spectrum given
in (3.19). The modification for the MIMO case is the same.
Thus we have a input design method which is not only very simple, but
also most economical: the nearly optimal input signal can be determined by
only measuring the output of the existing uncontrolled process. This makes the
method very applicable.
3.3 On the Use of the Upper Bound Matrix
When a nominal model and an upper bound matrix are obtained by the methods in
the previous subsection, the real process which is not completely
known, can be described in the following class:
= +
I I S K .(00) Vi,j VOl }
1 J 1 J
(3.20)
The upper bound matrix K(Ol) = (K..(Ol) J gives a structured description of the
1J
model uncertainty, because it preserves the multivariable nature of the
problem: the amplitudes of the errors of each transfer function estimate are
upper bounded by the elements of K(Ol) in the frequency domain, the only
missing information are the phase angles of the errors
It is known from linear algebra that
(3.21 )
This means that the maximum singular value of the upper bound matrix is an
upper bound of unstructured model uncertainty.
Robust Stability Analysis Kouvaritakis and Latchman (1985) have developed
a result which is most suitable for the stability test of the process given in
(3.20). Denote as the transfer function matrix of the feedback
13
controller, then the process defined in class (3.20) will be stabilized by the
controller if and only if
"VOl E (0, Ttl
where cr(-) denotes the maximum singular value, and is omitted to save
space. Here L and R are diagonal positive nonsingular scaling matrices which
need to be determined frequency by frequency by some optimization procedure.
Determining the Weighting Matrix for Il-Synthesis In Doyle's Il-synthesis
(Doyle, 1984), the robust stability and robust performance can be analyzed and
optimized simultaneously. In this framework, the model error matrix is
rearranged in the tollowing way:

W
1 2
where is a diagonal matrix:
=
"', ...
Notice that # . It can be shown that under this arrangement
IJ IJ
W, = [I In, W, = r K'K ], = b"]
m pl
Now we are ready to use the Il-synthesis.
Robust fault detection A robust fault detection method was proposes by
Emami-Naeini et al. (1988) where they use an upper bound of the unstructured
model uncertainty in the analysis and design of fault detection systems. The
upper bound given in (3.21) can serve their purpose.
4 Model Reduction of the Identified High Order ARX Model
In the procedure proposed in Section 3.1, the low order model is obtained by
model reduction. The existing popuiar methods, such as balanced model
reduction and model reduction by Hankel norm approximation, are numerically
simple, but they do not minimize a criterion .vhich is physically sensible. The
properties of these methods in the frequency domain are not clear, except some
upper bounds. In this section, a method of model reduction is proposed in
which the asymptotic properties of the high order model will be taken into
account
From (2.20)----(2.22) we know that for each high order transfer function
14
estimate:
(4.1)
and for each disturbance model estimate:
(4.2)
We can view the high order ARX model as the noisy observations of the true
process (Wahlberg, 1989). Because we know the asymptotic distribution of the
high order estimates, it is natural to apply the maximum likelihood method to
these observations to find the low order model. Denote d(q) and HI q) as the
reduced order process model and disturbance model respectively, then using
(4.1) and (4.2), it can be shown (Wahlberg, 1989) that the asymptotic maximum
likelihood loss function of the process model is:
(4.3)
and the loss function of the disturbance model is
(4.4)
Minimizing these loss functions is equal to the frequency weighted Frobenius
norm model reduction. Note that the weights in (4.3) and (4.4) are the
inverses of asymptotic variances of the high order process model and
disturbance model respectively. Hence frequencies where the high order model
has small variance (error) will have a large weight in the estimation
criterion; this is physically appealing. The direct minimization of these loss
functions involves nonlinear optimization which is numerically difficult; we
have not yet seen any efficient algorithm for MlMO models.
Conventional model reduction methods calculate the parameters of the
reduced model from the parameters of the high order model. The method proposed
here takes another approach. We first simulate the high order model to
generate the input/output data, and then calculate the reduced order model
from the data. This can be called an identification approach to model
reduction and the advantage of it will become clear soon.
15
_-----
A Reduction of the Process Model
Given the high order ARX model of the process:
= [An(q)r
I
8
n
(q)
where
An(q) = diag[An (q) ... An (q)]
II pp
and
[
8n (q) ... 8
n
(q) ]
... 111m
n .
B (q) = :.. :
An 'n
B (q) ... B (q)
. pi pm
As mentioned :!;is is a diagonal form representation of the process
model in which the model is decomposed into p MISO sub-models.
Given the low order model of the process, also in a diagonal form:
where
,
, I " 'I [II (q) .. -B 1m (q)]
A (q) = diag [A (q) ... A (q)] and B (q) = ... .
II pp . .
.... . .....
(q) .. .s (q)
pI pm
Notice that for the low order model there is a degree, denoted by Ii, for each
(4.5)
(4.6)
MISO sub-model; in general the degrees of various MISO sub-models are
different. Here I only means the low order model which is not the degree of
it.
We propose the following
Procedure 4.1
For the i-th MISO sub-model:
Stage 1 Simulation. Collect the inputs, (u(t), t = 1, ... ,N), which have
been used in the identification experiment. Filter the inputs by the inverse
of the disturbance model ljHn(q) = An. (q). Then simulate the i-th MISO high
, ,
order sub-model using the filtered inputs:
; (t) = 1 [8
n
(q) ... jJ" (q)][An (q)u(t)] (4.7)
, An () II 1m ..
ii q
This is equivalent to
; (t) = [8
n
(q) ... 8
n
(q)] u(t)
I 111m
(4.8)
16
which is simpler. Thus we obtain the input/output data of the i-th MISO
sub-model:
where
Z; := ;i(I), u
r
,I(1), ... , u
r
,m(1), ... , ... ,y(N), Ur,I(N), ur,m(N)
U (f) = A n(q)U(f)
r "
Stage 2 SteigIitz-McBride iteration. The algorithm was proposed by
Steiglitz and McBride (1965), which can be seen as an approximation of the
Ak Ak Ak
output error method. Denote A .. (q) and B .I(q), ... , B. (q) as the estimates of
11 1 1m
the low order i-th MISO sub-model at iteration k. First filter the inputs
Ak
U (t),,u
r
(t) and the output y.(t) by IjA .. (q). Then refined estimates
A'.t4- ,m A k+ A k+ 1 "
A . . I(q) and B. \q), ... , B. I(q) are obtained by minimizing
1 I 111m
0'+1 = I N {Ak+1 [y i (f) ]
. N L .. (q) -kr---
1 f=1 11 A ()
ii q
A U (t)}2
_ ... _ Bk+l(q) [ k r ,m ]
1m Ak ( )
ii q
Denote e
k
+ I as the parameter vector of Ak+l(q) and iJk+l(q), ... , iJk+l(q):
1 II 111m
Ak+1 _ [Ak+l Ak+1 i/+I Ak+1 Ak+1 Ak+1 T
8. - a .. I,,a .. /., 1 1,,b
1
I,,.,b. I,,b. I.]
1 II, 11,1 1, 1,1 1m, 10l,1
and denote the data vector as
A
Ak _ [-y ,<t-I) -y i (t-li)
<p(t.9) - k , .. " k
1 Aii(q) A)q)
U (f-I) U (f-li)
k];=' '--- ,"', ki;" --
A(q) A .(q)
I I I 1
U
r
... U
r
]
,Ak ,'Ak
Aii(q) Aii(q)
Then (4.9) is a linear least squares problem and its solution is
Ak+1
9.
1
If i < p go to Stage I, else stop.
The initial estimate
(4.9)
(4.10)

The most simple way of initialization is to perform a normal equation error
17
least squares on the data. This implies the prefilter = 1 in (4.9).
"k 11"
The filter If A .. (q) will be mostly a lowpass filter, so is 1/:4
n
.(q) where
'" 11 11
A n.(q) is from the high order model. Hence we can initialize the iteration by
11... ...
setting AO.(q) = An.(q).
11 11
It has been experienced that a (frequency weighted) balanced model
reduction or a (frequency weighted) Hankel norm approximation can deliver
"good" reduced order models. Hence the third choice of initialization is to
use the A..(q) polynomial from one of these models.
u
The convergence properties
The model reduction nroblem (4.9) and (4.10) is an identification problem in
which the data are noise free, but the model order is lower than the order of
the process. In the literature there is no theoretical result on the
convergence properties of the Steiglitz-McBride method when the model order is
lower than that of the process. Our experience shows that the algorithm always
converges when used in model reduction. The same finding was reported by Fan
and Jenkins (1986) who use the method for adaptive filtering. We feel that
these are not just coincidences. Before any theoretical result can be
developed, let us propose the following
Conjecture 4.1 Assume that the inputs are persistently exciting, that is,
is nonsingular, and
t = 1 I 1
"k
If A .. (q) is stable for all k. Then the
"
iteration (4.9) and (4.10) converges.
We welcome interested researchers to prove this, or to find a counter example.
The loss function in the frequency domain
Assume that the algorithm converges at iteration k+ 1. Then the loss function
in (4.9) becomes
_It liN { m" "['" }2
= N L L [Gn(q) - G(q)]An(q)u(t)
t = 1 j = 1 . ) 'J .. J
Letting N 00, applying Parseval's identity and remembering that ut(t),
u (t) are mutually independent, one can show (see Ljung, 1987) that
m
18

0+
1
= r 1
i j=l2lt
-It
This is precisely the loss function of the asymptotic maximum likelihood
estimation (4.3) for the i-th MISO sub-model, except that has been
... ",. 11
replaced by the high order estimate =
11 11
B Reduction of the Disturbance Model
Given the high order disturbance model
iIn(q) = [An(q)rl = diagrrAn (q)r
l
[An (q)rl]
Ll II pp
(4.11)
which is diagonal, we want to find the reduced order model in the following
form
C (q)
pp ]
D (q)
pp
where
,
-Ii
+ clq
D, ,
,
D .. (q) = 1 + d q.1 + ... + d q.li
u ii,l fi,ii
,
with Ii being the order of the i-th disturbance model which can be different
from Ii, the order of the i-th sub-model of the process. Because iIn(q) and
iii (q) are diagonal, we have p SISO model reduction problems.
For the i-th sub-model:
(4.12)
Stage 1 Simulation. Generate a zero mean white noise sequence (e(t),
I = 1, "', N}. Filter e(l) by the inverse of the disturbance model
IjiIn(q) = An. (q). Then simulate the i-th SISO high order sub-model of the
..
disturbance using the filtered white noise
v. (t) = n 1 (q)e(l)]
, A () ..
ii q
This is equivalent to
v. (t) = e(t)
,
which is simpler. Thus we obtain the input/output data of the i-th SISO
sub-model:
z!" := e
f
(1), ... ef(N)
" ,
19
where
ef(t) = An.(q)e(t)
11
Stage 2 Steiglitz-McBride iteration. This is similar to the iteration for
the process model reduction, and we will not repeat it.
Following the same argument as for (4.11), we can show that the proposed model
reduction method for the disturbance model minimizes the following loss
function
1
Z1t
when N ~ 00. This is precisely the loss function of the asymptotic maximum
likelihood estimation (4.4) for the i-th SISO sub-model of the disturbance,
except H"..(e
i
"1 being replaced by the high order estimate
A 11 A
Hn(e
l
"1 = IjAn(e
l
"1.
11 11
In this section an asymptotic maximum likelihood model reduction technique is
proposed, where the low order models of the process and the disturbance are
calculated from the simulated data of the high order models. The advantages of
the technique are: (1) it minimizes a criterion which is physically sensible,
(2) it is numerically simple, and (3) it can cope with MIMO models. This makes
the method suitable not only for model reduction of identified high order
models as in this work, but also for other purposes such as model reduction of
theoretical (Physical) models and controller reduction.
5 Model Structure Determination
Before performing model reduction as proposed in the previous section, the
orders of the p MISO sub-models, {It, ... , lpl, need to be determined. This
set of indices defines the model structure of the MIMO process model (4.6).
Model order and structure selection is a central issue in system
identification. Many researchers have put their attention on this topic, and
there exists a multitude of methods for model order/structure selection; see
Stoica et al. (1986) and Janssen (1988) for recent overviews. Most of the
existing methods assume that the true process belongs to a set of candidate
model structures and try to find a 'right' structure. This assumption,
however, is neither likely to be fulfilled in practice, nor is it necessary
20
for applications such as simulation, prediction and control. What we want is a
model structure which gives a suitable approximation of those process features
in which we are interested for the underlying application. Other difficulties
of many proposed methods are: they are computationally costly, especially for
MIMO processes; and they are numerically not reliable, they can fail to find
the best or a 'right' structure when some minimization algorithm stops at a
local minimum. Some numerically simple methods exist, but usually they do not
use a criterion which is physically sensible.
Here we will propose a method of model structure selection for the low
order model (4.6), based on the asymptotic theory in Section 2 and some
physical intuition. The advantages of the method are: (1) it uses frequency
domain measures in determining the model structure, and (2) it is numerically
simple and reliable.
Our idea of order/structure selection is: chose the order such that the
variance part and bias part of the errors (see (3.6 are approximately equal.
The reasoning behind this idea can be explained as follows: (loosely
speaking) suppose that the high order model has 10% error, it is only possible
to find the true model if the low order model can deviate from the high order
model with 10%. If we let the variance part and the bias part of the errors to
be approximately equal, we get the following:
A MIMO model structure selection rule
For the i-th sub-model: Perform model reduction with various orders and select
the lowest order, Ii, such that in the frequency range which is important for
control system design:
(5.1)
The relation (5.1) can be called bias/variance equivalence principle. This
selection can be done simply by visual inspection. Applying the order
selection rule (5.1) to each MISO high order sub-model; then we can determine
the structure of the low order MIMO model.
The same idea can be applied for determining the orders of the
disturbance model.
Remarks-In this work we look at the identification problem as approximate
modelling. Hence we do not intend to find the 'right' model structure; instead
21

we want to find a model structure such that the reduced order model will have
smallest errors in the important frequency range. The consequence of this
philosophy is that the selected order can be either lower or higher than the
, true' order of the process.
In this selection rule, the order to be selected is related to the noise-
to-signal ratio, to the experiment time and to the order of the high order
model. This is again physically sensible.
If one finds that the visual inspection according to (5.1) is not
scientific enough, we suggest to use the weighted Frobenius norm; this can be
done by integrating the both sides of (5.1) over the frequency range which is
important for control system design.
The final model obtained by our method is given in polynomial matrix
description in a diagonal form; and the process is decoupled into MISO
sub-processes. This is a Box-Jenkins model. This model can be directly used
for controller design by using the polynomial method; see Kucera (1979). The
model can also be converted to a state space realization for the use in
control system analysis and design. Generally speaking, the McMillan degree of
the model equals the sum of the degrees of the MISO sub-models; in this case
the diagonal form description is called irreducible. It possible, however,
that a diagonal form polynomial matrix description is not irreducible, i.e.,
the sum of the degrees of the MISO sub-models is greater than the McMillan
degree of the model. If this happens, the minimal realization technique can be
used to eliminate the extra states.
6 A Simulation Study
In this section, a simulation study is performed in order to validate the
identification method proposed in the previous sections. The model of a glass
tube production process is used for the the simulation study. Two variables
need to be controlled for this process: diameter of the _ tube and the wall
thickness of the tube; two variables are chosen as the control inputs: drawing
speed and pressure of the blowing gas. Thus this is a 2-input 2-output
process. The identification and control of this process has been studied in
Backx (1987); see also Backx and Damen (i 3 9 ) . The model is given in a state
space realization with McMillan degree 6. In this study that model is used as
the true process and the input/output data are generated .by simulation. Fig.
6. I shows the transfer functions (amplitudes) of the process. The output
disturbances are colored noises which are generated by filtering two white
noises by two lowpass filters. The power of the disturbances IS 10% of the
22
power of the correponding outputs The inputs are generated according to the
input design method (3.17). 2000 data samples are used in the identification.
High order ARX model estimation. An ARX model with degree 30 is estimated in
this step. In Fig. 6.1the transfer functions (amplitudes) are compared with
those of the process.
As mentioned before, the errors of this high order model will be called
the variance part of the modelling errors. The 30 bound of the variance part
of the errors are calculated according to (3.4), and they are plotted together
with the errors in Fig. 6.2. The error is defined as the absolute value of the
difference between each true transfer function and that of the model. We see
that the 30 bounds cover the errors almost entirely except some frequency
points.
Model order reduction and model structure selection. The model reduction
method in Section 4 is used for calculating the low order models. The
Steiglitz-McBride iteration converges in the calculations of all the low order
models. For the 4th degree model, the iteration converges at iteration 3; for
the 5th degree model, the iteration converges at iteration 15.
According the model structure selection rule (5.1), the structure (4, 4)
is the best choice. In Fig. 6.3, the errors of the models of structure (4, 4)
and (5, 5) are compared. We find that the model quality of the structure
(4, 4) is indeed better than that of structure (5, 5).
In Fig. 6.4 the errors of the low order model with structure (4, 4) are
compared with the errors of the 30th degree ARX model. We note that the low
order model has a better quality.
In Fig. 6.5 the errors of the low order model with structure (4, 4) are
compared with the errors of the low order model with structure (4,4) that are
obtained by performing the balanced model reduction on each 25th degree MISO
models. There is a sharp distinction between the performances of the two
methods of model reduction.
Upper bound matrix. Formula (3.11) is used to calculate the upper bound
matrix for the model with structure (4, 4); and the upper bounds are plotted
together with the errors of the model in Fig. 6.6. We see that the bounds
become loser than the 30 bound for the variance; compare Fig. 6.2. As remarked
in Section 3, if a good model structure is selected and a proper model
reduction is performed, then the errors of low order model will be in general
23
smaller than the errors of the high order model; in this case the 3cr bounds
can be used as the total bounds. In Fig. 6.7 the errors of the low order model
with structure (4, 4) are plotted together with the 3cr bounds. We find that
indeed the 3cr bounds can be used as the total bounds, except the transfer
function (1, 1) at very low frequencies where the model reduction does not
function very well; see Fig. 6.4.
7 Conclusions
Several fundamental problems in MIMO process identification, such as input
design, model order and structure selection, and model uncertainty have been
treated in this paper. Also an identification approach to model reduction has
been proposed. The method of input design presented addresses directly the
intended use of the model, and at same time it is very simple and economical.
A method has been proposed for identifying a nominal model of a MIMO process,
together with an upper bound matrix of the modelling errors in the frequency
domain. This is motivated by the development in robust control theory. The
identification method is based on recent results on identification combined
with our physical intuition. The applicability of our method relies on the
fact that in all the steps the method uses criteria which are physically
sensible; at the same time the computations needed are simple and reliable.
The method has been validated by a simulation study. Note that this method IS
not only valid for finite dimensional processes. It can also be used for the
identification of infinite dimensional processes, if we start with a
sufficiently high order model. Now we are ready for real applications.
References
Backx, A.C.P.M. (1987). Identification of an Industrial Process: A Markov
Parameter Approach. Dr. Dissertation, Dept. EE., Eindhoven University of
Technology, The Netherlands.
Backx, A.C.P.M. and A.A.H. Damen (1989). Identification of industrial MIMO
processes for fixed controllers-part 1 and part 2. Journal A, Vol. 30, pp.
3-12, and pp. 33-43.
Correa, C.O. (1989). A system identificl!tion problem motivated by robust
control. Int. J. Control, Vol. 50, No.2, pp 575-602.
24
Cloud, D.J. and B. Kouvaritakis (1986). Statistical bounds on muItivariable
frequency response: an extension of the generalized Nyquist criterion. lEE
Proc., Vol. 133, Pt.D., pp. 97-110.
Doyle, J.C. (1984). Lecture Notes - ONRIHoneywell Workshop on Advances on
Multivariable Control.
Emami-Naeini, A., M.M. Akhter and S.M. Rock (1988). Effect of model
uncertainty on failure detection: the threshold selector. IEEE Trans. Autom.
Control, Vol. AC-33, No. 12, pp. 1106-1115.
Fan, H. and W.K. Jenkins (1986). A new adaptive IIR filter. IEEE Trans. Autom.
Control, Vol. AC-33, No. 10, pp. 939-947.
Gevers, M. and L. Ljung (1986). Optimal experiment design with respect to the
intended model application. Automalica, Vol. 22, No.5, pp. 543-554.
Goodwin, G.C. and M.E. Salgado (1989). A stochastic embedding approach for
quantifying uncertainty in the estimation of restricted complexity models.
Int. J. Adaptive Control and Signal Processing, Vol.3, pp. 333-356.
Glover, K. (1984). All optimal Hankel-norm approximations of
multivariable systems and their Leo-error bounds. Int. J. Control, Vol.
pp. III 5-1193.
linear
39,
Helmicki, AJ., C.A. Jacobson and C.N. Nett (1989). H identification of
eo
stable LSI systems: a scheme with direct application to controller design.
Proc. Amer. of Contr. Con!, Pittsburg, pp. 1428-1434.
Janssen, P. (1988). On Model Parametrization and Model Structure Selection for
Identification of MIMO-Systems. Dr. Dissertation, Dept. EE., Eindhoven
University of Technology, The Netherlands.
Kosut, R.L. (1986). Adaptive calibration: an approach to uncertainty modelling
and on-line robust control design. Proc. 25th Con! Decision and Control,
Athens, Greece, December, 1986, pp. 455-461.
Kouvaritakis, B. and H. Latchman (1985). Necessary and sufficient stability
criterion for systems with structured uncertainties: the major principal
direction alignment principle. Int. 1. Control, Vol. 42, No.3, pp. 575-598.
Kucera, V. (1979). Discrete Linear Control. John Wiley & Sons, New York.
Lenssen, M. (1988). Optimal Input Design for MIMO Processes. M.Sc. thesis,
Dept. EE., Eindhoven University of Technology, The Netherlands.
Ljung, L. (1985). Asymptotic variance expressions for identified black-box
transfer function models. IEEE Trans. Autom. Control, Vol. AC-30, pp.
834-844.
Ljung, L. (1987). System Identification: Theory for the User. Prentice-Hall,
Englewood Cliffs, N.J ..
Ljung, L. and Z.D. Yuan (1985). Asymptotic properties of black-box
identification of transfer functions. IEEE Trans. Autom. Control, Vol.
25
AC-30, pp.514-530.
Hehra, R.K. (1974). Optimal input signals for parameter estimation in dynamic
systems-A survey and new results. IEEE Trans. Autom. Control, Vol. AC-19,
pp. 753-768.
Morari, M. and E. ZafIriou (1989). Robust Process Control. Prentice-Hall, New
Jersey.
Steiglitz, K. and L.E. McBride (1965). A technique for the identification of
linear systems. IEEE Trans. Autom. Control, VoL AC-lO, pp. 461-464.
Stoica, P., P. Eykhoff, P. Janssen and T. SOderstrom (1986). Model-structure
selection by cross-validation. Int. J. Control, Vol. 43, pp. 1841-1878.
Van den Boom, T , Klompstra and A. Damen (1991). System identification for
H -robust control. Submitted to the 9th IFACIIFORS Symposium on
00
Identification and System Parameter Estimation, Budapest, July, 1991.
Vidyasagar, M. (1985). Control System Synthesis: A Factorization Approach. MIT
press.
Wahlberg, B. (1989). Model reduction of high-order estimated models: the
asymptotic ML approach. Int. J. Control, Vol. 49, No.1, pp 169-192.
Yuan, Z.D. and L. Ljung (1984). Black-box identification of multivariable
transfer functions: asymptotic properties and optimal input design. Int. J.
Control, Vol. 40, No.2, pp. 233-256.
Yuan, Z.D. and L. Ljung (1985). Unprejudiced optimal open loop input design
for identification of thansfer functions. Automatica, Vol. 21, No. 6 pp
697-708.
Zames, G. (1981). Feedback and optimal model reference
transformations, multiplicative seminorms, and approxImate inverses. IEEE
Trans. Autom. Control, Vol. AC-26, pp. 301-320.
Zhu, Y.C. (1987a). On a bound of modelling errors of black-box transfer
function estimates. EUT Report 87-E-173, EE Dept., Eindhoven University of
Technology, The Netherlands.
Zhu, Y.c. (1987b). Black-box identification of MlMO transfer functions:
asymptotic properties of prediction error models. EUT Report 87-E182, EE
Dept., Eindhoven University of Technology, The Netherlands.
Zhu, Y.C. (1987c). On the bounds of modelling errors of black-box MIMO
transfer function estimates. EUT Report 87-E-183, EE Dept., Eindhoven
University of Technology, The Netherlands.
Zhu, Y.C. (1989a). Estimation of transfer functions: asymptotic theory and a
bound of model uncertainty. Int. J. Control, Vo1.49, pp. 2241-2258.
Zhu, Y.C. (1989b). Black-box identification of MIMO transfer functions:
asymptotic properties of prediction error models. Int. J. Adaptive Control
and Signal Processing, Vol.3, pp. 357-373. .
26
Zhu, Y.C. (1990). Identification and Control of MIMO Industrial Processes: An
Integration Approach. Dr. Dissertation, Dept. EE., Eindhoven University of
Technology, The Netherlands.
Zhu, Y.C. and A.C.P.M. Bacia. (1991). MlMO Process identification for
controller design: test signals, norminal model and error bounds. Submitted
to the 9th IFACIIFORS Symposium on Identification and System Parameter
Estimation, Budapest, July, 1991.
27
Fig. 6.1 Transfer functions (amplitudes) of the process (solid line) and
transfer functions (amplitudes) of the high order model (dashed
line).
28
Fig. 6.2 The errors of the high order model (variance part) (dashed
line) and the 3cr bounds (solid line).
29
Errors of G5 11 and G4 11
O. 2
0.15
,
0.1
Frequency [rls)
Errors of G5 21 and G4 21
0.2
0.15 r
" '
. "
\:\ \'
o '''----- -------- -
o 1 2 3
Frequency [rls)
4
Errors of G5 12 and G4 12
0.15
0.1
0.05
"
J\
IE "
,
,
,
, ,
" ,
III \
I 1\
. . '
,

o 1 2 3 4
Frequency [rls)
Errors of G5 22 and G4 22
0.15
0.1
"
"
II \
0.05 \
III \
'Ii \
'
"
, '--::-::---=----=

o 1 2 3 4
Frequency [rls)
Fig. 6.3 The errors of the model with structure (5,5) (solid line) and
the errors of the model with structure (4,4) (dashed line).
30
Errors of 04 11 and 030 11
0.2
0.15
0.1
4
Frequency [rls]
Errors of 04 21 and 030 21
0.2
0.15
0.1
0.05
"
" ..
, ,
, ,
, I "
, , '
, r
: ,.
\I \, I ," I
" ... ' \, ..... \,'
o
012 3 4
Frequency [rls]
Errors of 04 12 and 030 12
0.1 r-==r-=-=-=----T"'-="'-""P-='--,
0.08
0.06
",
I, "
:,: I
I t : "
" ,'" : :: I:
\ I"'" I \, ',' :: "
I I \ I I, /'

:', " I, :,'\ I

" I ' " .' I I
, I,' I

0.04
0.02
o 1 2 3 4
Frequency [rls]
Errors of 04 22 and 030 22
0.2
0.15
0.1
0.05
1 2 3 4
Frequency [rls]
Fig. 6.4 The errors of the model with structure (4,4) (solid line) and
the errors of the high order model (dashed line).
31
ERR 11 ERRbal 11
0.4 ,----"'f=-"'-"->""F="'-'r------,
,
,
,
0.3 :
,
,
..
..
0.2
,
,
,
,
,
0.1
,

o ------ -- -
o 2 3
-
4
Frequency [rls]
ERR 21 ERRbal 21
0.25 ,----"'f=-=-'-""F="'--'r---,
Frequency [rls]
ERR 12 ERRbal 12
0.25
0.2
,
,
'.
0.15 ':
,
,
0.1
,
,
,
0.05 :
o -==--::::;:--=--=--::--=--i:::::--=--::=--::=-
o 1 2 3 4
Frequency [rls]
ERR 22 ERRbal 22
0.3
-
1 2 3 4
Frequency [rls]
Fig. 6.5 The errors of the model with structure (4,4) (solid line) and
the errors of the model with structure (4,4) which is obtained
by balanced model reduction (dashed line).
32
ERR 11 BN011
0.25
0.2
0.15
0.1




0.05




o 1 2 3 4
Frequency [rls]
ERR 21 BN021
0.4
0.3
0.2
..
.
0.1 : ..
r. '
.. ,
.,
--
,
__ ____
o 1 2 3 4
Frequency [rls]
ERR 12 BND12
0.3
0.2
0.1
.' .,
.'
.'

o I '..- .. ---- -------------------
o 1 2 3
Frequency [rls]
4
ERR 22 BN022
0.4 ..-------.'="-"'''1-'''-'-'-''=.----,
0.3
0.2
0.1 .. \
III ,
n. ,
III ,
II ..... _
O' --------- -------- -
o 1 2 3
Frequency [rls]
4
Fig. 6.6 The errors of the model with structure (4,4) (dashed line) and
the upper bounds according to (3.11) (solid line).
33
ERRll 3SGMll
0.2
0.15
0.1
0.05
,
,
,
1 3 4
Frequency [rls)
ERR21 3SGM21
0.3,------r==<..p==,---,
0.2
0.1
---

o 1 2 3 4
Frequency [rls)
ERR12 3SGM12
0.2
, ,
I ', ...
...

o 1 2 3 4
Frequency [rls)
ERR22 3SGM22
0.3 ,------r==<..p==,---,
0.2
0.1
1 2 3 4
Frequency [rls)
Fig. 6.7 The errors of the model with structure (4,4) (dashed line) and
the 30" upper bounds according to (3.llb) (solid line).
34
Eindhoven University of Technology Research Reports
Faculty of Electrlcal Englneering
I SSN 0167-9708
Coden: TEUEOE
(222) Jozwiak, L.
THE FULL-DECOMPOSITION OF SEQUENTIAL MACHINES WITH THE SEPARATE REALIZATION
OF THE NEXT-STATE AND OUTPUT FUNCTIONS.
EUT Report 89-E-222. 1989. ISBN 90-6144-222-2
(223) Jozwiak, L.
THE BIT FULL-DECOMPOSITION OF SEQUENTIAL MACHINES.
EUT Report 89-E-223. 1989. ISBN 90-6144-223-0
(224) Book of abstracts of the first Benelux-Japan Workshop on Information and
Communication Theory, Eindhoven, The Netherlands, 3-5 September 1989.
Ed. by Han Vinck.
EUT Report 1989. ISBN 90-6144-224-9
(225)
(226)
M.J.
A PO SIBILITY TO INCORPORATE SATURATION IN THE
OF A SYNCHRONOUS MACHINE WITH RECTIFIER.
EUT Report 89-E-225. 1989. ISBN 90-6144-225-7
SIMPLE, GLOBAL MODEL
RoP. and E.M. van Veldhuizen, W.R. L.H.Tho Rietl'ens
XP MENTS DN INITIAL BEHAVIoUR OF CORONA WITH ELECTR CAL
PULSES SUPERIMPDSED ON DC BIAS.
EUT Report 89-E-226. 1989. ISBN 90-6144-226-5
(227) Bastings, R.H.A.
TowARD THE DEVELOPMENT OF AN INTELLIGENT ALARM SYSTEM IN ANESTHESIA.
EUT Report 89-E-227. 1989. ISBN 9D-6144-227-3
(228) Hekker, J.J.
ANIMATED GRAPHICS AS A TEACHING TDDL
SIMULATOR .
EUT Report 89-E-228. 1989. ISBN 90-6144-228-1
FOR THE ANESTHESIA MACHINE
(229) Oostrom, J.H.M. van
INTELLIGENT ALARMS IN ANESTHESIA: An implementation.
EUT Report 89-E-229. 1989. ISBN 90-6144-229-X
(230) Winter, M.R.M.
DESIGN OF A UNIVERSAL PROTOCOL SUBSYSTEM ARCHITECTURE: Specification of
functions and services.
EUT Report 89-E-230. 1989. ISBN 90-6144-230-3
(231) Schemmann, MoF.C. and HoC. Heyker, J.J.M. Kwaspen, Th.C. van de Roer
MOUNTING ANO DC TO 18 GHz CHARACTERISATION OF DOUBLE BARRIER RESONANT
TUNNELING DEVICES.
(232)
(233)
(234)
EUT Report 89-E-231. 1989. ISBN 90-6144-231-1
Sarma, A.D. and M.H.A.J. Herben
!)A'fAACQU I S IT I ON AND S I I NGt ANALYS I S OF SC I NT I LLA TI ON EVENTS
FOR THE OLYMPUS PROPAGATION EXPERIMENT.
EUT Report 89-E-232. 1989. ISBN 90-6144-232-X
J.A.
DESIGN A D IMPLEMENTATION OF A SECONO PROTOTYPE OF THE
SYSTEM IN ANESTHESIA.
EUT Report 90-E-233. 1990. ISBN 90-6144-233-8
E.H.J.
DESIG ING DEBUGGING TOOLS FOR SIMPLEXYS EXPERT SYSTEMS.
EUT Report 90-E-234. 1990. ISBN 90-6144-234-6
INTELLIGENT ALARM
(235) Heffels, J.J.M.
A PATIENT SIMULATOR FOR ANESTHESIA TRAINING: A mechanical lung model and a
physiological software model.
EUT Report 90-E-235. 1990. ISBN 90-6144-235-4
(236) Lammers,
KNOWLEDGE BASED ADAPTIVE BLOOD PRESSURE CONTROL: A Simplexys expert system
application.
EUT Report 90-E-236. 1990. ISBN 90-6144-236-2
(237) Ren Qingchang
PR<DICTION ERROR METHOD FOR IDENTIFICATION OF A HEAT EXCHANGER.
EUT Report 90-E-237. 1990. ISBN 90-6144-237-0
Ei
ISSN 0167-9708
Coden: TEUEDE
(238) Lammers, J.D.
[HE USE OF PETRI NET THEORY FOR SIMPLEXYS EXPERT SYSTEMS PROTOCOL CHECKING.
EUT Report 90-E-238. 1990. ISBN 90-6144-238-9
(239)
(240)
(241 )
X.
LIMINARY INVESTIGATIONS ON TACTILE PERCEPTION OF GRAPHICAL PATTERNS.
[UT Report 90-[-239. 1990. ISBN 90-6144-239-7
J.M.A.
BASE CORRECTNESS CHECKING FOR SIMPLEXYS EXPERT SYSTEMS.
EUT Eeport 90-E-240. 1990. ISBN 90-6144-240-0
Brinker, A.C. den
A MEMBRANE MODEL FOR SPATIOTEMPORAL COUPLING.
EUT Report 90-E-241. 1990. ISBN 90-6144-241-9
(242) Kwaspen, J.J.M. and H.C. J.I.M. van de Roer
MIcROWAVE NOISE DOUB'.E II TUNNElING!ITODES.
(243)
(244)
EUT Report 90-E-242. 1990. ISBN 90-6144-242-7
Massee
j
1-'. ana rl.A.L.M. de Graaf, W.J.M. Balemans, H.G. Knoopers, H.H.J.
ten --
PRELco,GN OF AN EXPERIMENTAL (5-10 MWt) DISK MHO FACILITY AND PROSPECTS OF
COMMERCIAL (1000 MWt) MHO/STEAM SYSTEMS.
EUT Report 90-E-243. 1990. ISBN 90-6144-243-5
Klompstra, Martin and Ton van den Boom, Ad Damen
A COMPARISON OF CLASSICAL AND MODERNICoNTROII11f"DESIGN: A case study.
EUT Report 90-E-244. 1990. ISBN 90-6144-244-3
(245) P.H.G. van de
N HE ACCURACY OF RADIOWAVE PROPAGATION MEASUREMENTS: Olympus propagation
(246)
(247)
(248)
experiment.
EUT Report 90-E-245. 1990. ISBN 90-6144-245-1
P.J.L. de
YNTHESIS METHOD FOR COMBINED OPTIMIZATION OF MULTIPLE ANTENNA PARAMETERS
AND ANTENNA PATTERN STRUCTURE.
EUT Report 90-E-246. 1990. ISBN 90-6144-246-X
Jozwiak, L. and T.
STAT ASSiGNMENT WITH REUSE OF STANDARD DESIGNS: Using
counters as sub-machines and using the method of maximal adjacensies to
select the state chains and the state codes.
EUT Report 90-E-247. 1990. ISBN 90-6144-247-8
Hoeiimakers, M.J. and J.M. Vleeshouwers
DERIVATION AND VERIFICATION OF A MODEL OF THE SYNCHRONOUS MACHINE WITH
RECTIFIER WITH TWO DAMPER WINDINGS ON THE DIRECT AXIS.
EUT Report 90-E-248. 1990. ISBN 90-6144-248-6