Centre d'Ing

enierie des Syst

emes,

d'Automatique et de M
ecanique Appliqu
ee

Avenue Georges Lema^

tre 4

Universit
e Catholique de Louvain B-1348 Louvain-la-Neuve

Facult
e des Sciences Appliqu
ees Belgique

Experiment Design Issues in

Modelling for Control
Beno^
t Codrons

These presentee en vue

de l'obtention du titre
de Docteur en Sciences Promoteur: Michel Gevers
Appliquees

Juillet 2000
 Centre d'Ing
enierie des Syst
emes,

d'Automatique et de M
ecanique Appliqu
ee

Avenue Georges Lema^

tre 4

Universit
e Catholique de Louvain B-1348 Louvain-la-Neuve

Facult
e des Sciences Appliqu
ees Belgique

Experiment Design Issues in

Modelling for Control
Beno^
t Codrons

Jury:

Prof. B.D.O. Anderson Australian National University, Canberra, Australia

Prof. G. Bastin (president) Universite Catholique de Louvain, Louvain-la-Neuve, Belgium
Prof. G. Campion Universite Catholique de Louvain, Louvain-la-Neuve, Belgium
Prof. B. De Moor Katholieke Universiteit Leuven, Leuven, Belgium
Prof. M. Gevers (supervisor) Universite Catholique de Louvain, Louvain-la-Neuve, Belgium
Prof. H. Hjalmarsson Kungl Tekniska Hogskolan, Stockholm, Sweden
 But the world is confusing, and just because we invent
myths and theories to explain away the chaos we're still
going to live in a world that's older and more complicated
than we'll ever understand. So many religious and polit-
ical and scientic and social systems fail in that they try
to impose a rigid structure onto what is an inherently am-
biguous world. I'm not suggesting that we stop trying to
understand things. Trying to understand the world can
be fun and, at times, helpful. But if we base our belief
systems on the humble assumption that the complexities
of the world are ontologically beyond our understanding,
then maybe our belief systems will make more sense : : :

Moby
(excerpt from essay #1 in
the booklet of his CD \Play")

I don't want knowledge. I want certainty.

David Bowie, \Law"

That is what this thesis is all about: imperfect knowledge, imperfect

modelling, and uncertainty (and still condence).
To the memory of my grandfather
Maurice Codrons (y 1995)
 Abstract

This thesis presents several results regarding experiment design issues in

modelling for control. Three aspects of modelling are essentially consid-
ered: closed-loop identication, model validation, and model reduction.
Although they may seem disconnected from one another, those three
aspects grow up on a common root: the fact that closed-loop identica-
tion with a controller that is `close' to the optimal one delivers models
with bias and variance errors that are ideally tuned for control design.
Our rst contribution is to show that problems arise when the controller
that is used during the closed-loop identication experiment has unstable
poles and/or nonminimum-phase zeros, and that the resulting model
may unexpectedly be unsuitable for control design. More precisely, we
show that
 the model that results from such an identication experiment may
or sometimes will be destabilised by the current controller if some
precautions are not taken regarding the approach that is used (indi-
rect, coprime-factor or direct approach, or the Hansen scheme) and
the choice of the excitation point of the closed-loop system (injection
of the excitation signal at the input of the plant or at the input of
the controller). We carry out a thorough analysis of these closed-loop
identication approaches regarding this issue;
 a model presenting this instability problem is ill-suited for control
design, in that sense that a controller that stabilises this model cannot
be guaranteed to stabilise the true system.
Our analysis results in experiment design guidelines for the choice of the
method and the choice of the excitation source in the loop.
The second contribution of this thesis concerns model validation; more
precisely, the computation of uncertainty regions guaranteed to contain
the unknown true system at some specied probability level, and which
i
ii Abstract

can be used for robust control design. The procedure that is used for the
computation of such uncertainty regions is prediction error identication
using an unbiased model structure. The resulting uncertainty sets can
be used
 for model validation: every model that lies within such an uncertainty
region is called validated since it is not possible, on the basis of the
realised experiment, to prove that it is dierent from the underlying
true system;
 for controller validation: it is possible to verify if a given controller
stabilises or achieves a specied level of performance with all systems
contained in this set. If it is the case, the controller is called validated
since it will stabilise and achieve the required level of performance
with the true system as well.
Our contribution is to show that this validation procedure can be applied
in closed loop, and that the uncertainty sets that result from closed-loop
validation are better tuned for control design (i.e. they allow less conser-
vative designs) than uncertainty sets obtained by open-loop validation.
The third major topic of this thesis is model and controller reduction.
Using the same heuristic reasoning that motivates the use of closed-loop
identication when the nal objective is control design, we propose a new
closed-loop performance-based coprime-factor reduction criterion that
can be used either for the reduction of a high-order model before control
design, or for the reduction of a high-order controller. The eÆciency of
this criterion is assessed by means of a realistic application example.
 Preface

This thesis consists of eight chapters:

Chapter 1: Introduction
Chapter 2: Modelling tools, closed-loop setup and generali-
ties
Chapter 3: A motivating discussion of issues in identica-
tion for control
Chapter 4: Identication in closed loop with an unstable or
nonminimum phase controller
Chapter 5: Model validation for control and controller vali-
dation
Chapter 6: A performance-based approach to control-
oriented model reduction and to controller re-
duction
Chapter 7: Future issues on variance and on model order
selection
Chapter 8: Conclusions

The material in Chapter 3 has been partly published in

Gevers, M., B.D.O. Anderson and B. Codrons (1998). Is-
sues in modeling for control. In: Proc. of the 1998 Ameri-
can Control Conference. Philadelphia, Pennsylvania, USA.
pp. 1615{1619.
The material in Chapter 4 can be found in
Codrons, B., B.D.O. Anderson and M. Gevers (2000b).
Closed-loop identication with an unstable or nonminimum
phase controller. Accepted for publication in Automatica.
iii
iv Preface

An edited version of the latter can be found in

Codrons, B., B.D.O. Anderson and M. Gevers (2000a).
Closed-loop identication with an unstable or nonminimum
phase controller. In: CD-ROM Proc. of the IFAC System
Identication Symposium (SYSID 2000). Santa Barbara,
California, USA. Paper ThPM1-3.
The basic concepts of closed-loop model validation, as described in
Chapter 5, have been published in
Gevers, M., B. Codrons and F. De Bruyne (1999a). Model
validation in closed loop. In: Proc. of the 1999 American
Control Conference. San Diego, California, USA. pp. 326{
330.
The application of model and controller validation techniques to a fer-
rosilicon production process can be found in
Codrons, B., X. Bombois, M. Gevers and G. Scorletti
(2000d). A practical application of recent results in model
and controller validation to a ferrosilicon production pro-
cess. Accepted for presentation at the 39th Conference on
Decision and Control, Sydney, Australia, December 2000.
The remainder of Chapter 5 (including the notions of worst-case  -gap
and the controller validation tests, for which the credit must be essen-
tially given to Michel Gevers and Xavier Bombois, respectively) has been
published in
Bombois, X., M. Gevers and G. Scorletti (2000a). A mea-
sure of robust stability for an identied set of parametrized
transfer functions. Accepted for publication in IEEE Trans-
actions on Automatic Control,
Bombois, X., M. Gevers, G. Scorletti and B.D.O. Anderson
(2000c). Controller validation for stability and performance
based on an uncertainty region designed from an identied
model. Submitted to Automatica,
 Preface v

Gevers, M., X. Bombois, B. Codrons, F. De Bruyne and

G. Scorletti (1999b). The role of experimental conditions in
model validation for control. In: Robustness in Identica-
tion and Control (A. Garulli, A. Tesi and A. Vicino, Eds.).
Vol. 245 of Lecture Notes in Control and Information Sci-
ences. pp. 72{86. Springer-Verlag. London, U.K.
and
Gevers, M., X. Bombois, B. Codrons, G. Scorletti and
B.D.O. Anderson (2000b). Model validation for control and
controller validation: a prediction error identication ap-
proach. Submitted to Automatica.
Preliminary versions of several parts of Chapter 6 can be found in
Bendotti, P., B. Codrons, C.-M. Falinower and M. Gevers
(1998). Control-oriented low-order modelling of a complex
PWR plant: a comparison between open-loop and closed-
loop techniques. In: Proc. of the 37th IEEE Conference on
Decision and Control. Tampa, Florida, USA. pp. 3390{3395
and
Codrons, B., P. Bendotti, C.-M. Falinower and M. Gevers
(1999). A comparison between model reduction and con-
troller reduction: Application to a PWR nuclear plant. In:
Proc. of the 38th IEEE Conference on Decision and Control.
Phoenix, Arizona, USA. pp. 4625{4630.

As granted by Electricit e de France, the material in this chapter
has also been reported on (in French) in
Codrons, B., P. Bendotti, C.-M. Falinower and M. Gevers

(2000c). Etude comparative des methodes d'identication et
de reduction pour la commande { application a la synthese
d'un regulateur d'ordre reduit pour un modele lineaire com-
plexe de REP. Study report. Centre for Systems En-
gineering and Applied Mechanics (CESAME), Universite
Catholique de Louvain. Louvain-la-Neuve, Belgium.
vi Preface
 Acknowledgements

From the very beginning, many people have contributed directly or indi-
rectly, voluntarily or involuntarily, continuous-timely or discrete-timely,
to the realisation of this thesis.
First of all, my thanks certainly go to my supervisor, Professor Michel
Gevers, for the condence he has put in me and for all the good and
enjoyable work we have done together (especially the writing of the big
paper, despite all the eorts Mr Smarty { one of our most regular and
critical readers { has done to discourage us).
Next, I would like to express my sincere gratitude to Professor Brian
Anderson, Franky De Bruyne and (future Professor?) Xavier Bombois,
with whom I have had an invaluable collaboration during these last four
years. I am especially grateful to Brian, who came every year in Louvain-
la-Neuve with an incredible amount of ideas and open problems (such
as that of closed-loop identication with a `bad' controller addressed in
Chapter 4).
I would also like to thank Pascale Bendotti and Clement-Marc Falinower

of Electricit
e de France for the work we did together on model and
controller reduction, and for the opportunity they gave me to spend
four months at EDF before the beginning of my Ph.D. research. This
stay has aroused my interest for industrial control problems, and I would
otherwise probably not have started a Ph.D. thesis.
I also gratefully thank Professor Georges Bastin, Professor Guy Cam-
pion, Professor Bart De Moor and Professor H akan Hjalmarsson for ac-
cepting to be members of my thesis jury together with Michel and Brian,
and for their enlightened reading of the rst version of this manuscript.
This thesis would not have been carried out without the nancial support

of Electricit
e de France and of the Belgian State (through the Belgian
Programme on Interuniversity Poles of Attraction).
vii
viii Acknowledgements

The time I have spent at Cesame has been enjoyable, and this has
certainly been a major motivation for the work I have done. It is a
great pleasure for me to thank our secretaries and technicians, namely,
Michele Termolle, Isabelle Hisette, Lydia De Boeck, Laurence Lamisse,

Dominique Pigeon, Michel De Wan, Guido Donders, Etienne Huens and
Victor Vermeulen (who prepares such a tasty coee), for their avail-
ability and sympathy. I also thank my roommates Vincent Vermaut,
Franky De Bruyne and Luc Moens, as well as all members of Cesame,
for the good time we have had together. I especially remember my

earliest years at Cesame, when I used to play whist at noon with Eric
Brasseur, Pierre Carrette, Gary Dantinne, Franky De Bruyne, Xavier
Gallez, Pierre Halin, Olivier Houppertz, Ingrid Jaumain, Cecile Jeggy,
Vincent Legat, Gregory Lielens, Olivier Magotte, Elisabetta Olivari,
Renaud Sizaire, Natasha Van Rutten, Vincent Vermaut, and probably
many others whose names are now buried in the unfathomable depths
of my mind.
Finally, I would like to thank my ancee Nathalie for her enthusiastic
support in hard times, and for her continuous love and condence in my
ability to bring this to a good end. I owe her so much!

Beno^t Codrons,
Louvain-la-Neuve,
July 6, 2000.
 Contents
Abstract i

Preface iii

Acknowledgements vii

Notation glossary xv
Abbreviations xv
General symbols xvi
Operators and functions xx

Chapter 1. Introduction 1
1.1. The historical framework 1
1.2. Where are we today? 2
1.3. The contribution of this thesis 3
1.4. Synopsis 6

Chapter 2. Modelling tools, closed-loop setup and

generalities 9
2.1. Introduction 9
2.2. Prediction error identication 10
2.2.1. Assumptions about the system 10
2.2.2. The identication method 10
2.2.3. Classical model structures 12
2.2.4. Computation of the estimate 12
2.2.5. Asymptotic properties of the estimate 14
2.3. Balanced truncation 16
2.3.1. The concepts of controllability and observability 17
ix
x Contents

2.3.2. Balanced realisation of a system 19

2.3.3. Balanced truncation 20
2.3.4. Frequency-weighted balanced truncation 21
2.3.5. Balanced truncation of discrete-time systems 23
2.4. General representation of a closed-loop system and
closed-loop stability 24
2.4.1. General closed-loop setup 24
2.4.2. Closed-loop stability 25
2.5. LFT-based representation of a closed-loop system 28
2.6. Coprime factorisations 30
2.7. The  -gap metric 34
2.7.1. Denition 34
2.7.2. The winding number condition and comparison
with the standard gap metric 35

Chapter 3. A motivating discussion of issues in

identication for control 37
3.1. Introduction 37
3.2. The role of feedback 38
3.3. The eect of feedback on the bias error 41
3.4. The eect of feedback on the variance error 44
3.5. Simulation 47
3.6. Conclusions 48

Chapter 4. Identication in closed loop with an unstable

or nonminimum phase controller 51
4.1. Introduction 51
4.2. Some reminders about properties of systems and
identication in closed loop 52
4.2.1. Poles and zeros of a system 52
4.2.2. Identication in closed loop 53
4.3. Destruction of the nominal stability in the case of unstable
poles or nonminimum phase zeros in the controller 54
4.3.1. The indirect approach 55
4.3.2. The coprime-factor approach 58
 Contents xi

4.3.3. The direct approach 61

4.3.4. Guidelines for an optimal experiment design 62
4.4. Consequences for robust control design 63
4.5. Stability-preserving closed-loop identication with an
unstable and nonminimum phase controller 65
4.5.1. Using a tailor-made parametrisation 65
4.5.2. A two-stage tailor-made parametrisation approach 67
4.5.3. The dual Youla parametrisation 68
4.6. Numerical illustration 71
4.6.1. Problem setting 71
4.6.2. The indirect approach 73
4.6.3. The coprime-factor approach 78
4.6.4. The direct approach 85
4.6.5. The Hansen scheme 91
4.6.6. Comments on the numerical example 105
4.7. Conclusions 106

Chapter 5. Model validation for control and controller

validation 111
5.1. Introduction 111
5.2. Open-loop model validation 112
5.3. Closed-loop model validation 115
5.3.1. An indirect approach 115
5.3.2. A direct approach 119
5.4. The role of the model error model, how it can be dropped,
and a generic form for the uncertainty region 121
5.5. Model set validation for control and controller validation 123
5.5.1. Model set validation for control: a design problem 123
5.5.2. Controller validation for stability: a verication
problem 125
5.5.3. Controller validation for performance: a verication
problem 126
5.6. Illustration I: exible transmission system 126
5.6.1. Problem setting 127
5.6.2. Open-loop validation experiment 128
xii Contents

5.6.3. Closed-loop validation experiment 130

5.6.4. Model set validation for control: comparison of the
validated sets 131
5.6.5. Controller validation for stability 133
5.6.6. Controller validation for performance 133
5.7. Illustration II: ferrosilicon production process 134
5.7.1. Problem setting 134
5.7.2. Open-loop validation experiment 137
5.7.3. Closed-loop validation experiment 137
5.7.4. Model set validation for control: comparison of the
validated sets 138
5.7.5. Controller validation for stability 140
5.7.6. Controller validation for performance 140
5.8. Conclusion 144

Chapter 6. A performance-based approach to

control-oriented model reduction and to
controller reduction 145
6.1. Introduction 145
6.2. Model order reduction 147
6.2.1. Open-loop reduction of the nominal model coprime
factors 147
6.2.2. Performance-preserving closed-loop reduction of
the nominal model coprime factors 149
6.2.3. Stability-preserving closed-loop reduction of the
nominal model coprime factors 152
6.2.4. Preservation of stability and performance by
closed-loop reduction of the nominal model coprime
factors 153
6.3. Controller order reduction 155
6.3.1. Open-loop reduction of the optimal controller
coprime factors 156
6.3.2. Performance-preserving closed-loop reduction of
the optimal controller coprime factors 157
6.3.3. Stability-preserving closed-loop reduction of the
optimal controller coprime factors 159
 Contents xiii

6.3.4. Other closed-loop controller reduction methods 159

6.4. Application to the design of a low-order controller for a
complex Pressurised Water Reactor model 162
6.4.1. Description of the system 162
6.4.2. Control objective and design 163
6.4.3. System identication 166
6.4.4. Model reduction 167
6.4.5. Controller reduction 173
6.4.6. Comparison of the performance of the designed
controllers 179
6.4.7. Comparison of the new performance-based
closed-loop model and controller reduction criteria
with the stability-based closed-loop reduction
criteria 181
6.4.8. Comparison of performance-based closed-loop
controller coprime-factor reduction with closed-loop
balanced truncation 182
6.5. Conclusions 191

Chapter 7. Future issues on variance and on model order

selection 195
7.1. Introduction 195
7.2. The eect of overmodelling on the variance of estimated
transfer functions 196
7.2.1. The eect of additional poles and zeros 196
7.2.2. The choice of the model structure 199
7.2.3. Conclusion 201
7.3. Variance issues in system identication and model
reduction 202

Chapter 8. Conclusions 207

8.1. Contribution of this thesis 207
8.2. Open questions 209

Bibliography 211

Index 217
xiv Contents
 Notation glossary

Abbreviations
ARX Auto-Regressive model structure with eXogenous
inputs
ARMAX Auto-Regressive Moving-Average model structure
with eXogenous inputs
BJ Box-Jenkins model structure
EDF 
Electricit
e de France
FIR Finite Impulse Response model structure
GPC Generalised Predictive Control
LFT Linear Fractional Transformation
LMI Linear Matrix Inequality
LQG Linear Quadratic Gaussian
LTI Linear Time Invariant
MIMO Multi-Input Multi-Output
OE Output Error model structure
PE Prediction Error
PI controller controller with Proportional and Integral actions
PID controller controller with Proportional, Integral and Deriva-
tive actions
PRBS Pseudo-Random Binary Signal
PWR Pressurised Water Reactor
QFT Quantitative Feedback Theory
SISO Single-Input Single-Output
SNR Signal-to-Noise Ratio
w.p. with probability
w.r.t. with respect to

xv
xvi Notation glossary

General symbols
A>0 the matrix A is positive denite
A>0 the matrix A is positive semi-denite
aij i-th row and j -th column entry of the matrix A
Aij i-th row and j -th column block entry (or subma-
trix) of a block (or partitioned) matrix A
Amn indicates that the matrix A has dimension m  n
AsN (m; P) x 2 AsN (m; P) means that the sequence of ran-
dom variables x converges in distribution to the
normal distribution with mean m and covariance
matrix P
C controllability matrix  
 ~
C , C (z ), C (s), C , C two-degree-of-freedom controller, C = F K
C space of complex numbers or complex plane
C0 C n f0g
C + closed right half plane: C + = fs 2 C j Re(s) > 0g
C+0 open right half plane: C +0 = fs 2 C j Re(s) > 0g
Cn complex vector space of dimension n  1
C n  m complex matrix space of dimension n  m
{Y X complement of X in Y  X : {Y X = Y n X
{D complement of D in C (enlarged domain of insta-
bility): {D = fz 2 C j jz j > 1g
{D c complement of D c in C (restricted domain of insta-
bility): {D c = fz 2 C j jz j > 1g
D uncertainty region in transfer function space
D open unit disc (restricted domain of stability): D =
fz 2 C j jzj < 1g
Dc closed unit disc (enlarged domain of stability):
D c = fz 2 C j jz j 6 1g
@D unit circle: @ D = fz 2 C j jz j = 1g
~ek ~ek 2 Rn , 1 6 k 6 n, is the k-th orthonormal basis
vector of Rn . Its entries are all 0 except the k-th
one which is 1.
E uncertainty region in closed-loop transfer function
space
F , F (z ), F (s), F , F~ feedforward part of a two-degree-of-freedom con-
troller C , C or C~
G Model set for the input-output dynamics
 General symbols xvii

G, G(z ), G(s) system or model (input-output dynamics)

Gn , Gn (z ), Gn (s) system or model of order n (input-output dynam-
ics)
G n balanced realisation of a system Gn
G~ (z; ) model error model
G(z; ), G (z; ) parametrised plant model
G^ , G plant model (In case of a parametrised model, G^ ,
G(^).)
G^ r approximation of order r of a system or model Gn
of order n > r
G0 , G0(z ), G0 (s) true input-output plant dynamics
@G model error @G = G0 G^
H (z; ), H~ (z; ),
H (z; ) parametrised noise model
H^ noise model (In case of a parametrised model, H^ ,
H (^).)
H0 , H0 (z ), H0 (s) true noise dynamics
H2 Hardy space: closed subspace of L2 with transfer
functions (or matrices) P (s) analytic in C +0 (contin-
uous case) or with transfer functions (or matrices)
P (z ) analytic in {D c (discrete case)
H1 Hardy space: closed subspace of L1 with transfer
functions (or matrices) P (s) analytic in C +0 (contin-
uous case) or with transfer functions (or matrices)
P (z ) analytic in {D c (discrete case)
I identity
p matrix (of appropriate dimension)
j 1
jR set of imaginary numbers
K , K (z ), K (s),
K , K~ feedback controller or feedback part of a two-
degree-of-freedom controller C , C or C~
Kid feedback controller present in the loop during iden-
tication or model validation
L2 Hilbert space: set of transfer functions (or matri-
ces) P (s) (continuous case) or P (z ) (discrete case)
such
R 1 that the following integral is bounded:
trace[ P ? (j! )P (j! )]d! < 1 (continuous case)
1
or
R
 trace[P (e )P (e )]d! < 1 (discrete case)
? j! j!
xviii Notation glossary

L1 Banach space: set of transfer functions (or matri-

ces) P (s) bounded on j R (continuous case), or set
of transfer functions (or matrices) P (z ) bounded
on @ D (discrete case)
M model set for the input-output and noise dynamics
M , M (z ), M (s) denominator of a right coprime factorisation of a
system G = NM 1 ; M 2 RH1
M~ , M~ (z ), M~ (s) denominator of a left coprime factorisation of a sys-
tem G = M~ 1 N~ ; M~ 2 RH1
M() particular model corresponding to the parameter
value 
N , N (z ), N (s) numerator of a right coprime factorisation of a sys-
tem G = NM 1 ; N 2 RH1
N~ , N~ (z ), N~ (s) numerator of a left coprime factorisation of a sys-
tem G = M~ 1 N~ ; N~ 2 RH1
N (G; K ) closed-loop transfer matrix (noise dynamics)
O observability matrix
P controllability Gramian
P , P (z ), P (s) generic notation for a transfer function (or transfer
matrix) of any system or controller
P (possibly estimated) covariance matrix of the pa-
rameter vector 
Q observability Gramian
R space of real numbers
R0 R n f 0g
Rn Euclidean space of dimension n
R nm real matrix space of dimension n  m
Ry (k) autocorrelation matrix of the signal y(t): Ry (k) =
 (t)yT (t k)
Ey
Ryx (k) correlation matrix of the signals y(t) and x(t):
 (t)xT (t k)
Ryx (k) = Ey
RH1 real rational subspace of H1: ring of proper and
real rational stable transfer functions or matrices
S the true system (G0 ; H0 )
S (G; K ) closed-loop sensitivity function: S (G; K ) = 1+1GK
T (G; K ) generalised closed-loop transfer matrix
Ti (G; K ) generalised closed-loop transfer matrix
To (G; K ) generalised closed-loop transfer matrix
 General symbols xix

T~ij closed-loop model error model

@Tij closed-loop model error @Tij = Tij (G0 ; K )
^ K)
Tij (G;
u(t) input variable at time t
U uncertainty region in parameter space
U , U (z ), U (s) numerator of a right coprime factorisation of a con-
troller K = UV 1 ; U 2 RH1
U~ , U~ (z ), U~ (s) numerator of a left coprime factorisation of a con-
troller K = V~ 1 U~ ; U~ 2 RH1
V , V (z ), V (s) denominator of a right coprime factorisation of a
controller K = UV 1 ; V 2 RH1
V~ , V~ (z ), V~ (s) denominator of a left coprime factorisation of a con-
troller K = V~ 1 U~ ; V~ 2 RH1
Wl left-hand side frequency weighting lter
Wr right-hand side frequency weighting lter
X cont controllable state subspace of a given system Gn :
X cont  Rn
X obs observable state subspace of a given system Gn :
X obs  Rn0
y(t) output variable at time t
y^(t j t 1; ) predicted output at time t using a model M() and
based on data Z t 1
ZN data set fu(1); y(1) : : : ; u(N ); y(N )g
2 (n) 2 -distributed random variable with n degrees of
freedom
"(t; ) prediction error y(t) y^(t j t 1; ) or residuals
(simulation errors y(t) y^(t))
! frequency or normalised frequency [rad=s]
'(t) regression vector
 y (! ) power spectral density of the signal y(t): Fourier
transform of Ry (k)
xy (!) part of the spectrum of y(t) due to x(t)
yx (!) cross-spectrum of y(t) and x(t): Fourier transform
of Ryx (k)
 open-loop model error model parameter vector in
Rq
^ estimate of 
x standard deviation of the random variable x
xx Notation glossary

 diagonal observability and controllability Gramian

of a balanced system G n:  = diag(&1 ; : : : ; &n ) =
P=Q
 parameter vector  2 D  Rn
^ estimate of 
 asymptotic estimate of 
 closed-loop model error model parameter vector in
Rm
^ estimate of 

Operators and functions

AT transpose of matrix A
A? complex conjugate transpose of matrix A
A 1 inverse of matrix A
arg minx f (x) minimising argument of f (x)
bG;K generalised stability margin
bt (P; r) r-th order system obtained by unweighted balanced
truncation of a higher-order system P
col(A) column vector of length mn obtained by stacking
the columns of the matrix Amn
col(x; y; : : : ) column vector of length nx + ny + : : : obtained by
stacking x, y, : : : , which are column vectors of re-
spective lengths nx , ny , : : :
cov x covariance matrix of the random variable x
diag(a1 ; : : : ; an ) n  n diagonal matrix with entries a1 : : : an
Ex mathematical expectation of the random variable
x P
E f (t) limN !1 N1 Nt=1 E f (t)
f 0 (x0 ), f 00 (x0 ), : : : dx jx=x0 , dx2 jx=x0 , : : :
df (x) d2 f (x)

Im(x) imaginary part of x 2 C

im A range space (image) of the real matrix Amn :
im A = fy 2 Rm j y = Ax; x 2 Rn g
J (G^ r ) model reduction criterion
J (K^ r ) controller reduction criterion
JW C (D; K; !) worst-case performance of the controller K over all
systems in the uncertainty set D
 Operators and functions xxi

ker A null space (kernel) of the real matrix Amn :

ker A = fx 2 Rn j Ax = 0g
Fl ( ; Q) lower LFT of h and Qi: with appropriately parti-
tioned as = 11 12
21 22
, Fl ( ; Q) = 11 + 12 Q(I
1
22 Q) 21
fwbt (P; Wl ; Wr ; r) r-th order system obtained by frequency-weighted
balanced truncation of a higher-order system P
with left and right lters Wl and Wr
Pr (x 6 ) probability that the random variable x is less than

s dierential
R operator: s f (t) = @f@t(t) , s 1 f (t) =
f (t) dt, or Laplace variable
Re(x) real part of x 2 C
trace A
sum of the diagonal entries of the matrix A
VN ; Z N identication criterion
var x variance of the random variable x
@x(t)
x_ (t) @t
z shift operator: z f (t) = f (t +1), z 1 f (t) = f (t 1),
or Z-transform variable
Æ (G1 ; G2)  -gap between two systems G1 and G2
~Æg (G1 ; G2 ) directed gap between two systems G1 and G2
ÆW C (G; D) worst-case  -gap between a system G and a set of


systems D
 G1 ; G2
chordal distance between two systems G1 and G2
W C G; D worst-case chordal distance between a system G
and a set of systems D
i (A) i-th eigen value of the matrix A
max (A) largest eigen value of the matrix A
(M ) stability radius of the loop (M; ), 8 : kk2 < 1
 (A) largest singular value of the (transfer) matrix A
(A) smallest singular value of the (transfer) matrix A
&i (P ) i-th Hankel singular value of the transfer function
(or matrix) P
&(P ) largest Hankel singular value of the transfer func-
tion (or matrix) P : &(P ) = &1 (P ) = kP kH
jxj absolute value of x 2 C
kP kH Hankel norm of P : kP kH = &(P )
xxii Notation glossary

k P k2 2-norm of PR 2 L2 :
kP k22 = 21 11 trace[P ?(j!)P (j!)]d! (continuous
case) or R
kP k22 = 21  trace[P ?(ej! )P (ej! )]d! (discrete
case)
k P k1 1-norm of P 2 L1:
kP k1 = sups2jR  (P (s)) (continuous case) or
kP k1 = supz2@D  (P (z)) (discrete case)
A
B Kronecker product of the 2matrices A and B3:
a11 B : : : a1n B
6 .. ... .. 7
Amn
Bpq = Cmpnq = 4 . . 5
am1 B : : : amn B
 CHAPTER 1
Introduction

1.1. The historical framework

Historically, the leading industry has always required more and more
accurate tools for delivering very valuable products while keeping the
production costs as low as possible. In this context, automatic control
has become one of the most popular latest technologies. Recent advances
in numerical computation and in digital electronics have allowed it to
permeate everybody's life: it is present in your kitchen, in your car, in
your CD player, etc.
In the early days, the control engineer essentially used his/her knowledge
of the process and his/her understanding of the underlying physics to
design very simple controllers, such as proportional-integral-derivative
(PID) controllers. Analysis methods were combined with an empirical
approach of the problem to design controllers with acceptable perfor-
mance.
As the performance specications became more stringent with the ad-
vent of new technologies, the need of precise models from which complex
controllers could be designed became a major issue, which resulted in the
theory of system identication. Quickly, the theorists of system iden-
tication oriented their research towards the computation of the `best'
estimate of a system, from which an optimal controller could be de-
signed using the so-called certainty equivalence principle: the controller
was designed as if the model was an exact representation of the system.
However, as good as it can be, a model is never perfect, with the result
that a controller designed to achieve some performance with this model
1
2 Introduction

may fail to meet the minimum requirements when applied to the true
system. Robust control design is an answer to this problem. It uses
bounds on the modelling error to design controllers with guaranteed
stability and/or performance. Such bounds are generally dened using
some prior knowledge of the physical system (e.g. an a priori description
of the noise process acting on the system such as a hard bound on its
magnitude, or condence intervals around physical parameters of the
system). As the identication theorists did not spend much eort in
trying to produce such bounds in complement to their models, a huge
gap appeared, at the end of the eighties, between robust control and
system identication.
During the last decade, many eorts have been accomplished in order
to bridge this gap, and a new sub-discipline emerged, which has been
called identication for control. However, the earlier works on control-
oriented identication only aimed at producing uncertainty descriptions
that were compatible with those required for doing robust control design,
without paying much attention to the interaction between identication
and model-based control design and to the production of uncertainty
descriptions that would be ideally tuned towards control design (i.e.
allowing the computation of high-performance controllers). It later be-
came clear that a major issue is the interplay between identication and
control, and that identication for control should be seen as a design
problem. This has been highlighted in (Gevers, 1991, 1993).

1.2. Where are we today?

During the nineties, identication for control became a eld of tremen-
dous research activity. Some notable results were produced, all concern-
ing the tuning of the bias and variance errors in a way that would be
suited for control design.
The fact that closed-loop identication could be helpful for tuning the
variance distribution of a model aimed at designing a controller for the
true system was shown for the rst time in (Gevers and Ljung, 1986) for
the case of an unbiased model and a minimum-variance control law. On
the other hand, the publication of (Bitmead et al., 1990) has initiated a
line of research consisting in producing models with a bias error that is
small at frequencies that are critical with respect to closed-loop stability
 The contribution of this thesis 3

or performance. A major result has been to show that closed-loop iden-

tication with appropriate data lters is generally required to deliver
such models.
On the basis of the ideas expressed in these two publications, much re-
search has been carried out in order to combine the identication and
control design in a mutually supportive way. These eorts have resulted
in iterative schemes where steps of closed-loop identication (with the
latest designed controller operating in the loop during data collection)
alternate with steps of control design (using the latest identied model),
the identication criterion matching the control design criterion; see e.g.
(
Astrom, 1993), (Liu and Skelton, 1990), (Schrama, 1992a, 1992b) and
(Zang et al., 1991, 1995). Other very interesting recent results on the in-
terplay between identication and control can be found in (Hjalmarsson
et al., 1994, 1996), (Van den Hof and Schrama, 1995), (De Bruyne, 1996),
and references therein.
For a few years, another line of research has emerged in the eld of model
validation. Until recently, model validation (in the case of prediction
error identication) essentially consisted in computing some statistics
about the residuals (i.e. the simulation errors) of the model. However, as
stressed in (Ljung and Guo, 1997) and (Ljung, 1997, 1998), there is more
information in these residuals than used by classical validation tests. In
particular, the residuals can be used to estimate the modelling error and
to build a frequency domain condence region guaranteed to contain the
true system with some specied probability. The claimed advantage of
this approach is that such a condence region can be used for robust
control design. However, it completely ignores the experimental design
issue of the interplay between identication and control.

1.3. The contribution of this thesis

The contributions of this thesis are in the line of research in identication
for control and model validation. They are essentially focused on the
experimental design aspects of control-oriented system modelling.
A rst contribution concerns the experiment design in closed-loop iden-
tication. As we just mentioned, the recent results on identication for
control have promoted the use of closed-loop identication for producing
models that are supposedly better suited for control design. This method
has gained popularity in the last ten years and is abundantly used in
4 Introduction

its numerous variants (e.g. the so-called direct, indirect and coprime-
factor approaches). Our contribution is to show that, although most
variants of closed-loop identication can be seen as the basic prediction
error method with particular parametrisations of the input-output and
noise models (Forssell and Ljung, 1999), they all have dierent proper-
ties regarding the closed-loop stability of the obtained model with the
controller used during identication, if this controller has unstable poles
or nonminimum phase zeros. We carry out a thorough analysis of this
problem and we give guidelines, regarding both the choice of the method
and that of the excitation input location in the loop, to avoid its emer-
gence. This issue, which has been completely ignored until now, is of
crucial importance because a model that is not stabilised by the current
(to-be-replaced) controller actually proves to be ill-suited for control de-
sign.
The motivation for doing closed-loop identication when the ultimate
goal is control design is the fact that the bias and variance errors at-
tached to the obtained model are tuned towards the control design ob-
jective (contrary to the errors attached to models obtained by open-loop
identication with untuned excitation spectra). On the other hand, we
mentioned the fact that a new line of research has recently emerged in
the eld of model validation for control, the goal being the computation
of condence regions that can be used for the design of controllers with
guaranteed performance and/or stability. The second major contribu-
tion of this thesis is to link these two aspects of modelling for control by
proposing a new closed-loop validation procedure, based on closed-loop
prediction error identication and adapted from the method of (Ljung
and Guo, 1997) and (Ljung, 1997, 1998), in order to produce tuned un-
certainty regions. The relevance of our approach with respect to the
control objective is established by means of analysis tools developed in
(Bombois et al., 1999a, 1999b, 2000c) and is illustrated in two realistic
simulation examples.
A third contribution of this thesis is the proposition of a new control-
oriented criterion for model or controller order reduction, based on the
same reasoning that motivates the use of closed-loop identication for
control. The necessity for such a criterion is motivated by the fact that,
in many practical applications, a high-order model of the plant is ob-
tained by methods like physical (white-box) modelling, nite-element
 The contribution of this thesis 5

modelling, etc. Such a model is generally very accurate, but its high-
orderness makes it ill-suited for control design. Indeed, low-order con-
trollers are generally preferred over high-order ones for numerous rea-
sons (e.g. easiness of implementation or maintenance), but the order of
a controller is generally determined by that of the design model. Two so-
lutions can be considered: model reduction followed by control design, or
design of a high-order controller followed by controller reduction. It is a
common practice to perform the reduction step in open loop, i.e. without
frequency weightings that would re ect the interconnection of the object
subject to reduction within the closed-loop setup. However, in order not
to spoil the energy that has been put in the modelling and/or the control
design steps, the reduction step should be carried out in such a way that
the closed-loop properties of the reduced-order model or controller be
the same as those of the initial high-order one. Our reduction criterion,
which is based on the preservation of the closed-loop transfer function,
fulls this requirement. Since it is based on the reduction of the coprime
factors of the model (resp. of the controllers) with frequency weightings
that depends on those of the controller (resp. of the model), all ob-
jects involved in the reduction procedure are stable, allowing the use
of frequency-weighted balanced truncation as reduction method. Our
model and controller reduction approach is successfully applied to the
design of a low-order controller for a complex model of a pressurised

water reactor (PWR) nuclear power plant supplied by Electricit e de
France (EDF), the national French electricity company, and is com-
pared to other methods of model and controller reduction (namely, un-
weighted coprime-factor reduction, stability-oriented reduction based on
the Bezout identity, and closed-loop balanced truncation). The viability
of closed-loop identication of a low-order model as an alternative to
model reduction is also illustrated by means of the same example.
To summarise this section, we can say that this thesis is based on the now
widely accepted fact that closed-loop identication is generally better
than open-loop identication when the goal is to obtain good models for
control design. Our three major contributions consist in
 showing that closed-loop identication will (perhaps unexpectedly)
deliver models that are ill-suited for control design if the controller
used during identication is unstable and/or nonminimum phase and
if some guidelines we give are not strictly followed;
 proposing a closed-loop validation procedure, based on closed-loop
prediction error identication, which produces tuned uncertainty re-
gions that can be used for robust control design;
6 Introduction

 transposing the heuristic reasoning that motivates the use of closed-

loop identication to the case of model or controller reduction, when
the ultimate goal is the computation of a low-order controller for a
high-order system.

1.4. Synopsis
This thesis is organised as follows.
Chapter 2: Modelling tools, closed-loop setup and generalities.
This chapter reviews the modelling and analysis tools that are used in
this thesis, namely prediction error identication, balanced truncation,
coprime factorisations and the  -gap metric. It also contains a descrip-
tion of the closed-loop setup that is considered in this thesis, as well as
the denitions of closed-loop stability and of generalised stability mar-
gin.
Chapter 3: A motivating discussion of issues in identication
for control. This chapter contains a summary of the basic motivations
for performing the identication of a system in closed loop when the
objective is to use the model for control design.
Chapter 4: Identication in closed loop with an unstable or
nonminimum phase controller. It often happens that the controller
used during identication is unstable or nonminimum phase. This chap-
ter shows that, in this case, the closed-loop stability of the resulting
model (when connected to the same controller) cannot always be guar-
anteed (sometimes, instability is guaranteed) if precautions are not taken
regarding the identication method and the excitation source. Further-
more, it is shown that a model that is destabilised by the controller used
during identication is not good for control design. Guidelines are given
to avoid this problem.
Chapter 5: Model validation for control and controller vali-
dation. This chapter explains how closed-loop data can be used to
compute a parametrised set of transfer functions that is guaranteed to
contain the true system with some probability, and that is better tuned
towards robust control design than a similar set that would be computed
from open-loop data. Such a set can be used for the validation of a given
controller. Our validation procedure is illustrated by means of two re-
alistic application examples: a exible transmission system (resonant
 Synopsis 7

mechanical system subject to step disturbances) and a ferrosilicon pro-

duction process (chemical process subject to stochastic noise); in both
cases its supremacy over open-loop validation is established.
Chapter 6: A performance-based approach to control-oriented
model reduction and to controller reduction. This chapters shows
how the motivations for performing the identication of a system in
closed loop can be transposed to model or controller reduction. A
performance-based reduction criterion is proposed and successfully ap-
plied to the design of a low-order controller for a complex PWR plant
model.
Chapter 7: Future issues on variance and on model order se-
lection. Our validation procedure of Chapter 5 is based on the identi-
cation of an unbiased model. This means that undermodelling has to be
avoided at all cost. But what about overmodelling? This chapter aims
at giving some insight about the eect of parameters in excess in the
identied model on the variance distribution over frequency. Another
question tackled in this chapter is that of the choice, with respect to the
variance of the nal estimate, between direct identication of a low-order
model and identication of a full-order model followed by model reduc-
tion. Only preliminary results are given, which could serve as starting
points for future research.
Chapter 8: Conclusions. This chapter concludes this thesis and
proposes some possible further research topics.
8 Introduction
 CHAPTER 2
Modelling tools, closed-loop setup and
generalities

2.1. Introduction
This chapter describes modelling and analysis tools that will be used in
this thesis.
In the rst part of this chapter we review two modelling tools commonly
used in industrial practice and which we shall consider throughout this
thesis. The rst one, described in Section 2.2, is prediction error (PE)
system identication. It uses data measured on the actual plant to com-
pute the best estimate in some parametrised model set with respect to
a criterion that minimises the H2 norm of the prediction errors. The
second one, described in Section 2.3, is model reduction by balanced
truncation. It uses knowledge of a precise linear high-order model of
the plant (obtained, for instance, by identication, rst-principle-based
physical modelling, nite-element modelling, or linearisation of a non-
linear simulator) to derive a lower-order model by discarding the least
controllable and observable modes of the high-order model.
In the second part of this chapter we brie y dene two possible repre-
sentations of a system G0 in closed-loop with some stabilising controller
K , namely a general representation based on a standard block-diagram
description of the closed loop (Section 2.4) and on which we shall base
the notions of closed-loop stability and generalised stability margin used
throughout this thesis, and a linear fractional transformation (LFT)-
based representation (Section 2.5), which is more convenient for manip-
ulating multivariable closed-loop systems. We show in Section 2.5 that
9
10 Modelling tools, closed-loop setup and generalities

any closed-loop system represented by means of a LFT can be put in the

general form dened in Section 2.4, allowing to use standard stability
analysis tools for closed-loop systems described by LFT's.
Finally, the third part of this chapter concerns two useful tools which
will often be used in this thesis, namely coprime factorisations and the
 -gap metric. The notions of left and right coprime factorisations of a
system are dened in Section 2.6. Procedures are given to build such
(possibly normalised) factorisations. The  -gap between two transfer
matrices is dened in Section 2.7.

2.2. Prediction error identication

2.2.1. Assumptions about the system

We make the assumption that the true system is the possibly multi-input
multi-output (MIMO) linear time-invariant (LTI) system described by
(
y(t) = G0 (z )u(t) + v(t)
S: v(t) = H0 (z )e(t)
(2.1)

where G0 (z ) and H0 (z ) are rational transfer functions or matrices. G0 (z )

is strictly proper and has p outputs and m inputs. H0 (z ) is a stable and
inversely stable, minimum phase, proper and monic lter. Here z 1 is
the backward shift operator (z 1 u(t) = u(t 1)), u(t) 2 Rm is the control
input signal, y(t) 2 Rp is the observed output signal and e(t) 2 Rp is a
white noise process with zero mean and covariance matrix 0 .
The assumption is made that all signals are quasi-stationary (Ljung,
1999), hence we can dene spectra and cross-spectra of signals, which
are denoted u (!), ue (!), etc.

2.2.2. The identication method

The objective of system identication is to compute a parametrised

model M() for the system:
M() : y^(t) = G(z; )u(t) + H (z; )e(t): (2.2)
This model lies in some model set M selected by the designer:
M , fM() j  2 D  Rn g ; (2.3)
 Prediction error identication 11

i.e. M is the set of all models with the same structure as M(). The
parameter vector  ranges over a set D  Rn which is assumed to be
compact and connected.
We say that the true system is in the model set, which is denoted S 2 M,
if
90 2 D : G(z; 0) = G0; H (z; 0) = H0 : (2.4)
Otherwise, we say that there is undermodelling of the system dynamics.
The case where the noise properties cannot be correctly described within
the model set, but where
90 2 D : G(z; 0) = G0; (2.5)
will be denoted G0 2 G.
The prediction error identication procedure uses a nite set of N input-
output data
Z N = fu(1); y(1); : : : ; u(N ); y(N )g (2.6)
to compute at each time sample t 2 [1; N ] the one-step ahead prediction
according to the model that corresponds to the parameter vector :


y^(t j t 1; ) = H 1 (z; )G(z; )u(t) + 1 H 1 (z; ) y(t) (2.7)
(observe that, due to the fact that G(z; ) is strictly proper and that
H (z; ) is monic, the two terms of the right-hand side only depends on
past u's and y's). The prediction error at time t is


"(t; ) = y(t) y^(t j t 1; ) = H 1 (z; ) y(t) G(z; )u(t) : (2.8)
Given the chosen model structure (2.2) and measured data (2.6), the
prediction error estimate of  is determined through


^ = arg min VN ; Z N (2.9)
2D
N
N
1X
VN ; Z = "T (t; ) 1 "F (t; ) (2.10)
N t=1 F
"F (t; ) = L(z; )"(t; ) (2.11)
where  is a symmetric positive denite weighting matrix and L(z; )
is any possible linear, stable, monic, and possibly parametrised prelter
(observe that, if L(z; ) = H (z; ), "F (t; ) is the simulation error at
time t). Since


"F (t; ) = L(z; )H 1 (z; ) y(t) G(z; )u(t) ; (2.12)
12 Modelling tools, closed-loop setup and generalities

this lter can be included in the noise model structure and, without
loss of generality, we shall make the assumption that L(z; ) = I in the
sequel.
The following notation will often be used for the estimates G(z; ^),
H (z; ^), etc.:
G^ (z ) = G(z; ^) and H^ (z ) = H (z; ^): (2.13)

2.2.3. Classical model structures

Some commonly used model structures are the following.

 FIR (Finite Impulse Response model structure):
y(t) = B (z )u(t) + e(t); (2.14)
 ARX (Auto-Regressive model structure with eXogenous inputs):
A(z )y(t) = B (z )u(t) + e(t); (2.15)
 ARMAX (Auto-Regressive Moving-Average model structure with eX-
ogenous inputs):
A(z )y(t) = B (z )u(t) + C (z )e(t); (2.16)
 OE (Output Error model structure):
y(t) = F 1 (z )B (z )u(t) + e(t); (2.17)
 BJ (Box-Jenkins model structure):
y(t) = F 1 (z )B (z )u(t) + D 1 (z )C (z )e(t): (2.18)

B (z ) is a polynomial or a polynomial matrix in z 1 with delay larger

or equal to 1 (i.e. the coeÆcient of z 0 is 0), while A(z ), C (z ), D(z )
and F (z ) are monic polynomials or polynomial matrices in z 1 (i.e. the
coeÆcient of z 0 is 1).

2.2.4. Computation of the estimate

For the sake of simplicity, we only consider the single-input single-output

(SISO) case in this subsection. The derivations can easily be extended
to the MIMO case.
 Prediction error identication 13

A. The FIR and ARX cases. With a FIR or ARX model structure,
the model is linear in the parameters (here, na is the order of A(z ), nb
is the order of B (z ) and nk 6 nb is the delay):
M() : y^(t) = 'T (t) + e(t); (2.19)
hence
y^(t j t 1; ) = 'T (t); (2.20)
where
 T
 = a1 : : : ana bnk : : : bnb (2.21)
and
'(t) =
 T
y(t 1) : : : y(t na ) u(t nk ) : : : u(t nb ) (2.22)
are respectively a parameter vector

and a regression vector. The min-
imising argument of VN ; Z N is then obtained by the standard least-
squares method:
" # 1
N N
1 X 1 X
^ = '(t)' (t)
T
'(t)y(t): (2.23)
N t=1
N t=1

B. Other cases. Other model structures require some numerical opti-

misation routine to nd the estimate. A standard search routine is the
following:
 

^i+1 = ^i i RNi 1 VN0 ^i ; Z N ; (2.24)
where ^i is the estimate at iteration i,
N

1 X


VN0 ^i ; Z N = t; ^i " t; ^i (2.25)
N t=1


is the gradient of VN ; Z N with respect to  evaluated at ^i , RNi
is a matrix that determines the search direction, i is a factor that
determines the step size and
d d
(t; ) = "(t; ) = y^(t j t 1; ) (2.26)
d d
14 Modelling tools, closed-loop setup and generalities

is the negative gradient of the prediction error. A common choice for

the matrix RNi is
N
1 X


RNi = t; ^i + Æi I T
t; ^i (2.27)
N t=1
which gives the Gauss-Newton direction if Æi = 0; see e.g. (Dennis
and Schnabel, 1983). Æi is a regularisation parameter chosen such that
RNi > 0.

2.2.5. Asymptotic properties of the estimate

A. Consistency. Under mild conditions there holds (Ljung, 1978)

VN ; Z N ! V () = E"
 T (t; ) 1 "(t; ) w.p. 1 as N ! 1 (2.28)
and
^ !  = arg min V () w.p. 1 as N ! 1: (2.29)
2D

Note that we can write V () as

V () = E trace  1 "(t; )"T (t; )
Z
1 
(2.30)
= trace  1 " (!) d!
2 
where " (!) is the power spectral density of the prediction error "(t; ).
The second equality comes from Parseval's relationship. As a result,
there holds1
Z    
   u ue
 = arg min trace G0 G() H0 H () 
2D  eu 0

?  
G G ( )
 H H ()
? H ()H () d! (2.31)
0 ?

1
0
or
Z  h


?
 = arg min trace G0 G() + B () u G0 G() + B ()
2D 



? i
+ H0 H () 0 euu 1 ue H0 H ()


1
 H ()H () ?
d! (2.32)
1
The ej! or ! arguments of the transfer functions or spectra will often be deleted to
simplify the notation.
 Prediction error identication 15

where


B (ej! ; ) = H0 (ej! ) H (ej! ; ) eu(!)u 1 (!) (2.33)
is a bias term that will vanish only if eu = 0, i.e. if the data are
collected in open loop so that u and e are uncorrelated, or if the noise
model H () is exible enough so that S 2 M; see (Forssell, 1999).
B. Asymptotic variance in transfer function space. The asymp-
totic covariance of the estimate, as N and n both tend to innity, is
given by
  
 n u (!) ue (!) T
cov col G(e ) H (e ) 
^ j! ^ j!
v (!) (2.34)
N eu (!) 0
where
denotes the Kronecker product. In open loop, ue = 0 and

n
cov col G^ (ej! )  u T (!)
v (!); (2.35)
N

n
cov col H^ (ej! )  0 1
v (!): (2.36)
N
In the SISO case (in open loop), this becomes

n v (!)
cov G^ (ej! )  ; (2.37)
N u (!)

n 
cov H^ (ej! )  H0 (ej! )2 : (2.38)
N
These results were established in (Ljung, 1985) for the SISO case and in
(Zhu, 1989) for the MIMO case.
Remark. These asymptotic variance expressions are widely used in
practice, although they are not always reliable. It has been shown in
(Ninness et al., 1999) that their accuracy could depend on choices of
xed poles or zeros in the model structure, and alternative variance ex-
pressions with greatly improved accuracy, and which make explicit the
in uence of any xed poles or zeros, are given. Observe for instance
that the use of a xed prelter L(z ) during identication amounts to
impose xed poles and/or zeros in the noise model. In this case, the
extended theory of (Ninness et al., 1999) should be used, the more so if
the number of xed poles and zeros is large with respect to the model
order.

C. Asymptotic variance in parameter space. The parameter vec-

tor estimate tends to a random vector with normal distribution as the
16 Modelling tools, closed-loop setup and generalities

number of data N tends to innity (Ljung, 1999):

^ !  w.p. 1 as N ! 1; (2.39a)
p ^ 
N (  ) 2 AsN (0; P ); (2.39b)

where

P = R 1 QR 1 ; (2.39c)
R = V 00 ( ) > 0; (2.39d)


 
Q = lim N
E VN0  ; Z N VN0 ; Z N T : (2.39e)
N !1

2.3. Balanced truncation

Here, we make the assumption that a strictly stable high-order
continuous-time model Gn(s) of the system is available (see Subsec-
tion 2.3.5 for some remarks on the discrete-time case), and that it has
the following state-space representation:
 
A B ;
Gn = C (2.40a)
D

i.e.

Gn (s) = C (sI A) 1 B + D (2.40b)

where s is the Laplace variable or the time-derivative operator (sx(t) =

x_ (t)). The balanced truncation procedure, that was rst proposed by
B.C. Moore (Moore, 1981), consists in computing a state-space represen-
tation of Gn(s) in which the most controllable modes coincide with the
most observable ones. The least observable and controllable modes are
then discarded. This procedure can be extended to the case where fre-
quency weightings are used (Enns, 1984a, 1984b). We shall rst brie y
review the notions of controllability and observability, and then describe
the balanced truncation procedure without and with frequency weight-
ings.
 Balanced truncation 17

2.3.1. The concepts of controllability and observability

A. Controllability.
Definition 2.1. (Controllability) A state x0 2 R n of the system
Gn is called controllable if there exists a control input signal u(t), t 2
[0; T ], that brings the system from initial state x(0) = x0 to x(T ) = 0 in
a nite time T . The controllable subspace of Gn, X cont  Rn , is the set
of all controllable states of Gn. Gn is called controllable if X cont = Rn .

The controllability matrix of Gn is

 
C , B AB : : : An 1 B : (2.41)
The columns of C generate X cont . As a result, Gn is fully controllable
if and only if rank C = n.
The controllability Gramian of Gn at time 0 6 t < 1 is dened as
Z t
T
P(t) , eA BB T eA  d; t 2 R+ : (2.42)
0
It has the following properties:
 8t 2 R+ P(t) = PT (t) > 0, i.e. it is symmetric and positive semi-
denite;
 8t 2 R+ im P(t) = im C, i.e. the columns of P(t) generate X cont
for all positive t.
For a strictly stable system, the innite controllability Gramian is de-
ned as
Z 1
T
P, eA BB T eA  d: (2.43)
0
It has the same properties as P(t). As a result, Gn is fully controllable if
and only if P > 0, i.e. if it is positive denite. This innite controllability
Gramian, which we shall simply call the controllability Gramian in the
sequel, can be computed by solving the following Lyapunov equation:
AP + PAT + BB T = 0: (2.44)

The controllability Gramian P gives the minimum input energy that is

necessary to bring the system from the free equilibrium x( 1) = 0 to
18 Modelling tools, closed-loop setup and generalities

a state x(0) = x0 :
Z 0 
min u (t)u(t)dt j x(0) = x0 = xT0 P 1 x0
T
u 1
when x( 1) = 0: (2.45)
B. Observability.
Definition 2.2. (Observability) A state x0 2 Rn
of the system Gn
is called unobservable if the free response of the output of Gn to this state
is identically zero for all t > 0, i.e. if this state cannot be distinguished
from the zero state. The unobservable subspace of Gn, {Rn X obs  Rn ,
is the set of all unobservable states of Gn . Gn is called observable if
{Rn X obs = f0g or, equivalently, if X obs = Rn0 .
The observability matrix of Gn is
2 3
C
6 CA 7
O,6 6 ..
7
7: (2.46)
4 . 5
CAn 1
The null space of O is {Rn X obs , hence Gn is fully observable if and only
if rank O = n.
The observability Gramian of Gn at time 0 6 t < 1 is dened as
Z t
T (t  ) T
Q(t) , eA C CeA(t  ) d; t 2 R+ : (2.47)
0
It has the following properties:
 8t 2 R+ Q(t) = QT (t) > 0, i.e. it is symmetric and positive semi-
denite;
 8t 2 R+ ker Q(t) = ker O, i.e. the null space of Q(t) is {Rn X obs for
all positive t.
For a strictly stable system, the innite observability Gramian is dened
as
Z 1
T
Q, eA  C T CeA d: (2.48)
0
It has the same properties as Q(t). As a result, Gn is fully observable if
and only if Q > 0, i.e. if it is positive denite. This innite observability
 Balanced truncation 19

Gramian, which we shall simply call the observability Gramian in the

sequel, can be computed by solving the following Lyapunov equation:
AT Q + QA + C T C = 0: (2.49)

The observability Gramian Q gives the total energy of the free output
response of the system to an initial state x(0) = x0 :
Z 1
yT (t)y(t)dt = xT0 Qx0
0
when u(t) = 0 8t > 0 and x(0) = x0 : (2.50)

2.3.2. Balanced realisation of a system

The following observations can be made regarding the controllability

and observability Gramians:
 the controllability and observability Gramians P and Q depend on
the state-space representation of the system Gn;
 their eigenvalues give information about the `level' of observability or
controllability of the state variables;
 depending on the chosen representation (2.40), some state variables
(i.e. some dynamics) can be very observable but little controllable,
or vice-versa.
If the representation (2.40) is minimal, i.e. if it is both controllable and
observable (P > 0 and Q > 0), it is possible to nd a state transforma-
tion that puts the system in a form where the most observable dynamics
are also the most controllable ones. This is called a balanced realisation.
When the system is in balanced form, its Gramians are diagonal and
equal:
P = Q =  , diag(&1 ; : : : ; &n ); (2.51)
where the &i 's are the Hankel singular values of Gn in decreasing order:
p
&i = i (PQ) > 0; &i > &i+1 : (2.52)
Observe that, although the Gramians depend on the state realisation
of the system, the Hankel singular values are independent of that re-
alisation. Once a balanced realisation has been obtained, it is possible
to reduce the order of the system by simply discarding the state vari-
ables that correspond to the least observable and controllable dynamics.
20 Modelling tools, closed-loop setup and generalities

Indeed, these are the dynamics that contribute the least to the global
input-output behaviour of the system.
Starting from any realisation (2.40) of Gn , one can compute a balanced
state-space representation as follows.

1. Compute the controllability and observability Gramians P and Q by

solving the Lyapunov equations (2.44) and (2.49).
2. Compute  = diag(&1 ; : : : ; &n ) where the &i 's are given by (2.52).
p
3. Compute R such that P = RT R. It can be obtained as R =  T
where  is a diagonal matrix containing the eigenvalues of P and 
is a matrix whose columns are the eigenvectors of P associated to the
entries of .
4. Compute U such that RQRT = U 2 U T (singular value decomposi-
tion).
5. Then T = 1=2 U T R T is the balancing transformation matrix and
there holds T PT T = T T QT 1 = .
6. Compute
A = T AT 1 ; (2.53a)
B = T B; (2.53b)
C = CT 1; (2.53c)
D = D: (2.53d)
Then,
" #
A B
G n = (2.54)
C D
is a balanced realisation of Gn (s).

2.3.3. Balanced truncation

The diagonal of  contains the Hankel singular values of Gn in decreasing

order. The state variables of the balanced realisation G n follow the same
order: xi is more observable and controllable than xj for 1 6 i < j 6 n.
More precisely, G n is more observable and controllable in the direction
of ~ei than in the direction of ~ej for 1 6 i < j 6 n. Let us then consider
 Balanced truncation 21

the following partition:

 
 = 011 0 ; (2.55a)
22
11 = diag(&1 ; : : : ; &r ); (2.55b)
22 = diag(&r+1 ; : : : ; &n ): (2.55c)
The corresponding partition of G n is
2 3
A11 A12 B1
G n = 6 4 A21 A22 B2 7 5: (2.56)
 
C1 C2 D 
Let us dene
" #
11 B1
A
G^ r = 
C1 D (2.57)
, bt (Gn; r)
which is obtained by truncating the least observable and controllable
dynamics of G n . The obtained reduced-order system G^ r is a stable
suboptimal solution to the following minimisation problem:
min
Gr (s)
kGn(s) Gr (s)k1 ;
of order r
(2.58)
which is hard to solve exactly. Upper and lower bounds of the reduction
error in the H1 norm can easily be computed from the Hankel singular
values of Gn (Glover, 1984), (Enns, 1984a, 1984b):
n
X
&r 6 Gn (s) G^ r (s) 62 &i : (2.59)
1 i=r+1
The level r of truncation can be chosen by plotting the &i 's: it is best
to have &r >> &r+1 , which means that the dynamics corresponding to
&r+1 ; : : : ; &n can really be neglected because of their relatively poor de-
gree of observability and controllability. A tighter upper bound in (2.59)
can be obtained by counting the &i 's of multiplicity larger than 1 only
once in the summation.

2.3.4. Frequency-weighted balanced truncation

A very common case is when the reduction criterion contains stable

input and/or output frequency weighting lters Wr and/or Wl . The
22 Modelling tools, closed-loop setup and generalities

minimisation problem becomes

min Wl (s) Gn (s) Gr (s) Wr (s) ; (2.60)
Gr (s) of order r 1
to which a suboptimal solution can be computed by frequency-weighted
balanced truncation.
Let us consider any minimal realisation of the stable system Gn (s) de-
ned in (2.40), and realisations of the two lters as follows:
   
A
Wl = C D ; l Bl
Wr = C D :A r Br
(2.61)
l l r r

Then, a realisation of Wl (s)Gn (s)Wr (s) , G~ n(s) = C~ (sI A~) 1 B~ + D~

is given by
2 3
" # Al Bl C Bl DCr Bl DDr
A~ B~ 6 0 A BCr BDr 7
G~ n = ~ ~ = 6 4 0 0 Ar Br 5 :
7 (2.62)
C D
Cl Dl C Dl DCr Dl DDr
The controllability and observability Gramians of this input-output
frequency-weighted system are respectively the solutions of the following
Lyapunov equations:
A~P~ + P~ A~T + B~ B~ T = 0; (2.63)
A~T Q
~ +Q ~ A~ + C~ T C~ = 0: (2.64)
These Gramians can be partitioned similarly to A~ in the state realisation
(2.62) of G~ n :
2 3 2 3
P11 P12 P13 Q11 Q12 P13
~ = 4P21 P22 P23 5 ;
P Q~ = 4Q21 Q22 P23 5 : (2.65)
P31 P32 P33 Q31 Q32 P33
P22 and Q22 are then the frequency-weighted controllability and observ-
ability Gramians of Gn (s). If Wr (s) 6= I , then P22 6= P given by (2.44),
which means that the input weighting lter modies the controllability
Gramian of the system. Similarly, if Wl (s) 6= I , then Q22 6= Q given
by (2.49), which means that the output weighting lter modies the
observability Gramian of the system.
The rest of the procedure consists in nding a transformation matrix T
such that T P22 T T = T T Q22 T 1 =  = diag(&1 ; : : : ; &n ) where the &i 's
are the frequency weighted Hankel singular values of Gn(s), and to apply
this transformation to the realisation (2.40). This produces a frequency-
weighted balanced realisation G n of the system Gn(s). Its order can
 Balanced truncation 23

then be reduced, as in unweighted balanced truncation, by discarding

the modes corresponding to the smallest Hankel singular values. In the
sequel, we shall denote
G^ r = fwbt (Gn ; Wl ; Wr ; r) (2.66)
the r-th order system produced by frequency-weighted balanced trunca-
tion of Gn(s). An upper bound of the approximation error in the H1
norm is given by (Kim et al., 1995)


Wl (s) Gn(s) Gr (s) Wr (s) 1
n q
X
62 &i2 + (i + i )&i3=2 + i i &i (2.67)
i=r+1
where

i = ki 1 (s)k1 Cr r (s)P133=2 ; (2.68a)
1
i = Q111=2 l (s)Bl k (s)k ;
1 i 1 1
(2.68b)
i 1 (s) = Ai21 1 (sI Ai 1 ) 1 Bi 1 + bi ; (2.68c)
i 1 (s) = Ci 1 (sI Ai 1 ) 1 Ai12 1 + ci ;(2.68d)
r (s) = (sI Ar ) 1 ; (2.68e)
l (s) = (sI Al ) ; 1 (2.68f)
 i 1 
Ai = Aii 11 A12 ; (2.68g)
A21 aii
 
Bi = Bbi 1 ; (2.68h)
i
 
Ci = Ci 1 ci ; (2.68i)
and bi and ci are the i-th row of Bi and the i-th column of Ci , respec-
tively, and An = A, Bn = B , Cn = C .
Let us nally remark that the reduced-order model G^ r is guaranteed to
be stable only if frequency weighting is used on one single side of the
reduction criterion.

2.3.5. Balanced truncation of discrete-time systems

The balanced truncation and frequency-weighted balanced truncation

procedures are exactly the same in the discrete-time case, except that the
24 Modelling tools, closed-loop setup and generalities

controllability and observability Gramians are respectively the solutions

of the discrete Lyapunov equations
APAT P + BB T = 0 (2.69)
and
AT QA Q + C T C = 0: (2.70)

Observe that, since the H2 norm of a stable system is upper bounded

by its H1 norm in the discrete-time case (Boyd and Doyle, 1987), the
H1 upper bound of the approximation error, computed from the Hankel
singular values (both in the unweighted and one-side weighted cases), is
also an H2 upper bound of it.

2.4. General representation of a closed-loop system

and closed-loop stability
2.4.1. General closed-loop setup

v(t)
r2 (t) - f u(t)
+ - G0 - ?f+ y(t-
)
6 +

f (t)
K  g ( t) + ?f r1 (t)

Figure 2.1. General representation of a system in

closed loop

We shall often consider the representation of Figure 2.1 for a system

G0 in closed loop with a controller K . In this representation, u(t) is
the input of the plant, y(t) its output, v(t) an output disturbance that
can be described by v(t) = H0 (z )e(t) (discrete time case) or v(t) =
H0 (s)e(t) (continuous time case) where e(t) is white noise and H0 some
stable, minimum phase, monic transfer function. r1 (t) and r2 (t) are two
possible sources of exogenous signals (references). g(t) and f (t) denote
respectively the input and the output of the controller.
 General representation of a closed-loop system and closed-loop stability 25

2.4.2. Closed-loop stability

Let us consider the closed-loop setup of Figure 2.1. In mainstream robust

control, the following generalised closed-loop transfer matrices are often
considered2:
" #
G0 (I + KG0) 1 K G0 (I + KG0 ) 1
Ti (G0 ; K ) = (2.71)
(I + KG0 ) 1 K (I + KG0) 1
and
" #
K (I + G0K ) 1 G0 K (I + G0 K ) 1
To (G0 ; K ) = : (2.72)
(I + G0 K ) 1 G0 (I + G0 K ) 1
The entries of Ti (G0 ; K ) are the transfer functions between the exoge-
nous reference signals and the input and output signals of the plant:
   
y(t) = T (G ; K ) r1 (t) + N (G ; K )H e(t) (2.73)
u(t) i 0 r2 (t) i 0 0

where
 1 
( I + G 0
Ni (G0 ; K ) = (I + KG ) 1 K ; K ) (2.74)
0
while those of To (G0 ; K ) are the transfer functions between the exoge-
nous reference signals and the output and input signals of the controller:
   
f (t) = T (G ; K ) r2 (t) + N (G ; K )H e(t) (2.75)
g(t) o 0 r1 (t) o 0 0

where
 1H 
K ( I + G
No (G0 ; K ) = (I + G K ) 1 0 :
0 K ) (2.76)
0

In the SISO case, we dene more simply

2 G K G0 3
0
 
6 1 + G0 K 1 + G0 K 7 T T
T (G0 ; K ) = 6 7, 11 12 (2.77)
4 K 1 5 T21 T22
1 + G0K 1 + G0 K
2
All transfer functions and matrices must be understood as rational functions of the
Z-transform variable z (discrete-time case) or of the Laplace variable s (continuous-
time case). However, when no confusion is possible, we shall often omit these symbols
in order to ease the notations.
26 Modelling tools, closed-loop setup and generalities

and
2 1 3
   
6 1 + G0 K 7
N (G0 ; K ) = 4
6 7
5 = T , N
T22
N
1 (2.78)
K 21 2
1 + G0 K
such that
   
y(t) r1 (t)
u(t) = T (G0 ; K ) r2 (t) + N (G0 ; K )H0 e(t): (2.79)

Definition 2.3. (Internal stability) The closed loop (G0 ; K ) of

Figure 2.1 is called `internally stable' if all four entries of Ti (G0 ; K ) or,
equivalently, all four entries of To (G0 ; K ), are stable, i.e. if they belong
to H1.

The generalised stability margin bG0 ;K is an important measure of the

internal stability of the closed loop. It is dened as
(
bG0 ;K ,
kTi(G0; K )k11 if (G0; K ) is stable, (2.80)
0 otherwise.
Note that kTi (G0 ; K )k1 = kTo (G0 ; K )k1 , as shown in (Georgiou and
Smith, 1990). An alternative denition, in the SISO case, is the follow-
ing:
 
1
bG0 ;K = min  G0 (e ); j!
(2.81)
! K (ej! )


where  G0 (ej! ); K (e1j! ) is the chordal distance at frequency ! between
G0 and K (e1j! ) , as dened in (Vinnicombe, 1993a): see Section 2.7.
The margin bG0 ;K plays an important role in robust optimal control
design. The following results hold in the SISO case (Vinnicombe, 1993b):
1 + bG0 ;K
gain margin > (2.82)
1 bG0 ;K
and
phase margin > 2 arcsin (bG0 ;K ) : (2.83)
 General representation of a closed-loop system and closed-loop stability 27

Note that there is a maximum attainable value of the generalised stabil-

ity margin bG0 ;K over all controllers stabilising G0 (Vinnicombe, 1993b):
s  
q N0 2
sup bG0 ;K = 1 max (Prcf Qrcf ) = 1
M0 (2.84)
K H

and (Georgiou and Smith, 1990)

 
sup bG0 ;K 6 inf  N0 : (2.85)
K z 2{D c M0
or
s2C +0

h i
Here, M N0
0
is a normalised right coprime factorisation of G0 (see Sec-
tion 2.6). Prcf and Qrcf are, respectively, the controllability
h i and ob-
N0
servability Gramians of the right coprime factors M0 . It should be
clear that it is easier to design a stabilising controller for a system with
a large supK bG0 ;K than for a system with a small one.
We refer the reader to (Zhou and Doyle, 1998, Chapter 16 and Sec-
tion 6.1) and references therein for more details about the links between
the generalised stability margin and controller performance.
The following proposition will be used several times in the sequel of this
thesis.
^ K ) be
Proposition 2.1. (Anderson et al., 1998) Let (G0 ; K ) and (G;
two stable closed-loop systems such that

^
Ti (G; K ) Ti (G0 ; K ) < " (2.86a)
1
or, equivalently,

^
To (G; K) To (G0 ; K ) <" (2.86b)
1
for some " > 0. Then,
 

bG;K
^ bG0 ;K  < ": (2.87)
28 Modelling tools, closed-loop setup and generalities

2.5. LFT-based representation of a closed-loop

system
In the MIMO case, it is often easier to represent a closed-loop system us-
ing Linear Fractional Transformations, as depicted in Figure 2.2, where
0 is called the standard plant and Q the standard controller.

z (t)     w(t)
= 11 12
0
21 22 
h(t) f (t)
- Q

Figure 2.2. LFT representation of a system in closed loop

In this representation, all exogenous signals are contained in w(t). f (t)

is the control signal dened above and h(t) is the input of the controller
Q, e.g. h(t) = g(t) if Q = K . z (t) contains all inner signals that are
useful for the considered application. For instance, if the objective is to
design a control law for the tracking of the reference signal r1 (t), z (t)
will typically contain the tracking error signal g(t) = y(t) r1 (t) and,
possibly, the control signal f (t) if the control energy is penalised by the
control law.
The closed-loop transfer function between w(t) and z (t) is given by

Tzw = Fl ( 0 ; Q) , 11 + 12 Q(I 22 Q)
1
21 : (2.88)

We now distinguish two possible cases corresponding to two dierent

controller structures.
 LFT-based representation of a closed-loop system 29

Case 1: A one-degree-of-freedom controller K is used. Consider

Figure 2.2. We set
Q = K; (2.89a)


w(t) = col r2 (t); r1 (t); v(t) ; (2.89b)
g(t) = y(t) r1 (t); (2.89c)


z (t) = col f (t); g(t) ; (2.89d)
h(t) = g(t): (2.89e)
In this case, Figure 2.2 is equivalent to Figure 2.1 and Tzw (G0 ; K ) =
To (G0 ; K ) No (G0 ; K ) if
   
= 0 0 0 ; = I ; (2.90a)
11 G0 I I 12 G0
   
21 = G0 I I ; 22 = G0 : (2.90b)
 
Case 2: A two-degree-of-freedom controller C = F K is used.
We make the assumption that r1 (t) and y(t) do not necessarily have the
same size. The control law is given by
f (t) = Ky(t) F r1 (t) (2.91)
and thus
 
Q=C= F K ; (2.92)
and we replace the tracking error g(t) = y(t) r1 (t) in z (t) by a linear
combination of r1 (t) and y(t):


g(t) = W  col r1 (t); y(t) ; (2.93)
where W is some (possibly dynamic) weighting matrix. The other signals
in Figure 2.2 are


w(t) = col r2 (t); r1 (t); v(t) ; (2.94a)


z (t) = col f (t); g(t) ; (2.94b)


h(t) = col r1 (t); y(t) ; (2.94c)
and the entries of 0 are
2 3 2 3
0 I
6  7 ; 6  7 ;
11 = 4 0 I 0 5 12 = 4 0 5 (2.95a)
W G 0 I W G
0 0
   
0 I 0 0 :
21 = G 0 I ; 22 = G0 (2.95b)
0
30 Modelling tools, closed-loop setup and generalities

With these 0 and Q, we have

Tzw (G0 ; C ) = Fl ( 0 ; Q)
 
= I0 W
0

2
3
K (I + G0 K ) 1 G0 I K (I + G0 K ) 1 G0 F K (I + G0 K )
 64 0
 
I
 
0
7
5
(I + G0 K ) 1 G0 (I + G0 K ) 1 G0 F (I + G0 K ) 1

    
= I0 W
0 T
o
0 ; F K  ;
G0
(2.96)
hence
h ai standard
 
representation with closed-loop transfer matrix
0
To G0 ; F K can be attached to the LFT representation of the
closed loop.

2.6. Coprime factorisations

We shall often use left and right coprime factorisations of systems and
controllers in the sequel. Here, we dene formally the notion of coprime-
ness, and we explain how to compute the coprime factors of a dynamic
system. Details can be found in, e.g., (Francis, 1987), (Vidyasagar, 1985,
1988) and (Varga, 1998).
Definition 2.4. (Coprimeness) Two matrices M ~ and N~ in RH1 are
left coprime over RH1 if they have the same number of rows and if
there exist matrices Xl and Yl in RH1 such that
 
  Xl
~ ~
M N Y = MX ~ l + NY
~ l = I: (2.97)
l
Similarly, two matrices M and N in RH1 are right coprime over RH1
if they have the same number of columns and if there exists matrices Xr
and Yr in RH1 such that
 
  N
Yr Xr M = Yr N + Xr M = I: (2.98)
Definition 2.5. (Left coprime factorisation) Let P be a proper
real rational transfer matrix. A left coprime factorisation of P is a
factorisation P = M ~ 1 N~ where M~ and N~ are left coprime over RH1 .
 Coprime factorisations 31

 
If P is detectable, a left coprime factorisation N~ M~ can be con-
structed according to the following proposition.
Proposition 2.2. Let (A; B; C; D ) be a detectable realisation of a
transfer matrix P , let L be any constant injection matrix stabilising the
output of P , and let Z be any nonsingular matrix. Dene
 

~ ~  A + LC B + LD L
N M = ZC ZD Z ; (2.99)
i.e. (continuous-time case)
 
N~ (s) M~ (s)

1   
= ZC sI (A + LC ) B + LD L + ZD Z (2.100a)
or (discrete-time case)
 
N~ (z ) M~ (z )

   
= ZC zI (A + LC ) 1 B + LD L + ZD Z : (2.100b)

Then, N~ M ~  2 RH1 and
P = M~ 1 N:
~ (2.101)

A normalised left coprime factorisation, i.e. such that N~ N~ ? + M~ M~ ? = I ,

can be obtained as follows.
Proposition 2.3. Let (A; B; C; D ) be a detectable realisation of a
transfer matrix P (s). Dene
" #
  A + LC B + LD L
N~ M~ = ; (2.102)
JC JD J
where


L = XC T + BDT R~ 1 ; (2.103)
J = R~ 1=2 ; (2.104)
X is the stabilising solution of the following Riccati equation:



A BDT R~ 1 C X + X A BDT R~ 1 C T


X C T R~ 1 C X + BR 1B T = 0; (2.105)
and
R = I + DT D; (2.106)
R~ = I + DDT : (2.107)
32 Modelling tools, closed-loop setup and generalities

 
Then, N~ (s) M
~ (s) 2 RH1,
8! N~ (j!)N~ ? (j!) + M~ (j!)M~ ? (j!) = I; (2.108)
and
P (s) = M~ 1 (s)N~ (s): (2.109)
Remark. The same result holds in the discrete-time case with X the
stabilising solution of the following discrete algebraic Riccati equation:
X = AXAT + BB T



AXC T + BDT R~ + CXC T 1 CXAT + DB T ); (2.110)
and L and J respectively given by



L = AXC T + BDT R~ + CXC T 1 (2.111)
and


J = R~ + CXC T 1=2 : (2.112)
Definition 2.6. (right coprime factorisation) Let P be a proper
real rational transfer matrix. A right coprime factorisation of P is a
factorisation P = NM 1 where N and M are right coprime over RH1 .

If P is controllable, such a right coprime factorisation can be constructed

as follows.
Proposition 2.4. Let (A; B; C; D) be a controllable realisation of a
transfer matrix P , let F be any constant feedback matrix stabilising P ,
and let Z be any nonsingular matrix. Dene
2 3
 
N A + BF BZ
M = C +FDF DZ
4 5; (2.113)
Z
i.e. (continuous-time case)
     
N (s) C + DF sI (A + BF )

1
BZ + DZ
M (s) = F Z (2.114a)
or (discrete-time case)
     
N (z ) C + DF zI (A + BF )

1
BZ + DZ
M (z ) = F Z : (2.114b)
h i
Then, N
M 2 RH1 and
P = NM 1 : (2.115)
 Coprime factorisations 33

The following proposition gives the procedure to build a normalised right

coprime factorisation, i.e. such that N ? N + M ? M = I .
Proposition 2.5. Let (A; B; C; D) be a controllable realisation of a
transfer matrix P (s). Dene
2 3

N
 A + BF BH
M =
4 C + DF DH 5 (2.116)
F H
where


F = R 1 B T X + DT C ; (2.117)
H = R 1=2 ; (2.118)
X is the stabilising solution of the following Riccati equation:



A BR 1 DT C T X + X A BR 1DT C


+ X BR 1B T X + C T R~ 1 C = 0; (2.119)
and
R = I + DT D; (2.120)
R~ = I + DDT : (2.121)
h i
Then, N (s)
M (s) 2 RH1,
8! N ? (j!)N (j!) + M ? (j!)M (j!) = I; (2.122)
and
P (s) = N (s)M 1 (s): (2.123)
Remark. The same result holds in the discrete-time case with X the
stabilising solution of the following discrete algebraic Riccati equation:
X = AT XA + C T C


1

AT XB + C T D R + B T XB B T XA + DT C ; (2.124)
and F and H respectively given by

1

F= R + B T XB B T XA + DT C (2.125)
and

1=2
H = R + B T XB : (2.126)
34 Modelling tools, closed-loop setup and generalities

 
In
h i
the sequel, we shall generally use the notations N~ M~ (respectively
N
M ) for the left (respectively right) coprime factorisations of a system
  h i
G = M~ 1 N~ = NM 1 , and U~ V~ (respectively VU ) for the left (re-
spectively right) coprime factorisations of a controller K = V~ 1 U~ =
UV 1 .

2.7. The  -gap metric

2.7.1. Denition

The  -gap metric between two transfer matrices G1 and G2 is a mea-

sure of distance between these two systems. It was rst introduced by
G. Vinnicombe in (Vinnicombe, 1993a, 1993b) .
Definition 2.7. ( -gap metric) The  -gap metric between two trans-
fer matrices G1 and G2 is dened as
8

>
>
<
 G1 (j!); G2 (j!) 1 if det (j!) 6= 0 8!
Æ (G1 ; G2 ) = and wno det (s) = 0,
>
>
:1
otherwise,
(2.127)
where

 (s) , N2?(s)N1 (s
) + M2? (s)M1 (s);

  G1(j!); G2(j!) , N~2(j!)M1 (j!) + M~ 2 (j!)N1 (j!) is called the
chordal distance between G1 and G2 at frequency !;
 G1 (s) = N1 (s)M1 1 (s) and G2 (s) = N2 (s)M2 1 (s) = M~ 2 1 (s)N~2 (s)
are normalised


of G1 and G2 ;
coprime
factorisations
 wno P (s) =  P 1 (s)  P (s) is called the winding number of
the transfer function P (s) and is dened as the number of counter-
clockwise encirclements (a clockwise encirclement counts as a nega-
tive encirclement) around the origin of the Nyquist contour of P (s)

around the right of any imaginary axis pole of P (s);
indented
  P (s) denotes the number of poles of P (s) in C +0 .

The denition also holds in the discrete-time case via the use of the
bilinear transformation s = zz+11 .
 The  -gap metric 35

It is shown in (Vinnicombe, 1993a) that

Æ (G1 ; G2 ) = Æ (G2 ; G1 ) = Æ (GT1 ; GT2 ): (2.128)
An alternative denition of the  -gap is the following:
8

>  G1 (j! ); G2 (j! )
>
> 1
>
> if det( I + G ? (j! )G (j! )) 6= 0 8! and
>
< 2 1


Æ (G1 ; G2 ) = wno det I + G ? (s)G (s) +  G (s)
2
1
1 (2.129)
>
>
>
>  G2 (s) 0 G2 (s) = 0,
>
>
:
1 otherwise,


where 0 P (s) is
the number of imaginary axis poles of P (s) and where
 G1 (j!); G2 (j!) can be written as



 G1 (j!); G2 (j!) = I + G2 (j!)G?2 (j!) 1=2



 G1(j!) G2(j!)  I + G?1(j!)G1 (j!) 1=2
: (2.130)
In the SISO case, the  -gap metric has a nice geometric interpreta-

tion. Indeed, the chordal distance at frequency !,  G1 (j!); G2 (j!) ,
is the distance between the projections onto the Riemann sphere of the
points of the Nyquist plots of G1 and G2 corresponding to that frequency
(hence the appellation `chordal distance'). The Riemann sphere is a unit
sphere tangent at its south pole to the complex plane at its origin, and
the points of the Nyquist plots are projected onto the sphere using its
north pole as centre of projection. Due to this particular projection, the
chordal distance has a maximal resolution at frequencies where jG1 j  1
and/or jG2 j  1, i.e. around the crossover frequencies of G1 and G2,
since the corresponding points are projected onto the equator of the Rie-
mann sphere. This property makes the  -gap a control-oriented measure
of distance between two systems.

2.7.2. The winding number condition and comparison with the stan-
dard gap metric

The denition of the  -gap metric, and in particular the winding num-
ber condition it involves, are based on a robust stability argument. In
the theory of robust control, coprime-factor uncertainties are often con-
sidered. Let G1(s) = N1 (s)M1 1 (s) dene the normalised
h i right coprime


factorisation of a nominal system G1, and let
=
M be a coprime-
N

factor perturbation. If K is a feedback controller that stabilises G1 , then

36 Modelling tools, closed-loop setup and generalities

K also stabilises all G2 in the set

n o
G = G2 = N1 +
N
M1 +
M
1 j
2 H1; k
k1 6  (2.131)
if and only if (Georgiou and Smith, 1990)
bG1 ;K > : (2.132)
An alternative denition of this set is the following:
n o
G = G2 j ~Æg (G1 ; G2) <  ; (2.133)
where ~Æg (G1 ; G2 ) is the directed gap and can be computed by
   
N1 (s) N (s )
~Æg (G1 ; G2 ) = inf 2 Q (s ) ; (2.134)
Q(s)2H1 M1 (s) M2 (s)
1
G2 (s) = N2 (s)M2 1 (s) dening the normalised right coprime factorisa-
tion of G2 . However, it can be shown (Vinnicombe, 1993b) that the
largest class of systems that can be guaranteed to be stabilised a priori
by K consists of those G2 satisfying
   
N1 (s) N2 (s) Q(s) ; wno det (s) = 0; (2.135)
inf M
Q(s)2L1 1 (s) M2 (s)
1
which is precisely the set
n

1
G2 = N1 +
N M1 +
M j
2 L1; k
k1 6 

o n o
and (G2 ) = wno M1 +
M = G2 j Æ (G1 ; G2 ) 6  : (2.136)

Thus, one can dene a larger set of plants that are guaranteed to be sta-
bilised by a given controller K with the  -gap metric than with directed
gap, since the  -gap allows coprime-factor perturbations in L1 rather
than in H1. Another serious advantage of the  -gap over the directed
gap is the fact that the former is much easier to compute than the lat-
ter. However, in order to be valid, the robust stability theory with the
 -gap metric requires the verication of the winding number condition
(which can take various forms depending on the chosen denition of the
 -gap). The demonstration of the necessity of the winding number con-
dition is very complicated and is outside the scope of this thesis, and
the interested reader is referred to (Vinnicombe, 1993a, 1993b, 1999).
 CHAPTER 3
A motivating discussion of issues in
identication for control

3.1. Introduction
This chapter discusses a number of issues in the problem of modelling
and identication for control design. The goal is to provide insights and
partial answers to the following central question: `How should we iden-
^ H^ ) that is good for control design?'. These insights will
tify a model (G;
constitute the red line of this thesis, and each of the following chapters
will investigate more deeply some of the questions raised here.
^ H^ ) for control design is
A reasonable qualication of a good model (G;

1. simultaneous stabilisation: the controller K (G; ^ H^ ) designed from

this model stabilises the model G^ and the plant G0 , and
2. similar performance: the achieved performance, on the
^ H^ )
loop, is close to the designed performance, on the
G0 ; K (G;
^ H^ )
loop.
^ K (G;
G;

The characterisation of all models that satisfy (1), for a model reference
control design criterion, was studied in (Blondel et al., 1997) and (Gevers
et al., 1997).
Observe that the problem of modelling for control involves three players:
^ H^ ), and the to-be-designed controller
the plant (G0 ; H0 ), the model (G;
^ H^ ). (For the sake of simplicity, the latter will simply be denoted
K (G;
K in the sequel of this chapter.)
37
38 A motivating discussion of issues in identication for control

Identication for control often involves one or several steps of closed-loop
identication, by which we mean identication of a model (G; ^ H^ ) of the
plant (G0 ; H0 ) using data collected on the closed-loop system formed by
the feedback connection of G0 and and some to-be-replaced controller
Kid . Thus, closed-loop identication also involves three players: the
plant (G0 ; H0 ), the model (G;^ H^ ), and the current controller Kid . The
third player (the controller) is dierent as before. As a result, closed-
loop identication for control involves four players: the plant (G0 ; H0 ),
the model (G;^ H^ ), the controller Kid applied during identication, and
the to-be-designed controller K . The focus of our discussion is on the
interplay between these four players. The typical context is one in which
a plant (G0 ; H0 ) is presently under closed-loop control with some (low-
order) controller Kid , and where it is desired to estimate a model (G;^ H^ )
with the view of designing a new (low-order) controller K that would
achieve better performance on the plant (G0 ; H0 ) while guaranteeing
stability robustness.
In particular we address the following issues.

 We examine the role of the controller Kid in changing the experimen-

tal conditions.
 We compare the eects of open-loop and closed-loop identication in
terms of bias and variance errors in the context of identication for
control.
 We compare iterative design using closed-loop data with the alterna-
tive of rst identifying a high-order model (preferably in open loop)
that is validated by the data, and of then using this model for the
design of a low-order controller using model or controller reduction
techniques. The claimed advantage of this alternative is that the
design model comes with an uncertainty description.
 We motivate the need for cautious controller adjustment in iterative
design.

3.2. The role of feedback

It is well known that two plants that present similar open-loop be-
haviours (in terms of their impulse responses or Nyquist diagrams, for
instance) may present enormously dierent closed-loop behaviours for
some controller. Conversely one can attach a stabilising controller to
 The role of feedback 39

two plants and observe what appears to be identical closed-loop be-

haviours, when the open-loop behaviours of the plants are enormously
dierent. These facts are illustrated in (Skelton, 1989) where it is shown,
for instance, that the plants G(1) (s) = s+1 1 and G(2) (s) = 1 have re-
s
markably dierent open-loop behaviours, while a static output feedback
u(t) = r(t) ky(t) will make their closed-loop behaviours almost indis-
tinguishable for large k (at least if one considers the closed-loop step
responses): see Figure 3.1. It is also interesting to observe that
Æ (100G(1) ; 100G(2) ) = 0:01 << Æ (G(1) ; G(2) ) = 0:71; (3.1)
where Æ (
;
) denotes the  gap (see Section 2.7), which shows quan-
titatively that G(1) and G(2) are much `closer' to one another in closed
loop with the static controller k = 100 than in open loop.

OPEN LOOP: G CLOSED LOOP: GK/(1+GK) CLOSED LOOP: G/(1+GK)

6 1 0.01

5
0.8 0.008

4
0.6 0.006
3
0.4 0.004
2

0.2 0.002
1

0 0 0
0 1 2 3 4 5 6 0 0.010.020.030.040.050.06 0 0.010.020.030.040.050.06
Time [s] Time [s] Time [s]

CLOSED LOOP: K/(1+GK) CLOSED LOOP: 1/(1+GK)

100 1

80 0.8

60 0.6

40 0.4

20 0.2

0 0
0 0.010.020.030.040.050.06 0 0.010.020.030.040.050.06
Time [s] Time [s]

1 (|) and
Figure 3.1. Step responses of G(1) (s) = s+1
of G(2) (s) = 1s ( ) in open loop and in closed loop with
k = 100

We return now to the modelling game with the three players described in
the introduction. Since the model (G; ^ H^ ) is used for the design, what is
^ H^ ) is that the stability of the (G;
required of (G; ^ K ) loop implies that of
40 A motivating discussion of issues in identication for control

the (G0 ; K ) loop (stability robustness), and that the closed-loop trans-
fer functions of these two loops are close to one another (performance
robustness).
Now one of the main aims of feedback (historically) is to reduce the
eects of model uncertainty on the open-loop plant. Feedback often has
a sensitivity reduction objective. Therefore any sensible feedback design
will have the eect that the closed-loop systems (G; ^ K ) and (G0 ; K )
behave much closer to one another than the open-loop systems G^ and
G0 .
What are the consequences of these observations on identication for
control? The issue here is change of experimental conditions. Models
can only have their quality evaluated for a particular set of experimen-
tal conditions. Changing from open-loop to closed-loop operation with
a specied controller, or changing the controller, is of course one change
of experimental conditions. Unless a plant model is exact, high accu-
racy under one set of experimental conditions (e.g. open loop) does not
guarantee its eÆcacy under changed experimental conditions. It is also
intuitively clear that small changes of controller should probably avoid,
or ameliorate, the problem of possible loss of eÆcacy of a model under
changed experimental conditions.
As a consequence, the best way to evaluate the quality of the model
^ H^ ) is to test it under the experimental conditions under which the
(G;
plant (G0 ; H0 ) is due to operate, i.e. in closed-loop with the to-be-
designed controller K . For the same reason, it should ideally be iden-
tied under those same feedback conditions; see e.g. (Hjalmarsson et
al., 1994, 1996). This is of course impossible since knowledge of the
model is required to design the controller K . The philosophy behind
iterative design of models and controllers is to approximate these exper-
imental conditions in successive steps. This allows one to successively
reduce the uncertainty in the frequency bands of importance for the de-
sign of the next controller, even with reduced-order models (i.e. models
such that S 2= M and/or G0 2= G).
The alternative, advocated in (Ljung, 1997, 1998), is to identify in open
loop a model of suÆciently high order that is validated by the data, and
to subsequently perform a step of model or controller reduction. This
will lead to much more conservative designs because such a validated
model will have an uncertainty distribution that is not shaped for control
 The eect of feedback on the bias error 41

design. In addition, this uncertainty region cannot be reduced by order

reduction, whether frequency shaped or not.

3.3. The eect of feedback on the bias error

For the purpose of analysis, we consider that there exists a true LTI
system that can be described by (2.1):
y(t) = G0 (z )u(t) + v(t): (3.2)
We consider that this system is possibly connected to a feedback con-
troller K :
u(t) = r(t) K (z )y(t): (3.3)

For simplicity, we restrict our analysis to the SISO case.

If we denote by y^(t) the output of the feedback connection of a model G^
with the same controller K , the closed-loop expressions for the outputs
y(t) and y^(t) are
G0 (z ) 1
y(t) = r(t) + v(t); (3.4)
1 + G0 (z )K (z ) 1 + G0 (z )K (z )
G^ (z )
y^(t) = r(t): (3.5)
1 + G^ (z )K (z )
The mean square error between the actual output and the nominal out-
put is therefore given by
Z   ^
2

1 G G
E jy(t) y^(t)j2 =  0

^

 jS (G0; K )j2 r d!
2   1 + GK 
Z
1 
+
2 
jS (G0; K )j2 v d!; (3.6)
where S (G0 ; K ) = 1 is the actual sensitivity function.
1+G0 K

Consider now that a parametrised model G(z; ^) is identied from N

input-output data Z N obtained on the true plant, possibly connected to
a feedback controller Kid , using a prediction error method. Dene, as
usual (see Section 2.2),


"F (t; ) = L(z; ) y(t) G(z; )u(t) (3.7)
42 A motivating discussion of issues in identication for control

(which is equivalent to equation (2.12) since the noise model H (z; )

can be included in the prelter L(z; )),
N

1 X
VN ; Z N
= "2F (t; ); (3.8)
N t=1
and


^ = arg min VN ; Z N : (3.9)

It follows easily that
^ !  = arg min V () as N !1 (3.10)

where
Z 
1
V () =
2
jG0 G()j2 jL()j2 rud!

Z
1 
+
2 
j1 + G()Kidj2 jS (G0 ; Kid )j2 jL()j2 v d!: (3.11)
Here ru is the spectral density of that part of the control signal that orig-
inates from the reference input. Indeed, since r and v are uncorrelated,
one can split the input spectrum into
u (!) = ru (!) + vu (!) (3.12)
where
 2
 1 
u (!) = 
r    (! ): (3.13)
1 + G0 (e )Kid (e )  r
j! j!

Observe now that if one takes a parametrised prelter L() = 1+G(1)Kid ,
the second term in (3.11) becomes independent of  and the minimising
argument of V () is the same as the minimising argument of
Z   
1  G0 G( ) 2
V~ () =   jS (G0 ; Kid )j2 r d!: (3.14)
2   1 + G()Kid 
As a result, the identication will tend to a model that realises the best
compromise between tting G() to G0 and making the model sensitivity
1
1+G()Kid small. Furthermore, the weighting factor in (3.14) contains the
actual sensitivity S (G0 ; Kid ). Thus, the t between G(^) and G0 will
be tight where either the actual sensitivity or the model sensitivity is
large, i.e. around the corresponding crossover frequencies.
This last observation has been the main heuristic motivation for per-
forming closed-loop identication when low-order models are used for
 The eect of feedback on the bias error 43

the purpose of designing new controllers with higher performance. The

heuristic reasoning is that, since the model is necessarily biased due to
its low orderness, one should tune the bias distribution such that it be
low around the crossover frequency of the present control loop. A small
uncertainty around this crossover frequency then allows one to design a
new controller that will expand the the existing bandwidth.
This heuristic motivation can be given more solid support by consid-
ering our observations of Section 3.2 on the assessment of the quality
of models. Ideally the model G(^) should be evaluated by how well it
mimics the behaviour of the actual system G0 when both are connected
in feedback with the to-be-designed controller K . In other words, a
model G(^) , G^ is judged good if the quantity E jy(t) y^(t)j2 in (3.6)
is small. We now observe that, if Kid = K , then the minimisation of
the asymptotic criterion (3.14) also minimises the performance crite-
rion (3.6). Thus, when low-order models are used for control design,
the bias should be minimised in the closed-loop sense of the criterion
(3.14). This is achieved by closed-loop identication with the to-be-
designed controller in the loop. Of course, this is impossible since the
to-be-designed controller is unknown.
The concept of iterative identication and control design is an attempt
to approach these ideal experimental conditions by slow adjustments
from one controller to the next one. By slowly increasing the required
performance, the to-be-designed controller is never very dierent from
the controller Kid that is presently operating, and therefore the experi-
mental conditions are always close to the optimal ones. This philosophy
has been called the `windsurfer approach' in (Lee et al., 1993). The rea-
son for introducing small steps of controller adjustment is also justied
by stability robustness arguments: see e.g. (Anderson et al., 1998).
In this section we have justied the use of closed-loop identication for
control on the basis of bias considerations. We observe that the match-
ing of the criteria (3.6) and (3.14) is impossible if the data are collected
in open loop, even if the to-be-designed controller K were known, be-
cause both criteria contain the actual sensitivity S which depends on the
unknown true system. This bias justication only holds when reduced-
order models are used. Full-order models can be identied without bias,
and there is then no theoretical reason to prefer closed-loop or open-loop
identication.
44 A motivating discussion of issues in identication for control

In the context of modelling for control, it has therefore been argued that
a (wiser?) alternative to closed-loop identication of low-order models
is to identify the plant in open loop using higher-order unbiased models
that can be validated, and to then apply a model or controller reduction
step if a low-order controller is desired. This approach has recently
been strongly advocated in (Ljung, 1997). In a statistical framework a
model is called validated if the uncertainty region around the estimated
model contains the true system with some prescribed probability level
(say 95%). For validated models, the variance error dominates the bias
error (Ljung and Guo, 1997). In the next section, we explain why such
a strategy may lead to very conservative controller designs because it
produces uncertainty regions that are essentially equally distributed over
the frequency range, and which are larger around the crossover frequency
than those obtained with models identied in closed loop.

3.4. The eect of feedback on the variance error

In this section we focus on the variance errors in the estimated transfer
functions. It was shown in (Gevers and Ljung, 1986) for minimum vari-
ance control design and in (Hjalmarsson et al., 1996) for other control
design criteria that closed-loop identication is optimal for the min-
imisation of the variance error on a to-be-designed controller when the
system is in the model set (i.e., the variance of the dierence between the
controller designed from the identied model and the optimal controller
that would be designed if the true system were known is minimised
when the model is obtained by closed-loop identication). For some cri-
teria (e.g. minimum variance) the optimal design consists of closed-loop
identication with the to-be-designed controller in the loop during data
collection (i.e. Kid = K ), a similar conclusion of that reached in the
previous section on the basis of bias considerations. Needless to say,
this optimal design is again not applicable because computation of the
optimal to-be-designed controller requires knowledge of the system be-
ing identied. However, this result gives once again heuristic support to
iterative schemes.
There are two important limitations to the relevance of the optimal de-
sign results of (Gevers and Ljung, 1986) and (Hjalmarsson et al., 1996):
1. they assume the system to be in the model set, i.e. only variance
errors are considered;
 The eect of feedback on the variance error 45

2. these results were based on performance considerations only, without

regard for robust stability considerations.
Here we focus on the robust stability issue, and we explain why, on the
basis of variance considerations (but bias errors are allowed too), closed-
loop identication with a controller that is as close as possible to the
to-be-designed controller is to be preferred over open-loop identication.
Recall that, in prediction error identication, the asymptotic variance in
transfer function space (i.e. as the number of data an the model order
both tend to innity) is given by (2.34):

G^ (ej! )  n  u (!) ue (!) T
cov ^ j!  v (!): (3.15)
H (e ) N eu (!) 0
In open loop, ue (!) = eu (!) = 0 and the variance of the input-output
model is

n v (!)
var G^ (ej! )  : (3.16)
N u (!)
In closed-loop, if we also focus on the asymptotic variance of the input-
output model, we get

n v (!)
var G^ (ej! ) 
N ru (!)
(3.17)
n  j! 2 v (! )

= 1 + G0 (e )Kid (e )
j!
N r (!)
by noting that
2 3
1 eu (!)
 
u (!) ue (!) T = 6 6 u (! )
r 0 ru (!) 7
7
6 7 : (3.18)
eu (!) 0 4 ue (! ) 1 5
0 u (!) 0 jue (!)j =u (!)
r 2

Thus, if white noise is used to excite the system in closed loop, i.e. if
r (!) is uniform, the variance of the resulting estimate will be smaller
around the crossover frequency, i.e. where the actual sensitivity func-
tion 1+G10 Kid is large (at least if v(t) is white). This is not the case in
open loop unless more excitation power is put in the frequency range
of the desired closed-loop bandwidth, i.e. if a tailor-made input spec-
trum u (!) is chosen. This increased condence in the model around
the crossover frequency results for instance in a tighter upper bound
on the gain margin, which allows to compute less conservative robust
controllers. This will be illustrated in the next section. Note however
46 A motivating discussion of issues in identication for control

that this potential advantage of closed-loop identication only holds if

the current controller Kid is not too dierent from the to-be-designed
controller K , i.e. if the desired closed-loop bandwidth is in the frequency
range where the current sensitivity function is large.
Sometimes, a (possibly low-order) model G^ is delivered without an un-
certainty region attached to it. To assess the quality of such a model,
L. Ljung has suggested a validation procedure based on open-loop pre-
diction error identication (Ljung, 1997). We consider again that there
exists a true LTI system described by (3.2). The evaluation of a given
model G^ is done by constructing the residuals formed by the dierence
between the measured and simulated outputs:
"(t) = y(t) y^(t)
= G0 (z )u(t) + v(t) G^ (z )u(t) (3.19)
, @G(z)u(t) + v(t):
The residuals contain two contributions: one comes from the input and
contains information on the model error, and one comes from the noise
and contains no such information. The problem of model validation is to
decide which part of the residuals can be explained by model errors, and
which can be attributed to the noise. The interesting observation made
in (Ljung, 1997) is that the data contain information about this split
between the two contributions to ". Indeed, the elements of the cross-
correlation function between "(t) and a nite vector of past u's can be
taken as a FIR estimate of @G. The model validation strategy proposed
in (Ljung, 1997) is to accept the model G^ if the uncertainty region
around the corresponding estimate of @G contains 0 at all frequencies
(at a specied probability level). This means that, at that probability
level, the true G0 is contained in the uncertainty region around G^ , i.e.
G^ is not falsied.
This strategy is very appealing as a model validation strategy. However,
it leads to models with a variance error that is { roughly speaking {
equally distributed over the frequency range if the input signal-to-noise
ratio is more or less equally distributed. It is easy to understand that, if
a controller is computed on the basis of such uncertainty regions, then
stability robustness guarantees will require conservative designs because
of the unfocused uncertainty regions. Instead, an identication using
similarly noisy data collected in closed loop with a controller that is
close to the to-be-designed controller will produce low variance errors
 Simulation 47

in the frequency range of interest for the control design, leading to less
conservative designs.
Two issues are thus involved in this validation procedure:
1. the uncertainty distribution is not tuned towards the control design
objective;
2. a model that is validated in all frequency regions is unbiased (at
least, the bias error is smaller than the variance error) and thus of
high order, which can be a problem when the objective is to design a
low-order controller.

3.5. Simulation
These insights on variance errors are still preliminary and will receive
more theoretical attention in the sequel of this thesis. However, we have
developed a simulation in order to test their validity.
Assume that the true system (G0 ; H0 ) is described by the following ARX
model:
S : 1 1:4z 1 + 0:45z 2
y(t) = z 1 1 + 0:25z 1
u(t) + e(t) (3.20)
where e(t) is unit-variance white noise. The experiment consists in iden-
tifying an unbiased model
M() : 1 + a1z 1 + a2z 2
y(t) = z 1 b1 + b2z 1
u(t) + e(t) (3.21)
 
where  = a1 a2 b1 b2 T is the parameter vector and, using the
variance of the parameter estimates, in computing the maximal gain
of a proportional output feedback controller that leaves the closed-loop
poles of all possible models within the stability region, with a probability
level of 95%. The 95% uncertainty region in parameter space is dened
as n o
U =  2 R4 j ( ^)T P 1( ^) <  (3.22)

P is its covariance matrix,

where ^ is the estimated parameter vector,
and  = 9:49 is given by Pr 2 (4) <  < 95% where 2 (4) is a 2 -
distributed random variable with 4 degrees of freedom.
We performed the two identication experiments described below, and
for each of these we computed the maximum gain kmax that would guar-
antee closed-loop stability for all models with parameters in U, i.e. such
48 A motivating discussion of issues in identication for control

that the characteristic


polynomial 1+(a1 +b1 kmax )z 1 +(a2 +b2 kmax )z 2
is stable for any a1 a2 b1 b2 T 2 U.
The two experiments were as follows.
Experiment 1: open-loop identication. The system was simulated
in open loop with unit-variance white noise as input signal u(t). 1000
input-output data were collected and used to estimate ^ and P .
Experiment 2: closed-loop identication. A proportional con-
troller u(t) = r(t) y(t) was used. The reference signal r(t) was chosen
as white noise with variance 18.38, leading to the same output variance
as in the open-loop experiment1. 1000 input-output data were collected
and used to estimate ^ and P .

Each of these two experiments was run 100 times, in order to get conclu-
sions that would not depend on a particular noise realisation. For the
true system, the smallest destabilising gain for the proportional con-
troller is kmax = 2:2. The estimated limit gains k^max obtained by av-
eraging the 100 Monte-Carlo runs with the two successive experiments
were 1.36 and 2.04, respectively. This shows that the variance error is
better tuned towards robust control design when the model is identied
in closed loop. In addition, the experimental variances around the esti-
mated limit gains were respectively 0.12 and 0.02, i.e. much smaller for
the closed-loop experiments than for the open-loop experiments, which
conrms the results of (Hjalmarsson et al., 1996).

3.6. Conclusions
In this chapter, we have shown that it is generally preferable to iden-
tify a (possibly low-order) model in closed loop when the purpose is to
use the model for control design. Indeed, the bias and variance errors
attached to the estimate are better tuned for control design when the
model is identied in closed loop rather than in open loop, at least if
similar spectra are used to excite the system in both cases, and if the
to-be-designed controller is close to the one used during identication.
1
Imposing the same output variance is a fair requirement when the objective is to
compare dierent open-loop and/or closed-loop congurations of a given system since,
in many practical applications, the output variance is constrained and the control
objective is to reduce it.
 Conclusions 49

One could of course tune the open-loop excitation spectrum in order

to put more energy around the desired closed-loop crossover frequency.
However, the claimed advantage of closed-loop identication is the fact
that this tuning happens `naturally' due to the ltering of the reference
spectrum through the closed-loop sensitivity function.
An alternative, regarding bias, is to identify an unbiased (and thus
higher-order) model in open loop. Since the ultimate goal is often the
computation of a low-order controller for the system, a reduction step
is then generally required. Clearly, the discussion of Section 3.3 remains
valid when a low-order (biased) model has to be computed via order
reduction of a high-order (unbiased) model. Therefore, the reduction
step should also be performed in closed loop, i.e. with frequency weight-
ings that take account of the current controller (if such a controller is
available), in such a way that the reduction criterion would be equiva-
lent to the performance degradation criterion (3.6) if the to-be-designed
controller were known. Such a procedure is proposed in Chapter 6, for
an alternative performance degradation criterion of the form

J (G^ r ) = To (Gn ; K ) To (G^ r ; K ) ; (3.23)
1
where Gn is a high-order unbiased model obtained by open-loop identi-
cation or any other appropriate technique, G^ r is the (biased) reduced-
order model, and K is the current controller.
Thus, the closed-loop reduction method of Chapter 6 reconciles the two
approaches (closed-loop identication of a low-order model and open-
loop identication of an unbiased model followed by order reduction)
from a bias distribution point of view. However, as stressed in Sec-
tion 3.4, the uncertainty distribution attached to a model obtained by
open-loop identication is not ideally tuned for control design if no pre-
cautions are taken regarding the excitation spectrum; furthermore this
uncertainty cannot be reduced by model reduction. Therefore, we sug-
gest to drop this uncertainty region (if any) and to build a new one
using closed-loop techniques. This is closed-loop model validation: see
Chapter 5.
Another point, which has not been discussed here, is the choice between
low-order model identication or full-order model identication followed
by model reduction when both approaches are carried out either in open
loop or in closed loop: what dierence does it make in terms of the
variance and bias errors attached to the nal estimates? This question
50 A motivating discussion of issues in identication for control

is tackled in Chapter 7. This chapter also deals with the question of

overmodelling.
 CHAPTER 4
Identication in closed loop with an
unstable or nonminimum phase controller

4.1. Introduction
In the previous chapter, we have given several reasons for performing
the identication of a system in closed loop when the objective is the
computation of a model that is good for control design. Before pursuing
these ideas, we show in this chapter that closed-loop identication may
fail to reach this objective if there are unstable poles or nonminimum
phase zeros in the controller1 used during the identication experiment
and if some precautions are not taken. We show that the presence of such
poles and zeros can be the source of identiability problems generally
ignored by the practitioner, leading to a model which may appear to
be input-output stable in closed loop, but which is actually internally
unstable. In more technical terms, the use of a nonminimum phase or
unstable controller (e.g. a controller containing an integrator) during
identication may lead to a nominal closed-loop system with a zero
generalised stability margin, although the true system is stabilised by
this controller2 . Such a model will prove unsuitable for the design of a
new controller.

1
This is likely to happen for controllers designed by optimal and/or robust techniques
such as LQG or H1 synthesis, since the designer has no control on the poles and
zeros of these controllers.
2
Of course, it is impossible to know for sure that a controller stabilises an imper-
fectly known plant. However, we can reasonably consider that closed-loop stability is
achieved if no unstable behaviour has been detected during a suÆciently long period
of closed-loop operation.
51
52 Identication in closed loop with an unstable or nonminimum phase controller

For the sake of simplifying the notations, we shall restrict our analysis
to the SISO case. However, our results can easily be extended to the
multivariable case.
The outline of this chapter is as follows. In Section 4.2, we brie y re-
view the notions of poles and zeros of a discrete-time system, as well as
some approaches used in closed-loop identication. Section 4.3 describes
how the use of an unstable or nonminimum phase controller during the
identication may lead to a model that is unstable in closed loop, even
though the actual closed-loop system is stable. Some experiment design
guidelines are given in order to avoid this problem. The consequences of
this instability problem in terms of robust control design are addressed
in Section 4.4. Section 4.5 explains how to deal with a controller that
is both unstable and nonminimum phase, using coprime plant and con-
troller factorisations and the dual Youla-Kucera parametrisation of all
plants that are stabilised by a given controller. A realistic numerical
simulation based on the Landau benchmark (Landau et al., 1995b) is
given as an illustration in Section 4.6. Finally, conclusions are drawn in
Section 4.7.

4.2. Some reminders about properties of systems

and identication in closed loop
In this section, we rst review some important notions about poles and
zeros of a discrete-time SISO system. Then, we introduce the dierent
closed-loop identication techniques that will be studied in Section 4.3.

4.2.1. Poles and zeros of a system

Definition 4.1. (Minimality) A state-space realisation (A; B; C; D )

of a system is called minimal if A has the smallest possible dimension,
which is true if and only if the pair (A; B ) is controllable and the pair
(C; A) is observable.
Definition 4.2. (Pole) Let P (z ) = C (zI A) 1 B + D be the transfer
function of a discrete-time system P with minimal state-space realisation
(A; B; C; D). z0 2 C is a pole of P if P (z0 ) = 1.
 Some reminders about properties of systems and identication in closed loop 53

Definition 4.3. (Transmission zero) Let P (z ) = C (zI A) 1 B + D

be the transfer function of a discrete-time system P with minimal state-
space realisation (A; B; C; D). z0 2 C is a transmission zero of P if
P (z0 ) = 0.

A zero at z = z0 is refered to as a `transmission zero' because P (z0 )ez0 t =

0 8t > 0 provided the initial state x(0) is unobservable, i.e. the response
of the system to the input signal u(t) = ez0 t is zero. We shall also
often talk of `blocking zeros' and `blocking poles', which we dene as a
particular class of zeros and poles.
Definition 4.4. (Blocking zero or pole) A zero (resp. a pole)
z0 2 C of a discrete-time system P (z ) will be called a blocking zero
(resp. a blocking pole) if it is located
 on the unit circle (z0 2 @ D ), i.e.
if z0 = e for some !0 2 0 2 .
j! 0

A `blocking zero' is thus a transmission zero with respect to a real si-

nusoidal excitation at frequency !0 : P (ej!0 ) sin(!0 t) = 0 8t > 0 if the
initial state x(0) is unobservable. A unit-circle pole of a controller is
called a `blocking pole' because it becomes a blocking zero of the closed-
loop transfer functions T12 and T22 dened in (2.77).
Finally, we dene the notions of (strict) instability and (strict) nonmin-
imum phaseness as follows.
Definition 4.5. (Instability) A pole z0 2 C of a discrete-time system
P (z ) will be called `unstable' (resp. `strictly unstable') if z0 2 {D , i.e. if
jz0 j > 1 (resp. if z0 2 {D c , i.e. if jz0j > 1). P (z) will be called (strictly)
unstable if it has at least one (strictly) unstable pole.
Definition 4.6. (Nonminimum phaseness) A zero z0 2 C of a discrete-
time system P (z ) will be called `nonminimum phase' (resp. `strictly non-
minimum phase') if z0 2 {D , i.e. if jz0 j > 1 (resp. if z0 2 {D c , i.e. if
jz0 j > 1). P (z) will be called (strictly) nonminimum phase if it has at
least one (strictly) nonminimum phase zero.

4.2.2. Identication in closed loop

The main approaches for prediction error closed-loop identication are:
 the indirect approach (Soderstrom and Stoica, 1989), which uses mea-
surements of r1 (t) or r2 (t), and y(t) or u(t) to identify one of the four
54 Identication in closed loop with an unstable or nonminimum phase controller

entries of the matrix T (G0 ; K ) of (2.77). From this estimate, a model

G^ for G0 is derived, using knowledge of the controller, which has to
be LTI. Such knowledge is required for this method;
 the coprime-factor approach (Van den Hof and Schrama, 1995), which
uses measurements of r1 (t) or r2 (t), y(t) and u(t) to identify the two
entries of a column of T (G0 ; K ). A model G^ for G0 is then given by
the ratio of these two entries. This method requires that the controller
be LTI;
 the direct approach, which consists in identifying a model G^ for G0
on the basis of measurements of u(t) and y(t) collected in closed loop.

This is of course only a rough description of these three possible ap-

proaches. Variants exist (e.g. indirect identication with a tailor-made
parametrisation (van Donkelaar and Van den Hof, 1997)), as well as
other methods (e.g. the Hansen scheme (Hansen, 1989; De Bruyne et
al., 1998), which is described in Section 4.5). It has recently been shown
that the qualitative properties of the dierent closed-loop identication
methods are essentially equivalent, by observing that these methods can
be seen as dierent parametrisations of one and the same prediction er-
ror method (Forssell and Ljung, 1999). However, we shall see in this
chapter that these dierent methods cope rather dierently with diÆ-
culties created by the presence of unstable poles or nonminimum phase
zeros in the controller.

4.3. Destruction of the nominal stability in the

case of unstable poles or nonminimum phase
zeros in the controller
Let us consider the identication of an unknown plant G0 in closed loop
with some known stabilising controller K as in Figure 2.1. We now
consider the three PE identication methods described in Section 4.2
and show that, whatever the method we use, identiability and nominal
stability problems may occur if the controller is nonminimum phase or
unstable. Our analysis will lead to experiment design guidelines for the
case of a nonminimum phase or unstable controller.
 Destruction of the nominal stability in the case of unstable poles or nonminimum : : : 55

4.3.1. The indirect approach

In this approach, a single entry of the generalised closed-loop transfer

matrix T (G0 ; K ) of (2.77) is identied, and used to derive a model G^
of the plant using knowledge of the controller K . We can thus consider
four dierent cases, depending on which of the four entries is identied.
Remember that K is assumed to stabilise G0 , so that all four entries of
T (G0 ; K ) are stable.

Case 1: estimate T11 . A stable estimate T^11 of T11 in (2.77) is obtained

from measurements of r1 (t) and y(t). A model G^ of the plant G0 is then
reconstructed according to

T^11
G^ = ; (4.1)
K K T^11

using the knowledge of K . The nominal generalised closed-loop transfer

matrix can be expressed as a function T^11 :
2 3
^
GK G^ 2 3
6 7 ^ T^11
^ K) = 6 ^
1 + GK ^ 7 6 T11
1 + GK K 7
T (G; 6 7=4 5 : (4.2)
4 K 1 5 ^ ^
K (1 T11 ) 1 T11
^
1 + GK ^
1 + GK

This shows that the closed loop made of G^ and K will be internally
unstable if K has nonminimum phase zeros or unstable poles. Indeed,

T^11
 if K has nonminimum phase zeros, will be unstable since, in
K
practice, there will never be a perfect matching of these zeros in T^11 ,
because of the possible bias error and of the unavoidable variance
error in the estimated T^11 . Furthermore, if such a zero is located
on the unit circle, meaning that K (ej!0 ) = 0 at some frequency !0 ,
then K will
not transmit
2
the component of r1 (t) at that frequency:
y (!0 ) = T11 (e ) r1 (!0 ) = 0. Hence the output Signal-to-Noise
r 1  j! 0 

Ratio will be 0 at !0 and very low around it, yielding a bad estimate
of T11 at !0 (this is why we call such a zero a blocking zero with
56 Identication in closed loop with an unstable or nonminimum phase controller

^ K ) is
respect to r1 (t))3 . The stability of all other entries of T (G;
unaected by nonminimum phase zeros in K ;
 if K has unstable poles, K (1 T^11 ) will be unstable for the same
reason of imperfect identication. The stability of the other entries
^ K ) is not aected by these poles.
of T (G;

Case 2: estimate T12 . A stable estimate T^12 of T12 in (2.77) is obtained

from measurements of r2 (t) and y(t). A model G^ of the plant G0 is then
reconstructed according to
T^12
G^ = ; (4.3)
1 K T^12
yielding the following nominal generalised closed-loop transfer matrix:
2 3
T^12 K T^12
^ K) = 4
T (G; 5: (4.4)
(1 T^12 K )K 1 T^12 K
The following comments can be made about its stability:
 the presence of nonminimum phase zeros in K does not aect the
^ K );
stability of T (G;
 if K has unstable poles, however, the three reconstructed entries of
^ K ) will be unstable. In addition, notice that if K (ej!0 ) = 1 at
T (G;
some frequency !0 , that is if K has a pole on the unit circle (which
is very common, since many 
controllers

contain an integrator), then
T12 (ej!0 ) = 0, ry2 (!0 ) = T12 (ej!0 )2 r2 (!0 ) = 0, and the output
SNR will be 0 at that frequency and very low around it, yielding
once again a bad estimate at that frequency. Thus this is a blocking
pole with respect to r2 (t).

Case 3: estimate T21 . A stable estimate T^21 of T21 in (2.77) is obtained

from measurements of r1 (t) and u(t). A model G^ of the plant G0 is then
3
It is clear that, if a nonparametric model is identied using spectral analysis methods,
this model will be undetermined at the frequencies of the blocking zeros. However,
in the case of a nite dimensional parametric model, we could learn the value of the
parameters by taking measurements at other frequencies, and infer the value of the
transfer function at the frequencies where there was zero input energy. Thus, we
could conclude that, for a nonparametric model, zeros in the input spectrum will
cause major diÆculties while, for a parametric model, in principle they will not.
However, given the existence of bias error and variance error, we can expect that the
estimated value of the transfer function at frequencies close to a blocking zero could
well have a signicant error attached to it.
 Destruction of the nominal stability in the case of unstable poles or nonminimum : : : 57

reconstructed according to
1 1
G^ = ^ ; (4.5)
T21 K
yielding the following nominal generalised closed-loop transfer matrix:
2  3
T^21 1 T^21
61 1
^ 6
T (G; K ) = 6 K K K 7 7
7: (4.6)
4 T^21 5
T^21
K
 Its stability is not aected by the presence of unstable poles in K ;
 nonminimum phase zeros in K make the three reconstructed entries
^ K ) unstable. In addition, zeros located on the unit circle have
of T (G;
once again a blocking eect on r1 (t) at the corresponding frequencies.

Case 4: estimate T22 . A stable estimate T^22 of T22 in (2.77) is obtained

from measurements of r2 (t) and u(t). A model G^ of the plant G0 is then
reconstructed according to
 
^ 1 1
G= 1 ; (4.7)
K T^22
yielding the following nominal generalised closed-loop transfer matrix:
2 3
^ 1 ^
1 T22 (1 T22 )7
^ K) = 6
T (G; 4 K 5: (4.8)
^
K T22 ^
T22

 If K has nonminimum phase zeros, K1 (1 T^22 ) will be unstable. The

^ K ) is not
stability of the other two reconstructed entries of T (G;
aected;
 if K has unstable poles, K T^22 will be unstable. The stability of the
other two reconstructed transfer functions is not aected. Poles on
the unit circle have a blocking eect on r2 (t).

Conclusion: With the indirect approach, one must

 identify G^ from T^12 using fy(t); r2(t)g data if K has nonminimum
phase zeros;
 identify G^ from T^21 using fu(t); r1 (t)g data if K has unstable poles.
58 Identication in closed loop with an unstable or nonminimum phase controller

All other strategies lead to unstable closed loops for the estimated
^ K ). If K is both unstable and nonminimum phase, the nomi-
T (G;
^ K ) will be unstable whichever entry of T (G0 ; K ) is
nal closed loop (G;
identied, hence bG;K
^ = 0.

4.3.2. The coprime-factor approach

Here the choice is between estimating G^ as

G^ = T^11 T^211 (4.9)
using stable estimates of T11 and T21 obtained from fy(t); u(t); r1 (t)g
data or
G^ = T^12 T^221 (4.10)
using stable estimates of T12 and T22 obtained from fy(t); u(t); r2 (t)g
data. Indeed, observe that, for a stable closed loop (G0 ; K ), there holds
G0 = T11 T211 = T12 T221 which denes two coprime factorisations of G0
since T11 , T12 , T21 and T22 are all stable.

Case 1: estimate T11 and T21 . When G^ is computed from (4.9), the
following nominal closed-loop transfer matrix is obtained:
2 3
2 3 ^11 T^11
^
GK ^
G T K
6 7
6 7 6 T^21 + T^ T^21
+ ^
T 7
^ K) = 6 ^
1 + GK 1 + GK 7^ 6 11 11 7
T (G; 6 7=6 K K
T^21
7 : (4.11)
4 K 1 5 6 ^
T 7
4 21 K 5
^
1 + GK ^
1 + GK T^21 ^ T^21 ^
K + T11 K + T11
Notice that
2 3
T^11
6T^11
^ K ) 6= 6
T (G; K7 7
(4.12)
6 7
4 T^ 5
T^21 21
K
^
because T^11 6= 1+GK
GK
^ and T21 6= 1+GK
^ K ^
^ with G computed via (4.9). In
view of this rather unpromising observation, it is obvious that, for the
identication procedure to be half-way acceptable, there will need to
hold
T^21 ^
+ T11  1 (4.13)
K
 Destruction of the nominal stability in the case of unstable poles or nonminimum : : : 59

if T^11 and T^21 are fair estimates of T11 and T21 , respectively. Incidentally,
observe that TK21 = 1+G1 0 K and T11 = 1+GG0 K0 K are respectively the actual
closed-loop sensitivity and complementary sensitivity functions, hence
K + T11  1, which also motivates the fullment of condition (4.13).
T21

If K has a strictly nonminimum phase zero, then (4.13) cannot hold

because the left-hand side is unstable. In addition, the (1; 2) and (2; 2)
^ K ) in (4.11) will be unstable. Thus, Case 1 will deliver
entries of T (G;
an unstable nominal closed-loop system if K has strictly nonminimum
phase zeros. Unstable poles of K do not pose a problem.
Now, let us assume that K has a blocking zero in z0 = ej!0 , say. Then,
T11 (z0 ) = 0, T21 (z0 ) = 0 and, as we have already remarked, this zero will
make the SNR in y(t) and u(t) equal to 0 at !0 , and very low around
it. As a result, T^11 and T^21 will typically be very bad estimates of T11
and T21 around !0 ; in particular, T^11 (z0 ) 6= 0, T^21 (z0 ) 6= 0, and, instead
of (4.13), it is then required that
 
^ 
 T21 (z0 )
 + T^11 (z0 ) =1 (4.14)
 K (z0 ) 

^ K ) given by (4.11) to be close to the true T (G0 ; K ) around

for T (G;
z = z0 . This will indeed be the case since K (z0 ) = 0. Furthermore,
(4.13) should hold at frequencies where the blocking zero in z = z0 does
not aect the quality of the estimates.
As a conclusion, the approach described here will deliver a stable matrix
^ K ) only if K has no strictly nonminimum phase zeros. In this
T (G;
case, condition (4.13) must hold, except at possible blocking zeros of K
where (4.14) will hold. Condition (4.13) serves as a way of validating
the quality of the estimates T^11 and T^21 . Finally, unstable poles or
unit-circle zeros of K do not pose any problem in Case 1.

Case 2: estimate T12 and T22 . A model G^ of the plant G0 is recon-

structed from estimates T^12 and T^22 according to (4.10), and the nominal
generalised closed-loop transfer matrix is
2 3
T^12 K T^12
6^ 7
6 T22 + T^12 K T^22 + T^12 K 7
^
T (G; K ) = 66 7:
7 (4.15)
4 ^
T22 K ^
T22 5
T^22 + T^12 K T^22 + T^12 K
60 Identication in closed loop with an unstable or nonminimum phase controller

Evidently, we need
T^22 + T^12 K  1 (4.16)
if T^12 and T^22 are fair estimates of T12 and T22 , respectively, which may
not be the case if K has poles on the unit circle. Also, observe that
T22 and T12 K are respectively the actual closed-loop sensitivity and
complementary sensitivity functions, hence T22 + T12 K  1.
The same reasoning as before leads to the conclusion that T (G; ^ K ) will
be stable in Case 2 only if K has no strictly unstable poles. In this
case, condition (4.16)
 must hold except at possible poles of K on the
unit circle, where T22 + T^12 K  = 1. Condition (4.16) serves as a way of
^ 

validating the quality of the estimates T^12 and T^22 . Finally, nonminimum
phase zeros or unit-circle poles of K do not pose any problem in Case 2.

Conclusion: With the coprime-factor approach, one must

 identify G^ from G^ = T^11T^211, if K has strictly unstable poles and
no strictly nonminimum phase zeros, using fy(t); u(t); r1 (t)g data.
^ K ) will be stable if T^K + T^11  1 except at possible unit-circle
T (G; 21

zeros of K ;
 identify G^ from G^ = T^12 T^221, if K has strictly nonminimum phase
zeros and no strictly unstable poles, using fy(t); u(t); r2 (t)g data.
^ K ) will be stable if T^22 + T^12 K  1 except at possible unit-
T (G;
circle poles of K .
Notice that condition (4.13)4 , respectively condition (4.16)5 , is inherent
in the coprime-factor approach, whatever the stability of the poles and
zeros of the controller. Thus, it is not possible to guarantee the stability
^ K ) a priori for the coprime-factor approach, for instance by
of T (G;
choosing stable structures for T^11 and T^21 , respectively T^12 and T^22 .
If K has both poles and zeros outside the unit circle, the coprime-factor
approach cannot be used to obtain a model G^ stabilised by K , since
^ K ) will be unstable whatever the reference signal being used, hence
T (G;

This condition implies that TK21 + T^11 must be minimum phase and cannot have other
4 ^

unstable poles than the blocking zeros of K .

5
This condition implies that T^22 + T^12 K must be minimum phase and cannot have
other unstable poles than the blocking poles of K .
 Destruction of the nominal stability in the case of unstable poles or nonminimum : : : 61

bG;K
^ = 0. Possible blocking poles and zeros of K do not put a constraint
on the reference signal to be used with this approach.

4.3.3. The direct approach

In this approach, the model G^ of the system is directly identied using

measurements of u(t) and y(t). It has the advantage that the structure
and the order of the model are easily tunable and that the controller
can be unknown, nonlinear and time-varying. However, it also suers
the following drawbacks:

 the input signal used for the identication is correlated with the out-
put noise, with the result that a correct noise model is required to
avoid bias (Forssell and Ljung, 1999);
 the bias error distribution is diÆcult to tune since it is in uenced by
the unknown closed-loop sensitivity function (Ansay et al., 1999).

The nominal closed-loop transfer matrix obtained by the direct approach

is simply
2 3
^
GK G^
6 ^ ^ 7
^ K) = 6
T (G; 6
1 + GK 1 + GK 7
7: (4.17)
4 K 1 5
^
1 + GK ^
1 + GK
Its stability does not hinge on the cancellation of unstable poles or zeros
in K by the transfer function that is identied (G^ here), contrary to
what happens in the indirect and coprime-factor approaches, where this
constraint can make these methods a priori unusable in some cases.
Thus, the direct approach can be used with any external excitation,
r1 (t) or r2 (t), even if K has unstable poles and/or nonminimum phase
zeros. However, there is no guarantee that the obtained nominal closed-
loop model will be stable, and the stability of T (G;^ K ) will have to be
checked a posteriori.
Notice that if K has a blocking zero at frequency !0 , say, and if r1 (t)
is used (r2 (t) = 0), there will hold ry1 (!0 ) = 0. Thus, the output SNR
will be 0 at !0 , yielding a model G^ that can be very dierent from G0
around !0 , which could result in a nominal closed loop (G; ^ K ) that can
62 Identication in closed loop with an unstable or nonminimum phase controller

be unstable or close to instability6. Therefore, r2 (t) should be used to

excite the system in this case. Conversely, any pole of K on the unit
circle will have a blocking eect on r2 (t), and it is then better to use r1 (t)
to excite the system. Notice that the direct method makes it possible
to use both reference signals simultaneously (in the other two methods,
the signal that is not used for identication would act as a disturbance).

4.3.4. Guidelines for an optimal experiment design

The observations made above lead to the following experiment design

guidelines, summarised in Table 4.1.

Indirect Cop.-fac. Direct

Singularities of K r1 ; y r2 ; y r1 ; u r2 ; u r1 r2 r1 r2 r1 + r2
Strictly unstable poles { + { { 0 { 0 0 0
Unit-circle poles { + { { 0 0 0 0 0
Strictly nonmin. ph. zeros { { + { { 0 0 0 0
Unit-circle zeros { { + { 0 0 0 0 0
Table 4.1. Stability of the nominal closed-loop model
w.r.t. the identication method, the excitation signal and
the singularities of K : stability is guaranteed if a stable
model structure is used (+); stability has to be checked
a posteriori (0); instability is guaranteed ({). When K
has several listed singularities, the most unfavourable one
outclasses the others.

In order to guarantee the stability of the nominal closed-loop model, and

if the indirect or coprime-factor approach is used, one must
 excite the system with r2(t) if K has nonminimum phase zeros;
 excite the system with r1(t) if K has unstable poles.
6
Indeed, the signals used for identication are y (t) = 1+GG0 (0s()sK)K(s()s) r1 (t) + N1 (s)v (t)
and u(t) = 1+GK0 ((ss))K (s) r1 (t) + N2 (s)v (t). It is clear that the resulting model G^ (ej! )
is allowed to take any value at ! = !0 if K (ej!0 ) = 0, since this value has no eect
^ j! j! (ej! )
on the closed-loop transfer functions 1+GG(^e(ej!)K)K(e(ej!) ) and 1+G^ (Kej! )K (ej! )
which are
identically zero at that frequency. However, the (1,2) entry of the nominal closed-
^ K ), ^ G^j! (ej! )
loop transfer matrix T (G; 1+G(e )K (ej! )
, depends on G^ (ej! ) even at ! = !0 ,
with the result that, since G^ (ej!0 ) can be anything, this entry might be unstable.
 Consequences for robust control design 63

The indirect approach then guarantees the stability of T (G; ^ K ) a pri-

ori, provided the adequate cross-diagonal entry of T (G0 ; K ) is identied
and its estimate is stable (which can be guaranteed by the choice of a
model structure with enforced stability7 ). However, this method cannot
be used if the controller has both unstable poles and nonminimum phase
zeros (even if these poles and/or zeros are on the unit circle).
With the coprime-factor approach, the stability of the nominal closed-
loop system can only be checked a posteriori (after computing T (G; ^ K )).
This method cannot be used if the controller has both poles and zeros
outside the unit circle (but poles and zeros on the unit circle are allowed).
With the direct approach, either of the excitation signals can be used
(they can also be used simultaneously). The direct approach is the
only method that can be used if the controller has both zeros and poles
outside the unit circle. However, the closed-loop stability of the resulting
model can also be checked only a posteriori.
None of the three classical closed-loop prediction error identication
methods guarantees nominal closed-loop stability if the controller is both
unstable and nonminimum phase. This problem can be overcome by
means of the Hansen scheme (Hansen, 1989), which is based on the
dual Youla parametrisation of all systems that are stabilised by a given
controller K . This will be shown in Section 4.5. But rst we describe
the eects of an unstable nominal closed-loop model on robust control
design, which motivates the need for a reliable solution to this instability
problem.

4.4. Consequences for robust control design

In the previous section, we have seen that, if the controller K present
during identication is unstable or nonminimum phase and unless some
precautions are taken, the nominal model G^ delivered by the identica-
tion experiment could be destabilised by this controller, i.e. bG;K
^ = 0.
We now show that this is a serious drawback in terms of robust control
design.
First, this may be a handicap when trying to check a priori the stability
of a new controller K (possibly designed using G^ , but this is not a
7
Enforcing stability of the estimate is not always possible with some methods like,
e.g., subspace identication methods.
64 Identication in closed loop with an unstable or nonminimum phase controller

requirement) connected to the true system G0 . This prior stability check

is often based on the use of a `small-gain theorem'-based result such as
Proposition 2.1. A suÆcient condition for K to stabilise G0 is that
(Vinnicombe, 1993a)
Æ (K; K ) < bG0 ;K (4.18)
where Æ (K; K ) is the  -gap between K and K (see Section 2.7). Since
bG0 ;K is unknown, a possible solution is to replace it by an estimate
bG;K
^ . This procedure is justied by Proposition 2.1 and has for instance
been suggested in (De Bruyne and Kammer, 1999) as a tool for checking
stability in an iterative feedback tuning scheme. However, this method
fails if bG;K
^ is a poor estimate of bG0 ;K and, a fortiori, if bG;K
^ = 0 while
the true closed-loop system (G0 ; K ) is stable.
A second problem, perhaps more fundamental, is related to the inability
to design a controller achieving high performance with G0 (indepen-
dently of the design method being used). Assume indeed that one uses
G^ to design a controller K that stabilises G^ with stability margin bG; ^ K .

This new K is guaranteed to stabilise G0 if (Vinnicombe, 1993a)
Æ (G0 ; G^ ) < bG;
^ K : (4.19)
However, since K does not stabilise G^ , we know that
bG0 ;K 6 Æ (G0 ; G^ ): (4.20)
This implies that, in order to guarantee the stabilisation of G0 by K
using (4.19), we need to design a controller K with a nominal stability
margin bG;^ K that is necessarily larger than the existing margin bG0 ;K . If
K is a stabilising controller achieving small closed-loop bandwidth (that
is, a controller that one would generally like to replace), then bG0 ;K will
often already be large, meaning that Æ (G0 ; G^ ) is large. We are thus
faced with the problem of computing a controller K with a very large
nominal stability margin bG; ^ ) is large, there
^ K . However, since Æ (G0 ; G
is a risk that
sup bG; ^ K 6 Æ (G0 ; G
^ ): (4.21)

K

In this case, no controller K stabilising G^ is guaranteed to stabilise G0 ,

and it may even be impossible to nd a controller that stabilises both
G^ and G0 .
One might criticise the fact that our reasoning is based on condition
(4.19) which is only suÆcient. However, when doing robust control
 Stability-preserving closed-loop identication with an unstable and nonminimum : : : 65

design, one does not only take the nominal model G^ into account, but
also an uncertainty region D around it, typically computed (in case of
prediction error identication) from the covariance of its parameters,
and guaranteed to contain the true system G0 (with some probability).
We shall see in Chapter 5 that
^ D) , sup Æ (G;
ÆW C (G; ^ G) (4.22)
G2D

should be as small as possible when G^ is used for control design. How-

ever, since
^ D) > Æ (G;
ÆW C (G; ^ G0 ) (4.23)
(because G0 2 D), it is of extreme importance to have Æ (G; ^ G0) as
small as possible; at least, it should not be larger than bG0 ;K , hence G^
should be stabilised by K .
We have shown in Section 4.3 that the presence of an unstable or non-
minimum phase controller may lead to an identied model that makes
the nominal closed-loop system unstable. In this section, we have shown
that this results in a loss of tools for the design of a controller with guar-
anteed stability margins.

4.5. Stability-preserving closed-loop identication

with an unstable and nonminimum phase
controller
While the analysis of Section 4.3 oers no satisfactory solution for the
case where the controller has both nonminimum phase zeros and unsta-
ble poles, it suggests that in such a case one should adopt a method that
uses both reference signals, r1 (t) and r2 (t), while guaranteeing that the
identied closed-loop transfer function is stable.

4.5.1. Using a tailor-made parametrisation

One might think of identifying the four entries of T (G0 ; K ), i.e. to nd

stable estimates of T11 , T12 , T21 and T22 in
      
y(t) = T11 T12 r1 (t) + N1 v(t) (4.24)
u(t) T21 T22 r2 (t) N2
66 Identication in closed loop with an unstable or nonminimum phase controller

using measurements of r1 (t), r2 (t), y(t) and u(t). Since the nal ob-
jective is to obtain a single model G^ for the plant G0 , it is neces-
sary to choose a parametrisation from which G^ can easily be recovered.
The most natural one is the so-called tailor-made parametrisation (van
Donkelaar and Van den Hof, 1997) applied to the whole matrix T :
2 3
G()K G()

6 6 1 + G( )K 1 + G( )K 7
7
T () , T G(); K = 6 7: (4.25)
4 K 1 5
1 + G()K 1 + G()K
An estimate ^ of the parameter vector is then obtained by minimising
the following criterion:
^ = arg min V ()

N   (4.26)
1 X
"(t; ) (t; )  "((t;t; ))
 
, arg min
 N
t=1

where  is some positive denite weighting matrix and "(t; ) and (t; )
are the (possibly ltered) closed-loop prediction errors:
     
"(t; ) = L()D 1 () y(t) T () r1 (t) : (4.27)
(t; ) u(t) r2 (t)
Here, D() is the closed-loop noise model, i.e.
2 3
H ( )
6 1 + G( )K 0 7
D() = 6
4 KH () 5 ;
7 (4.28)
0
1 + G()K
and L() is any stable and possibly parametrised prelter that can of
course be incorporated in D().
The method is based on a direct parametrisation of the open-loop model
^ H^ ) ,
(G; H ) with
a parameter vector . It directly delivers a model (G;
^ ^
G(); H () , thereby comparing with the direct approach, but it has
the advantage that consistency can be guaranteed even if the noise model
^ K ) can again
is wrong. However, the stability of the four entries of T (G;
not be guaranteed (except if the optimisation over  is carried out under
the constraint that the denominator polynomial of the four entries of
the transfer matrix be stable), and the tuning of the structure and of
^ H^ ) can be quite diÆcult.
the order of (G;
 Stability-preserving closed-loop identication with an unstable and nonminimum : : : 67

4.5.2. A two-stage tailor-made parametrisation approach

Here we propose an original two-stage alternative to the tailor-made

parametrisation, which has the advantage of delivering a model G^ that
is guaranteed to be stabilised by K .
The rst stage consists in identifying the four entries T11 , T12 , T21 and
T22 of the closed-loop transfer matrix T (G0 ; K ) as it is done in the four
cases of the indirect closed-loop identication approach, i.e. without a
tailor-made parametrisation. Alternatively, one can consider T (G0 ; K )
as an open-loop multivariable system and obtain an estimate of it using
measurements of the four signals r1 (t), r2 (t), u(t) and y(t). Whatever
approach is chosen (separate identication of the four entries or MIMO
identication of the whole matrix), the objective is to obtain a stable
estimate T^ of T (G0 ; K ). The stability can be guaranteed by an appro-
priate choice of the model structure, at least if the chosen identication
method is prediction error identication.
The second stage consists in nding a model G^ , G(^) that minimises
the following criterion:

02


G()K G() 3 1


B6 1 + G( )K 1 + G()K 7 ^C
7 C
^ = arg min Wl B6
5 T A Wr ; (4.29)
2D @4 K 1

1 + G()K 1 + G()K 2

where Wl and Wr are possible stable and minimum-phase weighting

lters that can be used to re ect the uncertainty distribution around
T^ (making G^ the maximum-likelihood estimate of G0 ), or the spectral
distribution of the signals used during identication, etc. Observe that
the tailor-made parametrisation is introduced in this second stage.
The resulting model G^ is guaranteed to
be stabilised by K since oth-
erwise the H2 norm of Wl T (G; ^ K ) T^ Wr would be innite. Other
criteria could of course be considered, such as the minimisation of the
H1 norm of Wl T (G; ^ K ) T^
Wr ; however, in this case, nominal sta-
bility is no longer guaranteed since an unstable system can have a nite
H1 norm.
68 Identication in closed loop with an unstable or nonminimum phase controller

4.5.3. The dual Youla parametrisation

The Hansen scheme (Hansen, 1989) for closed-loop identication uses
the dual Youla parametrisation of all LTI plants that are stabilised by
a given controller.
Consider any auxiliary system Gaux stabilised by K (i.e. such that
bGaux;K > 0), typically a known but imperfect model of the true plant
G0 (but this is not required), and let Gaux and K have the following
coprime factorisations:
N U
Gaux = and K= ; (4.30)
M V
where N , M , U and V belong to RH1 and are such that the following
Bezout identity holds:
V M + UN = 1: (4.31)
This equation expresses both the fact that the factors are coprime and
that the feedback loop formed by Gaux and K is internally stable.
Then, the set of all LTI plants stabilised by K is given by (Vidyasagar,
1985)
 
N +VR
 K = G (R ) =
M UR
j R 2 RH1 : (4.32)
Using this result, there exists R0 2 RH1 such that the true feedback
system (G0 ; K ) of Figure 2.1 can be redrawn as depicted in Figure 4.1.
Let us now dene the following auxiliary signals, as shown in Figure 4.1:
r(t) = Uy(t) + V u(t) = Ur1 (t) + V r2 (t); (4.33)
z (t) = My(t) Nu(t): (4.34)
They can be reconstructed from the data, and r(t) is uncorrelated with
the noise v(t) since it is the sum of ltered versions of r1 (t) and r2 (t).
We have
z (t) = R0 r(t) + (M UR0 )H0 e(t): (4.35)
A model G^ of G0 can be obtained through the identication of a model
R^ of the `true' dual Youla parameter R0 using data
Z N = fr(1); z (1); : : : ; r(N ); z (N )g: (4.36)
This is an open-loop identication problem since r(t) and e(t) are un-
correlated.
 Stability-preserving closed-loop identication with an unstable and nonminimum : : : 69

r2
r1- f- -+ f?u- -+ f -M 1 y -
6+ K N
6
z f+

+ M UR0 v
+6

R0
r 6
- V - f U 
++

Figure 4.1. Alternative representation of the closed-

loop system of Figure 2.1 using coprime factors and dual
Youla parametrisation of the plant.

Due to the properties of the dual Youla parametrisation, a necessary

and suÆcient condition for a model of the form
N^ (N + V R^ )
G^ = ^ , (4.37)
M (M U R^ )
to be stabilised by K is that R^ be stable, which can be guaranteed by
an appropriate choice of its structure and of the identication method.
As a result, the nominal generalised closed-loop transfer matrix is given
by

^ ^
NU ^ 
NV
T (G; K ) = ^ ^ (4.38)
MU MV
which is stable if and only if N^ and M^ (or, equivalently, R^ ) are stable.
Thus, with the Hansen scheme, the stability of the nominal closed-loop
system is guaranteed by the structure of the solution, whatever the pos-
sible unstable poles and nonminimum phase zeros of K .
The last point we have to address is the choice of the signals to be used
for identication. Notice that
 if K has blocking zeros, these blocking zeros will be in U . Therefore,
in this case, r2 (t) must be nonzero in order for r(t) to be excited
70 Identication in closed loop with an unstable or nonminimum phase controller

at the corresponding frequencies and avoid a poor estimation of R0

around these frequencies: see (4.33);
 if K has blocking poles, these will become blocking zeros in V . There-
fore, in this case, r1 (t) must be nonzero in order for r(t) to be excited
at the corresponding frequencies and avoid a poor estimation of R0
around these frequencies.

If there are no blocking poles or zeros, either of the two reference signals
can be used even if the controller is unstable and/or nonminimum phase;
this is a serious advantage in comparison with the indirect and coprime-
factor approaches, since in most practical cases only one reference signal,
r1 (t) or r2 (t), is available.
The factorisation of K into coprime factors also puts constraints on the
design of r1 (t) and r2 (t). Indeed, the coprime factors U and V may have
a low gain at frequencies that do not necessarily correspond to a blocking
pole or zero. If they are normalised (i.e. such that jU j2 + jV j2 = 1), a low
gain of one of these factors at some frequency always corresponds to a
high gain of the other factor at the same frequency. Consider for instance
this case with r1 (t) and r2 (t) chosen as mutually independent Gaussian
white noise sequences with variance 1: then, the power spectral density
of r(t), r(!), will be 1 at all frequencies. As a result, it is always better
to use both reference signals if possible { all the more so since U and V
may present a low gain over a large frequency range, which could result
in an estimation error that cannot be compensated for by interpolation.
Hence, a locally low gain of U , respectively V , should lead to the same
choice of reference signal as a blocking zero, respectively a blocking pole,
of K . This can be summarised as follows:

 If U has a low gain at some frequencies (due either to the factorisation

of K or to a blocking zero of K ), r2 (t) must be nonzero in order
to have a good SNR and avoid a poor estimation of R0 at these
frequencies;
 If V has a low gain at some frequencies (due either to the factorisation
of K or to a blocking pole of K ), r1 (t) must be nonzero in order
to have a good SNR and avoid a poor estimation of R0 at these
frequencies.
 Numerical illustration 71

4.6. Numerical illustration

In this section, we illustrate the eects of an unstable and nonminimum
phase controller on closed-loop identication by means of a numerical
simulation.

4.6.1. Problem setting

We consider as `true system' the following ARX model:

0:1028z + 0:1812
G0 (z ) = 4 ; (4.39a)
z 1:992z + 2:203z 2 1:841z + 0:8941
3

z4
H0 (z ) = 4 : (4.39b)
z 1:992z + 2:203z 2 1:841z + 0:8941
3
It describes a exible transmission system that was used in (Landau
et al., 1995b) and references therein as a benchmark for testing various
control design methods. The system is normally excited using r2 (t) with
the following control law:
u(t) = F (z )r2 (t) K (z )y(t); (4.40)
where F (z ) is the feedforward part of the controller. Here, we consider
the optimal controller that was obtained via an iterative feedback tuning
scheme in (Hjalmarsson et al., 1995):
0:104z 4 0:0857z 3 0:04274z 2 + 0:03793z + 0:03612
F (z ) = ;
z 3 (z 1)
(4.41a)
0:5517z 4 1:765z 3 + 2:113z 2 1:296z + 0:4457
K (z ) = : (4.41b)
z 3 (z 1)
As F (z ) does not in uence the stability of T (G0 ; K ), we shall replace it
by 1 in the sequel. For the purpose of illustration, we shall also consider
that it is possible to excite the system with reference signal r1 (t)8 .
8
In (Landau et al., 1995b), it was demanded that the controllers produced in the
benchmark study had an R-S-T structure, meaning that F (z ) and K (z ) had to share
the same denominator. In the present case, they both have an unstable pole at z = 1,
and hence the feedforward F (z ) cannot actually be implemented. However, since
this pole also appears in K (z ), it would be easy to replace the equation for u(t) by
u(t) = K (z )(L(z )r1 (t) y (t)), where L(z ) would be a stable feedforward lter. This
justies, if necessary, the replacement of F (z ) by 1, or the use of r1 (t) to excite the
system, in our experiments.
72 Identication in closed loop with an unstable or nonminimum phase controller

With this controller, the achieved generalised stability margin is

bG0 ;K = 0:2761: (4.42)
The maximum value that could be reached for this system is
sup bG0 ;K = 0:4621: (4.43)

K

Now, notice that K (z ) has a unit-circle pole located in z = 1, which

will have a blocking eect on r2 (t). K (z ) also has a pair of strictly
nonminimum-phase complex zeros in z = 1:2622  0:2011j , which may
pose problems if r1 (t) is used alone.
We now test the various closed-loop identication techniques described
in this chapter and we analyse their results in the light of Section 4.4.
Therefore, each of the models G^ obtained by these methods will be
used to compute a two-degree-of-freedom controller C stabilising G^ and
presenting good nominal performance (we want a zero static error and
a faster response than with the current controller K , namely a closed-
loop bandwidth located between the two resonant peaks of the open-loop
system, while keeping the step response overshoot at less than 10%. Be-
cause of the increased closed-loop bandwidth, the new controllers will
normally have smaller stability margins than the current one). The
suitability of a model G^ for control design will be assessed by check-
ing the suÆcient conditions of Section 4.4 and the performance of the
corresponding designed controller with the true system G0 .
The signals r1 (t) and/or r2 (t), and e(t) used to excite the closed-loop
system (G0 ; K ) for identication are chosen as mutually independent
Gaussian sequences with zero mean and variances 1, 1 and 0:05, respec-
tively. No prelter is used for identication (i.e. L(z; )=1), and the
model structures we use are all unbiased9 in order to show that the vari-
ance error that is always committed during identication is suÆcient to
produce the instability problems described in this chapter. It should be
clear that biased models would also present at least the same problems.
The control design method used with each identied model G^ is the fol-
lowing. First, coprime-factor based H1 control design is used to com-
pute a one-degree-of-freedom controller K for an augmented nominal
9
Observe that the indirect and coprime-factor methods, as well as the Hansen scheme,
are `open-loop' methods since they use the reference signals (r1 (t), r2 (t) or r(t)),
which are not correlated with the noise, as input signals. Therefore, output error
model structures can be used with these methods.
 Numerical illustration 73

model G^ aug (z ) = W2 (z )G^ (z )W1 (z ), where W1 and W2 are loop shap-

ing lters aimed at producing good nominal performance, such that K
stabilises G^ aug and satises the following H1 constraint:

^  ) 6 1
T (Gaug ; K +" (4.44)
1 supK bG^ aug ;K
 
where " = 10 6 . Then, a two-degree-of-freedom controller C = F K
for the nominal model G^ is obtained by setting K (z ) = W2 (z )K (z )W1 (z )
and F (z ) = W2 (1)K (1)W1 (z ). W1 will typically contain an integrator
for ensuring zero static error, while W2 will be adjusted in order to
reduce the eect of the second resonant mode.
Sometimes, the order of G^ will be reduced using coprime-factor balanced
truncation10 before control design, the goal being the cancellation of the
nearly nonminimal modes of high-order models produced by coprime-
factor or Hansen scheme identication methods. Indeed, because they
are little observable and controllable, these modes can pose problems for
control design. The reduced-order models are denoted G in the sequel,
and the two-degree-of-freedom
  controllers computed from such models
are denoted C~ = F~ K~ .
For the sake of illustration, the transfer functions of all obtained models
and controllers are given. However, it should be noted that the lim-
ited number of displayed digits may signicantly reduce the numerical
precision of high-order ones.

4.6.2. The indirect approach

Here, we consider two alternatives: identifying T12 using r2 (t) or T21

using r1 (t).
A. Estimating T12 . The closed-loop system was simulated with r1 (t)
set to zero, and r2 (t) and e(t) as dened above. Using 1000 measure-
ments of r2 (t) and y(t), an output error model with exact structure
T^12 (OE[3,8,3]11 ) was obtained, from which a model G^ of the plant was
10
This procedure is explained in Subsection 6.2.1.
11
The notation OE[nb ,nf ,nk ] is the standard notation used in the Matlabr Identi-
cation Toolbox (Ljung, 1995). It represents an output error model with numerator
polynomial B (z ) of order nb 1, denominator polynomial F (z ) of order nf , and delay
nk .
74 Identication in closed loop with an unstable or nonminimum phase controller

derived using (4.3):

G^ (z ) = z9
0:09402z 6 +0:1506z 5
3:274z 8 +4:149z 7 2:489z 6 +0:1145z 5
0:5658z 4 +0:3211z 3
(4.45)
+1:655z 4 1:69z 3 +0:2889z 2 +0:3935z 0:1491 :
Its generalised closed-loop transfer matrix is given by (4.4).
40

20

Magnitude [dB]
0

−20

−40

−60

200

0
Phase [deg]

−200

−400

−600

−800
−3 −2 −1 0
10 10 10 10

Normalised frequency [rad/s]

Figure 4.2. Indirect approach using T12 : Bode dia-

grams of G0 (|) and G^ ( )

Figure 4.2 shows the Bode diagrams of G0 and G^ . As expected, the

estimate is very bad at low frequencies, i.e. around the blocking pole of
K in z = 1. Furthermore, the three reconstructed entries of the nominal
closed-loop transfer matrix (4.4) all have a pole in z = 1, meaning that
the nominal closed-loop transfer matrix is unstable and bG;K^ = 0. This
is because the zero at z = 1 of T12 is imperfectly estimated as a zero at
z = 0:9590 in T^12 , which does not cancel the integrator in K . Finally,
note that
Æ (G0 ; G^ ) = 1 > sup bG;
^ K = 0:2389; (4.46)

K
where the rst equality comes from the fact that the winding number
condition in the computation of the Æ metric is not satised for G0 and
G^ (see (Vinnicombe, 1993a) and Section 2.7). Thus, whatever controller
K stabilises G^ (even the most conservative one), there is no guarantee
 Numerical illustration 75

that this controller will also stabilise the true system G0 . Clearly, ac-
cording to this suÆcient condition, the model we have obtained here is
not suitable for control design.
The following loop shaping lters were used to compute a stabilising
controller for G^ :


W1 (z ) = z z 1 2 and W2 (z ) = 2z:44 z 2 1:017z +0:167
2 0:1519z +0:7423 : (4.47)
Here, a double integrator was necessary to ensure zero static error, be-
cause of the derivator contained in G^ (zero at z = 1). The resulting
two-degree-of-freedom controller is
F (z ) = 0z:205517
2z +1 ;
z2 (4.48a)

K (z ) = 7:z397 z 14 19:55z 13 +26:77z 12 25:83z 11 +13:65z 10 +2:18z 9 8:791z 8

14 1:648z 13 +2:644z 12 3:111z 11 +0:1979z 10 +1:809z 9 2:786z 8

+7:677z 7 3:373z 6 0:5953z 5 +1:125z 4 0:3579z 3 +0:04374z 2 (4.48b)

+2:804z 7 1:178z 6 +0:3108z 5 0:02285z 4 0:04408z 3 +0:03586z 2
8:76210 7 z +6:73810 8 :
0:01274z +0:001813
Its performance with G^ and G0 is depicted in Figure 4.3, where it can
be seen that the nominal closed loop is stable and has very good perfor-
mance while, as expected, the achieved closed loop is unstable.
B. Estimating T21 . The closed-loop system was simulated with r2 (t)
set to zero, and r1 (t) and e(t) as dened above. Using 1000 measure-
ments of r1 (t) and u(t), an output error model with exact structure T^21
(OE[9,8,0]) was obtained, from which a model G^ of the plant was derived
using (4.5):
G^ (z ) = 0:001555 z 12 0:002111z 11 +0:02023z 10 +0:09467z 9 0:1902z 8 0:343z 7
z 12 5:266z 11 +11:02z 10 11:09z 9 +3:195z 8 +6:702z 7 11:34z 6
+0:7336z 6 +0:1266z 5 0:6701z 4 +0:1296z 3 +0:05091z 2 (4.49)
+7:879z 5 0:8743z 4 3:292z 3 +3:238z 2
0:04299z +0:1106 :
1:441z +0:2919
Its generalised closed-loop transfer matrix is given by (4.6).
Figure 4.4 shows the Bode diagrams of G0 and G^ . It also shows the
true T21 , and one can see that its modulus is very low around 0:4 rad=s,
which results in a low SNR around that frequency and explains the bad
quality of the estimate around the rst resonance peak. In the present
case, the error is even worse distributed than in the previous case, since
it is essentially located around the bandwidth of the closed-loop system
rather than in the low-frequency range.
76 Identication in closed loop with an unstable or nonminimum phase controller

Bode diagrams Step responses

50

1.5

Magnitude [dB]
0

−50 1

Amplitude
−100
0.5

500

Phase [deg]
0 0

−500

−0.5
−1000

−1500
−2 −1 0
10 10 10 −1
0 10 20 30 40 50

Normalised frequency [rad/s] Time [samples]

Figure 4.3. Control design via indirect identication

using T12 : Bode diagrams (left) and step responses
(right) of 1+FKG^G^ (|) and 1+FKG
 0 (
G0
)

The three reconstructed entries of the nominal closed-loop transfer ma-

trix (4.6) all have a pair of complex unstable poles in z = 1:2622 
0:2011j , meaning that the nominal closed-loop transfer matrix is unsta-
^ = 0. This is because the zeros at z = 1:2622  0:2011j of
ble and bG;K
T21 are imperfectly estimated as zeros at z = 1:2650  0:1757j in T^21 ,
which do not cancel the nonminimum phase zeros of K .
Finally, note that

Æ (G0 ; G^ ) = 1 > sup bG;

^ K = 0:0032: (4.50)

K

The extremely small value of supK bG; ^ K indicates that, even if one can
compute a controller that stabilises G^ , it will have an extremely small
stability margin and, given the inequality (4.50), it will be very unlikely
that this controller also stabilises G0 . Clearly, the model we have ob-
tained here is not suitable for control design.
 Numerical illustration 77

40

20

Magnitude [dB] 0

−20

−40

−60

200

0
Phase [deg]

−200

−400

−600

−800
−1 0
10 10

Normalised frequency [rad/s]

Figure 4.4. Indirect approach using T21 : Bode dia-

grams of G0 (|), G^ ( ) and T21 (


)

^ K is very small, we could use H1 control design to

Although supK bG;
compute a stabilising controller for G^ . Therefore, the following loop
shaping lters were used:
W1 (z ) = z z 1 and W2 (z ) = 2:44z 2 1:017z +0:167 : (4.51)
z 2 0:1519z +0:7423

The resulting two-degree-of-freedom controller is

F (z ) = 0:001936z ;
z 1 (4.52a)

K (z ) = z94 :06z 14 344z 13 +562:9z 12 546:3z 11 +230z 10 +241:2z 9 510:9z 8

14 0:1143z 13 +0:3503z 12 10:03z 11 +0:7116z 10 +26:38z 9 14:68z 8

+442:9z 7 175:6z 6 74:08z 5 +159:4z 4 121:6z 3 +50:87z 2 10:96z (4.52b)

3:557z 7 +5:736z 6 20:9z 5 +6:221z 4 +7:391z 3 0:3101z 2 +4:127z
1:91910 5
+0:006065 :

Its performance with G^ and G0 is depicted in Figure 4.5, where it can

be seen that the nominal closed loop is stable (but the performance is
poor, due to the small value of supK bG;
^ K ), while the achieved closed
loop is unstable as expected.
78 Identication in closed loop with an unstable or nonminimum phase controller

Bode diagrams Step responses

50

1.5
0

Magnitude [dB]
−50 1

Amplitude
−100
0.5

1000

500

Phase [deg]
0 0
−500

−1000

−1500
−0.5
−2000

−2500

−3000
−2 −1 0
10 10 10 −1
0 20 40 60 80 100
Normalised frequency [rad/s] Time [samples]

Figure 4.5. Control design via indirect identication

using using T21 : Bode diagrams (left) and step responses
(right) of 1+FKG^G^ (|) and 1+FKG
 0 (
G0
)

4.6.3. The coprime-factor approach

Since the controller has a pole on the unit circle and zeros outside the
unit circle, we can expect that the model G^ obtained via the rst alter-
native, i.e. from estimates of T11 and T21 , will be destabilised by K and
unsuitable for control design, while that obtained via the second alter-
native, i.e. from estimates of T12 and T22 , will hopefully be stabilised by
K and suited for control design.
A. Estimating T11 and T21 . Here, the closed-loop system was sim-
ulated with r2 (t) set to zero, while the other signals were chosen as
above. T^21 was obtained as in the indirect approach, while an output
error model with exact structure (OE[6,8,3]) was used to estimate T^11
from 1000 samples of r1 (t) and y(t). A model G^ for the plant was then
 Numerical illustration 79

reconstructed according to (4.9):

G^ (z ) = z16 4:713z015:1772 z 13 0:183z 12 0:3782z 11 +0:5265z 10
+9:325z 14 10:17z 13 +5:534z 12 +2:289z 11 7:844z 10
0:1402 0:1487 +0:4054 0:2058 0:091 +0:06638z 4
z9 z8 z7 z6 z5 (4.53)
+7:117z 9 2:809z 8 0:5269z 7 +1:569z 6 1:035z 5 +0:4535z 4
0:05503z 3 +0:009577z 2 +0:007219z +0:01205 :
0:5123z 3 +0:5407z 2 0:2707z +0:05111
Its generalised closed-loop transfer matrix is given by (4.11). Figure 4.6
shows the Bode diagrams of G0 and G^ .
40

20
Magnitude [dB]

−20

−40

−60

200

0
Phase [deg]

−200

−400

−600

−800
−1 0
10 10

Normalised frequency [rad/s]

Figure 4.6. Coprime-factor approach using r1 (t): Bode

diagrams of G0 (|) and G^ ( )

As expected, the nominal closed-loop transfer matrix (4.11) is unstable

and
Æ (G0 ; G^ ) = 1 > sup bG;
^ K = 0:0299: (4.54)

K
Once again, the model we have obtained here is not suitable for con-
trol design. A H1 controller stabilising G^ could be obtained with the
following weightings:
W1 (z ) = z z 1 and W2 (z ) = 1:09 z 2 0:3616z +0:2327
z 2 0:3072z +0:2683 : (4.55)
The resulting two-degree-of-freedom controller is
F (z ) = 0:z01426
1 ;
z (4.56a)
80 Identication in closed loop with an unstable or nonminimum phase controller

K (z ) = 6:z756 z 22 0:6695z 21 30:26z 20 +49:76z 19 37:37z 18 11:47z 17

22 0:8277z 21 +1:006z 20 0:2377z 19 5:505z 18 +0:9871z 17

+68:08z 16 73:07z 15 +29:78z 14 +15:07z 13 35:11z 12 +25:33z 11

+3:798z 16 1:64z 15 +1:972z 14 +3:642z 13 3:127z 12 0:806z 11
6:7z 10 1:089z 9 +1:279z 8 1:048z 7 +1:829z 6 2:099z 5
(4.56b)
+0:7612z 10 1:709z 9 +0:1477z 8 +0:3441z 7 0:1012z 6 +0:2118z 5
+1:517z 4 0:6833z 3 +0:1967z 2 0:03229z 1:47210 12
+0:02572z 4 +0:05004z 3 0:0005392z 2 +0:007289z +4:69110 5 :

Its performance with G^ and G0 is illustrated in Figure 4.7. The nominal

closed loop is stable (but the performance is poor, due to the small value
of supK bG;
^ K ) while, as expected, the achieved closed loop is unstable.

Bode diagrams Step responses

50

0 1.5
Magnitude [dB]

−50

1
−100

Amplitude
−150
0.5

500

0
Phase [deg]

−500

−1000
−0.5

−1500

−2000
−2 −1 0
10 10 10 −1
0 20 40 60 80 100

Normalised frequency [rad/s] Time [samples]

Figure 4.7. Control design via coprime-factor identi-

cation using r1 (t): Bode diagrams (left) and step re-
sponses (right) of 1+FKG^G^ (|) and 1+FKG
 0 (
G0
)

^ via balanced
Remark. We made an attempt to reduce the order of G
truncation of its normalised coprime factors, the objective being the
cancellation of its nearly nonminimal modes as explained in Subsec-
tion 4.6.1. A 6th-order model G was produced, from which a new con-
troller C~ was produced using the same technique as with the 16th-order
model G^ . However, this new controller also destabilised the true system
G0 .
 Numerical illustration 81

B. Estimating T12 and T22 . Here, the closed-loop system was simu-
lated with r1 (t) set to zero, while the other signals were as above. T^12
was obtained as in the indirect approach, while an output error model
with exact structure (OE[6,8,0]) was used to estimate T^22 from 1000 sam-
ples of r2 (t) and u(t). A model G^ for the plant was then reconstructed
according to (4.10):
G^ (z ) = z13 5:0570:09205 z 10 0:01733z 9 0:6329z 8 +1:456z 7
z 12 +12z 11 18:04z 10 +18:88z 9 13:09z 8 +4:373z 7
1:837z 6 +1:616z 5 0:9527z 4 +0:3649z 3 0:167z 2 +0:1172z 0:03932 :
(4.57)
+1:142z 1:335z 1:118z +2:899z 2:602z +1:179z 0:2212
6 5 4 3 2

Its generalised closed-loop transfer matrix is given by (4.15).

Figure 4.8 shows that, although T^12 and T^22 are very bad estimates at
low frequency because of the integrator in K , the way they deviate from
the true T12 and T22 is the same, with the result that their ratio, G^ , is
close to G0 , as shown in Figure 4.12.

20

10

0
Magnitude [dB]

−10

−20

−30

−40

−50

−60

200
Phase [deg]

−200

−400

−600

−800
−2 −1 0
10 10 10

Normalised frequency [rad/s]

Figure 4.8. Coprime-factor approach using r2 (t): Bode

diagrams of T12 (|), T22 ( ), T^12 (
) and T^21 (


)

Figure 4.9 shows that

T^22 + T^12 K  1 (4.58)
82 Identication in closed loop with an unstable or nonminimum phase controller

60

50

40

Magnitude [dB]
30

20

10

−10

−20

50

Phase [deg]
0

−50

−100
−3 −2 −1 0
10 10 10 10

Normalised frequency [rad/s]

Figure 4.9. Coprime-factor approach using r2 (t): Bode

diagram of T^22 + T^12 K

except near z = 1, i.e. ! = 0, where it goes to innity. Recall that this is

a necessary condition to ensure the stability of the nominal closed-loop
transfer matrix (4.15) when the controller contains an integrator. Here,
the nominal closed-loop transfer matrix is indeed stable and the nominal
stability margin is close to the actual one:
^ = 0:2517  bG0 ;K = 0:2761:
bG;K (4.59)
Furthermore,
Æ (G0 ; G^ ) = 0:1881 < sup bG;
^ K = 0:4560; (4.60)

K

which means that G^ can be used for control design, since every controller
K such that bG;
^ K > 0:1881 will stabilise G0 .

We computed a stabilising controller for G^ . The following loop shaping

lters were used:
W1 (z ) = z z 1 and W2 (z ) = 1:54 z 2 0:5912z +0:208
z 2 0:2672z +0:424 : (4.61)
The resulting two-degree-of-freedom controller is
F (z ) = 0:z16221 z ; (4.62a)
 Numerical illustration 83

K (z ) = 3:143 z 20 17:86z 19 +46:3z 18 72:41z 17 +71:13z 16 30:17z 15

z 20 4:772z 19 +10:89z 18 16:01z 17 +14:77z 16 3:614z 15
32:39z 14 +76:14z 13 75:38z 12 +40:94z 11 4:123z 10 11:61z 9
12:44z 14 +22:81z 13 21:31z 12 +10:26z 11 +1:721z 10 7:256z 9
+6:138z 8 +5:036z 7 10:12z 6 +8:398z 5 4:384z 4 +1:494z 3
(4.62b)
+6:168z 8 2:669z 7 +0:04282z 6 +0:8556z 5 0:7075z 4 +0:3595z 3
0:3179z 2 +0:03537z +1:21610 16 :
0:1307z 2 +0:03059z 0:003873
Its performance with G^ and G0 is depicted in Figure 4.10. Although the
actual performance is not as good as expected, the nominal and achieved
closed loops are both stable, with respective stability margins
bG;
^ K = 0:1175 and bG0 ;K = 0:0106: (4.63)

Bode diagrams Step responses

1.2

10

0
1
−10
Magnitude [dB]

−20

−30
0.8
−40

−50

−60 0.6
Amplitude

−70

200
0.4
0
Phase [deg]

−200
0.2
−400

−600

−800 0

−1000

−1200
−2 −1 0
10 10 10 −0.2
0 20 40 60 80 100

Normalised frequency [rad/s] Time [samples]

Figure 4.10. Control design via coprime-factor identi-

cation using r2 (t): Bode diagrams (left) and step re-
sponses (right) of 1+FKG^G^ (|) and 1+FKG
 0 (
G0
)

As we did in Case A, we also made an attempt to cancel the quasi-

nonminimal modes of G^ by model reduction. Therefore, we computed
^ ^ ^ 1 and we reduced the
h i coprime factorisation G = N M
a normalised
N^
order of M^ by balanced truncation. The plot of the Hankel singular
h i
^
N
values of ^
M (see Figure 4.11) led to the selection of the order 4 for
84 Identication in closed loop with an unstable or nonminimum phase controller

the reduced-order model:

    
N = bt N^ ; 4 ;
M M^
G (z ) = N (z )M 1 (z ) (4.64a)
= 0z:4004326 z 3 0:01012z 2 +0:1422z +0:1666
1:965z 3 +2:154z 2 1:792z +0:8864 :

0
10

−1
10

−2
10

−3
10
0 2 4 6 8 10 12 14

^ obtained by coprime-
Figure 4.11. Reduction of G
h iidentication using r2 (t): Hankel singular values
factor
N^
of M ^

The reduced-order model G has properties close to those of G^ : bG;

 K=
0:2846 and
Æ (G0 ; G ) = 0:2016 < sup bG;
 K = 0:4585: (4.65)

K
Its Bode diagram is plotted in Figure 4.12.
A stabilising controller was computed for G using loop shaping lters
W1 ( z ) = z z 1 and W2 ( z ) = 2:382z 2 1:068z +0:1816 : (4.66)
z 2 0:2556z +0:7513
Its transfer functions are:
F~ (z ) = 0:1588z ;
z 1 (4.67a)
 Numerical illustration 85

40

20

Magnitude [dB] 0

−20

−40

−60

200

0
Phase [deg]

−200

−400

−600

−800
−1 0
10 10

Normalised frequency [rad/s]

Figure 4.12. Coprime-factor approach using r2 (t):

Bode diagrams of G0 (|), G^ ( ) and G (
)

K~ (z ) = 3:447 z 9 9:067z 8 +14:68z 7 18:15z 6 +15:92z 5 10:32z 4

z 9 1:43z 8 +2:668z 7 3:182z 6 +2:318z 5 1:856z 4
+5:138z 3 1:341z 2 +0:1613z +2:16710 7 :
(4.67b)
+0:6727z 3 0:277z 2 +0:1092z 0:02435
In this case, the actual performance is much closer to the nominal one
(see Figure 4.13), and the nominal and achieved closed loops are both
stable with respective stability margins
bG;
 K~ = 0:0700 and bG0 ;K~ = 0:0105: (4.68)

4.6.4. The direct approach

Because of the blocking pole in K , we can expect better results if ref-

erence signal r1 (t) is used to excite the system than if it is not. We
consider three dierent cases, depending on the reference signals used.
In all cases, 1000 samples of u(t) and y(t) were used to identify a model
G^ with the correct structure (ARX[4,2,3]12 ).
12
ARX[na ,nb ,nk ] is the Matlabr notation for an ARX model with denominator poly-
nomial A(z ) of order na , numerator polynomial B (z ) of order nb 1, and delay nk
(Ljung, 1995).
86 Identication in closed loop with an unstable or nonminimum phase controller

Bode diagrams Step responses

1.2

20

0 1

Magnitude [dB]
−20

0.8
−40

−60
0.6

Amplitude
−80

500
0.4

Phase [deg]
0

0.2

−500

−1000 0

−1500
−2 −1 0
10 10 10 −0.2
0 20 40 60 80 100

Normalised frequency [rad/s] Time [samples]

Figure 4.13. Control design via coprime-factor identi-

cation using r2 (t) followed by model reduction: Bode
diagrams (left) and step responses (right) of 1+F~KG~G (|)
and 1+F~KG
~ 0 (
G0
)

A. Using r1 (t) as excitation. Here, r2 (t) was set to zero and the other
signals were as above. The model G^ obtained by direct identication is
close to the true system G0 as shown in Figure 4.14. Its transfer function
is
G^ (z ) = z4 1:987z03:+2
09708z +0:1778
:193z 2 1:838z +0:8806 : (4.69)

It is stabilised by K and the achieved nominal stability margin is close

to that of G0 :
^ = 0:2751  bG0 ;K = 0:2761:
bG;K (4.70)
Furthermore,
Æ (G0 ; G^ ) = 0:0620 < sup bG;
^ K = 0:4760: (4.71)

K

Clearly, this is a good model for control design.

 Numerical illustration 87

40

20

Magnitude [dB] 0

−20

−40

−60

200

0
Phase [deg]

−200

−400

−600

−800
0
10

Normalised frequency [rad/s]

Figure 4.14. Direct approach with r1 (t): Bode dia-

grams of G0 (|) and G^ ( )

We computed a stabilising controller for G^ . The following loop shaping

lters were used:
W1 (z ) = z z 1 and W2 (z ) = 2z:44 z 2 1:017z +0:167
2 0:1519z +0:7423 : (4.72)
The resulting two-degree-of-freedom controller is
F (z ) = 0:z16761 z ; (4.73a)

K (z ) = 3:z821 z 9 9:592z 8 +15:22z 7 18:6z 6 +15:96z 5 10:23z 4 +5:074z 3

9 1:412z 8 +2:666z 7 3:052z 6 +2:144z 5 1:786z 4 +0:6429z 3

1:281z 2 +0:1508z +9:44110 8 :

(4.73b)
0:3043z +0:1292z 0:02949
2

Its performance with G^ and G0 is illustrated in Figure 4.15. The actual

performance is as good as the designed one. The nominal and achieved
stability margins are respectively
bG;^ K = 0:0648 and bG0 ;K = 0:0665: (4.74)

B. Using r2 (t) as excitation. Here, r1 (t) was set to zero and the
other signals were as above. The obtained model is
G^ (z ) = z4 1:976z03:+2
09893z +0:1933
:187z 2 1:851z +0:9042 : (4.75)
88 Identication in closed loop with an unstable or nonminimum phase controller

Bode diagrams Step responses

1.2

10

0
1
−10

Magnitude [dB]
−20

−30
0.8
−40

−50

−60 0.6

Amplitude
−70

200
0.4
0

−200

Phase [deg]
0.2
−400

−600

−800 0

−1000

−1200
−2 −1 0
10 10 10 −0.2
0 20 40 60 80 100

Normalised frequency [rad/s] Time [samples]

Figure 4.15. Control design via direct identication us-

ing r1 (t): Bode diagrams (left) and step responses (right)
of 1+FKG^G^ (|) and 1+FKG
 0 (
G0
)

Because of the blocking pole of K , we could expect less satisfactory

results than in Case A. The  -gap distance between G^ and G0 has indeed
doubled, but the obtained model is still very close to the true system
(see Figure 4.16) and achieves a nominal stability margin bG;K
^ = 0:2725.

We have
Æ (G0 ; G^ ) = 0:1269 < sup bG;
^ K = 0:4603: (4.76)

K

This model is also good for control design. We can explain the { perhaps
surprisingly { good quality of G^ in the low frequency range by the fact
that not only T12 but also N1 has a blocking zero at z = 1, i.e. at ! = 0
(see Figure 4.16). Therefore, the output SNR does not tend to 0 but
remains constant as the frequency decreases. This would not be true if,
for instance, step disturbances were acting on the output of the system.
In this case, indeed, H0 would have a pole at z = 1 which would cancel
the corresponding zero in N1 .
 Numerical illustration 89

40

20
Magnitude [dB]

−20

−40

−60

200
Phase [deg]

−200

−400

−600

−800
−2 −1 0
10 10 10

Normalised frequency [rad/s]

Figure 4.16. Direct approach with r2 (t): Bode dia-

grams of G0 (|), G^ ( ), T12 (
) and N1 H0 (


)

We computed a stabilising controller for G^ . The same lters as in Case A

were used, and the resulting two-degree-of-freedom controller is
F (z ) = 0:z16211 z ; (4.77a)
K (z ) = 3:z738 z 9 9:364z 8 +14:84z 7 18:2z 6 +15:67z 5 10:07z 4 +5:042z 3
9 1:352z 8 +2:568z 7 2:892z 6 +1:948z 5 1:635z 4 +0:5301z 3

1:28z 2 +0:1514z +9:0110 8 :

(4.77b)
0:255z 2 +0:1146z 0:02709
Its performance with G^ and G0 is illustrated in Figure 4.17. The ac-
tual performance is as good as the designed one, and the nominal and
achieved stability margins are respectively
bG;^ K = 0:0650 and bG0 ;K = 0:0780: (4.78)

C. Using r1 (t) and r2 (t) as excitations. Here, all three signals r1 (t),
r2 (t) and e(t) dened above were used to excite the system. The result-
ing model is
G^ (z ) = z4 1:984z03:+2
1016z +0:1841
:195z 2 1:845z +0:896 : (4.79)
As expected, this case yields the best results with an achieved nominal
stability margin bG;K
^ = 0:2776 and an even smaller  -gap between G0
90 Identication in closed loop with an unstable or nonminimum phase controller

Bode diagrams Step responses

1.2

10

0
1
−10

Magnitude [dB]
−20

−30
0.8
−40

−50

−60 0.6

Amplitude
−70

200
0.4
0

Phase [deg]
−200
0.2
−400

−600

−800 0

−1000

−1200
−2 −1 0
10 10 10 −0.2
0 20 40 60 80 100

Normalised frequency [rad/s] Time [samples]

Figure 4.17. Control design via direct identication us-

ing r2 (t): Bode diagrams (left) and step responses (right)
of 1+FKG^G^ (|) and 1+FKG
 0 (
G0
)

and G^ than in Case A:

Æ (G0 ; G^ ) = 0:0507 < sup bG;
^ K = 0:4651: (4.80)

K
The Bode diagram of G^ cannot be distinguished from that of G0 (see
Figure 4.18).
Using the same loop shaping lters as in Cases A and B, we computed
a two-degree-of-freedom controller
F (z ) = 0:z16461 z ; (4.81a)
K (z ) = 3:z744 z 9 9:408z 8 +14:92z 7 18:28z 6 +15:73z 5 10:1z 4 +5:042z 3
9 1:371z 8 +2:601z 7 2:962z 6 +2:039z 5 1:709z 4 +0:5864z 3

1:278z 2 +0:151z +9:16710 8 :

(4.81b)
0:2774z 2 +0:1204z 0:02791

It has similar performance with G^ and G0 , as shown in Figure 4.19, and

the nominal and achieved stability margins are respectively
bG;
^ K = 0:0646 and bG0 ;K = 0:0688: (4.82)
 Numerical illustration 91

40

20
Magnitude [dB]

−20

−40

−60

200
Phase [deg]

−200

−400

−600

−800
0
10

Normalised frequency [rad/s]

Figure 4.18. Direct approach with r1 (t) and r2 (t):

Bode diagrams of G0 (|) and G^ ( )

4.6.5. The Hansen scheme

This method requires the use of an auxiliary model Gaux . We have

chosen a model that represents the plant with a dierent load (Landau
et al., 1995b) (0% instead of 50% load, which corresponds to G0):
0:2826z + 0:5067
Gaux (z ) = 4 : (4.83)
z 1:418z + 1:589z 2 1:316z + 0:8864
3

It also has two resonant peaks, but at dierent frequencies than G0 (see
Figure 4.31 on page 104), and it is stabilised by K .
The auxiliary model and the controller were factorised as in (4.30), in
such a way that the Bezout identity (4.31) held and that the coprime
factors of K were normalised. This makes this factorisation unique.
Figure 4.20 shows the Bode diagrams of these factors. Since V has
a blocking zero at ! = 0, using r2 (t) alone may not produce a good
estimate. But the same problem may occur when using r1 (t) alone,
because of the low gain of U between 0:1 rad=s and 2 rad=s. Hence,
it might be necessary to use both reference signals to obtain a model
92 Identication in closed loop with an unstable or nonminimum phase controller

Bode diagrams Step responses

1.2

20

0 1

Magnitude [dB]
−20

0.8
−40

−60
0.6

Amplitude
−80

500
0.4

Phase [deg]
0

0.2

−500

0
−1000

−1500
−2 −1 0
10 10 10 −0.2
0 20 40 60 80 100
Normalised frequency [rad/s] Time [samples]

Figure 4.19. Control design via direct identication us-

ing r1 (t) and r2 (t): Bode diagrams (left) and step re-
sponses (right) of 1+FKG^G^ (|) and 1+FKG
 0 (
G0
)

really suited for control design. We shall now consider the three possible
scenarios.
A. Using r1 (t) as excitation. The closed-loop system was simulated
with r1 (t) and e(t) dened above, r2 (t) being set to zero. An output
error model structure OE[14,16,3] was used for the Youla parameter R^ .
It was identied using the auxiliary signals dened in (4.33) and (4.34),
and a high-order model G^ of G0 was obtained using (4.37):

G^ (z ) = z28 0:05752z 25 0:04435z 24 0:3217z 23 +0:7451z 22

5:635z 27 +14:56z 26 22:38z 25 +20:58z 24 5:733z 23 14:19z 22
0:5936z 21 0:1538z 20 +0:8464z 19 0:8217z 18 +0:0939z 17 +0:5997z 16
+24:08z 21 15:92z 20 2:564z 19 +15:95z 18 15:39z 17 +4:57z 16
0:6031z 15 +0:116z 14 +0:2132z 13 0:2185z 12 +0:1533z 11 0:1489z 10 (4.84)
+6:224z 15 10:47z 14 +9:158z 13 6:054z 12 +3:087z 11 0:3418z 10
+0:1446z 9 0:07709z 8 0:03287z 7 +0:09886z 6 0:07331z 5 +0:04671z 4
2:177z 9 +3:829z 8 4:067z 7 +3:229z 6 2:143z 5 +1:3z 4
0:05254z 3 +0:03299z 2 0:007687z +0:0008467 :
0:7243z 3 +0:3384z 2 0:1137z +0:01977
 Numerical illustration 93

10

0
Magnitude [dB]

−10

−20

−30

−40

200
Phase [deg]

100

−100

−200

−300
−2 −1 0
10 10 10

Normalised frequency [rad/s]

Figure 4.20. Bode diagrams of the normalised coprime

factors of K : U (|) and V ( )

It achieves bG;K
^ = 0:0612 and

Æ (G0 ; G^ ) = 0:6472 > sup bG;

^ K = 0:5414: (4.85)

K

Although the obtained model is stabilised by K , it has a very low gen-

eralised stability margin and it is therefore not good for control design:
no controller that stabilises G^ would be guaranteed to stabilise the true
system G0 . The Bode diagram of G^ is shown in Figure 4.23. Clearly,
the quality of the estimate is poor in the frequency range where the gain
of U is low.
The following loop shaping lters were used to compute a stabilising
controller for G^ :
W1 (z ) = and W2 (z ) = 2:411 z 1:042z +0:1741
z 2 0:2037z +0:7468 : (4.86)
2
z
z 1
The resulting two-degree-of-freedom controller is
F (z ) = 0:1926z ;
z 1 (4.87a)
94 Identication in closed loop with an unstable or nonminimum phase controller

K (z ) = 3:539 z 33 22:18z 32 +66:82z 31 128:5z 30 +170:7z 29 148:4z 28

z 33 5:321z 32 +14:16z 31 25:11z 30 +31:24z 29 24:96z 28
+51:1z 27 +71:26z 26 140z 25 +111:8z 24 15:4z 23 73:92z 22
+6:413z 27 +14:15z 26 23:6z 25 +16:5z 24 +0:2738z 23 13:62z 22
+96:83z 21 52:25z 20 16:5z 19 +65:62z 18 78:63z 17 +64:18z 16
+15:25z 21 6:909z 20 3:998z 19 +11:34z 18 13:26z 17 +11:08z 16
38:04z 15 +10:64z 14 +12:67z 13 28:18z 12 +33:98z 11 31:37z 10
(4.87b)
6:797z 15 +1:962z 14 +2:283z 13 4:976z 12 +5:87z 11 5:475z 10
+24:11z 9 16:1z 8 +9:614z 7 5:132z 6 +2:369z 5 0:8895z 4
+4:331z 9 2:996z 8 +1:866z 7 1:023z 6 +0:5044z 5 0:2391z 4
+0:2455z 3 0:04199z 2 +0:003479z +3:14910 9 :
+0:1004z 3 0:03273z 2 +0:007215z 0:0008122
It stabilises G^ and presents good nominal performance but, as expected,
it destabilises G0 : see Figure 4.21.
Bode diagrams Step responses

20

0
1.5
Magnitude [dB]

−20

−40
1

−60

Amplitude
−80
0.5

500
Phase [deg]

0 0

−500

−0.5
−1000

−1500
−2 −1 0
10 10 10 −1
0 20 40 60 80 100

Normalised frequency [rad/s]

Time [samples]

Figure 4.21. Control design via Hansen scheme iden-

tication using r1 (t): Bode diagrams (left) and step re-
sponses (right) of 1+FKG^G^ (|) and 1+FKG
 0 (
G0
)

Since the model G^ has a very high order, we made once again an attempt
to reduce it as we did after coprime-factor identication. The inspection
of the Hankel singular values of the coprime factors of G^ (shown in
Figure 4.22) led to the selection of order 4 for the reduced-order model:
G = 0:0002484 z 3 0:01595z 2 +0:09327z +0:1189
z 4 2:053z 3 +2:384z 2 1:988z +0:9338 : (4.88)
 Numerical illustration 95

0
10

−1
10

−2
10

−3
10

−4
10

−5
10
0 5 10 15 20 25 30

Figure 4.22.h Hansen scheme with r1 (t): Hankel singu-

^i N
lar values of ^
M

40

20
Magnitude [dB]

−20

−40

−60

200

0
Phase [deg]

−200

−400

−600

−800
−3 −2 −1 0
10 10 10 10

Normalised frequency [rad/s]

Figure 4.23. Hansen scheme with r1 (t): Bode diagrams

of G0 (|), G^ ( ) and G (
)
96 Identication in closed loop with an unstable or nonminimum phase controller

The reduced-order model G appears to be more suited for control design

than G^ . Indeed, the nominal stability margin with the current controller
has increased to an acceptable value: bG;  K = 0:2795, and the  -gap
between the model and the true system has decreased:
Æ (G0 ; G ) = 0:5121 < sup bG;
 K = 0:5523: (4.89)

K

A stabilising controller was computed for G using the same loop shaping
lters as for the high-order model:
F~ (z ) = 0:z20661 z ; (4.90a)

K~ (z ) = 3:449 z 9 9:325z 8 +15:38z 7 19:1z 6 +16:78z 5 10:85z 4

z 9 1:791z 8 +3:249z 7 3:992z 6 +3:298z 5 2:631z 4
+5:355z 3 1:368z 2 +0:1621z +3:07210 7 :
(4.90b)
+1:279z 3 0:577z 2 +0:2081z 0:04244
In this case, the actual closed loop is stable, although the achieved per-
formance is not as good as the nominal one (see Figure 4.24). The
nominal and achieved closed loops have respective stability margins
bG;
 K~ = 0:0742 and bG0 ;K~ = 0:0558: (4.91)

B. Using r2 (t) as excitation. The closed-loop system was simulated

with r2 (t) and e(t) dened above, r1 (t) being set to zero. The same out-
put error model structure OE[14,16,3] was used for the Youla parameter
R^ . A model G^ with degree 28 was produced as before:
G^ (z ) = 28
z
0:1186z 25 0:3061z 24 +0:2341z 23 +0:07172z 22
5:952z +17:51z 34:88z +53:68z 68:32z +75:77z
27 26 25 24 23 22

0:3082z 21 +0:3768z 20 0:3797z 19 +0:5039z 18 0:8391z 17 +1:116z 16

77:4z 21 +76:3z 20 74:44z 19 +73:23z 18 73:88z 17 +76:67z 16
1:04z 15 +0:8064z 14 0:6629z 13 +0:6404z 12 0:7295z 11 +0:8613z 10 (4.92)
80:55z 15 +83:18z 14 82:16z 13 +76:91z 12 68:41z 11 +57:74z 10
0:985z 9 +1:105z 8 1:162z 7 +1:041z 6 0:7592z 5 +0:4429z 4
45:81z 9 +33:61z 8 22:34z 7 +13:19z 6 6:812z 5 +3:018z 4
0:1954z 3 +0:06108z 2 0:01244z +0:001232 :
1:105z 3 +0:3158z 2 0:06265z +0:006137
It achieves bG;K
^ = 0:1298 and

Æ (G0 ; G^ ) = 0:2366 < sup bG;

^ K = 0:4493: (4.93)

K

This is a better model for control design than the high-order one ob-
tained in Case A; it also matches G0 much better, as can be seen in
Figure 4.27.
 Numerical illustration 97

Bode diagrams Step responses

1.4

20

Magnitude [dB]
0 1.2

−20
1

−40

0.8
−60

Amplitude
−80
0.6

500

0.4
Phase [deg]

0.2
−500

−1000 0

−1500
−2 −1 0
10 10 10 −0.2
0 20 40 60 80 100

Normalised frequency [rad/s] Time [samples]

Figure 4.24. Control design via Hansen scheme iden-

tication using r1 (t) followed by model reduction: Bode
diagrams (left) and step responses (right) of 1+F~KG~G (|)
and 1+F~KG
~ 0 (
G0
)

A stabilising controller was computed for G^ using the following loop

shaping lters:
W1 (z ) = z z 1 and W2 (z ) = 1:558 z 0:5329z +0:1935 : (4.94)
2
z 2 0:1904z +0:409
The resulting two-degree-of-freedom controller is
F (z ) = 0:z16441 z ; (4.95a)

K (z ) = 3:z3233z 5:592
21:51z 32 +70:25z 31 157:5z 30 +274z 29 394:1z 28
33
z 32 +16:74z 31 36:19z 30 +61:92z 29 88:03z 28
+490:1z 27 548:3z 26 +572z 25 573:2z 24 +565:9z 23 562:2z 22
+108:8z 27 121:8z 26 +127:6z 25 128:6z 24 +127:1z 23 125:2z 22
+569:1z 21 587:1z 20 +607:8z 19 617:8z 18 +606z 17 568:9z 16
+125z 21 127:6z 20 +131:8z 19 134:5z 18 +132:7z 17 125:1z 16
+509z 15 432:2z 14 +346:1z 13 258:5z 12 +178z 11 111:8z 10
(4.95b)
+111:8z 15 94:18z 14 +74:12z 13 53:76z 12 +35:51z 11 21:16z 10
+63:33z 9 31:97z 8 +14:2z 7 5:449z 6 +1:762z 5 0:4627z 4
+11:34z 9 5:491z 8 +2:431z 7 0:9982z 6 +0:3812z 5 0:1315z 4
+0:09273z 3 0:01285z 2 +0:0009465z +6:46210 10 :
+0:03876z 3 0:008937z 2 +0:001418z 0:0001153
98 Identication in closed loop with an unstable or nonminimum phase controller

Although the actual performance is not as good as the nominal one (see
Figure 4.25), the nominal and achieved closed loops are both stable,
with respective stability margins
bG;
^ K = 0:1171 and bG0 ;K = 0:0041: (4.96)

Bode diagrams Step responses

1.2

20

0 1

Magnitude [dB]
−20

0.8
−40

−60
0.6

Amplitude
−80

500
0.4
Phase [deg]

0.2

−500

0
−1000

−1500
−2 −1 0
10 10 10 −0.2
0 20 40 60 80 100

Normalised frequency [rad/s] Time [samples]

Figure 4.25. Control design via Hansen scheme iden-

tication using r2 (t): Bode diagrams (left) and step re-
sponses (right) of 1+FKG^G^ (|) and 1+FKG
 0 (
G0
)

Once again, the nominal model G^ has a very high order and we reduced it
using coprime-factor balanced truncation. The inspection of the Hankel
singular values of the coprime factors of G^ (shown in Figure 4.26) led to
the selection of order 4 for the reduced-order model:
G = 0z:000999 z 3 0:007541z 2 +0:1345z +0:1515
4 1:992z 3 +2:164z 2 1:788z +0:8806 : (4.97)

The reduced-order model G has properties close to those of G^ : bG;

 K=
0:2512 and
Æ (G0 ; G ) = 0:2224 < sup bG;
 K = 0:4504: (4.98)

K
Its Bode diagram is shown in Figure 4.27.
 Numerical illustration 99

0
10

−1
10

−2
10

−3
10

−4
10

−5
10

−6
10

−7
10
0 5 10 15 20 25 30

Figure 4.26.h Hansen scheme with r2 (t): Hankel singu-

^i N
lar values of ^
M

40

20
Magnitude [dB]

−20

−40

−60

500
Phase [deg]

−500

−1000

−1500
−2 −1 0
10 10 10

Normalised frequency [rad/s]

Figure 4.27. Hansen scheme with r2 (t): Bode diagrams

of G0 (|), G^ ( ) and G (
)
100 Identication in closed loop with an unstable or nonminimum phase controller

A stabilising controller was computed for G using the following loop

shaping lters:
W1 (z ) = z z 1 and W2 (z ) = 2z:44 z 1:017z +0:167
2 0:1519z +0:7423 : (4.99)
2

The resulting two-degree-of-freedom controller is

F~ (z ) = 0:z16341 z ; (4.100a)

K~ (z ) = 3:809 z 9 9:595z 8 +15:08z 7 18:44z 6 +15:84z 5 10:13z 4

z 9 1:359z 8 +2:631z 7 3:117z 6 +2:248z 5 1:906z 4
+5:078z 3 1:291z 2 +0:1532z +1:99710 7 :
(4.100b)
+0:7239z 3 0:3196z 2 +0:1272z 0:0284
In this case, the actual closed loop is stable, and the achieved perfor-
mance is very close to the nominal one (see Figure 4.28). The nominal
and achieved closed loops have respective stability margins
bG;
 K~ = 0:0627 and bG0 ;K~ = 0:0133: (4.101)

Bode diagrams Step responses

1.2

20

0 1
Magnitude [dB]

−20

0.8
−40

−60
0.6
Amplitude

−80

500
0.4
Phase [deg]

0.2

−500

−1000 0

−1500
−2 −1 0
10 10 10 −0.2
0 20 40 60 80 100

Normalised frequency [rad/s] Time [samples]

Figure 4.28. Control design via Hansen scheme iden-

tication using r2 (t) followed by model reduction: Bode
diagrams (left) and step responses (right) of 1+F~KG~G (|)
and 1+F~KG
~ 0 (
G0
)
 Numerical illustration 101

C. Using r1 (t) and r2 (t) as excitations. The closed-loop system was

simulated with r1 (t), r2 (t) and e(t) dened above. The same output
error model structure OE[14,16,3] was used for the Youla parameter R^ .
Once again, a model G^ with degree 28 was produced:
G^ (z ) = z28 5:952z027:1023 z 25 0:212z 24 0:09333z 23 +0:7217z 22
+16:76z 26 29:68z 25 +36:36z 24 30:82z 23 +16:33z 22
0:973z 21 +0:5141z 20 +0:2219z 19 0:5213z 18 +0:1271z 17 +0:4553z 16
4:23z 21 +3:488z 20 12:06z 19 +20:45z 18 21:91z 17 +17:56z 16
0:575z 15 +0:2748z 14 0:02237z 13 +0:01458z 12 0:1386z 11 +0:2151z 10 (4.102)
12:98z 1 5z 15 +11:52z 14 11:9z 13 +11:58z 12 9:912z 11 +7:813z 10
0:1929z 9 +0:1322z 8 0:06708z 7 +0:003154z 6 +0:01672z 5 +0:01355z 4
6:097z 9 +4:763z 8 3:514z 7 +2:361z 6 1:49z 5 +0:9052z 4
0:03084z 3 +0:01751z 2 0:004096z +0:0004371 :
0:4937z 3 +0:2139z 2 0:063z +0:009152
It achieves bG;K
^ = 0:2847 and
Æ (G0 ; G^ ) = 0:1581 < sup bG;
^ K = 0:4574: (4.103)

K
As expected, this is the best model for control design obtained by ap-
plication of the Hansen scheme. Its Bode diagram is depicted in Fig-
ure 4.31; one can see that it is very close to the true G0 . For the purpose
of illustration, the Bode diagram of the initial auxiliary model Gaux is
also plotted in Figure 4.31, and one can see that it is very dierent from
G0 .
A stabilising controller was computed for G^ using the following loop
shaping lters:
W1 (z ) = z z 1 and W2 (z ) = 2:411 z 2 1:042z +0:1741
z 2 0:2037z +0:7468 : (4.104)
The resulting two-degree-of-freedom controller is
F (z ) = 0:z15961 z ; (4.105a)

K (z ) = 3:69zz33 524:18z 32 +77:87z 31 165:1z 30 +256:5z 29 303:2z 28

33
:403z 32 +15:06z 31 29:14z 30 +42:4z 29 47:3z 28
+275z 27 192:3z 26 +113:9z 25 89:12z 24 +119:7z 23 166:5z 22
+40:75z 27 27:69z 26 +17:4z 25 16:04z 24 +22:08z 23 28:57z 22
+189:2z 21 178:6z 20 +153:1z 19 132:4z 18 +121:6z 17 113:5z 16
+30:3z 21 27:61z 20 +23:78z 19 21:27z 18 +20:08z 17 18:76z 16
+101:4z 15 85:09z 14 +68:04z 13 52:59z 12 +39:23z 11 27:88z 10
(4.105b)
+16:55z 15 13:74z 14 +10:94z 13 8:529z 12 +6:597z 11 4:957z 10
+18:64z 9 11:65z 8 +6:744z 7 3:502z 6 +1:547z 5 0:5423z 4
+3:471z 9 2:195z 8 +1:215z 7 0:585z 6 +0:2629z 5 0:1162z 4
+0:1368z 3 0:02144z 2 +0:001618z +9:14410 10 :
+0:04566z 3 0:01377z 2 +0:002743z 0:0002705
Although the actual performance is not as good as expected (see Fig-
ure 4.29), the nominal and achieved closed loops are both stable, with
102 Identication in closed loop with an unstable or nonminimum phase controller

respective stability margins

bG;
^ K = 0:0663 and bG0 ;K = 0:0394: (4.106)

Bode diagrams Step responses

1.2

20

0 1

Magnitude [dB]
−20

0.8
−40

−60
0.6

Amplitude
−80

500
0.4
Phase [deg]

0.2

−500

0
−1000

−1500
−2 −1 0
10 10 10 −0.2
0 20 40 60 80 100

Normalised frequency [rad/s] Time [samples]

Figure 4.29. Control design via Hansen scheme identi-

cation using r1 (t) and r2 (t): Bode diagrams (left) and
step responses (right) of 1+FKG^G^ (|) and 1+FKG
 0 (
G0
)

Since the model G^ has a very high order, we have once again made an
attempt to reduce it. The inspection of the Hankel singular values of its
normalised coprime factors (shown in Figure 4.30) led to the selection
of order 4 for the reduced-order model G :
G (z ) = 0:z001704z 3 0:0004727z 2 +0:1052z +0:1861
4 1:988z 3 +2:197z 2 1:837z +0:8949 : (4.107)
As shown in Figure 4.31, the two models G^ and G are nearly indistin-
guishable from the plant G0 .
The reduced-order model G has properties as good as those of G^ : bG;
 K=
0:2761 and
Æ (G0 ; G ) = 0:0290 < sup bG;
 K = 0:4580: (4.108)

K
 Numerical illustration 103

Observe that
Æ (G0 ; G ) = 0:0290 << Æ (G0 ; G^ ) = 0:1581; (4.109)
which can be explained by the fact that the reduction procedure has
cancelled the quasi-nonminimal modes of the model, as shown in Fig-
ure 4.32.
0
10

−1
10

−2
10

−3
10

−4
10

−5
10
0 5 10 15 20 25 30

Figure 4.30. Hansenh scheme

i with r1 (t) and r2 (t): Han-
^
N
kel singular values of ^
M

A stabilising controller was computed for G using the same loop shaping
lters as for the high-order model:
F~ (z ) = 0:z16021 z ; (4.110a)

K~ (z ) = 3:667 z 9 9:479z 8 +15:22z 7 18:73z 6 +16:26z 5 10:47z 4

z 9 1:42z 8 +2:661z 7 3:058z 6 +2:138z 5 1:733z 4
+5:195z 3 1:332z 2 +0:1585z +2:14910 7 :
(4.110b)
+0:5923z 3 0:2694z 2 +0:1154z 0:02652
In this case, the actual performance is as good as the nominal one (see
Figure 4.33), and the nominal and achieved closed loops are both stable
with respective stability margins
bG;
 K~ = 0:0662 and bG0 ;K~ = 0:0645: (4.111)
104 Identication in closed loop with an unstable or nonminimum phase controller

40

20

Magnitude [dB]
0

−20

−40

−60

200

Phase [deg]
0

−200

−400

−600

−800
−2 −1 0
10 10 10

Normalised frequency [rad/s]

Figure 4.31. Hansen scheme with r1 (t) and r2 (t): Bode

diagrams of G0 (|), G^ ( ), G (
) and Gaux (


)
2 2

1.5 1.5

1 1

0.5 0.5
Imag Axis

Imag Axis

0 0

−0.5 −0.5

−1 −1

−1.5 −1.5

−2 −2
−2 −1.5 −1 −0.5 0 0.5 1 −2 −1.5 −1 −0.5 0 0.5 1

Real Axis Real Axis

Figure 4.32. Hansen scheme with r1 (t) and r2 (t): Poles

() and zeros (0) of G^ (left) and G (right)
 Numerical illustration 105

Bode diagrams Step responses

1.2

20

0 1
Magnitude [dB]

−20

0.8
−40

−60
0.6

Amplitude
−80

500
0.4
Phase [deg]

0.2

−500

−1000 0

−1500
−2 −1 0
10 10 10 −0.2
0 20 40 60 80 100

Normalised frequency [rad/s] Time [samples]

Figure 4.33. Control design via Hansen scheme identi-

cation using r1 (t) and r2 (t) followed by model reduction:
Bode diagrams (left) and step responses (right) of 1+F~KG~G
(|) and 1+F~KG
~ 0 (
G0
)

4.6.6. Comments on the numerical example

This example conrmed all the theoretical results of this chapter and
clearly showed that, unless an appropriate method is selected together
with an appropriate experiment design, the model that results from
closed-loop identication of a system with an unstable and nonminimum
phase controller can be a very poor estimate of the true system (see
Figures 4.2, 4.4 and 4.6). Furthermore, this model will often not be
stabilised by the controller used during identication, thereby causing
problems with a subsequent control design.
In the present situation, it was possible to use the coprime-factor ap-
proach with excitation r2 (t) because the unstable poles of the controller
were located on the unit circle. However, it should be clear that this
approach would not work with a strictly unstable controller, just as it
did not work with r1 (t) because the controller was strictly nonminimum
phase.
106 Identication in closed loop with an unstable or nonminimum phase controller

The direct approach delivered very good results; however, nominal sta-
bility cannot be guaranteed beforehand, and it should be noticed that
^ H^ ).
we have used an unbiased model structure for (G;
The Hansen scheme delivered models with guaranteed closed-loop sta-
bility; this is the main advantage of this method. However, an inap-
propriate choice of reference signal may yield a model with a very small
generalised stability margin, which is not ideal for control design. There-
fore, the design of the reference signals must take account of the gain
distribution over frequency of the coprime factors of the controller.
From a control design point of view, the direct approach and the Hansen
scheme (with r1 (t) and r2 (t)) delivered the best models, i.e. the mod-
els with generalised stability margin closest to the achieved one, and
the smallest  -gap distance to the true system. This is summarised in
Table 4.2.
Another important nding produced by this numerical illustration is
the fact that the models that are produced by the indirect and coprime-
factor approaches, as well as by the Hansen scheme, may not be ideal
for control design because they are of much higher order than the true
system, thereby having quasi-nonminimal modes. It can be very advan-
tageous to discard these modes by performing a step of model reduction
before control design13 (the order of the true system, to which the high-
order model should be reduced, can easily be determined by observing
that the states in excess have very small Hankel singular values attached
to them: see Figures 4.11, 4.22, 4.26 and 4.30). The resulting low-order
models are then generally much more suited for control design.

4.7. Conclusions
In order to be useful for control design, a model should have a stability
margin as close as possible to the stability margin of the plant with
the same controller, for the reasons explained in Section 4.4. Therefore,
most identication for control methods are based on the computation of
a model from closed-loop data, in such a way that this model is close to
the plant around its crossover frequency.
13
Here, we have used unweighted balanced truncation of the normalised coprime fac-
tors of the high-order models. Details about this technique can be found in Subsec-
tion 6.2.1. However, as explained in Chapter 6, it would be even more advantageous
to take account of the current controller in the reduction criterion.
 Conclusions 107

Method bG;K
^ Æ (G0 ; G^ ) supK bG;
^ K G0 ; C (G^ )
indirect: T12 0 1 0.2389 U
indirect: T21 0 1 0.0032 U
cop.-fac.: r1 (t) 0 1 0.0299 U
idem + reduction 0 1 0.0302 U
cop.-fac.: r2 (t) 0.2517 0.1881 0.4560 S
idem + reduction 0.2846 0.2016 0.4585 P
direct: r1 (t) 0.2751 0.0620 0.4760 P
direct: r2 (t) 0.2725 0.1269 0.4603 P
direct: r1 (t) + r2 (t) 0.2776 0.0507 0.4651 P
Hansen: r1 (t) 0.0612 0.6506 0.5414 U
idem + reduction 0.2795 0.5121 0.5523 S
Hansen: r2 (t) 0.1298 0.2366 0.4493 S
idem + reduction 0.2512 0.2224 0.4504 P
Hansen: r1 (t) + r2 (t) 0.2847 0.1581 0.4574 S
idem + reduction 0.2761 0.0290 0.4580 P
True system G0 0.2761 0.4621
Table 4.2. Summary of the numerical values attached
to the models obtained via dierent closed-loop identi-
cation procedures. The last column summarises the per-
formance of the new controllers C (resp. C~ ) computed
from the models G^ (resp. G ) when they are applied to
the true system G0 (U = unstable closed-loop, S = stable
closed-loop, P = stable closed-loop with good perfor-
mance).

This study has shown that, in the case where an unstable or nonmini-
mum phase controller is used during the closed-loop identication pro-
cedure, the nominal generalised stability margin could be zero although
the plant is stabilised by the controller. As a result, classical tools and
measures used in robust control (e.g. the  -gap metric and the gener-
alised stability margin) to guarantee robust stability and performance
become useless.
108 Identication in closed loop with an unstable or nonminimum phase controller

To cope with this problem, it is possible to use the indirect approach

of closed-loop identication if the controller is not simultaneously un-
stable and nonminimum phase. Depending on the singularities of the
controller, one of the cross-diagonal entries of the generalised closed-
loop transfer matrix must be identied in order to guarantee nominal
closed-loop stability. This is summarised in Table 4.3.

y(t) T12
nonminimum phase zeros
u(t) T21
unstable poles
- r1 (t) r2 (t)
Table 4.3. Entries of T (G0 ; K ) to be identied with
the indirect approach if K has nonminimum phase zeros
or unstable poles, in order to guarantee the stability of
^ K ). The indirect approach cannot be used if the
T (G;
controller is unstable and nonminimum phase.

If the controller is both unstable and nonminimum phase, one solution

is to choose a model parametrisation which ensures closed-loop stabil-
ity. Here, we have proposed to use a Youla-Kucera parametrisation,
as suggested in the so-called Hansen scheme. With that method, and
without poles or zeros of the controller on the unit circle, any reference
signal can be used. If the controller has poles or zeros on the unit circle,
it is best to choose the reference signal(s) as shown in Table 4.4 (not
doing so may produce a model with a huge variance and, hence, a very
large uncertainty region around the frequencies of these blocking poles
or zeros). However, as the signal-to-noise ratio is also enormously in u-
enced by the frequency responses of the normalised coprime factors of
the controller, their gain distribution over frequency should be checked
before choosing a reference signal. Not doing so may produce a model
with a very small generalised stability margin. It is always best to use
both reference signals if possible.
An alternative to the Hansen scheme, when the controller is both un-
stable and nonminimum phase, is the direct approach of closed-loop
identication; the reference signal should then also be chosen according
 Conclusions 109

nonminimum phase zeros: r1 (t) and/or r2 (t)

unstable poles: r1 (t) and/or r2 (t)
zeros on unit circle: r2 (t) (and possibly r1 (t))
poles on unit circle: r1 (t) (and possibly r2 (t))
Table 4.4. Reference signals to be used in the Hansen
scheme and in the direct approach in case of nonminimum
phase zeros and/or unstable poles in K .

to Table 4.4. The coprime-factor approach can also be used if the con-
troller does not have both poles and zeros outside the unit circle (poles
and zeros on the unit circle are allowed); the entries of T (G0 ; K ) that
must be identied are indicated in Table 4.5. However, precautions must
be taken when using these last two methods, since the stability of the
resulting models in closed loop cannot be guaranteed beforehand.

y(t) T11 and T21 T12 and T22

u(t)
strictly strictly
unstable poles nonminimum phase zeros
- r1 (t) r2 (t)
Table 4.5. Entries of T (G0 ; K ) to be identied with the
coprime-factor approach if K has poles or zeros outside
the unit circle. The coprime-factor approach cannot be
used if the controller has both poles and zeros outside
the unit circle.

Finally, it should be noted that a step of model reduction should ideally

be carried out before control design with high-order models produced
by indirect, coprime-factor and Hansen scheme closed-loop identication
approaches, in order to discard their nearly nonminimal modes.
110 Identication in closed loop with an unstable or nonminimum phase controller
 CHAPTER 5
Model validation for control and controller
validation

5.1. Introduction
In this chapter, we elaborate on the question of model validation for con-
trol, already tackled in Chapter 3. This question concerns the construc-
tion of an uncertainty region around a given (possibly low-order) model,
which contains the true system at some specied probability level. The
basics of our approach of model validation were developed by L. Ljung
in (Ljung, 1998). The main idea is that, when a model G^ of a system
G0 is given, one can use the (open-loop) simulation errors to identify a
model for the modelling error, that is for the dierence between G0 and
G^ . This model error model is obtained with an uncertainty region that
can be used either to decide whether G^ is validated, or as information
that one should use when doing robust control design. Although this
validation method is very appealing, it delivers uncertainty regions that
are not tuned for control design, possibly leading to very conservative
controllers.
A rst contribution of this chapter is to propose a closed-loop model
validation procedure, adapted from that of (Ljung, 1998). Its rel-
evance for control design is established by means of a control-oriented
measure of size of the produced uncertainty region: the worst-case  -
gap, proposed M. Gevers in (Gevers et al., 1999b, 2000a, 2000b) and the
technical aspects of which were developed by X. Bombois in (Bombois et
al., 2000a, 2000b). An uncertainty region obtained by closed-loop model
validation will typically be smaller, according to this measure, than an

111
112 Model validation for control and controller validation

uncertainty region obtained under similar conditions by open-loop vali-

dation.
A second contribution is to show that an uncertainty region delivered
by our closed-loop validation procedure will typically lead to less con-
servative robust control designs than an uncertainty region produced by
open-loop validation. More precisely we show, on the basis of two re-
alistic examples, that a given controller may fail to meet some robust
stability and performance specications when it is applied to all systems
contained in an uncertainty region obtained by an open-loop model vali-
dation experiment, while the same specications are met with all systems
contained in an uncertainty region resulting from closed-loop validation.
As a result, a sensible engineer would not accept to apply this controller
to the true system if open-loop validation is used to check its robustness,
while closed-loop validation would give full condence in this controller.
This is controller validation for stability and performance. It uses
tools developed by X. Bombois in (Bombois et al., 1999a, 1999b, 2000c)
for connecting prediction error identication to robust control design.
The outline of this chapter is as follows. In Section 5.2, the open-loop
validation method of (Ljung, 1998) is brie y reviewed. Section 5.3 shows
how it can be adapted to re ect the closed-loop aspects involved in con-
trol design. Some engineering aspects are tackled in Section 5.4. The
tools used for linking prediction error validation with robust control
design are presented in Section 5.5, as well as controller validation pro-
cedures for stability and performance. Finally, the closed-loop and open-
loop validation procedures are compared in realistic simulation examples
in Sections 5.6 and 5.7.

5.2. Open-loop model validation

Let us assume that the true system is the SISO LTI system described
by (2.1)
(
y(t) = G0 (z )u(t) + v(t)
S: v(t) = H0 (z )e(t)
(5.1)

where, as usual, G0 (z ) and H0 (z ) are unknown rational transfer func-

tions, with G0 (z ) strictly proper and H0 (z ) stable and inversely stable.
 Open-loop model validation 113

Remark. If the true system is not LTI, the theoretical derivations of

this chapter are no longer valid. Indeed, they are based on the identi-
cation of an unbiased estimate of the system using prediction error iden-
tication and on the construction, from the covariance of this estimate,
of a frequency-domain uncertainty region containing the true system.
However, the presence of a small amount of nonlinearities will probably
not destroy the ground rules and, as prediction error identication is of-
ten applied to nonlinear systems for computing linear estimates that are
valid around some operating point, so can our validation procedure be
used to build uncertainty regions that are approximately correct around
a chosen operating point. Finally, it should be noted that nonlinearities
in the controller do not pose any problem here.

We assume here that one has given us a model (G; ^ H^ ) that has to be
validated in a way that is relevant for control design, using a set of
validation data collected on the true system (5.1). It must be remarked,
here, that the way this model was obtained is immaterial; it has been
given to us as it is. In particular, we do not assume that this model
was obtained by prediction error identication, nor that it comes with a
condence region attached to it. The validation procedure will consist
in performing some experiment on the true system in order to build an
uncertainty region in which the model will be embedded.
The simulated output of the model (G; ^ H^ ) is denoted
y^(t) = G^ (z )u(t): (5.2)
Most of the model validation tests are based on some statistics over
the residuals (which are the dierence between the measured and the
simulated outputs resulting from the same input signal u(t)), possibly
ltered by some lter L(z ):
"(t) = y(t) y^(t) = G0 (z )u(t) + v(t) G^ (z )u(t); (5.3)
"F (t) = L(z )"(t): (5.4)
Here L(z ) is any frequency weighting prelter. To ease the notations,
we shall assume that L(z ) = 1. Note, however, that a common choice
of prelter is L(z ) = H^ 1 (z ), which makes the residuals "F (t) equal to
the model prediction errors.
Two sources can be considered for the residuals: one that originates from
the input u(t) and which would be zero if the model G^ was a perfect
representation of the plant G0, and another that originates from the
114 Model validation for control and controller validation

process noise. According to this separation, we can write

"(t) = @G(z )u(t) + v(t) (5.5)
where
@G(z ) = G0 (z ) G^ (z ) (5.6)
is the model error. The consideration of two distinct sources for the
model residuals is only relevant if the disturbance v has nothing to
do with the input u. This amounts to say that, in the probabilistic
framework we consider here, u and v should be mutually independent
sequences of random variables (Ljung and Guo, 1997). This requires
that the validation data be not collected in closed loop. Although this
is an important drawback of the standard validation method, we shall
see in Subsection 5.3.2 that a proper choice of the prelter L can help
solve that problem.
The validation procedure proposed in (Ljung and Guo, 1997) and (Ljung,
1998) consists in computing an unbiased estimate G~ ol (^) of @G using a
set of data
Z N = fu(1); "(1); : : : ; u(N ); "(N )g ; (5.7)
with G~ ol () in some model set
n o
Mol = G~ ol (z; ) j  2 Rq : (5.8)
G~ ol is called the model error model. The identication delivers ^ and an
estimate P of cov(^), from which an ellipsoidal uncertainty region in
parameter space can be computed:
Uol =  2 Rq j ( ^)T P 1 ( ^) <  (5.9)


where  is dened as Pr 2 (q) <  = p, with 2 (q) a 2 -distributed
random variable with q degrees of freedom and p the chosen probability
level, typically 0.95 or 0.99. This parametric uncertainty region Uol
denes a corresponding uncertainty region Dol in the space of transfer
functions: n o
Dol(!) = G^ (ej! ) + G~ ol (ej! ; ) j G~ ol(z; ) 2 Mol ;  2 Uol : (5.10)
We then have the following property.
Lemma 5.1. G0 (ej! ) 2 Dol with probability p.

Proof. This follows from the unbiasedness of the model error model
G~ ol : only a variance error is committed.
 Closed-loop model validation 115

The importance of this lemma is that the validation procedure has de-
livered a model set Dol in which the true system is guaranteed to lie, at
some probability level. We can now dene the notion of validation of
the model G^ .
Definition 5.1. (Model validation) The model G ^ is called validated
with the uncertainty region Dol if G(e ) 2 Dol (!) 8! or, equivalently,
^ j!

if 9 2 Uol : G~ (z;  ) = 0.

According to (Ljung, 1998), the uncertainty set Dol can be used for
control design even if the model G^ is falsied. However, if the model
G^ has failed a range of validation attempts, any sensible designer will
want to replace it by a model that is contained in the validated set Dol
(in order to design a robust controller, it is indeed quite reasonable to
choose a nominal model in Dol . A possible choice would be G^ + G~ ol , or
a low-order approximation of it in Dol ). Thus, the typical situation is
where the control design is based on a validated G^ .
We can make the following remarks about this validation procedure.
 A validation experiment is nothing but a prediction error identica-
tion experiment. As a result, the observations we have made about
the frequency distribution of the variance in Chapter 3 also apply to
the distribution of Dol (!) in the present case; hence using Dol (!) for
robust control may yield very conservative control designs.
 If G^ is not validated, one can always embed Dol in a set that also
contains G^ . However, such a procedure would lead to even more
conservative control designs.
Therefore, another way of validating a model should be considered, fo-
cusing on the closed-loop relevant dynamics and using closed-loop vali-
dation data.

5.3. Closed-loop model validation

5.3.1. An indirect approach

The motivation for closed-loop validation follows from the fact that, as
pointed out in Chapter 3, the model G^ should be evaluated by how
well it mimics the behaviour of the actual system G0 when both are
116 Model validation for control and controller validation

connected with the same controller; ideally the optimal to-be-designed

controller.
Consider once again the true system G0 and a model G^ , and suppose
that the input signal is determined by a known stabilising controller Kid
as in Figure 2.1:


u(t) = r2 (t) Kid (z ) y(t) r1 (t) : (5.11)
For the sake of generality, we make the assumption that r1 (t) and r2 (t)
are two possible sources of excitation (reference signals) that are quasi-
stationary and uncorrelated with v(t). The case where only one of these
two signals is available simply puts a constraint on the method to be
used.
The true and the nominal closed-loop transfer matrices are then respec-
tively given by (2.77) with K replaced by Kid and
2 3
^ id
GK G^

6 ^ id 1 + GK
6 1 + GK ^ id 7 ^11 T^12 
T
^ 7, ^
7
T (G; Kid ) = 6 : (5.12)
4 Kid 1 5 T21 T^22
^ id 1 + GK
1 + GK ^ id
The following closed-loop residuals are considered1 :
     
"(t) = y(t) y^(t)
(t) u(t) u^(t)
 
^ Kid )
r1 (t) + N (G0 ; Kid )v(t)
= T (G0 ; Kid ) T (G; r2 (t)
 
@GKid @G  
@GKid2 @GKid r 1 (t )
=
(1 + @GKid + GK ^ id ) r2 (t) + N (G0 ; Kid )v(t)
^ id )(1 + GK
  
, @T11 @T12 r1 (t) + N (G ; K )v(t)
@T21 @T22 r2 (t) 0 id
| {z }
^ Kid )
@T (G0 ; G;
(5.13)

1 ^ Kid ) be
Note that the computation of these closed-loop residuals requires that T (G;
stable.
 Closed-loop model validation 117

where N (G0 ; Kid ) is dened in (2.78). We have assumed that no prelter
is used, but we could use for instance
" #
N 1 ( ^ Kid )H^
1
G; 0
L= ^ Kid )H^
1 (5.14)
0 N2 (G;
which would make the closed-loop residuals equal to the closed-loop
prediction errors.
Since r1 and r2 are uncorrelated with v, we can apply the open-loop val-
idation procedure described in Section 5.2 to any of the four entries of
the closed-loop model error @T (G0 ; G; ^ Kid ), provided the experiment
design guidelines for indirect closed-loop identication given in Chap-
ter 4 are followed. Indeed, assume for instance that a model T~11 of @T11
is identied. Then, an estimate of @G is given by
T^11 + T~11
G~ ind = ^
G: (5.15)
Kid Kid (T^11 + T~11 )
G^ + G~ ind is a rened (i.e. asymptotically unbiased) estimate of G0 , and
a rened estimate of T (G0 ; Kid ) is given by
2 3
^ ~ T^11 + T~11
T11 + T11
T (G^ + G~ ind ; Kid ) = 6
4 Kid 7
5; (5.16)
Kid (1 T^11 T~11 ) 1 T^11 T~11
which presents the same problems of instability as (4.2) when the con-
troller Kid is unstable and/or nonminimum phase. Similar comparisons
^ Kid ) are identied, hence the iden-
hold when other entries of @T (G0 ; G;
tication of a model error model must follow the same guidelines as that
of a nominal model when an indirect closed-loop identication method
is used.
Thus, an unbiased estimate T~ij (^) of the closed-loop model error @Tij is
identied using an appropriate set of data (i.e. measurements of r1 (t) or
r2 (t) and measurements of "(t) or (t)), with T~ij ( ) in some parametrised
model set (with parameter vector  )
n o
Mind = T~ij (z; ) j  2 Rm : (5.17)
The identication delivers ^ and an estimate P of cov(^). The uncer-
tainty region around ^ is given by
n o
Uind =  2 Rm j ( ^)T P 1( ^) <  (5.18)
118 Model validation for control and controller validation

where  is dened as Pr 2 (m) <  = p, with 2 (m) a 2 -distributed
random variable with m degrees of freedom and p the chosen probability
level. The true closed-loop transfer function Tij (ej! ) is then contained
(with probability p) in a condence region
n o
E(!) = T^ij (ej! ) + T~ij (ej! ;  ) j T~ij (z;  ) 2 Mind ;  2 Uind : (5.19)

In accordance with Section 5.2, T^ij is validated if it lies in E. The

important observation now is that, if T^ij is validated with an uncertainty
set E(!), then G^ is validated with an uncertainty set Dind (!) that can
be computed from Uind . Indeed, observe that (5.13) gives the following
relations between the four possible T~ij 's and G~ ind:

T~11 (^)(1 + GK
^ id )2
G~ ind (^) = ^ id ) ; (5.20a)
Kid Kid T~11 (^)(1 + GK
or

T~12 (^)(1 + GK
^ id )2
G~ ind (^) = ^ id ) ; (5.20b)
1 Kid T~12 (^)(1 + GK
or

T~21 (^)(1 + GK
^ id )2
G~ ind (^) = ^ id ) ; (5.20c)
Kid2 + Kid T~21 (^)(1 + GK
or

T~22 (^)(1 + GK
^ id )2
G~ ind (^) = ^ id ) : (5.20d)
Kid + Kid T~22 (^)(1 + GK

An uncertainty region around G^ + G~ ind (^), guaranteed to contain G0

with probability p, is then simply given by

Dind(!) = G^ (ej! ) + G~ ind(ej! ; ) j G~ ind is computed

according to (5.20); T~ij (z;  ) 2 Mind ;  2 Uind : (5.21)
 Closed-loop model validation 119

A possible way of estimating Dind would be the following rst-order

approximation:

@ ~ ind (ej! ;  )
G @ ~ ind (ej! ;  ) 
G
Dind(!)  G(e ) +
^ j!
@1



@m  =^
 ?
~ ~
 Uind  @ Gind(e ; )


@ Gind(e ; )
j! j!
: (5.22)
@1 @m  =^

It delivers an ellipsoidal estimate of Dind (!) centred on G^ (ej! ) +

@ G~ ind (ej! ; ^) at each !. However, this approximation can be very bad;
if it is larger than the true Dind , it may lead to conservative control
designs while, on the contrary, it may not contain the true G0 if it is
underestimated. We therefore suggest to use an algorithm proposed in
(Bombois et al., 1999b), which is based on linear matrix inequalities
(LMI's) and allows to compute Dind very accurately.
This closed-loop validation method presents a serious drawback: it is
based on the indirect approach of closed-loop prediction error identi-
cation, which suers from severe constraints on the excitation source
(r1 or r2 ) if the controller is either unstable or nonminimum phase;
furthermore, it cannot be used if the controller is both unstable and
nonminimum phase. Therefore, another approach should be considered.

5.3.2. A direct approach

As stated in Section 5.2, the `open-loop' validation procedure can only

be used if the input signal is uncorrelated with the noise. Actually, since
we identify a strictly causal model error model, this requirement can be
relaxed, and all that is needed is that u(t) be uncorrelated with the
future noise v(t + 1), v(t + 2), etc. Observe that
 if closed-loop data are used and if v(t) is white, then u(t) is only
correlated with v(t), v(t 1), etc.;
 otherwise, if v(t) is not white, then it is correlated with v(t + k) for
some k > 0, and therefore u(t) is correlated with v(t + k), v(t + k 1),
etc.
To avoid this problem, a rst possible solution is to use a prelter L(z ) =
H^ 1 (z ) which will make L(z )v(t) (almost) white, if H^ is a good model
of the noise dynamics H0 . However, the quality of the result depends
on the accuracy of H^ .
120 Model validation for control and controller validation

A second possible solution is to use a parametrised noise model during

the validation, i.e. L(z; ) = H~ cl 1 (z; ). The closed-loop validation then
consists of minimising the experimental variance of the ltered residuals
prediction errors, i.e. in computing
N
1 X
^ = arg minq 2F (t; ) (5.23)
2R N
t=1

where


F (t; ) = L(z; ) "(t) "^(t j t 1; )


= H~ cl 1 (z; ) "(t) G~ cl (z; )u(t)

(5.24)
= H~ cl 1 (z; ) @G(z ) G~ cl (z; ) u(t)
+ H~ cl 1 (z; )H0 (z )e(t):
It is easy to see, using results of Section 2.2, that
Z   
2
^ !  = arg minq 
@G G~ cl () + B () u
2R 

+ jH H~ H(cl)()j
0
~ 2
0 u jue j2 d! (5.25)
j cl j 2 u

as N ! 1, where 0 is the variance of the white noise e(t) and

 (! )
B (ej! ; ) = H0 (ej! ) H~ cl (ej! ; ) eu (5.26)
u (!)
is a bias term which will be 0 only if the noise model H~ cl (z; ) is exible
enough to represent the true H0 (z ).
As a conclusion, this direct approach of closed-loop model validation will
deliver an unbiased estimate of the modelling error @G only if the model
error model has an unbiased
structure, i.e. if it lies in some model set
that contains @G(z ); H0 (z ) :
n
o

Mcl = G~ cl (z; ); H~ cl (z; ) j  2 Rq 3 @G(z); H0(z) : (5.27)
The validation procedure then works exactly as the open-loop validation
procedure of Section 5.2 and delivers uncertainty regions
Ucl =  2 Rq j ( ^)T P 1 ( ^) < 

(5.28)
 The role of the model error model, how it can be dropped, and a generic form : : : 121

and
n
Dcl(!) = G^ (ej! ) + G~ cl (ej! ; )

o
j G~ cl(z; ); H~ cl(z; ) 2 Mcl;  2 Ucl (5.29)
where P and  are dened as in Section 5.2.
The disadvantage of using the direct closed-loop validation procedure
rather than the indirect one is that an unbiased noise model must also
be identied. However, it has several advantages:
 it can be used even if the controller is unstable or nonminimum phase;
 it does not need to transform an uncertainty region around a closed-
loop transfer function into an uncertainty region around an open-loop
transfer function (therefore, it is also easier to use in the MIMO case);
 the order of the T~ij 's in the indirect approach will typically be greater
than that of G~ cl in the direct approach. This means that a larger num-
ber of parameter is used to identify G~ ind (^) than G~ cl (^), although
both are unbiased estimates of one and the same @G. As a result,
the uncertainty region Dind will generally be larger than Dcl , yield-
ing more conservative control designs. Furthermore, the smaller the
number of parameters, the easier the computation of the optimal es-
timate.

5.4. The role of the model error model, how it can

be dropped, and a generic form for the
uncertainty region
At this point, any sensible reader will have noticed that the role of the
model error model2 G~ only lies in the fact that the covariance matrix P
of its parameter vector ^ allows the computation of an uncertainty region
D containing the true G0 with some probability p. In most practical
applications, the knowledge of the model error model G~ itself is not
useful, since the design of a robust controller will typically be carried
out on the basis of the nominal model G^ and of the uncertainty region
D.
2
The results of this section hold for both the open-loop and the direct closed-loop
case. They also apply to the indirect closed-loop case with some changes in the
notations.
122 Model validation for control and controller validation

Actually, the identication of an unbiased model for @G = G0 G^ us-

ing fu; "g data can be seen as the identication of an unbiased model
for G0 parametrised as G^ + G~ () (i.e. using a model structure with a
xed part G^ and a exible part G~ ()) using fu; yg data. This partic-
ular parametrisation is not useful; on the contrary, the order of G~ ()
will typically be that of G0 plus that of G^ 3 . As a result, an interesting
alternative to the computation of a model error model G~ (^) is the com-
putation of a full-order model G (^) for the plant G0 itself. This is only
a reparametrisation of the problem, but the fact that G () has less pa-
rameters than G~ () makes the optimisation easier and delivers smaller
uncertainty regions.
The uncertainty regions then become
n o
Uol=cl =  2 Rn j ( ^)T P 1( ^) < (5.30)


where is dened as Pr 2 (n) < = p, and
Dol (!) = G ol (ej! ; ) j G ol (z; ) 2 Mol ;  2 Uol (5.31a)
or
Dcl(!) = G cl(ej! ; ) j G cl(z; ); H cl(z; )
2 Mcl ;  2 Ucl : (5.31b)
Here the chosen model set is either
Mol = G ol (z; ) j  2 Rn 3 G0(z) (5.32a)
or
Mcl =  G cl(z; ); H cl (z; )
j  2 Rn 3 G0(z); H0(z)
: (5.32b)

Finally, note that it has been shown in (Bombois et al., 2000c) and
(Gevers et al., 2000b) that, whatever the method being used (open-loop,
direct or indirect closed-loop identication) and the object being esti-
mated (a model G~ for @G or a model G for G0 ), the resulting uncertainty
region D always has the following generic form, which we call generic
prediction error (PE) uncertainty set :
 
(z ) + ZN (z )
D = G(z; ) j G(z; ) = 1 + Z (z) and  2 U  R n
(5.33)
D
where
 ZN (z) and ZD (z) are row vectors of size n of known transfer functions;
3
For example, if G0 (z ) = z2+1z3+z+24 has order 2 and G^ (z ) = z+12 has order 1, then
@G(z ) = G0 (z ) G^ (z ) = (1 31 )z2 +(1 2 +2 3 1 )z+2 2 4 4 has order 3.
z +(3 +2 )z2 +(3 2 +4 )z+4 2
 Model set validation for control and controller validation 123

 (z) is a known transfer function with a delay equal to that of G0(z).

5.5. Model set validation for control and controller

validation
In this section, we brie y describe tools that have been developed in
(Bombois et al., 1999a, 1999b, 2000a, 2000b) for linking PE model valida-
tion with robust control. These tools are, respectively, a control-oriented
measure of size of a generic PE uncertainty set D of the form (5.33)
(Subsection 5.5.1) and controller validation tools for stability (Subsec-
tion 5.5.2) and performance (Subsection 5.5.3). A global picture of
model set validation for control and controller validation using these
tools is presented in the survey papers (Gevers et al., 2000a, 2000b).

5.5.1. Model set validation for control: a design problem

The validated set D of (5.33) depends very much on the experimental

conditions used for the validation experiment: open-loop or closed-loop
setup, choice of signal spectra, etc. Thus, dierent validation experi-
ments yield dierent validated sets D(i) , each containing the true G0
with probability p. It is therefore useful to possess a measure of quality
of a validated model set D that is connected to the size of a controller
set that stabilises all models in D. Such a measure, called the worst-
case -gap, and its frequency-by-frequency counterpart, the worst-
case chordal distance, have been proposed in (Bombois et al., 1999a,
2000a, 2000b) and (Gevers et al., 2000b). They are extensions of Vin-
nicombe's  -gap and chordal distance between two transfer functions
(Vinnicombe, 1993a) and are dened as follows.
Definition 5.2. (Worst-case chordal distance) The worst-case
chordal distance at frequency ! between a nominal system G^ and all
systems in an uncertainty set D is dened as



W C G^ (ej! ); D , sup  G^ (ej! ); G(ej! ) : (5.34)
G2 D
Definition 5.3. (Worst-case  -gap) The worst-case  -gap between
^ and all systems in an uncertainty set D is given by
a nominal system G
^ D) , sup Æ (G;
ÆW C (G; ^ G)
G2D

(5.35)
= max
!
W C G^ (ej! ); D :
124 Model validation for control and controller validation

Here  G^ (ej! ); G(ej! ) and Æ (G;
^ G) are, respectively, the chordal dis-
tance at frequency ! and the  -gap between G^ and G, as dened in
(Vinnicombe, 1993a): see Section 2.7. The second equality in (5.35) is
proved in (Bombois et al., 2000a, 2000b).
LMI-based optimisation tools have been developed to compute the worst-
case chordal distance and the worst-case  -gap between a nominal model
G^ and a PE uncertainty set of the form (5.33): see (Bombois et al.,
1999a) or (Gevers et al., 2000b).
The following theorem gives suÆcient conditions for robust stabilisation
of all plants in a PE uncertainty set D by a given controller K (typically,
a new controller aimed at replacing the possible current controller Kid ).
Theorem 5.1. (Bombois et al., 1999a; Gevers et al., 2000b) Consider
a PE uncertainty set D containing the true G0 with probability p, with
^ D) < 1; and let K be a controller (typically computed from G^ )
ÆW C (G;
that stabilises G^ . Then K stabilises all models in D (and it therefore
stabilises G0 with probability p) if
 

1
W C G(e ); D <  G(e );
^ j! ^ j!
8!: (5.36)
K (ej! )
In particular (Min-Max version), K stabilises all models in D (and it
therefore stabilises G0 with probability p) if
^ D) < bG;K
ÆW C (G; ^ (5.37)
where bG;K
^ is the generalised stability margin of the nominal closed-loop
system, dened by (2.80) or (2.81).

The worst-case  -gap is essentially a control-oriented measure of the size

of D (Bombois et al., 2000a, 2000b). In other words, it is a tool for vali-
dating the experimental design under which the validation is carried out.
If two PE uncertainty sets D(1) and D(2) obtained from two dierent val-
idation experiments are such that ÆW C (G; ^ D(1) ) < ÆW C (G; ^ D(2) ), then
the set of G^ -based controllers that robustly stabilise D(2) is contained
within the set of controllers that robustly stabilise D(1) ; see (Bombois et
al., 2000b) and (Gevers et al., 2000b). Hence, the smaller ÆW C (G; ^ D),
the larger the set of controllers that stabilise D. It is therefore impor-
tant to design the validation experiment in such a way that the obtained
D has a worst-case  -gap (with respect to G^ ) that is as small as pos-
sible. The role of the experimental conditions on the properties of the
validated PE uncertainty set D is discussed in some detail in (Gevers et
 Model set validation for control and controller validation 125

al., 1999b). Here we just stress that, since a major property of the  -gap
is that its resolution is largest around the crossover frequency of the
plant, we expect that closed-loop validation experiments will deliver PE
uncertainty sets with smaller worst-case  -gap than open-loop validation
experiments. This will be illustrated in the two numerical examples of
Sections 5.6 and 5.7.

5.5.2. Controller validation for stability: a verication problem

The stability conditions of Theorem 5.1 are only suÆcient. One might
think that deriving conditions that are also necessary is like crushing a
mosquito with an elephant. However, as our application examples will
show, the suÆcient conditions may be quite conservative and lead to
reject a controller that, in fact, robustly stabilises D. The following
theorem (only valid in the SISO case) gives a necessary and suÆcient
condition for a given controller K (z ) to stabilise all models in a PE
uncertainty set D.
Theorem 5.2. (Bombois et al., 2000c) Consider a generic PE uncer-
tainty set D of the form (5.33) and a controller K (z ) = X (z )=Y (z )
that stabilises the centre of that set, G (z; ^). Then all models in D are
stabilised by K (z ) if and only if


max
!
 MD (ej! ) 6 1; (5.38)

where

 MD (z) is dened as

!
X (z ) ZN (z ) (z )ZD (z )
ZD (z ) + T 1
Y (z ) + (z )X (z )
MD (z ) =
! ; (5.39)
X (z ) ZN (z ) (z )ZD (z ) ^
1 + ZD (z ) + 
Y (z ) + (z )X (z )

 T is a square root of the matrix P  dening U: P  = T T T ;

1 1

  = T ( ^), whereby  2 U , kk2 < 1;

126 Model validation for control and controller validation

  MD(ej! ) is called the stability radius of the loop (MD (z ); ). For
a real vector  it is computed as follows:


 MD (ej! )
8s
2
>
> Re(MD )Im(MD )T
>
< kRe(MD )k 2 if Im(MD ) 6= 0,
=
2
kIm(MD )k22 (5.40)
>
>
>
: kM D k 2 if Im(MD ) = 0.

5.5.3. Controller validation for performance: a verication problem

When designing a controller for a plant G0 , one does generally not only
aim at stabilising this plant: there is most often a performance issue
(sometimes, the plant is even already stable in open loop). It is thus
important to be able to assess the worst-case performance a controller
will achieve over a validated PE uncertainty set. Here, we consider a
worst-case performance criterion JW C expressed in the frequency do-
main by weighting the four entries of the closed-loop transfer matrix
individually:
JW C (D; K; Wl ; Wr ; !) =
0 1
Wl (ej! ) Wr (ej! )
z
 }| { z }| {

B
max  B Wl1 0  T G(ej! ); K (ej! )  W0r1

0 C C
(5.41)
B
G2D @ 0 Wl2 Wr2 CA

where Wl1 (z ); Wl2 (z ) and Wr1 (z ); Wr2 (z ) are frequency weights that al-
low one to dene specic performance levels, and where  (
) denotes
the largest singular value. The frequency function JW C denes a tem-
plate. Any function that is derived from JW C can of course also be
handled, such as kJW C k1, for example. It has been shown in (Bombois
et al., 2000c) that JW C can be computed by solving an optimisation
problem involving LMI constraints.

5.6. Illustration I: exible transmission system

Our rst illustration is representative of control applications of exi-
ble mechanical structures subject to step disturbances. It is based on
 Application I: exible transmission system 127

a benchmark proposed by I.D. Landau in a special issue of the Euro-

pean Journal of Control (Landau et al., 1995b). We illustrate the fact
that, for control-oriented validation, a closed-loop experiment with a
at reference spectrum typically yields a smaller worst-case  -gap and
an uncertainty region that leads to less conservative control designs than
an open-loop experiment with a similar at input spectrum.

5.6.1. Problem setting

We consider as true system the half-load model of the exible transmis-

sion system described in (Landau et al., 1995b):
B (z ) 3
G0 (z ) = z
A(z )


0:10276 + 0:18123z 1 z 3
= : (5.42)
1 1:99185z 1 + 2:20265z 2 1:84083z 3 + 0:89413z 4
The output of the system is subject to step disturbances ltered through
1
H0 (z ) = : (5.43)
A (z )
One of the major specications of the benchmark was that the designed
controllers should ensure rejection of step output disturbances ltered by
1
A(z ) within 1:2 s (the sampling period being 0:05 s): see (Landau et
al., 1995b). This means that the plant can be seen as a nonstandard
ARX system described by
A(z )y(t) = B (z )z 3 u(t) + p(t) (5.44)
where u(t) is the input of the plant, y(t) its output and p(t) a sequence
of step disturbances with zero mean and variance p2 . From a stochastic
point of view, p(t) is equivalent, up to second order moments, with
1 1

(z ) e(t) where
(z ) = 1 z and e(t) is a sequence of Gaussian white
noise with zero mean and appropriate variance. Hence, a standard ARX
description of the plant is given by
A(z )
(z )y(t) = B (z )z 3
(z )u(t) + e(t); (5.45)
and the standard prediction error identication algorithm for ARX mod-
els can be used to identify the system, provided the data are preltered
by L(z ) =
(z ).
128 Model validation for control and controller validation

Our objective is to validate (or invalidate) the following model and con-
troller on the basis of data collected on G0 :
B^ (z )
G^ (z ) = ^ z 3
A(z ) (5.46)


0:0990 + 0:1822z 1 z 3
=
1 1:9878z 1 + 2:2004z 2 1:8456z 3 + 0:9038z 4
and
0:0355 + 0:0181z 1 18:8379 43:4538z 1 + 26:4126z 2
K (z ) =
1 z 1
 1 + 0:6489z 1 + 0:1478z 2
0:7492z 1 + 0:3248z 2 1:3571 1:0741z 1 + 0:4702z 2
 0:5626
1 1:4998z 1 + 0:6379z 2
 1 0:6308z 1 + 0:3840z 2
+ 0:5633z 2
 1 + 10::0461
4564z 1 + 0:1530z 2
: (5.47)
This controller was obtained in (Nordin and Gutman, 1995) using QFT
design. It is the controller that satises the largest number of speci-
cation of the benchmark (Landau et al., 1995b). It also presents very
good performance when applied to our nominal model G^ . Of course, for
the purpose of illustrating our theory, we pretend not to know if this
controller stabilises the true system.

5.6.2. Open-loop validation experiment

We rst performed an open-loop validation experiment. The input sig-

nal uol (t) applied to the true system G0 (z ) was chosen as a PRBS
with variance u2ol = 0:1 and at spectrum. The output step distur-
bances p(t) were simulated as a random binary sequence with vari-
ance p2 = 0:01 and cut-o frequency at 0.05 times the Nyquist fre-
quency; that is, the mean length of the steps of p(t) is about twenty
times longer than that p of the steps of uol (t) while the amplitude of
the steps of p(t) is 10 times smaller than that of uol (t). The spectra
and a realisation of uol (t) and p(t) are shown in Figure 5.1. p(t) was
ltered by A(1z) and added to the output of the system. 256 data sam-
ples fuol (1); yol (1)
; : : : ; uol (256); yol (256)g were measured, and a model
G ol (^ol ); H ol (^ol ) with the same ARX[4,2,3] structure as (G0 ; H0 ) was
identied after preltering these data by
(z ). Thus, here, the valida-
tion is based on the identication of an unbiased model for G0 rather
than on the identication of a model error model for G0 G^ . The model
 Application I: exible transmission system 129

G ol (z; ^ol ) is given by

0:1052 + 0:1774z 1 z 3
G ol (z; ^ol ) = : (5.48)
1 1:997z 1 + 2:23z 2 1:876z 3 + 0:9039z 4
The covariance matrix of the parameter vector is
2
0:2034 0:2970 0:2411
6
6 0:2970 0:5735 0:5136
Pol = 0:001 
6
6 0:2411 0:5136 0:5725



6
6 0:1150 0:2397 0:2962
4 0:0139 0:0119 0:0130
0:0027 0:0064 0:0008
3
0:1150 0:0139 0:0027
0:2397 0:0119 0:0064 7
7



0:2962 0:0130 0:0008 7
7: (5.49)
0:2013 0:0094 0:0020 7
7
0:0094 0:0392 0:0126 5
0:0020 0:0126 0:0391

0.4

0.2

−0.2

−0.4
0 50 100 150 200 250
data samples

0
10

−1
10

−2
10

−3
10

−4
10
−2 −1 0
10 10 10
normalised frequency [rad/s]

Figure 5.1. Open-loop validation. Top: realisations of

uol (t) ( ) and p(t) (|). Bottom: uol (!) ( ) and
p (!) (|)
130 Model validation for control and controller validation

The following observations were made:

(nom ^ol )T Pol1 (nom ^ol ) = 9:7168 < = 12:6 (5.50)
and
(0 ^ol )T Pol1 (0 ^ol ) = 5:7555 < = 12:6; (5.51)
where 0 and nom denote respectively the vectors of coeÆcients of G0 (z )
and G^ (z ) in the same order as those of G ol (z; ^ol ). This means that both
G0 and G^ lie within the 95% uncertainty region Dol around G ol (^ol ),
hence G^ is validated at this probability level. (For
p = 0:95 and n = 6
free parameters, we have = 12:6 in Pr 2 (n) < = p.)

5.6.3. Closed-loop validation experiment

Here, the validation experiment is carried out in closed loop, the con-
troller present in the loop during the validation procedure being the
one obtained in (Landau et al., 1995a) using a combined pole place-
ment/sensitivity function shaping method. Its feedback part is described
by
0:401602 1:079378z 1 + 0:284895z 2 + 1:358224z 3
Kid (z ) =
1 1:031142z 1 0:995182z 2 + 0:752086z 3 (5.52)
0:986549z 4 0:271961z 5 + 0:306937z 6
:
+0:710744z 4 0:242297z 5 0:194209z 6
It also has a feedforward part that we shall not consider here, since
we want an excitation signal with a at spectrum. According to the
benchmark specications, the performance of this controller is not as
good as that of the to-be-validated controller K . We use it here only for
the validation experiment.
The closed-loop system (G0 ; Kid ) was excited by means of a reference
signal r(t) injected at the input of G0 (r2 (t) in Figure 2.1). In order
to establish a fair comparison with the results obtained in open-loop
validation, r(t) was chosen as a PRBS with variance r2 = 0:5541, while
p(t) was the same as in the open-loop validation experiment. With these
settings, the total output variance was the same in closed loop than in
open loop: y2cl = y2ol = 0:8932.
Kid has unstable poles and nonminimum phase zeros. Therefore, the in-
direct closed-loop identication method could not be used for validation,
as it would systematically have delivered a model G cl (^cl ) that would
 Application I: exible transmission system 131

have been destabilised by Kid . We therefore used the direct closed-loop

identication method to compute a prediction error uncertainty set Dcl .
Once again, 256 data samples fucl (1); ycl (1)
; : : : ; ucl (256); ycl (256)g were
measured, and a model G cl (^cl ); H cl (^cl ) with the same ARX[4,2,3]
structure as (G0 ; H0 ) was identied after preltering these data by
(z ).
The model G cl (z; ^cl ) is given by
0: 1016 + 0:1782z 1
z 3
G cl (z; ^cl ) = : (5.53)
1 1:986z 1 + 2:187z 2 1:824z 3 + 0:8897z 4
The covariance matrix of its parameter vector is
2
0:0840 0:1166 0:1024
6
6 0:1166 0:2145 0:1966
Pcl = 0:001 
6
6 0:1024 0:1966 0:2184



6
6 0:0532 0:1009 0:1197
4 0:0062 0:0057 0:0074
0:0027 0:0008 0:0041
3
0:0532 0:0062 0:0027
0:1009 0:0057 0:0008 7 7
0: 1197 0 : 0074 0: 0041 7



0:0853 0:0063 0:0037 77 : (5.54)
7
0:0063 0:0064 0:0021 5
0:0037 0:0021 0:0061
As in the open-loop case, the nominal model is validated with respect to
the 95% uncertainty region Dcl around G cl (z; ^cl ), which also contains
the true system:
(nom ^cl )T Pcl1 (nom ^cl ) = 9:7590 < = 12:6; (5.55)
(0 ^cl ) Pcl (0 ^cl ) = 4:7050 < = 12:6:
T 1 (5.56)

5.6.4. Model set validation for control: comparison of the validated

sets

We computed the worst-case chordal distance and the worst-case  -gap

between G^ and all models in Dol and, respectively, Dcl , and we found
that
^ Dol ) = 0:3010 > bG;K
ÆW C (G; ^ = 0:2381
(5.57)
^ Dcl ) = 0:1423:
> ÆW C (G;
132 Model validation for control and controller validation

Hence, according to the Min-Max version of Theorem 5.1, the open-

loop validation experiment does not guarantee that K stabilises the true
system G0 , while the closed-loop validation experiment does. However,
since
 

1
W C G^ (ej! ); Dol <  G^ (ej! ); (5.58)
K (ej! )
and
 

1
W C G^ (ej! ); Dcl <  G^ (ej! ); (5.59)
K (ej! )
at all frequencies (see Figure 5.2), both validation experiments lead to
the conclusion that K stabilises G0 . Yet, the uncertainty region Dcl
delivered by the closed-loop experiment should be preferred over Dol
delivered by the open-loop experiment, since it leads to the largest set
of stabilising controllers: see (Bombois et al., 2000a, 2000b).

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
−2 −1 0
10 10 10
normalised frequency [rad/s]

Figure 5.2. W C G^ (ej! ); Dol
(
),
W C G^ (ej! ); Dcl
(|),
 G^ (ej! ); G0 (ej! ) ( ),
 G^ (ej! ); K (e1j! ) (


)
 Application I: exible transmission system 133

5.6.5. Controller validation for stability

Using the open-loop and closed-loop PE uncertainty sets Dol and Dcl , we
built the dynamic matrices MDol (ej! ) and MDcl (ej! ) with respect to the
candidate controller K and we computed their stability radii according
to Theorem 5.2. Their respective values are
j!

max
!
 M D ol ( e ) = 0:3227 < 1; (5.60)
and


max
!
 MDcl (ej! ) = 0:2384 < 1: (5.61)
According to this test, K stabilises all systems contained in both uncer-
tainty sets Dol and Dcl , hence both the open-loop and the closed-loop
approach validate the controller. This was predictable since such a con-
clusion had already been obtained from the more conservative test based
on the worst-case chordal distance.

5.6.6. Controller validation for performance

A major specication of the benchmark was that the designed controller
should ensure a maximum value of less than 6 dB of the output sen-
sitivity function (modulus margin greater than 0.5). Our performance
criterion is thus  
 1 
JW C (D; K; !) = max   (5.62)
G(ej! ; )2D  1 + G(ej! ;  )K (ej! ) 

obtained by setting Wl1 = Wr1 = 0 and Wl2 = Wr2 = 1 in (5.41), the

controller being validated if max! JW C (D; K; !) < 6 dB .
For the open-loop and closed-loop PE uncertainty sets Dol and Dcl , we
have respectively
max
!
JW C (Dol ; K; !) = 5:97 dB (5.63)
and
max
!
JW C (Dcl ; K; !) = 5:00 dB; (5.64)
which means that the open-loop validation procedure nearly leads to
reject the controller, while the closed-loop validation procedure leads
to its acceptance. The Bode diagrams of the worst-case and achieved
sensitivity functions are depicted in Figure 5.3.
134 Model validation for control and controller validation

10

dB
−5

−10

−15
−2 −1 0
10 10 10
normalised frequency [rad/s]

Figure 5.3. JW C (Dol ; K ) (
), JW C (Dcl ; K ) (|),

jS (G0 ; K )j ( ), jS (G;^ K )j (


),
6 dB -limit (
)

5.7. Illustration II: ferrosilicon production process

We now present an application that is representative of industrial process
control applications, in which the control objective is the rejection of
stochastic disturbances.

5.7.1. Problem setting

The plant model and the controllers used in this simulation example
are taken from (Ingason and Jonsson, 1998). Ferrosilicon is a two-phase
mixture of the chemical compound FeSi2 and the element silicon. The
balance between silicon and iron is regulated around 76% of the total
weight in silicon, 22% in iron and 2% in aluminium by adjusting the
input of raw materials to the furnace. Those are charged batchwise
to the top of the furnace, each batch consisting of a xed amount of
quartz (SiO2 ) and a variable quantity of coal/coke (C) and iron oxide
(Fe2 O3 ). The quantity of coal/coke which is burned in the furnace does
 Application II: ferrosilicon production process 135

not in uence the silicon ratio in the mixture, hence the control input is
the amount of iron oxide.
In (Ingason and Jonsson, 1998), the following ARX model has been
obtained for the process:
y(t) + ay(t 1) = bu(t 1) + d + e(t) (5.65)
where the sampling period is one day, y(t) is the percentage of silicon
in the mixture that must be regulated around 76%, u(t) is the quantity
of iron oxide in the raw materials (expressed in kilogrammes), d is a
constant and e(t) is a stochastic disturbance. The nominal values of the
parameters and their standard deviations are:
a = 0:44; b = 0:0028; d = 46:1; (5.66a)
a = 0:07; b = 0:001; d = 5:6: (5.66b)
Here, for the sake of illustrating our theory, we make the assumption
that the true system is
bz 1 0:0032z 1
G0 (z ) = 0 = ; (5.67a)
1 + a0 z 1 1 0:34z 1
1 1
H0 (z ) = = ; (5.67b)
1 + a0 z 1 1 0:34z 1
for which4 a0 2 [a 2a ; a + 2a ] and b0 2 [b 2b ; b + 2b ], while the
nominal model is the one of (Ingason and Jonsson, 1998):
bz 1 0:0028z 1
G^ (z ) = = ; (5.68a)
1 + az 1 1 0:44z 1
1 1
H^ (z ) = = : (5.68b)
1 + az 1 1 0:44z 1
This model was used in (Ingason and Jonsson, 1998) to compute a GPC
controller. The control law that minimises the cost function
2
X
2
JGP C (u) = E y(t + i) r(t + i)
i=1
!
2
X
2
+ 
(z )u(t + i 1) ; (5.69)
i=1

4
Since we have no access to the real plant, we have decided to randomly pick one
system in the two-standard-deviation condence interval around the nominal model
and to use it as the true system.
136 Model validation for control and controller validation

where
(z ) = 1 z 1 , is given by

 
1
u(t) = 1 0 H T H + F T F


 H T w(t) x(t) F T g(t) (5.70)
where
 
H = bab 0b ; (5.71a)
 
1
F= 1 1 ; 0 (5.71b)
 
x(t) = ay(t) + d a2 y(t) ad + d T ; (5.71c)
 
w(t) = r(t) r(t + 1) T ; (5.71d)
 
g(t) = u(t 1) 0 T (5.71e)
and
 = I: (5.71f)
 is a tuning parameter. The resulting controller is a two-degree-of-
freedom controller C (z ):
 
r
u(t) = C (z ) y(t)( t)



 r (t)
 (5.72)
= F (z ) K (z ) y(t)
where
b3 + 2b ab

F (z ) = 4
(5.73a)
b + 3b2  + a2 b2  + 2 2ab2  b2  + 2 z 1
and
ab3 + ab a2 b +
a3 b
K (z ) =
: (5.73b)
b4 + 3b2  + a2 b2  + 2 2ab2  b2  + 2 z 1

Experiments made by the authors of (Ingason and Jonsson, 1998) showed

that, for a constant reference r(t) = 76 and for  = 0:0007, a control
input variance u2 = 20 was obtained. This results from a white noise
variance 0 = 0:078 for the true system described by (5.66).
In this example, we shall compare open-loop and closed-loop experi-
ments for the validation of the GPC controller C (z ) with  = 0:0007.
 Application II: ferrosilicon production process 137

5.7.2. Open-loop validation experiment

The `true plant' model (G0 ; H0 ) was excited with u(t) chosen as a PRBS
with variance u2ol = 20, which is the maximum input variance admissi-
ble for this process. e(t) was chosen as a Gaussian white noise sequence
with variance 0 = 0:078, which corresponds to the noise acting on the
real process. The variance of the output was then y2ol = 0:0884. Recall
that the validation experiment, i.e. the construction of a validated set
Dol , consists of performing a PE identication using an unbiased model
structure. Therefore, 300 input-output data samples were collected,
corresponding approximately to one year of measurements. These data
were used to identify an ARX model with exact structure
z 1
G ol (z; ol ) = 2 ; (5.74a)
1 + 1 z 1
1
H ol (z; ol ) = : (5.74b)
1 + 1 z 1
We found
   
^ol = ^1 = 0:3763 ; (5.75)
^2 0:0073
 5
2 : 8131  10 3 1
Pol = 1:2784  10 5 1:4887  10 5 ;: 2784  10 (5.76)
and the nominal model is validated with respect to the 95% uncertainty
region Dol around G ol (z; ^ol ):
     
a ^ T P 1 a ^ = 2:6151 < = 5:99: (5.77)
b ol ol b ol

(For p = 0:95
and n = 2 free parameters, we have = 5:99 in

Pr 2 (n) < = p.) Observe that Dol , whose probability of containing
the true plant G0 is p = 95%, eectively contains it:
  T   
a0 ^ol Pol
1 a0 ^ol = 1:6744 < = 5:99: (5.78)
b0 b0

5.7.3. Closed-loop validation experiment

The closed-loop validation was performed with a suboptimal GPC con-

troller obtained by setting  = 0:001 in (5.73). Since this value of 
reduces the input variance (compared to the optimal one  = 0:0007),
138 Model validation for control and controller validation

it was possible to add a PRBS signal to the reference r(t) without vi-
olating the constraint u2 6 20. The variance of r(t) was chosen as
r2 = 0:014, e(t) having the same properties as in open-loop validation.
The input and output variances were then, respectively, u2cl = 20 and
y2cl = 0:0880. Observe that the input variance is the same as in open
loop, and that the output variance is very close to that of the open-loop
experiment. Again, 300 input-output data samples were collected and
used to identify an ARX model with the same structure as in open-loop
validation (5.74), using a direct prediction error method. We found
   
^cl = ^1 = 0:3575 ; (5.79)
^2 0:0067
 
Pcl = 28:8323  10 3 8:7845  10 6 : (5.80)
:7845  10 6 6:2416  10 6
As in open-loop validation, the nominal model is validated with respect
to the 95% uncertainty region Dcl around G cl (z; ^cl ):
  T   
a ^cl Pcl1 a ^cl = 4:5575 < = 5:99: (5.81)
b b
One can also verify that Dcl eectively contains the true G0 :
  T   
a0 ^cl Pcl
1 a0 ^cl = 2:1611 < = 5:99: (5.82)
b0 b0

5.7.4. Model set validation for control: comparison of the validated

sets

Observe that the nominal model G^ is closer to the centre of the validated
set Dol than to the centre of the validated set Dcl , when measured
by the normalised distance in parameter space : compare (5.77) with
(5.81). However, when measured in the control-oriented measures W C
or ÆW C , which are computed in transfer function space, the conclusions
are reversed. Figure 5.4 shows the worst-case chordal distance between

all models in Dol and all models
G^ and, respectively,
in Dcl . Clearly
W C G^ (ej! ); Dcl is smaller than W C G^ (ej! ); Dol at all frequencies,
and
^ Dcl ) = 0:0156 < ÆW C (G;
ÆW C (G; ^ Dol ) = 0:0225: (5.83)
 Application II: ferrosilicon production process 139

0
10

−1
10

−2
10

−3
10

−4
10
−4 −3 −2 −1 0
10 10 10 10 10
normalised frequency [rad/s]

Figure 5.4. W C G^ (ej! ); Dol
(
),
W C G^ (ej! ); Dcl
(|),
 G^ (ej! ); G0 (ej! ) ( ),
 G^ (ej! ); K=0:0007 (ej! ) (


)
1

We observe that the open-loop validation procedure does not guarantee

that the GPC controller C=0:0007 stabilises G0 , since
 

1
W C G(e ); Dol >  G(e );
^ j! ^ j!
(5.84)
K=0:0007 (ej! )
at some frequencies, while the closed-loop validation procedure does
guarantee such stability with G0 , since
 

1
W C G^ (ej! ); Dcl <  G^ (ej! ); (5.85)
K=0:0007 (ej! )
at all frequencies. The Min-Max version of this stability test yields the
same conclusion, since
^ Dol ) = 0:0225 > bG;K
ÆW C (G; ^ =0:0007 = 0:0169
(5.86)
> ÆW C (G;^ Dcl ) = 0:0156:
We conclude from this analysis that the uncertainty region Dcl delivered
by the closed-loop experiment should be preferred over Dol delivered by
140 Model validation for control and controller validation

the open-loop experiment, since it leads to the largest set of stabilising

controllers: see (Bombois et al., 2000a, 2000b).

5.7.5. Controller validation for stability

Using the open-loop and closed-loop PE uncertainty sets Dol and Dcl ,
we built the dynamic matrices MDol (ej! ) and MDcl (ej! ) with respect
to the candidate controller K=0:0007 and we computed their stability
radii. Their respective values are


max
!
 M D ol ( e j!
) = 0:6572 < 1; (5.87)


max
!
 MDcl (ej! ) = 0:2111 < 1: (5.88)
According to this test, K=0:0007 stabilises all systems contained in both
uncertainty sets Dol and Dcl , hence both the open-loop and the closed-
loop approach validate the controller. This shows that the necessary
and suÆcient condition for closed-loop stability is much less conservative
than the suÆcient condition based on the chordal distance or the  -gap
stability results.

5.7.6. Controller validation for performance

Since the control objective is to reject the noise v(t) = H0 (z )e(t) which is
essentially located at low frequencies (H0 (ej! ) is a rst order low-pass l-
ter; see the thick dotted line in Figure 5.5), a performance specication in
the frequency domain is that the sensitivity function S (G0 ; K ) = 1+G1 0 K
be low at low frequencies, since it lters v(t) in closed loop. However,
since S (G0 ; K ) cannot be made arbitrarily low at all frequencies, a trade-
o must be made between reducing the noise at low frequencies and
amplifying it at high frequencies. Actually, the ideal situation would
be that jH0 S (G0 ; K )j be uniform over frequency, i.e. that the noise be
whitened by the closed-loop system dynamics.
A possible performance criterion is thus
C (D; K=0:0007 ; ! )
(1)
JW
 
 1 
 ; (5.89)
= max 
G(ej! ; )2D  1 + G(ej! ;
)K=0:0007 (e ) 
j!

obtained by setting Wl1 = Wr1 = 0 and Wl2 = Wr2 = 1 in (5.41).

(1)
The controller is validated if JW C remains at all frequencies below some
 Application II: ferrosilicon production process 141

template, which is high-pass since we want the worst-case sensitivity to

be small in the frequency range where the noise is large. The amplitude
Bode diagrams of the worst-case and achieved sensitivity functions are
depicted in Figure 5.5.

1
dB

−1

−2

−3

−4
−4 −3 −2 −1 0
10 10 10 10 10
normalised frequency [rad/s]

Figure 5.5. JW(1)C (Dol ; K=0:0007 ) (
),

C (Dcl ; K=0:0007 ) (|),
(1)
JW
jS (G0; K=0:0007 )j ( ),
jS (G;
^ K=0:0007 )j (


),
jH0j (
)

Clearly, for most reasonable templates the controller is validated by the

closed-loop validation experiment yielding Dcl , but not by the open-loop
experiment yielding Dol .
Now observe that a noise model H (z; ^) = 1+^11 z 1 , sharing its param-
eter with G (z; ^), is also identied during each validation experiment,
due to the particular ARX model structure we use. As a result, each 
in Uol or Ucl denes not only a plant model G() in Dol or Dcl , but also
a noise model H () associated with this plant model. An alternative
142 Model validation for control and controller validation

performance criterion5 could thus be

 
 H (ej! ; ) 
JW C (U;
(2)
K=0:0007 ; !) = max  : (5.90)
2U 1 + G(ej! ;  )K =0:0007 (e ) 
j!

(2)
In this case, the distribution of JW C over frequency should be as close as
possible to a uniform distribution, and its amplitude should be as low as
possible. As shown in Figure 5.6, the worst-case performance attached to
the closed-loop uncertainty region is closer to the performance achieved
by the controller on the actual system, and to the uniform distribution
we would like to obtain, than the worst-case performance attached to
the open-loop uncertainty region.
8

4
dB

−2

−4
−4 −3 −2 −1 0
10 10 10 10 10
normalised frequency [rad/s]

Figure 5.6. JW(2)C (Uol ; K=0:0007 ) (
),

C (Ucl ; K=0:0007 ) (|),
(2)
JW
jH0S (G0 ; K=0:0007 )j ( ),
jHS
^ (G; ^ K=0:0007 )j (


)

In practice, it often happens that the noise model is not estimated but
is given a priori and assumed to be exact. If we make the assumption
that H0 is known exactly, a third possible performance criterion is the
5
This performance criterion is not one of the form (5.41). However, it only requires
a slight modication of the LMI-based algorithm used to compute (5.41).
 Application II: ferrosilicon production process 143

following:
JW(3)C (D; K=0:0007 ; H0 ; !)
 
 H 0 (e j! ) 
;
= max  (5.91)

G(ej! ; )2D 1 + G(ej! ;  )K=0:0007 (ej! ) 
obtained by setting Wl1 = Wr1 = 0, Wl2 = H0 and Wr2 = 1 in (5.41).
(3)
The controller is then validated if the amplitude of JW C is lower than
that of H0 at almost all frequencies, i.e. the worst-case sensitivity re-
duces the actual noise at almost all frequencies. The results are depicted
in Figure 5.7.
7

3
dB

−1

−2

−3
−4 −3 −2 −1 0
10 10 10 10 10
normalised frequency [rad/s]

Figure 5.7. JW(3)C (Dol ; H0 ; K=0:0007 ) (
),

C (Dcl ; H0 ; K=0:0007 ) (|),
(3)
JW
jH0S (G0; K=0:0007 )j ( ),
jH0S (G; ^ K=0:0007 )j (


),
jH0j (
)
Once again, the controller is validated when the worst-case performance
criterion is computed over Dcl , but not if it is computed over Dol .
The results we have obtained in this subsection conrm that the PE
uncertainty set Dcl obtained via a closed-loop experiment is to be pre-
ferred to Dol obtained via an open-loop experiment, as the worst-case
144 Model validation for control and controller validation

 -gap test of Subsection 5.7.4 already indicated, when the purpose is to

validate a given candidate controller.

5.8. Conclusion
The validation of a model should always be carried out with respect to
the intended use of this model. In this chapter, we have considered the
case where the model is used for control design. We have developed a
new closed-loop validation procedure for this case, based on an open-
loop validation procedure proposed in (Ljung, 1998), but taking the
considerations of Chapter 3 into account, in order to obtain uncertainty
regions that are better tuned for control design.
Using a control-oriented measure of size of frequency domain uncertainty
sets, we have shown that our closed-loop validation procedure is indeed
more suited for computing tuned uncertainty regions. Robustness anal-
ysis tools, applied to realistic numerical examples, have also shown that
these uncertainty regions yield less conservative control designs than
those obtained by open-loop model validation.
 CHAPTER 6
A performance-based approach to
control-oriented model reduction and to
controller reduction

6.1. Introduction
In many practical applications, low-order controllers are preferred over
high-order ones for numerous reasons (e.g. easiness of implementation
or maintenance, numerical robustness, etc.). However, it often happens
that the design has to be carried out on the basis of a high-order model
of the plant obtained by methods like physical (white-box) modelling,
nite-element modelling, linearisation of a high-order nonlinear simula-
tor, etc. In this case, three approaches can be considered for the design
of a low-order controller, as depicted in Figure 6.1. The direct approach,
which would evidently be the most appealing a priori, suers several ma-
jor problems: it involves very complicated equations with no intuitive
content and which may present a very great number of solutions; the
solution of these equations is far from straightforward; there is no com-
mercial software available for the direct computation of a model-based
low-order controller for a high-order plant. Examples of direct meth-
ods can be found in (Bernstein and Hyland, 1984) and (Gangsaas et
al., 1986). On the contrary, the two indirect methods, on which this
chapter focuses, involve a step of model or controller reduction and a
step of control design for both of which eÆcient, well understood and
very popular tools and software packages are available. Examples of
such tools are H1 and LQG control design, balanced truncation, Han-
kel norm approximation, etc.

145
146 A performance-based approach to control-oriented model reduction and to : : :

High-order - High-order
High-order control design
plant (LQG, H1, etc.) controller
HHH
Model HH H Controller
reduction HDirect
HH
design
reduction
HHH
? Hj ?
Low-order Low-order control design- Low-order
plant (LQG, H1, etc.) controller

Figure 6.1. Three approaches for the design of a low-

order controller for a high-order system

In this chapter, we propose a new criterion for the reduction of a model

or a controller in closed loop. This criterion is based on a `small-gain
theorem'-based result and on the heuristic results of Chapter 3, and is
aimed at preserving the closed-loop properties of the nominal high-order
model or controller. This criterion is applied to the design of a low-
order controller for a high-order model of a pressurised water reactor
(PWR) nuclear plant. The results it produces are compared to those
obtained by open-loop model and controller reduction, closed-loop model
and controller reduction using `standard' criteria, and direct closed-loop
identication of a low-order model.
The goal of this chapter is to give an answer to the following three
questions:
Question 1 (tuning of the bias error): Is closed-loop identication
a viable alternative to the classical system order reduction techniques,
when the goal is to use the resulting model for the design of a (new)
low-order controller for the high-order system?
Question 2 (experiment design): Does the incorporation of the con-
trol objective in the reduction criterion yield signicantly better results
than open-loop model or controller reduction?
Question 3 (methodology): When an order reduction step is re-
quired, should it be carried out on the model (i.e. before control design),
or on the controller (i.e. after controller design)?
 Model order reduction 147

6.2. Model order reduction

Here, we consider the reduction of a high-order plant model Gn to order
r. The resulting low-order model is denoted G^ r . The chosen reduction
method is balanced truncation1 . Since this method can only be applied
to a stable system, it is necessary to adapt it if Gn is unstable. Two
alternatives can be considered:
1. Separate the stable and unstable parts of Gn. Reduce the stable part,
keep the antistable part intact.
2. Reduce the coprime factors of Gn instead of Gn itself.
The second solution, initially proposed in (Meyer, 1988, 1990), is gen-
erally better because all modes of the system are taken into account
during the reduction. The application of the rst alternative is totally
ineÆcient if Gn has many unstable modes.

6.2.1. Open-loop reduction of the nominal model coprime factors

 
Let N~n M~ n be a left coprime factorisation of Gn (see Proposition 2.2
for the construction of such a factorisation). It is stable by hypothesis,
hence it can be reduced
 by classical balanced truncation. It is advanta-
geous to choose N~n M~ n normalised, i.e. such that N~ N~ ? + M~ M~ ? = I ,
because the H1 norm of the reduction error gives a bound of the dis-
tance between Gn and G^ r in the gap metric and in the  -gap metric:
see (Georgiou and Smith, 1990), (Meyer, 1990) and (Vidyasagar, 1984).
Such a normalised left coprime factorisation can be obtained following
Proposition 2.3.
 
Let us denote N~r M~ r the r-th order realisation of the reduced
coprime
factors obtained by balanced truncation of the normalised N~n M~ n
1
The motivation for using balanced truncation is the fact that it is a purely algebraic
method, very easy to implement and, therefore, very popular in the common practice.
Furthermore, it allows to choose the order of the reduced-order model a priori by
inspection of the Hankel singular values. One might criticise the fact that it does not
minimise the approximation error in any norm and that, in the continous-time case, an
upper bound on the approximation error is only available in the H1 norm. However,
it should be noted that the problem of minimising kGn G^ r k1 or kGn G^ r k2 over
all G^ r of order r is a very complicated problem. No very reliable solution exists, and
most algorithms are iterative and time consuming, and deliver suboptimal solutions
corresponding to local minima. The problem is even worse in case of a multivariable
system, and if frequency weightings are used.
148 A performance-based approach to control-oriented model reduction and to : : :

(note that balanced truncation preserves the normalisation, as shown in

(Meyer, 1990), i.e. N~r N~r? + M~ r M~ r? = I ):
   

N~r M~ r = bt N~n M~ n ; r : (6.1)
 
The reduction procedure, which considers N~n M~ n as a single object,
ensures that N~r and M~ r share the same state realisation. The recon-
struction of a reduced model G^ r = M~ r 1 N~r must then be carried out
without duplicating the states, such that G^ r has order r and not 2r.
This is possible by observing that G^ r can be obtained via a lower LFT:
  
I
Gr = Fl I N~r M~ r
^    

0 I ; I : (6.2)

Recall that the upper and lower bounds of the reduction error are given
by
n
X
 
&r+1 6
N~
M~ 1 6 2 &i (6.3)
i=r+1
 
where the &i 's are the Hankel singular values of N~n M~ n and
     

N~
M~ = N~n M~ n N~r M~ r : (6.4)

The interpretation of the approximation error in the gap metric and in

the  -gap metric is given by the following relation (Zhou and Doyle,
1998):
n


G = M~ n +
M~ 1
N~n +
N~ j
    o

N~
M~ 2 H1;
N~
M~ 1 6 "
n o n o
= G j ~Æg (Gn ; G) 6 "  G j Æ (Gn ; G) 6 " ; (6.5)
where ~Æg (
;
) is the directed gap dened in (2.134). As a result, the
following proposition holds, which gives an upper bound of the distance
between Gn and G^ r in the gap metric and in the  -gap metric.
Proposition 6.1. Let Gn = M ~ n 1 N~n be a normalised left coprime fac-
torisation of the n-th order system Gn , and let G^ r = M ~ r 1 N~r be a r-th

order system (r < n) obtained by balanced truncation of N~n M ~ n .
Then,
n
X
~Æg (Gn ; G^ r ) 6 2 &i (6.6)
i=r+1
 Model order reduction 149

and
n
X
Æ (Gn; G^ r ) 6 2 &i (6.7)
i=r+1
 
where the &i 's are the Hankel singular values of N~n M
~n .

According
 to this result, what we have called the `open-loop' reduction
of N~n M~ n , i.e. its reduction without frequency weighting, can how-
ever be considered as `closed-loop' or `control'-oriented, since it preserves
as much as possible the properties of Gn around its crossover frequency
(recall that the  -gap is a measure of distance between two transfer func-
tions that has a maximal resolution at the crossover frequency, because
of the properties of the projection of the Nyquist contour onto the Rie-
mann sphere). However, from a control design point of view, it is less
the crossover frequency of Gn than that of Gn K , where K is the optimal
controller (or, in an iterative framework, a controller that yields better
performance than open-loop operation) that should be considered.

6.2.2. Performance-preserving closed-loop reduction of the nominal

model coprime factors

When a reasonably good controller is already available, it is better to

take it into account for carrying out the reduction of the model order.
Indeed, as we just explained, GnK will generally be closer to GnKopt ,
where Kopt is the optimal to-be-designed controller, than Gn alone. The
heuristic motivations of Chapter 3 for performing the identication of a
low-order model in closed loop when the nal objective is control design
remain valid when the computation of the low-order model is carried
out by order reduction of a high-order (unbiased) model. Recall that
the heuristic reasoning was based on the fact that, if the optimal to-be-
designed controller were known and used during identication, then the
identication criterion would match the performance criterion that the
low-order model should minimise. This low-order model would then be
obtained with a bias error that is ideally tuned towards control design.
Since the optimal controller is unknown, it should be replaced during the
identication experiment by a controller that is not too dierent from
the optimal one.
In order to apply this reasoning to closed-loop model reduction, we rst
have to dene a performance criterion that can be used as model re-
duction criterion. Here, we suggest to use the following performance
150 A performance-based approach to control-oriented model reduction and to : : :

degradation index:

J (G^ r ) = To (Gn ; K ) To (G^ r ; K ) ; (6.8)
1
where K is the current to-be-replaced controller and where To (Gn ; K )
is dened in (2.72). Clearly, this criterion ensures that G^ r remains close
to Gn in a closed-loop sense.
The choice of this criterion is further motivated by the `small-gain
theorem'-based result of Proposition 2.1:

J (G^ r ) = To (Gn; K ) To (G^ r ; K ) < "
1  
=) bGn ;K bG^ r ;K  < " (6.9)

provided the closed loops (Gn ; K ) and (G^ r ; K ) are both internally sta-
ble. This means that the reduction criterion J (G^ r ) is a measure (upper
bound) of the performance degradation measured by the generalised
stability margin if the resulting low-order model G^ r is stabilised by K ,
which is easy to check (but which is not guaranteed by the reduction
procedure). One might criticise the fact that the minimisation of J (G^ r )
in (6.9) does not imply the minimisation of jbGn ;K bG^ r ;K j, hence our
approach may seem somewhat conservative. However, recall that our
primary concern is J (G^ r ) and not jbGn ;K bG^ r ;K j; (6.9) just tells us
that it is not a bad choice regarding robust stability, provided nominal
stability is maintained.
In order to be able to carry out the reduction, it is necessary to ex-
press To (Gn ; K ) as a function of the (normalised) coprime hfactors
i of
Gn . Therefore, we also consider a right coprime factorisation V of the
U

controller K . It is then easy to verify that

" 1#
K (I + GnK ) 1 Gn K (I + Gn K )
To (Gn ; K ) =
(I + GnK ) 1 Gn (I + Gn K ) 1
2 3
U  1 N~n U  1 M~ n (6.10)
=4 5
1 ~
V  N n V  Mn 1 ~
 
= VU  1 N 
~n M~ n
 Model order reduction 151

where
 
 = N~n M~ n VU
 
(6.11)

can be made equal to I by an appropriate choice of the coprime factori-

sation of K . In fact, two of the following three equalities can always be
satised simultaneously:

 normalisation of the left coprime factors of Gn:

N~n N~n? + M~ n M~ n? = I ; (6.12)
 normalisation of the right coprime factors of K :
U ?U + V ?V = I ; (6.13)
 Bezout identity:
 = N~n U + M~ n V = I: (6.14)

The reduction error is then given by

   
J (G^ r ) = U U ^
1 N  1 N

~ M~ n ~ M~ r
V  n V  r
1
(6.15)

where
 
^ = N~r M~ r U
 
V : (6.16)

In practice, we shall use the following zero-order approximation,

 
J (G^ r )  VU 
1 N  


~
n M~ n
N~r M~ r ; (6.17)
1
which allows to compute a suboptimal solution to the problem of nding
G^ r that minimises J (G^ r ) by frequency
 weighted
 balanced truncation:
the object being approximated is N~n M~ n , while the left frequency
weighting lter is
 
Wl = VU  1: (6.18)

A more general criterion can be dened by adding input and output

weighting lters Win and Wout if one desires to specically weight some
152 A performance-based approach to control-oriented model reduction and to : : :

inputs and outputs of To (Gn ; K ):

 

J (G^ r ) = Wout VU  1 N~n
   

M~ n N~r Win
M~ r |{z} : (6.19)

| {z } Wr

Wl 1
A suboptimal G^ r is then obtained by two-sided frequency weighted bal-
anced truncation:
   

N~r M~ r = fwbt N~n M~ n ; Wl ; Wr ; r ; (6.20a)
G^ r = M~ r 1 N~r : (6.20b)
 
Remark.
h i
If a two-degree-of-freedom controller C = F K is used,
 
U
V is a coprime factorisation of C while N~n M~ n is a coprime fac-
h i
torisation of G0n where the number of outputs of the 0 block is equal
to the number of inputs of F . Theh ireduction procedure then yields
 
Nr Mr of order r, such that G^0r = M~ r 1 N~r . It is then easy to
~ ~
recover G^ r .

6.2.3. Stability-preserving closed-loop reduction of the nominal model

coprime factors

Instead of focusing on the preservation of closed-loop performance during

the model reduction procedure, one can formulate the reduction criterion
in terms of closed-loop stability (recall that it is important to have a
nominal model that is stabilised by the current to-be-replaced controller:
see Section 4.4). Consider the following lemma.
Lemma 6.1. Let Gn = M ~ n 1 N~n and K = UV 1 be two coprime fac-
torisations satisfying the Bezout identity (6.14). Then, K stabilises any
system G = (M ~ n +
M~ ) 1 (N~n +
N~ ) such that
 

M~ VU
 

~
N < 1: (6.21)
1

Proof. This lemma is the dual of Lemma 6.2 for stability-based reduc-
tion of controller coprime factors (see page 159), for which a proof is
given in (Anderson and Liu, 1989).
 Model order reduction 153

It leads very naturally to the denition of the following model reduction

criterion:
 
   
U
J (G^ r ) = N~n M~ n N~r M~ r V : (6.22)
1
This criterion aims at preserving the Bezout identity (6.14). Since this
equality means that

 the factors N~n and M~ n are left coprime,

 the factors U and V are right coprime, and
 the closed-loop (Gn; K ) is stable,
this criterion is indeed based on closed-loop stability considerations.
 As
a result, a model G^ r obtained by minimising J (G^ r ) over all N~r M~ r
of order r would be optimal in that sense that the closed-loop stability
with controller K is maintained. However, G^ r may be very dierent
from Gn regarding closed-loop performance.
Finally, notice that, in this case, frequency weighted balanced truncation
can also be used to compute a suboptimal G^ r :
   
M~ n ; I; VU ; r ;
   
N~r M~ r = fwbt N~n (6.23a)
G^ r = M~ r 1 N~r : (6.23b)

6.2.4. Preservation of stability and performance by closed-loop re-

duction of the nominal model coprime factors

In Subsections 6.2.2 and 6.2.3, we have presented closed-loop coprime-

factor reduction methods based, respectively, on the preservation of the
closed-loop performance and on the preservation of the closed-loop sta-
bility. A question that naturally arises is: `Would it be possible to com-
bine both approaches?'. The answer, in case of coprime-factor reduction,
is generally `no'. However, in case of an observer-based controller, it is
possible to use two dierent methods, as we now show.
A. Unweighted balanced truncation of coprime factors satis-
fying the Bezout identity. This method was initially proposed in
(Anderson and Moore, 1989) for the case of a LQG controller. It later
154 A performance-based approach to control-oriented model reduction and to : : :

received an alternative justication, not based on LQG-performance con-

siderations, in (Zhou and Chen, 1995)2 .
Let K = UV 1 and Gn = M~ n 1 N~n be coprime factorisations satisfying
(6.13) and (6.14). Observe that, in this case, the normalised coprime
factors of the controller are normalised. Let G^ r = M~ r 1 N~r dene a
coprime factorisation of a reduced-order model G^ r . Then, if

N~n M~ n  N~r M~ r  < " < 1; (6.24)
1
it is possible to show that G^ r is stabilised by K and that

^ r ; K ) 6 kTo (Gn ; K )k1 " :
To (Gn ; K ) To (G (6.25)
1 1 "
Thus, the unweighted reduction of system coprime factors satisfying the
Bezout identity (when the controller coprime factors are normalised)
makes it possible to guarantee stability and  to overbound
 the perfor-
mance degradation. However, the order of N~n M~ n will generally be
twice that of Gn plus that of K if K is not an observer-based controller,
which
 makes the method useless except in that particular case. Indeed,
let Nn M n be any n-th order left coprime factorisation of Gn . Then,
   
1
N~n M~ n = Nn M n N~n U + M~ n V (6.26)
is a left coprime factorisation of Gn which satises the Bezout identity
h i
(6.14). It has a minimal realisation of order at most n if and only if VU
is of the form
   
U (s) = F (sI A BF ) 1 H (6.27)
V (s) I C (sI A BF ) 1 H
where (A; B; C ) is a minimal state-space realisation of Gn, and F and H
are such that the eigenvalues of A + BF and A + HC lie in the open left
half plane. This corresponds typically to an observer-based controller
where F is the stabilising feedback gain, and H is the Kalman lter
gain.
B. Multiplicative coprime-factor approximation. The method pre-
sented here follows an idea in (Gu, 1995). Let K = V~ 1 U~ = UV 1 and
2
Actually, in (Anderson and Moore, 1989) and (Zhou and Chen, 1995), controller
reduction is considered. Here, we present a dual version of this method, adapted to
model reduction.
 Controller order reduction 155

Gn = M~ n 1 N~n = Nn Mn 1 be coprime factorisations satisfying the fol-

lowing double Bezout identity:
  
M~ n N~n U Nn = I: (6.28)
U~ V~ V Mn
h i
Once again, UV M n has order at most n only if it can be put in the
Nn

form
     
U (s) Nn (s) = I 0 + C (sI A BF ) 1  H B  (6.29)
V (s) Mn (s) 0 I F
where (A; B; C ) is a minimal state-space realisation of Gn, and F and
H are such that the eigenvalues of A + BF and A + HC lie in the open
left half plane.
We can seek Ur , Vr , Nr and Mr of order r such that the H1 norm of

2 RH1 in
   
Ur Nr = (I
) U Nn (6.30)
Vr Mr V Mn
is minimised. This amounts to nd Ur , Vr , Nr and Mr such that
      1

k
k1 = U Nn Ur Nr U Nn
(6.31)

V Mn Vr Mr V Mn
1
is minimised. G^ r is then given by G^ r = Nr Mr 1 and the closed loop
(G^ r ; K ) is internally stable if
k
k1 kTo (Gn; K )k1 < 1: (6.32)
Furthermore, there holds


T (
o n G ; K ) T ( ^
G ; K )


6 kTo (Gn ; K )k21 k
k1
: (6.33)
o r
1 1 kTo (Gn ; K )k1 k
k1
Clearly, the minimisation of k
k1 aims at preserving closed-loop sta-
bility and performance, but the method is generally ineÆcient if K is
not an observer-based controller.

6.3. Controller order reduction

Instead of reducing the order of the nominal model Gn before designing
a (new) controller, it is possible to design a high-order controller Km on
the basis of Gn, and to then reduce it to a lower order r. We denote
this low-order controller K^ r . As in the case of model reduction, it is
156 A performance-based approach to control-oriented model reduction and to : : :

not possible to carry out the reduction of Km in a straightforward way

if it is unstable. Therefore, we shall also consider the reduction of the
coprime factors of Km , either in `open loop', i.e. without frequency
weightings, or in `closed loop', i.e. with frequency weightings that take
the interconnection with Gn into account.

6.3.1. Open-loop reduction of the optimal controller coprime factors

As for model reduction, it is stronglyhrecommended

i to apply the reduc-
tion to normalised coprime factors Vm of the controller Km . Such
Um

normalised factors can be constructed as explained in Proposition 2.5.

The reduced-order
h i controller can then be obtained by balanced trunca-
Um
tion of Vm :
    
Ur = bt Um ; r ; (6.34a)
Vr Vm
K^ r = Ur Vr 1 : (6.34b)

The motivation for using normalised coprime factors comes from the
following theorem.
Theorem 6.1. (Vinnicombe, 1993a, 1993b) Let Gn be a nominal plant
and Km be a stabilising controller such that bGn ;Km 6 supK bGn ;K . Let
Km = Um Vm 1 be a normalised coprime factorisation and let Ur ; Vr 2
RH1 be such that    
Um Ur 6 ":

Vm Vr 1 (6.35)

Then K ^ r , Ur Vr 1 stabilises Gn if " < bGn ;Km . Furthermore,

arcsin bG;Kr > arcsin bGn ;Km arcsin " arcsin  (6.36)
8G : Æ (G; Gn) 6  .
The rst part of this theorem concerns the stabilisation of the nominal
model Gn by the reduced-order controller K^ r . It says that nding the
best possible approximation of the normalised coprime factors of the op-
timal controller in the H1 norm is appropriate (although not optimal,
since the condition of the theorem is only suÆcient) regarding this sta-
bility issue (of course, balanced truncation only delivers a suboptimal
approximation). The second part of the theorem concerns a robustness
 Controller order reduction 157

issue: if the nominal model Gn has an uncertainty region of size  at-

tached to it (i.e. the  -gap between the underlying unknown true system
G0 and the nominal model Gn is known to be at most equal to  ), a
lower bound of the generalised stability margin that K^ r will achieve with
any plant in this uncertainty region is given. Observe that this
h itheorem
does not make any assumption about the computation of UVrr , which
means that it can be used to check closed-loop stability with a controller
K^ r computed by any method, e.g. frequency weighted coprime-factor
reduction, even though the criterion it induces is unweighted normalised-
coprime-factor reduction. Note also that this theorem can be very con-
servative, as the practical application of Section 6.4 will show.
As it was the case for the `open-loop' model reduction procedure, this
approach is somewhat closed-loop oriented, since it is based on a suÆ-
cient condition for maintaining closed-loop stability. However, it does
not take the nominal model Gn explicitly into account.

6.3.2. Performance-preserving closed-loop reduction of the optimal

controller coprime factors

The idea inherent in our closed-loop controller reduction procedure is

the same as for closed-loop model reduction: the goal is to maintain the
closed-loop properties of the controller with respect to the particular
model Gn for which it was designed. Our controller reduction criterion
is thus the following:

J (K^ r ) = To (Gn; Km ) To (Gn; K^ r ) ; (6.37)
1
where To (Gn; K ) is dened in (2.72). This criterion ensures that K^ r
remains close to Km in a closed-loop sense. Proposition 2.1 also applies
in the present case:

J (K^ r ) = To (Gn ; Km ) To (Gn ; K^ r ) < "
1 
=) bGn ;Km bGn ;K^ r  < " (6.38)
if the loops (Gn ; Km ) and (Gn ; K^ r ) are both internally stable. This
means that the reduction criterion J (K^ r ) is a measure (upper bound)
of the loss of robustness measured by the generalised stability margin if
the obtained reduced-order controller stabilises the nominal model.
158 A performance-based approach to control-oriented model reduction and to : : :

The reduction criterion

h i can be expressed as a function of the normalised
coprime factors Vm of Km as follows:
Um

   
J (K^ r ) = UV m  1 N~n M~ n Ur ^ 1  ~ ~ 
 
Vr Nn Mn (6.39)
m 1
 
where N~n M~ n is any left coprime factorisation of Gn ,
 
  Um
~
 = Nn Mn V ~ (6.40)
m
 
can be made equal to I by an appropriate choice of N~n M~ n , and
 
^ = N~n M~ n U
  r
Vr : (6.41)
In practice, we shall once again use a zero-order approximation,
   
Um U  
J (Kr )  V
^ 1 ~
Vr  Nn Mn 1 ;
r ~ (6.42)
m
which allows to compute a suboptimal solution to the problem of nd-
ing K^ r that minimises J (K^ r ) by frequency
h iweighted balanced trunca-
Um
tion: the object being approximated is Vm , while the right frequency
weighting lter is
 
Wr =  1 N~n M~ n : (6.43)
This criterion can be generalised by adding input and output weighting
lters Win and Wout if one desires to specically weight some inputs and
outputs of To (Gn ; Km ):

   

^
J (Kr ) = W U m Ur  
 1 N~n M~ n Win : (6.44)
out
| {z } Vm Vr | {z }
Wl
Wr 1
A suboptimal K^ r is then obtained by two-side frequency weighted bal-
anced truncation:     
Ur = fwbt Um ; W ; W ; r ; (6.45a)
Vr Vm l r

K^ r = Ur Vr 1 : (6.45b)
 
Remark.
h i
If a two-degree-of-freedom controller Cm = Fm Km is
 
used, Um
Vm is a coprime factorisation of Cm while N~n M~ n is a co-
h i
prime factorisation of G0n where the number of outputs of the 0 block
 Controller order reduction 159

is equalh toi the number of inputs of Fm . The reduction procedure then

yields UVrr of order r, such that C^r = Ur Vr 1 . It is then easy to recover
F^r and K^ r .

6.3.3. Stability-preserving closed-loop reduction of the optimal con-

troller coprime factors

The performance-based reduction method presented in the previous sub-

section is dierent from the traditional closed-loop controller coprime-
factor reduction procedure generally addressed in the literature (see e.g.
(Anderson and Liu, 1989), (Liu et al., 1990), (Zhou, 1995) and (Zhou
and Chen, 1995)), which is based on the following lemma.
Lemma 6.2. (Anderson and Liu, 1989) Let Gn = M ~ n 1 N~n and Km =
1
Um Vm be two coprime factorisations satisfying the Bezout identity (6.14).
^ r = Ur Vr 1 such that
Then, every K
   

N~ M~ n
 Um Ur
<1 (6.46)
n Vm Vr
1
stabilises Gn .

Hence, according to this method, the optimal

h i K^ r = Ur Vr 1 would be
obtained by minimising (6.46) over all UVrr of order r. This K^ r would
be optimal in that sense that the closed-loop stability with Gn is main-
tained. However, in practice, the closed-loop performance of the two
controllers can be very dierent. Once again, frequency weighted bal-
anced truncation can be used to compute a suboptimal K^ r :
    
Ur = fwbt Um ;  ~ ~  ; I; r ; (6.47a)
Vr Vm Nn Mn
K^ r = Ur Vr 1 : (6.47b)

6.3.4. Other closed-loop controller reduction methods

Here, we brie y describe a few alternative approaches for closed-loop

controller reduction. Other methods can be found in (Obinata and An-
derson, 2000).
160 A performance-based approach to control-oriented model reduction and to : : :

A. Performance-based approaches based on frequency weighted

balanced truncation. A performance-based frequency weighted ap-
proach for controller reduction has been proposed in (Goddard and
Glover, 1993). It is based on balanced truncation or Hankel norm ap-
proximation of the stable part of the high-order controller, the antistable
part being kept unchanged, which makes this method ineÆcient in case
of an unstable controller.
The same authors have proposed a coprime-factorisation approach for
performance-preserving frequency weighted controller approximation
(Goddard and Glover, 1994). Although the ultimate goal of this method
is essentially the same as in our method of Subsection 6.3.2, the techni-
cal aspects are much more involved (and much less intuitive), resulting
in a very complicated algorithm which makes the method less appealing
in practice.
B. Closed-loop balanced truncation. A very nice and intuitive
performance-based method, called closed-loop balanced truncation, has
been proposed in (Ceton et al., 1993) and (Wortelboer, 1994). Its princi-
ple is the following. Consider a closed-loop system represented by a LFT
as in Figure 2.2, where (s) is a standard plant that can be represented
by
2 3
  A B 1 B 2
(s) = 11 ((ss)) 12 ((ss)) = 4 C1 D11 D12 5 ; (6.48)
21 22 C D D 2 21 22
and where Q(s) is a stabilising high-order one or two-degree-of-freedom
controller (see Section 2.5) with the following state-space representation:
 
Q(s) = A Q BQ
CQ DQ : (6.49)
Without loss of generality, we make the assumption that D22 = 0, since
if 22 (s) is not strictly proper, the problem can always be transformed
into another one where 22 (s) is strictly proper by means of a loop
transformation: see (Glover and Doyle, 1988). The closed-loop transfer
matrix is then given by

1
Tzw (s) = 11 (s) + 12 (s)Q(s) I 22 (s)Q(s) 21 (s)
2 3
A + B2 DQ C2 B2 CQ B1 + B2 DQ D21 (6.50)
=4 BQ C2 AQ BQD21 5:
C1 + D12 DQ C2 D12 CQ D11 + D12 DQ D21
 Controller order reduction 161

Since this realisation is stable, its controllability and observability Grami-

ans can be computed and partitioned as
   
P P Q Q
P = PT P Q and Q = QT Q Q : (6.51)
Q Q Q Q

We can then perform balanced truncation on the controller Q(s) using

PQ and QQ, the parts of these Gramians corresponding to the controller
states.
Note that there is no theoretical reason for preferring our performance-
based controller reduction procedure over closed-loop balanced trunca-
tion, or vice-versa. Both are indeed based on the preservation of the
same closed-loop transfer matrix and accept additional input and/or
output frequency weightings; see e.g. (Wortelboer and Bosgra, 1994)
for a discussion on such lters. One might consider that the necessity
of computing coprime factorisations of the plant and the controller is a
drawback of our method. On the contrary, our method presents a nice
symmetry property, since the same criterion can be used for reducing
the order of the model or that of the controller: one has only to choose
which pair of coprime factors plays the role of the frequency weighting
and which plays the role of the object being approximated.
C. Performance and stability-preserving methods that can be
used with observer-based controllers. When the controller to be
reduced is an observer-based controller, two methods, dual of those pre-
sented in Subsection 6.2.4, can be used for preserving closed-loop per-
formance and stability.
The rst one, proposed by (Anderson and Moore, 1989) and (Zhou and
Chen, 1995),
h i is based on the unweighted reduction of the right coprime
Um
factors Vm of the controller. These factors are chosen such that the
Bezout identity (6.14) holds when the left coprime factors of the system
Gn are normalised. Then, if
   
Um Ur < " < 1;

Vm Vr 1 (6.52)

it is possible to show that K^ r = Ur Vr 1 stabilises Gn and that

"

To (Gn ; Km ) To (Gn ; K^ r ) 6 kTo (Gn ; Km )k1 : (6.53)
1 1 "
162 A performance-based approach to control-oriented model reduction and to : : :

The hsecond one

i is that of (Gu, 1995). The reduced-order transfer
h ma-i
Ur Nr Um Nn
trix Vr Mr is obtained by multiplicative error reduction of Vm Mn
(observe that the high-order controller Km has the same order as Gn in
this case) as explained in Subsection 6.2.4, x B. However, this time, Ur
and Vr are used to compute K^ r = Ur Vr 1 , and the closed loop (Gn ; K^ r )
is internally stable if
k
k1 kTo(Gn; Km )k1 < 1: (6.54)
Furthermore, there holds


To (Gn ; Km ) To (Gn ; K^ r )

6 kTo(Gn; Km)k1 k
k1 : (6.55)
2
1 1 kT (G ; K )k k
k
o n m 1 1

6.4. Application to the design of a low-order

controller for a complex Pressurised Water
Reactor model
6.4.1. Description of the system

A realistic nonlinear simulator, based on a rst-principle model describ-

ing both primary and secondary circuits of a pressurised water reactor
(PWR) nuclear power plant (see Figure 6.2), has been developed at

Electricit e de France (EDF). It includes all local controllers in-
volved in both primary and secondary circuits.
We focus on the behaviour of the plant around a xed operating point
corresponding to 95% of maximum operating power. This results in
a high (42nd) order model3 G42 , which includes the dynamics of the
primary and secondary circuits and of all local controllers, except some
specic controllers of the secondary circuit that we want to redesign;
these are denoted Ktb and Kcd in Figure 6.3. They control the electrical
power and the condenser water level, respectively, and their structures
are very simple: Ktb is a PI controller acting on the dierence between its
two inputs, while Kcd is a second order two-input one-output controller
which includes an integrator. For the sake of simplicity, Ktb and Kcd
will both be called `PID' controllers and collectively denoted CP ID in
the sequel.
3
As explained in (Adams, 1979, p. 135), the universal true system is of order 42.
 Application to the design of a low-order controller for a complex PWR model 163

turbine
control valve

secondary circuit
TURBINE
feedwater
control valve
STEAM
GENERATOR CONDENSER
control rod
system

to the pressurizer

feedwater pump

REACTOR primary circuit

reactor coolant pump

Figure 6.2. PWR plant description

In Figure 6.3, We (t) is the electrical power produced by the plant, con-
trolled to follow the demand Wref (t) of the network, and directly related
to the steam ow in the turbine, which highly depends on the high-
pressure turbine control valve aperture Ohp (t); Qex(t) is the extraction
water ow, and Ncd (t) the water level in the condenser (both depend
essentially on the feedwater extraction valve aperture Ucex(t)). dOhp(t)
and dUcex(t) represent disturbances acting on the control inputs (they
can also be used as excitation sources for identication), while cOhp(t)
and cUcex (t) are the computed values of the two valves apertures. The
strong couplings between We (t) and Ohp(t) on the one hand, and be-
tween Ncd (t) and Ucex (t) on the other hand, would explain the structure
of the present PID controllers. However, the control performance might
be enhanced by taking the cross-couplings into account.
Remark. For reasons of condentiality, upon EDF's request, we are not
allowed to publish the transfer functions of the models and controllers
used in this application.

6.4.2. Control objective and design

Our goal is to redesign controllers for the electrical power control in

the secondary circuit, i.e. to replace the present PID controllers Ktb
164 A performance-based approach to control-oriented model reduction and to : : :

dOhp - d Ohp - We-

dUcex- d 6 Ucex - PWR Qex-
Ncd-
6 -
Primary and Secondary
Circuits

local controllers 
G42

cOhp  Wref
Ktb 
cUcex 
Kcd 
CP ID

Figure 6.3. Interconnection of G42 with the PID controllers

and Kcd by a single multivariable controller in order to achieve a better

performance. The chosen control design is a linear quadratic Gaussian
(LQG) controller computed from a reduced-order model G^ r of the plant
G42 or from this plant itself.
The control objective is to use the feedwater tank, rather than the con-
trol rods in the primary circuit, to absorb the fast and medium range
variations in the power demands by acting on the valve apertures. The
controller will have to ensure that the electrical power supply We(t)
follows accurately the reference signal Wref (t), and to regulate the con-
denser water level Ncd (t) around its nominal value: see Figure 6.3. Also,
it will have to reject disturbances dOhp (t) and dUcex(t) acting on the sys-
tem. The verication of the controller performance will be done with
step signals applied at the Wref , dOhp and dUcex inputs of the system.
The extraction water ow Qex (t) is measured and can be used to com-
pute the control signals, but it does not have to be controlled.
 Application to the design of a low-order controller for a complex PWR model 165

In order to remain consistent in the comparative study, the same LQG

criterion will be used with each model:


JLQG Ohp (t); Ucex(t)
Z 1   
F (s)Werr (t) F (s)Ncd (t) Q FF((ss))W err (t)

= Ncd (t)
0  
  O hp (t)
+ Ohp (t) Ucex (t) R U (t) dt; (6.56)
cex

where Werr (t) = We (t) Wref (t) is the tracking error and F (s) =
1 + 20=s is a lter aimed at ensuring a zero static error. This loop
shaping lter is then connected to the corresponding inputs of the de-
signed controller. Since the main goal is to control We(t) (the regulation
of Ncd (t) being only a secondary requirement), more weight is put on
the electrical power tracking error than on the condenser water level in
JLQG . Furthermore, since the nominal value of Ucex (t) is 0.01 while it
is 1 (i.e. 100 times bigger) for Ohp (t), a proper scaling of the entries of
R requires that R22 = 1002 R11 . The chosen weighting matrices are
  2 
Q = 10 100 2 ; R = 100 1006 : (6.57)

These weightings have proved very satisfactory.

For the design of the Kalman lter, the external signals dOhp, dUcex and
Wref , which are in the low frequency range, are modelled as indepen-
dent Gaussian white noises ltered through a low-pass lter N (s) =
1=(s + 0:01). Since the external signals have a typical amplitude of 1 for
dOhp and Wref , and of 0.01 for dUcex, their covariance matrix is chosen
as Qn = diag(1; 0:0001; 1). In order to ensure a good roll-o at high
frequency, the measurement noise is parametrised as a Gaussian white
noise with large covariance Rn = diag(10; 10; 10) (remember that Qex
is measured and used for state estimation, although it is not regulated,
which is why Rn is 3  3).
The presence of the lters F (s) (twice) and N (s) (3 times) in the (aug-
mented) model used for the design will yield a controller with order equal
to that of the original model (G^ r or G42 ) plus 5. Furthermore, F (s) is a
loop shaping lter that has to be explicitly added to the designed con-
troller before it can be used with G^ r or G42 , and the total order of the
controller will therefore nally be that of G^ r or G42 plus 7. This will
determine the order of G^ r if a controller of some xed order is desired.
166 A performance-based approach to control-oriented model reduction and to : : :

6.4.3. System identication

The rst approach we have considered for the computation of a low-order

model for the PWR plant is closed-loop identication. The identication
experiments were carried out using the nonlinear simulator as the plant.
In the spirit of the identication for control literature (see e.g. (Gevers,
1993, 1995) and (Van den Hof and Schrama, 1995)), the identication
was conducted in closed loop with excitation signals dOhp and dUcex
in the frequency range of interest. This has the eect of producing
a model that is accurate in the frequency range that is important for
control design. Our goal was to obtain a reasonably low-order linear
model for the plant around the chosen operating point. We used an
approach that was already used in (Bendotti and Bodenheimer, 1994)
for the identication of a model of the primary circuit of a PWR.
A. MIMO state-space description. Consider the system depicted
in Figure 6.3. The physical system is described by a descrete-time LTI
system around an operating point by the following equations:

    
x(k + 1) = A B x(k) (6.58)
y(k) C D u(k)

   
where y(k) = We (k) Qex(k) Ncd (k) T , u(k) = Ohp (k) Ucex(k) T
and x(k) represent respectively the output, the input and the state at
sample time kTs . The sampling period was chosen as Ts = 0:2 s.
Physical insights and preliminary identication of several SISO and
MISO transfer functions were used to provide insight into an appro-
priate identiable parametrisation with a low number of parameters,
the other entries of A, B , C and D being set at zeros and ones. The
electrical power and the water ow are mainly related to the control
inputs by third-order systems, while the water level integrates the wa-
ter ow. Furthermore, Ohp aects essentially the electrical power, while
Ucex aects the water ow. Hence, the system is diagonal dominant with
some cross-couplings. These insights led to an identiable state-space
realisation of order 7 with a vector  of 19 free parameters:
 Application to the design of a low-order controller for a complex PWR model 167

 
A (  ) B ( )
C ( ) D ( )
2 3
0 1 0 0 0 0 0 b1 0
6
6 0 0 1 0 0 0 0 b2 0 7
7
6
6 a1 a2 a3 a4 0 a5 0 b3 0 7
7
6
6 0 0 0 0 1 0 0 0 b4 7
7
6
=6 0 0 0 0 0 1 0 0 b5 7
7: (6.59)
6
6 0 0 a6 a7 a8 a9 0 0 b6 7
7
6
6 0 0 a10 a11 0 0 1 0 0 7
7
6
6 1 0 0 0 0 0 0 d1 0 7
7
4 0 0 0 1 0 0 0 0 d2 5
0 0 0 0 0 0 1 0 0

B. MIMO identication. The output of the model is denoted

y(k; ) = C ()(zI A()) 1 B () + D() u(k): (6.60)
A set of 10000 input/output data collected on the closed loop system
made up of the nonlinear simulator with the PID controllers Ktb and Kcd
was used to determine the parameter estimates. A standard quadratic
prediction error criterion was minimised.
A continuous-time model of order 7, G^ id7 , could be obtained by rst-
order approximation of the identied discrete-time model, i.e. by setting
x_ (t) = (x(k) x(k 1))=Ts , since the sampling period Ts = 0:2 s of
the latter is much smaller than the fastest natural time constant of
the system. G^ id
7 is then validated by checking how well it matches the
behaviour of the nonlinear plant when simulated with a new data set
(Figure 6.4).
C. Control design. This continuous-time model G^ id 7 was used to com-
pute a LQG controller following the procedure of Subsection 6.4.2. This
controller has order 14 and is denoted C14 (G^ id
7 ). Its performance will be
studied in Subsection 6.4.6.

6.4.4. Model reduction

Our second approach for obtaining a low-order controller is model reduc-

tion applied to the high-order model G42 and followed by LQG control
design.
168 A performance-based approach to control-oriented model reduction and to : : :

0.05

W_e
0

−0.05
0 10 20 30 40 50 60
5

Q_ex
0

−5
0 10 20 30 40 50 60

N_cd 0 10 20 30 40 50 60
time
Figure 6.4. Time-domain responses of the true system
(nonlinear simulator) (|) and of G^ id
7 ( ) to validation
data
A. Open-loop coprime-factor reduction. Since the nominal model
G42 is unstable,the reduction
 procedure was applied to its normalised
coprime factors N~42 M~ 42 following the procedure of Subsection 6.2.1:
   

N~r M~ r = bt N~42 M~ 42 ; r ; (6.61a)
G^ olr = M~ r 1 N~r : (6.61b)
 
The Hankel singular values of N~42 M~ 42 are depicted in Figure 6.5.
Two models were produced:
 rstly, we computed a 7th-order model G^ ol7 for the sake of establishing
a comparison with the 7th-order model G^ id
7 obtained by identication.
It was used to compute a 14th-order LQG controller C14 (G^ ol7 ) follow-
ing the procedure of Subsection 6.4.2. This controller destabilises the
nominal system G42 ;
 secondly, we computed a 10th-order model G^ ol10 . This is the lowest-
order model computed by this method that produces a stabilising
LQG controller for G42 . This controller has order 17 and is denoted
C17 (G^ ol10 ).
The performance of C17 (G^ ol10 ) will be studied in Subsection 6.4.6.
 Application to the design of a low-order controller for a complex PWR model 169

1
10

0
10

−1
10

−2
10
Hankel singular values

−3
10

−4
10

−5
10

−6
10

−7
10

−8
10
0 5 10 15 20 25 30 35 40 45
# of states

Figure 6.5. Open-loop model reduction: Hankel singu-

~ ] () and upper bound of the
lar values &i of [N~42 M
P4242
approximation error 2 i=r+1 &i (|)

B. Performance-based coprime-factor reduction. According to

the block diagram of Figure 6.3, CP ID should be considered as a two-
degree-of-freedom controller:

 
CP ID = K0tb K0
cd
2 3
Ktb 11 Ktb 12 0 0 (6.62)
=4 0 0 Kcd 11 Kcd 12 5:
| {z } | {z }
FP ID KPID

This is Case 2 of Section 2.5. The PWR plant of Figure 6.3 can then be
put in the formalism of Figure 2.2 by setting

r1 (t) = Wref (t); (6.63a)

170 A performance-based approach to control-oriented model reduction and to : : :

 
r2 (t) = dOhp(t) dUcex (t) T ; (6.63b)
 
v(t) = 0 0 0 T since no noise acts on the outputs; (6.63c)
 
w(t) = dOhp(t) dUcex (t) Wref (t) 0 0 0 T ; (6.63d)
 
z (t) = cOhp(t) cUcex (t) Werr (t) Qex (t) Ncd (t) T ; (6.63e)
 
h(t) = Wref (t) We (t) Qex (t) Ncd (t) T ; (6.63f)
2 3
1 1 0 0
W = 4 0 0 1 05 ; (6.63g)
0 0 0 1
 
g(t) = W  h(t) = Werr (t) Qex (t) Ncd (t) T ; (6.63h)

where Werr (t) = We (t) Wref (t). This yields the following relation
between the exogenous signals and the internal signals of the loop:

2 3
cOhp(t)
6cUcex (t)7     
6 7 I 2 2 0 0 1 2
6 Werr (t) 7 =
z (t) = 6 0 W To G42 ; CP ID w(t)
7
4 Qex (t) 5
Ncd (t)
2 3
1 0 0 (6.64)
60 1 07 2 3
60 0 17 dOhp (t)
    6 7
I 2 2 0 0 1 2
= 0 W To G ; CP ID 6 60 0 07 dUcex (t) :
74 5
42 6 7 Wref (t)
40 0 05
0 0 0

In this application, the regulated outputs are only Werr and Ncd . There-
fore, we shall only put a penalty on these two outputs in the reduction
 Application to the design of a low-order controller for a complex PWR model 171

criterion, which is thus4

   

J (G^ r ) = Wl N42 M 42 Nr M r Wr 1 (6.65)
where
2 3
1 0 0
60 1 07
6 7
60 0 17
Wr = 6
60 0 07
7 (6.66)
6 7
40 0 05
0 0 0
and
   
Wl = 00 0 1 0 0 I22 0 UP ID  1
0 0 0 1 0 W VP ID
   (6.67)
= 00 0 1 1 0 0 UP ID  1 :
0 0 0 0 1 VP ID
 
(Observe that Werr (t)h Ncd (t) = Wl  col(cOhp(t); cUcex (t); h(t)).)
  U40 i    
Here,  = N42 M 42 V40 and N42 M 42 (respectively Nr M r )
h i
are the normalised coprime factors of 0G1422 (respectively coprime fac-
h i
012
tors of ^r
G
). They are given by
 
  012 I11 013
N42 M 42 = N~42 031 M~ 42 (6.68)

4
The choice of a H1 -based reduction criterion may seem inappropriate, since our
control design criterion is H2 -based. We make the following remarks regarding this
choice.
 In practice, the choice of a H1-based reduction criterion is immaterial since the
reduction is carried out via balanced truncation which does not minimise the H1
norm.
 Our objective is to discuss the potential interest that one can have in taking the
closed-loop interconnection into account during model or controller reduction, the
reduction method being imposed a priori. This is very similar to the problem
of computing an uncertainty region that can be used for H1 robust control de-
sign using prediction error identication with a H2 -based criterion, as we do in
Chapter 5.
 Most reduction methods are based on H1 robustness criteria.
 H2 norm approximation is, as we already mentioned, a very diÆcult problem for
which no reliable solution exists.
172 A performance-based approach to control-oriented model reduction and to : : :

and
 
  012 I11 013
Nr M r = N ~r 031 M~ r ; (6.69)

where ~42 M~ 42  are the normalised coprime factors of G42 and
N
 
N~r M~ r are coprime factors of G^ r .
A suboptimal G^ r can be obtained by two-side frequency weighted bal-
anced truncation:
   

Nr M r = fwbt N42 M 42 ; Wl ; Wr ; r ; (6.70a)
G^ r = M~ r 1 N~r : (6.70b)

The frequency response of the weighting lter Wl is shown in Figure 6.6.

Clearly, the frequencies around 0:1 rad=s are amplied, which corre-
sponds to the crossover frequency of KP ID G42 . The eect of this lter
is thus comparable to that of the sensitivity function in closed-loop iden-
tication, which becomes obvious by noticing that
       1
UP ID  1 = CPIID I + 0G12 CP ID M 421 ; (6.71)
VP ID 42
whereby
 
Wl = 01 10 00 01
 
I11 013
 So (G42 ; KP ID )G42 FP ID So (G42 ; KP ID )M~ 421 (6.72)
where
So (G42 ; KP ID ) = (I + G42 KP ID ) 1 : (6.73)

The Hankel singular

 values
 of the input-output frequency weighted co-
prime factors N~42 M~ 42 are shown in Figure 6.7.
Two models were produced:
 rstly, we computed a 7th-order model G^ cl7 for the sake of establishing
a comparison with the 7th-order model G^ id
7 obtained by identication.
It was used to compute a 14th-order LQG controller C14 (G^ cl7 ) follow-
ing the procedure of Subsection 6.4.2;
 Application to the design of a low-order controller for a complex PWR model 173

From: U(1) From: U(2) From: U(3) From: U(4)

200

Magnitude [dB]
100

−100
To: Y(1)

−200
1000
Phase [deg]

500

−500

−1000
200
Magnitude [dB]

100

−100
To: Y(2)

−200

0
Phase [deg]

−500

−1000
0 0 0 0
10 10 10 10

Frequency [rad/s]

Figure 6.6. Closed-loop model reduction: Bode dia-

gram of Wl

 secondly, we computed a 4th-order model G^ cl4 . This is the lowest-

order model computed by this method that produces a stabilising
LQG controller for G42 . This controller has order 11 and is denoted
C11 (G^ cl4 ).
The performance of these two controllers will be studied in Subsec-
tion 6.4.6.

6.4.5. Controller reduction

A. Design of the optimal high-order controller. The rst step

in this approach consists in designing the optimal LQG controller for
G42 following the procedure of Subsection 6.4.2. Therefore, a minimal
realisation of G42 was used. This realisation has order  33, and  the
corresponding optimal LQG controller, denoted C40 = F40 K40 , has
order 40. This controller has a faster and less oscillatory response than
CP ID (see Subsection 6.4.6), and also a generalised stability margin
quasi twice as big:
bG42 ;K40 = 0:0140 > bG42 ;KPID = 0:0073: (6.74)
174 A performance-based approach to control-oriented model reduction and to : : :

2
10

0
10

−2
10

Hankel singular values

−4
10

−6
10

−8
10

−10
10
0 5 10 15 20 25 30 35 40 45
# of states

Figure 6.7. Closed-loop model reduction: Hankel sin-

gular values &i of input-output frequency weighted
[N42 M 42 ] ()

B. Open-loop coprime-factor reduction. Since the optimal con-

troller C40 is unstable, the
h reduction
i procedure was applied to its nor-
U40
malised coprime factors V40 following the procedure of Subsection 6.3.1:
    
Ur = bt U40 ; r ; (6.75a)
Vr V40
C^rol = Ur Vr 1 : (6.75b)
h i
U40
The Hankel singular values of V40 are depicted in Figure 6.8.
Three low-order controllers were produced:

 a 14th-order controller C^14ol , for the sake of establishing a compari-

son with the 14th-order controllers C14 (G^ id ^ ol ^ cl
7 ), C14 (G7 ) and C14 (G7 )
obtained via identication or model reduction;
 a 11th-order controller C^11
ol , for the sake of establishing a compari-

son with the 11th-order controller C11 (G^ cl4 ) obtained via closed-loop
model reduction;
 Application to the design of a low-order controller for a complex PWR model 175

2
10

0
10

−2
10
Hankel singular values

−4
10

−6
10

−8
10

−10
10

−12
10
0 5 10 15 20 25 30 35 40
# of states

Figure 6.8. Open-loop

h icontroller reduction: Hankel
singular values &i of UV40
40
() and upper bound of the
P40
approximation error 2 i=r+1 &i (|)

 a 9th-order controller C^9ol, for the sake of establishing a comparison

with the 9th-order controller C^9cl that will be computed by closed-loop
controller reduction.

The performance of these controllers will be studied in Subsection 6.4.6.

Remark. A straightforward application of Theorem 6.1 yields

   
U40 U14 = 1:4  10 3

V40 V14 1
   
U40 U11 = 2:5  10 3
< V

V11 1
40 (6.76)
< bG42 ;K40 = 1:4  10 2
   
U40 U9 = 2:3  10 2 ;
< V

V9 1
40
176 A performance-based approach to control-oriented model reduction and to : : :

hence only C^14 ^11

ol and C ol are guaranteed to stabilise G . However, the
42
condition of this theorem is only suÆcient and all three controllers ac-
tually stabilise G42 with respective generalised stability margins
bG42 ;K^ 14ol = 0:0141 > bG42 ;K^ 11ol = 0:0138 > bG42 ;K^ 9ol = 0:0115: (6.77)
Even though C^9ol stabilises G42 , its generalised stability margin is smaller
than that of the other two controllers, and its performance is much worse,
as shown in Subsection 6.4.6.

C. Performance-based coprime-factor reduction. The same rea-

soning as in closed-loop model reduction yields the following reduction
criterion:
   
^
J (Cr ) = Wl V
U 40 U r
W

(6.78)
40 Vr r
1
h i
where U40
V40 is a normalised coprime factorisation of C40 ,
2 3
1 0 0
60 1 07
6 7
1 N  60 0 17
Wr =   42 
M42 660
7; (6.79)
6 0 077
40 0 05
0 0 0
and
 
Wl = 00 00 1 1 0 0
0 0 0 1 : (6.80)
  h i  
Here,  = N42 M 42 U40
V40 where N42 M 42 is any coprime factori-
h i
sation of 0G1422 . As in closed-loop model reduction, the chosen criterion
J (C^r ) penalises the degradation of the transfer function between Wref ,
dOhp and dUcex on the one hand, and Werr and Ncd on the other hand.
A suboptimal C^r can be obtained by two-side frequency weighted bal-
anced truncation:
    
Ur = fwbt U40 ; W ; W ; r ; (6.81a)
Vr V40 l r

C^rcl = Ur Vr 1 : (6.81b)
 Application to the design of a low-order controller for a complex PWR model 177

The frequency response of the weighting lter Wr is shown in Figure 6.9.

Clearly, more weight is put in the frequency range corresponding to the
bandwidth of the closed loop, i.e. between 0 and 0:1 rad=s.

From: U(1) From: U(2) From: U(3)

Phase [deg] Magnitude [dB]

500

0
To: Y(1)

−500
0

−500

−1000
500
Phase [deg] Magnitude [dB] Phase [deg] Magnitude [dB]

0
To: Y(2)

−500
0

−500

−1000
500

0
To: Y(3)

−500
500
0
−500
−1000
500
Phase [deg] Magnitude [dB]

0
To: Y(4)

−500
500

−500
0 5 0 5 0 5
10 10 10 10 10 10

Frequency [rad/s]

Figure 6.9. Closed-loop controller reduction: Bode di-

agram of Wr (rst three inputs only; the others are 0)

The Hankel singular

h i values of the input-output frequency weighted co-
U40
prime factors V40 are shown in Figure 6.10.
Three low-order controllers were produced:

 a 14th-order controller C^14cl that will be compared with the 14th-order

controllers C14 (G^ id ^ ol ^ cl ^ ol
7 ), C14 (G7 ), C14 (G7 ) and C14 ;
 a 11th-order controller C^11 cl that will be compared with with the 11th-

order controllers C11 (G4 ) and C^11

^ cl ol ;

 a 9th-order controller C^9cl , which is the lowest-order controller ob-

tained by this method that achieves closed-loop stability and accept-
able performance with G42 . It will be compared with C^9ol computed
by open-loop controller reduction.

The performance of these controllers will be studied in Subsection 6.4.6.

178 A performance-based approach to control-oriented model reduction and to : : :

2
10

0
10

−2
10

Hankel singular values

−4
10

−6
10

−8
10

−10
10
0 5 10 15 20 25 30 35 40
# of states

Figure 6.10. Closed-loop controller reduction: Han-

kel
h singular
i values &i of input-output frequency weighted
V40 ()
U40

Remark. The three controllers computed here violate the condition of

Theorem 6.1, since
   
U40
U14 = 1:8  10 2
bG42 ;K40 = 1:4  10 2
V14 1 < V

40
   
U40
U11 = 3:4  10 2
V11 1 < V (6.82)
40
   
U40
U9 = 1:6  10 1 :
V9 1 < V

40
However, they all stabilise G42 with respective generalised stability mar-
gins
bG42 ;K^ 14cl = 0:0140  bG42 ;K^ 11cl = 0:0136
(6.83)
 bG42 ;K^ 9cl = 0:0141
quasi identical to bG42 ;K40 = 0:0140. Hence, from a generalised stabil-
ity margin point of view, our closed-loop reduction method appears to
be more performant than the unweighted reduction of the normalised
coprime factors.
 Application to the design of a low-order controller for a complex PWR model 179

6.4.6. Comparison of the performance of the designed controllers

A. Comparison of controllers computed after low-order model

identication or model reduction. Here we compare the perfor-
mance with G42 of the controllers computed from low-order models ob-
tained via an identication or a model reduction step.
Performance of 14th-order controllers. We rst consider the three
7th-order models G^ id ^ ol ^ cl
7 , G7 and G7 . The corresponding 14th-order con-
trollers are, respectively, C14 (G7 ), C14 (G^ ol7 ) and C14 (G^ cl7 ). Recall that
^ id

C14 (G^ ol7 ) destabilises G42 . Figures 6.11 to 6.13 (see pp. 182 and next)
show the behaviour of C14 (G^ id ^ cl
7 ) and C14 (G7 ), as well as that of the orig-
inal PID controller CP ID , when steps are applied to the Wref , dOhp and
dUcex inputs. One can observe that C14 (G^ id ^ cl
7 ) and C14 (G7 ) have simi-
lar performance when applied to G42 . They also have nearly identical
generalised stability margins, which are much larger than that of KP ID :
bG42 ;K14 (G^ id7 ) = 0:0139  bG42 ;K14 (G^cl7 ) = 0:0141
(6.84)
> bG42 ;KPID = 0:0073:

The Bode diagrams of C14 (G^ id ^ cl ^ ol

7 ), C14 (G7 ) and C14 (G7 ) are shown in
Figure 6.14. Clearly, the rst two controllers are very close to one an-
other, while the third one, which destabilises G42 , has a very dierent
frequency response.
Performance of other controllers. Here, we show that it is possible
to compute a performant controller from a much lower-order model when
the reduction step is carried out in closed loop rather than in open loop.
We consider the three low-order models G^ ol10 , G^ cl7 and G^ cl4 . The corre-
sponding controllers are, respectively, C17 (G^ ol10 ), C14 (G^ cl7 ) and C11 (G^ cl4 ).
C17 (G^ ol10 ), which has order 17, is the lowest-order controller with ac-
ceptable performance that could be computed via open-loop model re-
duction. As shown in Figures 6.15 to 6.17, its performance with G42
when steps are applied to the Wref , dOhp and dUcex inputs is not better
than that of C14 (G^ cl7 ), which has order 14 and was computed via closed-
loop model reduction. Furthermore, with closed-loop model reduction,
the 11th-order controller C11 (G^ cl4 ) obtained by pushing further the re-
duction of the design model order only presents a small degradation of
performance (larger transient response to a step disturbance dUcex : see
Figure 6.17). The three controllers also have nearly identical generalised
180 A performance-based approach to control-oriented model reduction and to : : :

stability margins (but the smallest is that of C17 (G^ ol10 )):
bG42 ;K17 (G^ ol10 ) = 0:0134  bG42 ;K14 (G^ cl7 ) = 0:0141
(6.85)
 bG42 ;K11(G^cl4 ) = 0:0144:
B. Comparison between model reduction and controller reduc-
tion. Here we compare the performance with G42 of the low-order con-
trollers computed either via a model reduction step before control design,
or via a controller reduction step after control design.
Performance of 14th-order controllers. We rst consider the four
14th-order controllers C14 (G^ ol7 ), C14 (G^ cl7 ), C^14 ol and C ^14
cl . Recall that

C14 (G^ ol7 ) destabilises G42 . Figures 6.18 to 6.20 show the behaviour of
the other three stabilising 14th-order controllers, as well as that of the
optimal 40th-order controller C40 , when steps are applied to the Wref ,
dOhp and dUcex inputs. One can observe that the responses of C^14 ol and

C^14
cl are nearly indistinguishable from those of C . C (G
40 14 ^ 7 ) has a larger
cl
transient response amplitude and a larger settling time with respect to
a step disturbance dUcex (see Figure 6.20). All three controllers have
generalised stability margins very close to that of the optimal C40 :
bG42 ;K14 (G^ cl7 ) = 0:0141  bG42 ;K^ 14ol = 0:0141
 bG42 ;K^ 14cl = 0:0140  bG42 ;K40 = 0:0140: (6.86)
Performance of 11th-order controllers. We consider the three 11th-
order controllers C11 (G^ cl4 ), C^11
ol and C ^11
cl . Recall that it was not possible
to use open-loop model reduction to obtain a stabilising controller of
order 11 for G42 , and that a stabiling controller with order less than 11
could not be obtained via closed-loop model reduction. Figures 6.21 to
6.23 show the behaviour of these controllers, as well as that of the opti-
mal 40th-order controller C40 , when steps are applied to the Wref , dOhp
and dUcex inputs. The observations made for the 14th-order controllers
are conrmed here. The transient response to a step at the dUcex input
has an amplitude ve times larger with C11 (G^ cl4 ) than with either of the
other two 11th-order controllers (see Figure 6.23). The responses of the
latters remain almost undistinguishable from those of C40 . All three
controllers have generalised stability margins very close to that of the
optimal C40 :
bG42 ;K11 (G^ cl4 ) = 0:0144  bG42 ;K^ 11ol = 0:0138
 bG42 ;K^ 11cl = 0:0136  bG42 ;K40 = 0:0140: (6.87)
 Application to the design of a low-order controller for a complex PWR model 181

Performance of 9th-order controllers. The reduction is pushed

down to order 9. Only controller reduction (either in open loop or in
closed loop) yields controllers with such a low order that stabilise G42 .
Figures 6.24 to 6.26 show that the performance of C^9cl (obtained by
closed-loop controller reduction of C40 ) is very close to that of C40 ,
while that of C^9ol (obtained by open-loop controller reduction of C40 ) is
totally unacceptable. Closed-loop reduction also preserves the stability
margin of the controller, contrary to open-loop reduction:
bG42 ;K^ 9cl = 0:0141  bG42 ;K40 = 0:0140
(6.88)
> bG42 ;K^ 9ol = 0:0115:

6.4.7. Comparison of the new performance-based closed-loop model

and controller reduction criteria with the stability-based closed-
loop reduction criteria

As we have already remarked in Subsections 6.2.3 and 6.3.3, our closed-

loop model and controller reduction criteria are dierent from the
stability-based ones in that sense that they aim at preserving the closed-
loop transfer matrix, i.e. the dynamics of the system that are dominant
in closed loop or the closed-loop performance of the controller, rather
than closed-loop stability.
It is thus interesting to verify if our performance-based criteria are really
more powerful than the stability-based ones. Therefore, we have used
the criterion (6.22) to compute a suboptimal 7th-order approximation
G^ st
7 of G42 by frequency weighted coprime-factor balanced truncation.
This model was then used to compute a 14th-order controller C14 (G^ st 7)
following the procedure of Subsection 6.4.2. The performance of this
controller is compared to that of C14 (G^ cl7 ) in Figure 6.27 (response to
a step on dOhp). C14 (G^ st7 ) does not ensure a correct regulation of Ncd
(nonzero static error) and is therefore not acceptable.
We have also used the criterion (6.46) to compute a 9th-order controller
C^9st by frequency weighted balanced truncation of the coprime factors
of C40 . The performance of this controller is compared to that of C^9cl in
Figure 6.28 (response to a step on Wref ). The response is very oscilla-
tory, hence C^9st is unacceptable.
Note that if the reduction is limited at a higher order (both in the case
of model reduction and in the case of controller reduction), the standard
182 A performance-based approach to control-oriented model reduction and to : : :

stability-based criteria deliver results that are comparable with those of

our new performance-based criteria.

6.4.8. Comparison of performance-based closed-loop controller coprime-

factor reduction with closed-loop balanced truncation

We have applied the closed-loop balanced truncation procedure, brie y

described in Subsection 6.3.4, to the reduction of our optimal controller
C40 . Two controllers, respectively of order 9 and of order 10, were
produced. The 9th-order controller does not stabilise G42 , while the
10th-order controller has performance similar to that of C^9cl .

6 1.5

4 1
Ohp

We
2 0.5

0 0
0 100 200 300 0 100 200 300
time [s] time [s]
0.01 1

0.5
cex

cd

0.005
N
U

0 −0.5
0 100 200 300 0 100 200 300
time [s] time [s]
3

2
ex

1
Q

−1
0 100 200 300
time [s]

Figure 6.11. Responses of G42 to a step on Wref with

controllers C14 (G^ id ^ cl
7 ) (|), C14 (G7 ) ( ) and CP ID (


)
Application to the design of a low-order controller for a complex PWR model 183

1 0.3

0.2
0.5

Ohp

e
0.1

W
0
0

−0.5 −0.1
0 100 200 300 0 100 200 300
−3
x 10 time [s] time [s]
2 0.05

0
Ucex

cd
0

N
−2

−4 −0.05
0 100 200 300 0 100 200 300
time [s] time [s]
0.4

0.2
Qex

−0.2

−0.4
0 100 200 300
time [s]

Figure 6.12. Responses of G42 to a step on dOhp with

controllers C14 (G^ id ^ cl
7 ) (|), C14 (G7 ) ( ) and CP ID (


)

0.1 0.01

0.05 0.005
Ohp

0 0
W

−0.05 −0.005

−0.1 −0.01
0 100 200 300 0 100 200 300
time [s] time [s]
0.02 0.4

0.2
0.01
Ucex

Ncd

0
0
−0.2

−0.01 −0.4
0 100 200 300 0 100 200 300
time [s] time [s]
2

1
Qex

−1
0 100 200 300
time [s]

Figure 6.13. Responses of G42 to a step on dUcex with

controllers C14 (G^ id ^ cl
7 ) (|), C14 (G7 ) ( ) and CP ID (


)
184 A performance-based approach to control-oriented model reduction and to : : :

From: Qex From: We From: Ncd From: Wref

200

Magnitude [dB]
100

−100

To: Ohp
−200
1000

Phase [deg]
500

−500

−1000
200

Magnitude [dB]
100

−100

To: Ucex
−200
500

Phase [deg]
0

−500

−1000
0 0 0 0
10 10 10 10

Frequency [rad/s]

^ id
Figure 6.14. Bode diagrams of C14 (G ^ cl
7 ) (|), C14 (G7 )
( ) and C14 (G^ ol7 ) (
)
6 1.5

4 1
Ohp

We
2 0.5

0 0
0 100 200 300 0 100 200 300
time [s] time [s]
0.01 0.3

0.2
cex

cd

0.005 0.1
N
U

0 −0.1
0 100 200 300 0 100 200 300
time [s] time [s]
3

2
Qex

0
0 100 200 300
time [s]

Figure 6.15. Responses of G42 to a step on Wref with

controllers C17 (G^ ol10 ) (|), C14 (G^ cl7 ) ( ) and C11 (G^ cl4 )
(
)
Application to the design of a low-order controller for a complex PWR model 185

1 0.3

0.2
0.5

Ohp

e
0.1

W
0
0

−0.5 −0.1
0 100 200 300 0 100 200 300
−3
x 10 time [s] time [s]
2 0.06

0.04
0
Ucex

cd
0.02

N
−2
0

−4 −0.02
0 100 200 300 0 100 200 300
time [s] time [s]
0.4

0.2
Qex

−0.2

−0.4
0 100 200 300
time [s]

Figure 6.16. Responses of G42 to a step on dOhp with

controllers C17 (G^ ol10 ) (|), C14 (G^ cl7 ) ( ) and C11 (G^ cl4 )
(
)
0.05 0.02

0.01
hp

0
W
O

−0.05 −0.01
0 100 200 300 0 100 200 300
time [s] time [s]
0.01 0.2

0.005 0.1
cex

Ncd

0 0
U

−0.005 −0.1

−0.01 −0.2
0 100 200 300 0 100 200 300
time [s] time [s]
1

0.5
ex

0
Q

−0.5

−1
0 100 200 300
time [s]

Figure 6.17. Responses of G42 to a step on dUcex with

controllers C17 (G^ ol10 ) (|), C14 (G^ cl7 ) ( ) and C11 (G^ cl4 )
(
)
186 A performance-based approach to control-oriented model reduction and to : : :

6 1.5

4 1

Ohp

We
2 0.5

0 0
0 100 200 300 0 100 200 300
−3
x 10 time [s] time [s]
10 0.3

0.2
5

cex

Ncd
0.1

U
0
0

−5 −0.1
0 100 200 300 0 100 200 300
time [s] time [s]
3

ex
1
Q
0

−1
0 100 200 300
time [s]

Figure 6.18. Responses of G42 to a step on Wref with

controllers C14 (G^ cl7 ) (|), C^14
ol ( cl (
) and C
) , C^14 40
(


)
1 0.3

0.2
0.5
hp

e
0.1

W
O

0
0

−0.5 −0.1
0 100 200 300 0 100 200 300
−3
x 10 time [s] time [s]
2 0.15

0.1
0
cex

cd

0.05
N
U

−2
0

−4 −0.05
0 100 200 300 0 100 200 300
time [s] time [s]
0.4

0.2
ex

0
Q

−0.2

−0.4
0 100 200 300
time [s]

Figure 6.19. Responses of G42 to a step on dOhp with

controllers C14 (G^ cl7 ) (|), C^14
ol ( cl (
) and C
) , C^14 40
(


)
Application to the design of a low-order controller for a complex PWR model 187

−3
x 10
0.05 4

hp

We
0
O
0

−0.05 −2
0 100 200 300 0 100 200 300
time [s] time [s]
0.01 0.2

0.005 0.1
cex

Ncd
0 0
U

−0.005 −0.1

−0.01 −0.2
0 100 200 300 0 100 200 300
time [s] time [s]
1

0.5
ex

0
Q

−0.5

−1
0 100 200 300
time [s]

Figure 6.20. Responses of G42 to a step on dUcex with

controllers C14 (G^ cl7 ) (|), C^14
ol ( cl (
) and C
) , C^14 40
(


)
6 1.5

4 1
hp

e
W
O

2 0.5

0 0
0 100 200 300 0 100 200 300
time [s] time [s]
0.03 0.6

0.02 0.4
Ucex

Ncd

0.01 0.2

0 0

−0.01 −0.2
0 100 200 300 0 100 200 300
time [s] time [s]
4

2
Qex

−2
0 100 200 300
time [s]

Figure 6.21. Responses of G42 to a step on Wref with

controllers C11 (G^ cl4 ) (|), C^11
ol ( cl (
) and C
) , C^11 40
(


)
188 A performance-based approach to control-oriented model reduction and to : : :

1 0.3

0.2
0.5

hp

e
0.1

W
O
0
0

−0.5 −0.1
0 100 200 300 0 100 200 300
time [s] time [s]
0.01 0.15

0.005 0.1

cex

cd
0 0.05

N
U
−0.005 0

−0.01 −0.05
0 100 200 300 0 100 200 300
time [s] time [s]
1

0.5

ex
0

Q
−0.5

−1
0 100 200 300
time [s]

Figure 6.22. Responses of G42 to a step on dOhp with

controllers C11 (G^ cl4 ) (|), C^11
ol ( cl (
) and C
) , C^11 40
(


)
0.05 0.02

0.01
hp

e
0

W
O

−0.05 −0.01
0 100 200 300 0 100 200 300
time [s] time [s]
0.01 0.2

0.005 0.1
cex

cd

0 0
N
U

−0.005 −0.1

−0.01 −0.2
0 100 200 300 0 100 200 300
time [s] time [s]
1

0.5
ex

0
Q

−0.5

−1
0 100 200 300
time [s]

Figure 6.23. Responses of G42 to a step on dUcex with

controllers C11 (G^ cl4 ) (|), C^11
ol ( cl (
) and C
) , C^11 40
(


)
Application to the design of a low-order controller for a complex PWR model 189

6 1.5

4 1

hp

e
W
O
2 0.5

0 0
0 100 200 300 0 100 200 300
time [s] time [s]
0.04 0.5

0.02
Ucex

Ncd
0
0

−0.02 −0.5
0 100 200 300 0 100 200 300
time [s] time [s]
6

4
Qex

−2
0 100 200 300
time [s]

Figure 6.24. Responses of G42 to a step on Wref with

controllers C^9ol (|), C^9cl ( ) and C40 (
)
1 0.3

0.2
0.5
hp

We

0.1
O

0
0

−0.5 −0.1
0 100 200 300 0 100 200 300
time [s] time [s]
0.02 0.3

0.01 0.2
Ucex

Ncd

0 0.1

−0.01 0

−0.02 −0.1
0 100 200 300 0 100 200 300
time [s] time [s]
1

0
Qex

−1

−2
0 100 200 300
time [s]

Figure 6.25. Responses of G42 to a step on dOhp with

controllers C^9ol (|), C^9cl ( ) and C40 (
)
190 A performance-based approach to control-oriented model reduction and to : : :

−3
x 10
0.1 4

0.05 2

hp

e
0 0

W
O
−0.05 −2

−0.1 −4
0 100 200 300 0 100 200 300
time [s] time [s]
0.01 0.2

0.005 0.1

cex

cd
0 0

N
U
−0.005 −0.1

−0.01 −0.2
0 100 200 300 0 100 200 300
time [s] time [s]
2

ex
0

Q
−1

−2
0 100 200 300
time [s]

Figure 6.26. Responses of G42 to a step on dUcex with

controllers C^9ol (|), C^9cl ( ) and C40 (
)
1 0.3

0.2
0.5
hp

e
0.1

W
O

0
0

−0.5 −0.1
0 100 200 300 0 100 200 300
−3
x 10 time [s] time [s]
2 0.06

0.04
0
cex

cd

0.02
N
U

−2
0

−4 −0.02
0 100 200 300 0 100 200 300
time [s] time [s]
0.4

0.2
ex

0
Q

−0.2

−0.4
0 100 200 300
time [s]

Figure 6.27. Responses of G42 to a step on dOhp with

controllers C14 (G^ cl7 ) (|) and C14 (G^ st
7)( )
 Conclusions 191

6 1.5

4 1

hp

e
W
O
2 0.5

0 0
0 100 200 300 0 100 200 300
time [s] time [s]
0.03 0.5

0.02
Ucex

Ncd
0.01 0

−0.01 −0.5
0 100 200 300 0 100 200 300
time [s] time [s]
6

4
Qex

−2
0 100 200 300
time [s]

Figure 6.28. Responses of G42 to a step on Wref with

controllers C^9cl (|) and C^9st ( )

6.5. Conclusions
In this chapter we have compared three approaches for the computation
of a low-order controller for a complex system. The rst two are based
on the construction of a low-order design model, either via identication
of such a model based on data collected in closed loop, or by explicit
reduction of a precise high-order model of the system. In these cases the
last step of the procedure is the design of the controller, and its order is
directly linked to that of the design model. The third method, on the
contrary, starts by the computation of a high-order controller based on
a high-order model of the system. The order of this controller is reduced
in the second step of the design procedure.
The two methods based on an explicit order reduction step (model or
controller reduction) can both be considered either in open loop or in
closed loop, depending on the use of frequency weighting lters aimed
at modifying the reduction criterion in order to orient it towards the
preservation of the closed-loop performance.
These techniques have been tested on a realistic high-order linearised
model of a pressurised water reactor nuclear power plant, the goal being
192 A performance-based approach to control-oriented model reduction and to : : :

the replacement of two PID controllers by a multivariable controller for

the electrical power while ensuring an acceptable regulation of the water
level in the condenser.
Following the observations made in Section 6.4, two tables can be drawn
up. Table 6.1 classies the methods with respect to the performance
achieved by controllers of the same order (when this order is suÆciently
low to make dierences appear). The performance of each controller is
evaluated in terms of the discrepancy between its responses and those of
the optimal high-order controller (C40 in this application). On the other
hand, Table 6.2 classies the methods with respect to the lowest order
that can be reached while achieving a pre-specied level of performance.

Performance Method
MAX closed-loop controller reduction
& open-loop controller reduction
closed-loop system identication
&& closed-loop model reduction
MIN open-loop model reduction
Table 6.1. Classication of the methods w.r.t. the per-
formance that can be achieved with controllers of the
same order

Order Method
MIN closed-loop controller reduction
closed-loop model reduction
% open-loop controller reduction
%% closed-loop system identication
MAX open-loop model reduction
Table 6.2. Classication of the methods w.r.t. the or-
der that can be reached while achieving a pre-specied
level of performance

The comparative study has delivered three major ndings:

 Conclusions 193

1. For a pre-specied level of performance, a given method (model or

controller reduction) always makes it possible to compute a lower-
order controller if it is used in its closed-loop version rather than in
open loop. Alternatively, for a given order, the controllers computed
via closed-loop methods have generally better performance than those
computed via open-loop methods (only open-loop controller reduction
remains competitive versus closed-loop model reduction, but this is
due to the fact that the former is based on a robust stability criterion
if it is carried out as explained in Section 6.2.2, i.e. by reducing the
normalised coprime factors of the controller). This message is prob-
ably familiar to theoreticians, but still needs to permeate industrial
practice.
2. Closed-loop identication proves to be a viable alternative to model
or controller reduction for the computation of low-order controllers.
It constitutes in fact an implicit order reduction method that has the
added advantage that no high-precision model is necessary to start
with (such a model may sometimes be diÆcult to obtain). Only data
collected on the closed loop, generally during normal operation of
the plant, are required. On the other hand, the absence of a precise
model of the plant may pose problems when trying to assess the
performance of the controller before its application to the plant. As
a result, the identied low-order model should always be delivered
with an estimation of the committed modelling error in order to do
robust control design. Note also that the estimate will depend on the
controller operating in the loop during data collection, as the low-
order model obtained by closed-loop model reduction does depend on
the controller used during reduction. It is therefore important to use
a controller that is suÆciently close to the to-be-designed controller,
as stressed in Chapter 3.
3. Methods based on closed-loop (respectively open-loop) controller re-
duction appear to be potentially more powerful than those based on
closed-loop (respectively open-loop) model reduction. This result is
quite logical since any reduction step implies a loss of information: it
is best to keep all the information (i.e. the richness of the model) as
long as possible. In other words, it may be diÆcult to predict how
the error introduced during model reduction will propagate during
the subsequent steps of the control design procedure and aect the
resulting controller while, on the other hand, in case of controller re-
duction, the starting point is the optimal high-order controller and
the loss of performance is directly related to the approximation error
committed during the reduction. Another point in favour of controller
194 A performance-based approach to control-oriented model reduction and to : : :

reduction is the fact that the use of lters during control design (e.g.
loop shaping lters) generally yields a controller with higher order
than the design model. It is of course easier to compensate for this
increase in order after the design than before.
Finally, let us remark that the performance-based criterion we have sug-
gested for closed-loop model or controller reduction has delivered lower-
order controllers than the standard stability-based criterion for a given
level of performance.
 CHAPTER 7
Future issues on variance and on model
order selection

7.1. Introduction
In this chapter, we would like to open a door on variance issues related
to the model order in prediction error identication.
A rst point we address is the eect of overmodelling (i.e. the choice of
a model structure that is unbiased, but which has a higher number of
parameters than the true system) on the variance of the estimate in the
frequency domain. More precisely, we would like to give some insight
on the eect of poles or zeros in excess on the frequency distribution
of the variance. This issue is strongly related to our model validation
procedure, since the latter involves the identication of an unbiased
model (i.e. a model which has at least the exact structure of the true
system, but which could be overparametrised). Our reasoning is only
based on simulation results and, as it is, should only be considered as
the starting point of possible future theoretical research.
A second point concerns the choice between the identication of a low-
order model versus the identication of a full-order model followed by
model reduction. Here also, we shall essentially focus on variance as-
pects. This issue comes in complement to Chapter 6, where the com-
parison has been established regarding the bias distribution (through
the lens of control design) on the basis of a practical application.

195
196 Future issues on variance and on model order selection

7.2. The eect of overmodelling on the variance of

estimated transfer functions
A major issue in our model validation procedure (see Chapter 5) is
the computation of an unbiased estimate of the model error or of the
system itself. The covariance matrix of the parameter estimate is then
used to compute an uncertainty region in transfer function space that
is guaranteed to contain the true system (at some probability). In this
section, we try to give some insight on the choice of the number of
parameters and of the model structure.

7.2.1. The eect of additional poles and zeros

Since bias must absolutely be avoided for the model or the model error
model that is computed during the validation procedure, it may be very
appealing to voluntarily choose an overparametrised model structure.
The question that arises, in this case, is `what is the eect of poles
and/or zeros in excess on the frequency distribution of the uncertainty
region?'.
It is very easy to show that every parameter in excess results in an
increased Cramer-Rao lower bound for the estimate of the other param-
eters. Indeed, let
 
F (1 ; 2 ) = FF11 FF12 (7.1)
21 22

be the information matrix of the overparametrised model (1 denotes

the parameter vector of the minimal unbiased model, which we call the
full-order model, while 2 represents the parameters in excess). The
covariance matrix of the full-order model parameter vector is
Pfull
 1
= F111 (7.2)
while that of the overparametrised model parameter vector is
Ph  i = F
1
1 ( ;  )
1 2
2
" over #
P Pover (7.3)
, over over :
1 ; 1 2

P ;  P
2 1 2
 The eect of overmodelling on the variance of estimated transfer functions 197

where


Pover
1
= F111 + F111 F12 F22 F21 F111 F12 1 F21 F111 ; (7.4a)


Pover 1
 ;  = F11 F12 F22
1 2
F21 F111 F12 ;1
(7.4b)


Pover
 ; =
2 1
1
F22 F21 F111 F12 F21 F111 ; (7.4c)


Pover
2
= F22 F21 F111 F12 : 1
(7.4d)
It is easy to verify that the covariance matrix P of the subset 1 of
over
1
parameters, in the case of an overparametrised model, is given by
Pover
1 = F11 + F11 F12 P2 F21 F11 > P1 :
1 1 over 1 full
(7.5)
This expression clearly shows that adding parameters to a model in-
creases its uncertainty. The asymptotic expressions for the variance in
transfer function space (see Subsection 2.2.5, x B) yield the same con-
clusion, since the number of parameters n is a factor of the asymptotic
covariance:

n v ( ! )
cov G^ (ej! )  in open loop, and (7.6)
N u (!)

n v ( ! )
cov G^ (ej! )  in closed loop. (7.7)
N ru (!)
These expressions mean that the uncertainty on the parametric model
is proportional to that on the nonparametric estimate, but that the
parametric model has an averaging eect over frequency which results in
a reduction of the eect of the noise by a factor Nn . However, these results
are only valid asymptotically, i.e. for n tending to innity. Although it
is very obvious that any additional parameter will increase the variance
of the estimate, one can expect that this increase will not be uniform
over the frequency range.
Imagine for instance that a pole is added to an unbiased model struc-
ture M. Since S 2 M without this additional pole, it is clear that this
pole is not useful and that its estimate will tend to a value such that
the impact on the cost function V () be as small as possible in (3.11)
(where S (G0 ; Kid ) = 1, Kid = 0 and ru = u in case of open-loop iden-
tication). Hence, this pole will tend to frequencies where the SNR (or,
equivalently, the closed-loop sensitivity function) is small, i.e. typically
outside the frequency band of interest, and its impact on the overall
model variance will remain small at frequencies where the sensitivity
function is large, i.e. around the crossover frequency and the frequency
where the gain margin is determined. On the contrary, if a pole and a
zero are added, they will tend to cancel. As this cancellation can happen
198 Future issues on variance and on model order selection

at any frequency, this will probably result in a uniform increase of the

variance over the frequency range.
To illustrate this discussion, let us consider the following ARX `true'
system:
1 + 0:5z 1 1 1
G0 (z ) = z ; H0 (z ) = : (7.8)
1 0:1z 1 1 0:1z 1
This system is connected to the following feedback controller:
0:15z 1
K (z ) = (7.9)
1 z 1
and excited using Gaussian white noise with zero mean and unit variance
for r2 (t) in Figure 2.1. The white noise e(t) acting on the system has zero
mean and variance 100. 1000 input-output data samples are collected
and used to identify ve unbiased models using a direct closed-loop
identication method. These models have respectively the following
structures:
1. ARX[1,2,1], which has the exact structure of (G0 ; H0 );
2. ARX[5,2,1], which has four additional poles (total: 4 parameters in
excess);
3. ARX[1,6,1], which has four additional zeros (total: 4 parameters in
excess);
4. ARX[3,4,1], which has two additional poles and two additional zeros
(total: 4 parameters in excess);
5. ARX[9,2,1], which has height additional poles (total: 8 parameters
in excess).
100 Monte-Carlo runs of this experiment are carried out. During each
of these runs, the ve models mentioned above are identied, hence 100
realisations of each model are available. These realisations are used to
compute the experimental standard deviations of the estimates, which
are plotted in Figure 7.1 together with the closed-loop system sensitivity
function S (G0 ; K ).
For the ARX[3,4,1] model, the increase of variance with respect to the
full-order model ARX[1,2,1] is, as expected, nearly uniform over fre-
quency due to the pole-zero cancellations. For the two models with
additional poles only, ARX[5,2,1] and ARX[9,2,1], on the contrary, the
variance remains relatively small where S (G0 ; K ) is large, but is much
larger at low frequencies, where the sensitivity function is small. As a
result, pole-zero cancellations will typically lead to more conservative
 The eect of overmodelling on the variance of estimated transfer functions 199

−5
amplitude [dB]

−10

−15

−20

−25
−2 −1 0
10 10 10
frequency [rad/s]

Figure 7.1. Amplitude Bode plots of S (G0 ; K )

(|) and of the experimental standard deviations
of ARX[1,2,1] (+), ARX[5,2,1] (), ARX[1,6,1] (4),
ARX[3,4,1] () and ARX[9,2,1] ()

uncertainty regions than additional poles only when the purpose is to

make robust control design; however, since pole-zero cancellations are
also easier to detect, this should not be a major problem. Additional
zeros, as in the ARX[1,6,1] model, also lead to a quasi-uniform increase
of the variance, but with less eect than pole-zero cancellations.

7.2.2. The choice of the model structure

It has been argued in (Ljung, 1997) that, in the case of model vali-
dation by identication of a model error model, the elements of the
cross-correlation function R"u (t) between the residuals "(t) and a nite
vector of past inputs u(t) could be taken as a FIR estimate of the model
error @G.
More generally, any transfer function in RH1 can be approximated
arbitrarily well by a FIR model, provided a suÆcient number of Markov
parameters are identied. However, this FIR model will generally have
a larger number of parameters than an innite impulse response model.
200 Future issues on variance and on model order selection

Does this result in an increase of the variance? The answer is yes, as

the following example shows.
Consider the same system (G0 ; H0 ) as above. Figure 7.2 shows that G0
can be almost perfectly approximated by a FIR model with 4 nonzero
parameters, and that the noise process v(t) = H0 (z )e(t) can be repre-
sented by a moving-average process with 4 nonzero parameters (among
which the rst one is 1, since H0 is monic). As a result, the system
can be represented by an ARMAX[0,4,3,1] model (i.e. with denomina-
tor polynomial of order 0, input-output numerator polynomial of order
3 with 4 free parameters, noise numerator polynomial of order 3, and
system delay 1).

1.2

1
Amplitude

0.8

0.6

0.4

0.2

0
0 1 2 3 4 5 6 7 8 9 10

Time (sec.)

1.2

1
Amplitude

0.8

0.6

0.4

0.2

0
0 1 2 3 4 5 6 7 8 9 10

Time (sec.)

Figure 7.2. Impulse responses of G0 (top) and H0 (bottom)

During the Monte-Carlo runs described above, the following models were
also identied:

1. ARMAX[0,4,3,1], which is the lowest-order unbiased FIR representa-

tion of (G0 ; H0 ). It has the same number of parameters as ARX[5,2,1]
which is overparametrised;
2. ARMAX[0,8,3,1], which has four additional parameters. It has the
same number of parameters as ARX[9,2,1] which is overparametrised.
 The eect of overmodelling on the variance of estimated transfer functions 201

The experimental standard deviations of these models are compared to

those of ARX[1,2,1], ARX[5,2,1] and ARX[9,2,1] in Figure 7.3. Clearly,
the ARMAX[0,4,3,1] (FIR) model has a larger variance error attached
to it than the ARX[1,2,1] model, although both represent the true
system without overmodelling. Furthermore, the variance of the AR-
MAX[0,4,3,1] model has a distribution over frequency that is very dif-
ferent from that of the ARX[5,2,1] model, which has a smaller variance
where the closed-loop sensitivity function is large. The same conclusion
holds when comparing the ARMAX[0,8,3,1] and ARX[9,2,1] models.

−5
amplitude [dB]

−10

−15

−20

−25
−2 −1 0
10 10 10
frequency [rad/s]

Figure 7.3. Amplitude Bode plots of S (G0 ; K )

(|) and of the experimental standard deviations of
ARX[1,2,1] (+), ARX[5,2,1] (), ARX[9,2,1] (), AR-
MAX[0,4,3,1] (o) and ARMAX[0,8,3,1] ()

7.2.3. Conclusion

The eects of overmodelling on the uncertainty distribution would de-

serve more theoretical investigations. However, we can already formulate
the following preliminary conclusions. We insist on the fact that these
conclusions must be taken with caution, since they are only based on
simulation results.
202 Future issues on variance and on model order selection

 The presence of a parametrised denominator in the identied model

adds exibility to this model, which is then delivered with a tuned
uncertainty region. A model with no such denominator (e.g. a FIR
model or a model with xed poles) appears to have less exibility,
which results in an uncertainty region that is more equally distributed
over frequency and, therefore, less adapted to the control design ob-
jective.
 When an unbiased estimate is required, it can be tempting to over-
parametrise the model. However, cancelling pairs of poles and zeros
make the uncertainty grow uniformly over the frequency range. When
there are only poles in excess, the uncertainty grows essentially out-
side the frequency range of interest, with less consequences in terms of
control design. Zeros in excess seem to give less exibility than poles
and to induce a uniform increase of the variance, but in a smaller
amount than cancelling poles and zeros.

7.3. Variance issues in system identication and

model reduction
Several ways exist for obtaining a low-order model for some system. If
prediction error identication is used, one can choose between
 identifying the best estimate in some low-order (biased) model set, or
 identifying a full-order (unbiased) model and then applying a step of
model reduction to this model.
Recently, it has been shown in (Tjarnstrom and Ljung, 1999) that, for a
FIR system (but this result is also valid for other structures), the second
approach could lead to a low-order model with smaller variance than the
rst approach, if L2 reduction is used with
the exact input spectrum as
frequency weighting. Let G^ (z; ); H^ (z; ) denote the full-order model
with parameter vector , and let ^ be the parameter estimate obtained
by prediction error identication. In this case, ^ tends asymptotically
to
Z   

 = arg minn 
G0 (ej! ) G^ n (ej! ; )2  u (!) 2 d! (7.10)
2R  ^ (ej! ; )
H

as the number of data N tends to innity (H^ (ej! ; ) = 1 in the FIR

case). In the second step, the parameters  of the L2 reduced-order
 Variance issues in system identication and model reduction 203

model G^ r are determined through

Z   2
 (! )
 = arg minr ^
Gn (e ;
j!
^) G^ r (ej! ; )  uj! 2 d! (7.11)
 2R  ^ (e ; ^)
H

using knowledge of u (!).

The major limitation of this result is the fact that, in many practical
situations, the exact input spectrum is unknown. This will for instance
always be the case if the data are collected in closed loop, since in this
case the input depends on the noise. Even in open loop, an imper-
fect knowledge of the actuator dynamics coupled with the presence of
measurement noise will always result in an imperfect knowledge of the
spectrum of the input signal that is actually applied to the plant.
When the exact input spectrum is unknown, the maximum-likelihood
L2 reduced-order estimate of the high-order model is obtained by least-
square approximation (in the FIR case). In this case, the variance of the
low-order estimate is the same for the two approaches considered here,
as we now show.
Assume that the true FIR system is described by the following equation:
n
X
y(t) = bk u(t k) + e(t) , 0T '(t) + e(t)
k=1
r n
X X (7.12)
= bk u(t k) + bk u(t k) + e(t)
k=1 k=r+1
, 0 '1 (t) + 0 '2 (t) + e(t);
T T

where
 
0 = b1 : : : br T ; (7.13a)
 
0 = br+1 : : : bn T ; (7.13b)
 
0 = 0 ; (7.13c)
0
 
'1 (t) = u(t 1) : : : u(t r) T ; (7.13d)
 
'2 (t) = u(t r 1) : : : u(t n) T ; (7.13e)
 
'(t) = u(t 1) : : : u(t n) T ; (7.13f)
and e(t) is white noise with zero mean and variance 0 .
204 Future issues on variance and on model order selection

If a low-order model of order r is directly identied using N input-output

data, the estimate is given by
" N
# 1 N
1 X 1 X
^dir = '1 (t)'1 (t)
T
'1 (t)y(t)
N t=1
N t=1
" # 1
N N
1 X 1 X
(7.14)
= 0 + '1 (t)'T1 (t) '1 (t)'T2 (t)0
N t=1
N t=1
" N
# 1 N
1 X 1 X
+ ' (t)'T (t) ' (t)e(t):
N t=1 1 1 N t=1 1
Its expectation (with respect to the noise) is
N
" # 1 N
1X 1 X
E ^dir = 0 + ' (t)'T (t) '1 (t)'T2 (t)0 (7.15)
N t=1 1 1 N t=1
and its covariance matrix is
" N # 1
X
P^dir = '1 (t)'T1 (t) 0 : (7.16)
t=1

If a full-order model is rst identied and then reduced by least-square

approximation, we have succesively
" # 1
N N
1 X 1 X
^ = '(t)' (t)
T
'(t)y(t) (7.17)
N t=1
N t=1

and
" N
# 1
N
1 X 1 X
^red = '1 (t)'1 (t)
T
'1 (t)^y (t; ^) (7.18)
N t=1
N t=1
where
y^(t; ^) = 'T (t)^ (7.19)
is the simulated output of the full-order estimate. It follows easily that
" # 1" #
N N
1 X 1 X
^red = '1 (t)'T1 (t) '1 (t)'T (t)
N t=1
N t=1
" # 1 (7.20)
N N
1 X 1 X
 N
'(t)'T (t)
N
'(t)y(t)
t=1 t=1
 Variance issues in system identication and model reduction 205

where
N
X N
X N
X
'(t)y(t) = '(t)'T (t)0 + '(t)e(t); (7.21)
t=1 t=1 t=1
and nally
" # 1
N N
1 X 1 X
^red = 0 + '1 (t)'1 (t)
T
'1 (t)'T2 (t)0
N t=1
N t=1
" # 1" #
N N
1 X 1 X
(7.22)
+ '1 (t)'1 (t)
T
'1 (t)' (t)
T
N t=1
N t=1
" N
# 1 N
 N1 1
X X
'(t)'T (t) '(t)e(t):
t=1
N t=1
Its expectation and covariance matrix are respectively
" # 1
N N
1 X 1 X
E ^red = 0 + '1 (t)'1 (t)T
'1 (t)'T2 (t)0 (7.23)
N t=1
N t=1
and
" N # 1
X
P^red = '1 (t)'T1 (t) 0 ; (7.24)
t=1
which are equal to those of ^dir . The same bias and variance errors
are thus committed with both approaches if the exact noise spectrum is
unknown and the reduction is carried out on the basis of the same data
samples that are used for identication.
It should also be noted that L2 model reduction is not an easy prob-
lem. Most of the time, it requires the use of iterative algorithms, with
all the problems that such algorithms involve: local minima, huge com-
putational cost, etc. Therefore, balanced truncation (or other analytic
methods) is generally preferred. The way the variance of a low-order
model computed by balanced truncation is related to the variance of the
initial high-order model is still an open problem.
Finally, we remark that the variance of the low-order model is not a
very important issue, at least with respect to a robust control design
objective. Indeed, this variance cannot be used for the computation
of an uncertainty region containing the true system, since the model is
biased. The only role it plays is that it determines the variance of the
206 Future issues on variance and on model order selection

resulting controller: the higher the variance of the model, the higher the
variance of the optimal controller designed for this model. Yet, from a
robustness point of view, one must take into account the variance of an
unbiased (high-order) model, and it is clear that this variance cannot be
reduced by model reduction.
 CHAPTER 8
Conclusions

8.1. Contribution of this thesis

In this thesis, we have elaborated on the question of `How should we
design the modelling of a system when the ultimate goal is control de-
sign?'. Several important ndings have been produced; all are related to
the in uence of the experimental conditions on the resulting modelling
errors.
A rst important result concerns the identication of a system in closed
loop when the operating controller is unstable and/or nonminimum
phase. It has been shown that the presence of poles and/or zeros of
the controller outside the unit circle leads to a model with guaranteed
closed-loop instability, when this model is connected to the same con-
troller, if some precautions are not taken during the identication pro-
cedure. Guidelines have been given in order to avoid the emergence of
this problem. They concern both the choice of the identication method
(direct, indirect or coprime-factor approaches, or Hansen scheme) and
the choice of the excitation source (excitation signal injected at the in-
put of the plant or at the input of the controller). It has also been shown
that poles or zeros of the controller on the unit circle may have a block-
ing eect on the excitation signals, possibly leading to huge modelling
errors at some frequencies because of the zero signal-to-noise ratio at
these frequencies.
The consequences of this instability problem for control design are very
important. Indeed, we have shown that the destabilisation of the nom-
inal model G^ by the current (to-be-replaced) controller has a direct

207
208 Conclusions

interpretation in terms of the  -gap distance between this model and

the underlying unknown true system G0 . Two issues are involved here:
 this distance, which should ideally be as small as possible, is lower
bounded by the stability margin of the actual closed loop, with the
consequence that suÆcient conditions for the stabilisation of the true
system by a new G^ -based controller may systematically be falsied;
 although these are only suÆcient conditions, the size of the set of
G^ -based controllers that stabilise G0 decreases as the  -gap between
G^ and G0 increases.
As a result, closed-loop identication should always be carried out ac-
cording to our guidelines when the purpose is to obtain a good model
for control design.
A second important result concerns the validation of a given model or
controller. One way of validating a model G^ is to use data collected on
the true system G0 to build an uncertainty region guaranteed to contain
this true system (at least with some specied probability). Therefore,
an unbiased model of G0 is identied and the covariance matrix of its
parameters is used to construct this uncertainty region in the frequency
domain. G^ is then called validated if it is contained in this uncertainty
region, while a given controller is called validated if it stabilises and
achieves a specied level of performance with all systems in this uncer-
tainty region. Our contribution has consisted in proposing a closed-loop
validation procedure based on the heuristic motivation that the identi-
cation criterion should match the performance criterion. Our validation
procedure has proved to produce uncertainty regions that are better
tuned towards robust control design than uncertainty regions produced
by open-loop validation.
In many real-life situations, a high-order model is obtained by linearisa-
tion of a nonlinear plant simulator, or by nite-element modelling, etc.
When a low-order controller is desired, a possible solution is to apply
an order reduction technique to that model before designing the con-
troller; an alternative is to rst design a high-order controller and to
then reduce its order. The third main contribution of this thesis has
been the application of the heuristic motivations for performing closed-
loop identication to the case of model or controller reduction. We have
proposed a new performance-oriented criterion for model or controller
reduction based on the preservation of the closed-loop transfer function
(in that, our method compares with closed-loop balanced truncation).
 Open questions 209

This criterion is symmetric, in that it can be used for the reduction of

the left coprime factors of the model with a frequency weighting that
depends on the right coprime factors of the controller, or to the reduc-
tion of the latters with a frequency weighting that depends of the form-
ers (in that, our criterion compares with the standard stability-oriented
criterion based on the preservation of the Bezout identity). Since the
degradation of the closed-loop transfer function (measured in the H1
norm) is an upper bound of the degradation of the generalised stability
margin (provided the closed loop remains internally stable after reduc-
tion), our performance-based method is also oriented towards stability
robustness.

8.2. Open questions

At the end of this thesis, several interesting open problems remain and
mark out the path for possible future research. Here we point out two
major questions which, in our opinion, would deserve attention in the
near future.
The rst issue concerns our model validation procedure. An important
feature of this procedure is the fact that an unbiased model (or model
error model) has to be identied in order to build an uncertainty re-
gion in the frequency domain. However, it seems important to avoid
overmodelling as much as possible to save the variance from blowing
up. In-depth theoretical study of the eect of overmodelling on the vari-
ance of estimated transfer functions should therefore be considered as a
possible research activity.
The second point, also related to our model validation procedure, con-
cerns control design. We have seen how an uncertainty set delivered
by a validation experiment could be used for the validation of a given
controller. However, it would be even more interesting to be able to use
such an uncertainty set for the design of this controller. The problems
that are involved here are thus

 the denition of a control design criterion over a parametrised set

of transfer functions produced by a prediction error identication or
validation experiment;
 the computation of the optimal controller with respect to this crite-
rion.
210 Conclusions

Other interesting work would concern the extension of our results on

model and controller validation to the case of nonlinear systems.
 Bibliography
Adams, D. (1979). The Hitch-Hikers Guide to the Galaxy: Life, the Universe and
Everything. Pan Books. London, UK.
Anderson, B.D.O. and J.B. Moore (1989). Optimal Control: Linear Quadratic Meth-
ods. Prentice-Hall. Englewood Clis, New Jersey, USA.
Anderson, B.D.O. and Y. Liu (1989). Controller reduction: Concepts and approaches.
IEEE Trans. on Automatic Control 34(8), 802{812.
Anderson, B.D.O., X. Bombois, M. Gevers and C. Kulcsar (1998). Caution in iterative
modeling and control design. In: Proc. of the 1998 IFAC Workshop on Adaptive
Systems in Control and Signal Processing. Glasgow, Scotland. pp. 13{19.
Ansay, P., M. Gevers and V. Wertz (1999). Closed-loop or open-loop models in identi-
cation for control. In: CD-ROM Proc. of the 5th European Control Conference.
Karlsruhe, Germany. Paper AP3-2 (F544).

Astrom, K.J. (1993). Matching criteria for control and identication. In: Proc. of the
2nd European Control Conference. Groningen, The Netherlands. pp. 697{701.
Bendotti, P. and B. Bodenheimer (1994). Identication and H1 control design for a
Pressurized Water Reactor. In: Proc. of the 33th IEEE Conference on Decision
and Control. Orlando, Florida, USA. pp. 1072{1077.
Bendotti, P., B. Codrons, C.-M. Falinower and M. Gevers (1998). Control-oriented
low-order modelling of a complex PWR plant: a comparison between open-loop
and closed-loop techniques. In: Proc. of the 37th IEEE Conference on Decision
and Control. Tampa, Florida, USA. pp. 3390{3395.
Bernstein, D.S. and D.C. Hyland (1984). The optimal projection equations for xed-
order dynamic compensation. IEEE Trans. on Automatic Control AC-29, 1034{
1037.
Bitmead, R.R., M. Gevers and V. Wertz (1990). Adaptive Optimal Control { The
Thinking Man's GPC. Series in systems and control engineering. Prentice-Hall.
Englewood Clis, New Jersey, USA.
Blondel, V., M. Gevers and R.R. Bitmead (1997). When is a model good for control
design? In: Proc. of the 36th IEEE Conference on Decision and Control. San
Diego, California, USA. pp. 1283{1288.
Bombois, X., M. Gevers and G. Scorletti (1999a). Controller validation based on an
identied model. In: Proc. of the 38th IEEE Conference on Decision and Control.
Phoenix, Arizona, USA. pp. 2816{2821.
Bombois, X., M. Gevers and G. Scorletti (1999b). Controller validation for a vali-
dated model set. In: CD-ROM Proc. of the 5th European Control Conference.
Karlsruhe, Germany. Paper 869.
Bombois, X., M. Gevers and G. Scorletti (2000a). A measure of robust stability for
an identied set of parametrized transfer functions. Accepted for publication in
IEEE Trans. on Automatic Control.
211
212 Bibliography

Bombois, X., M. Gevers and G. Scorletti (2000b). A measure of robust stability

for an uncertainty set obtained by prediction error identication. Accepted for
presentation at the 39th Conference on Decision and Control, Sydney, Australia,
December 2000.
Bombois, X., M. Gevers, G. Scorletti and B.D.O. Anderson (2000c). Controller val-
idation for stability and performance based on an uncertainty region designed
from an identied model. In: CD-ROM Proc. of the IFAC System Identication
Symposium (SYSID 2000). Santa Barbara, California, USA. Paper WePM1-6;
also submitted to Automatica.
Boyd, S.P. and J. Doyle (1987). Comparison of peak and RMS gains for discrete time
systems. Systems and Control Letters 9, 1{6.
Ceton, C., P.M.R. Wortelboer and O.H. Bosgra (1993). Frequency-weighted closed-
loop balanced truncation. In: Proc. of the 2nd European Control Conference.
Groningen, The Netherlands. pp. 697{701.
Codrons, B., B.D.O. Anderson and M. Gevers (2000a). Closed-loop identication
with an unstable or nonminimum phase controller. Accepted for publication in
Automatica.
Codrons, B., B.D.O. Anderson and M. Gevers (2000b). Closed-loop identication
with an unstable or nonminimum phase controller. In: CD-ROM Proc. of the
IFAC System Identication Symposium (SYSID 2000). Santa Barbara, Califor-
nia, USA. Paper ThPM1-3.
Codrons, B., P. Bendotti, C.-M. Falinower and M. Gevers (1999). A comparison be-
tween model reduction and controller reduction: Application to a PWR nuclear
plant. In: Proc. of the 38th IEEE Conference on Decision and Control. Phoenix,
Arizona, USA. pp. 4625{4630.

Codrons, B., P. Bendotti, C.-M. Falinower and M. Gevers (2000c). Etude comparative
des methodes d'identication et de reduction pour la commande { application
a la synthese d'un regulateur d'ordre reduit pour un modele lineaire complexe
de REP. Study report. Centre for Systems Engineering and Applied Mechanics
(CESAME), Universite Catholique de Louvain. Louvain-la-Neuve, Belgium.
Codrons, B., X. Bombois, M. Gevers and G. Scorletti (2000d). A practical application
of recent results in model and controller validation to a ferrosilicon production
process. Accepted for presentation at the 39th Conference on Decision and Con-
trol, Sydney, Australia, December 2000.
De Bruyne, F. (1996). Aspects on System Identication for Robust Process Control.
PhD thesis. Universite Catholique de Louvain. Louvain-la-Neuve, Belgium.
De Bruyne, F. and L. Kammer (1999). Iterative Feedback Tuning with guaranteed
stability. In: Proc. of the 1999 American Control Conference. San Diego, Cali-
fornia, USA. pp. 3317{3321.
De Bruyne, F., B.D.O. Anderson and N. Linard (1998). The Hansen scheme revisited.
In: Proc. of the 37th IEEE Conference on Decision and Control. Tampa, Florida,
USA. pp. 706{711.
Dennis, J.E. and R.B. Schnabel (1983). Numerical Methods for Unconstrained Opti-
mization and Nonlinear Equations. Prentice-Hall.
Enns, D. (1984a). Model Reduction for Control System Design. PhD thesis. Stanford
University. Stanford, California, USA.
Enns, D. (1984b). Model reduction with balanced realizations: an error bound and
a frequency weighted generalization. In: Proc. of the 23rd IEEE Conference on
Decision and Control. Las Vegas, Nevada, USA. pp. 127{132.
 Bibliography 213

Forssell, U. (1999). Closed-Loop Identication: Methods, Theory, and Applications.

PhD thesis. Linkoping University. Linkoping, Sweden.
Forssell, U. and L. Ljung (1999). Closed-loop identication revisited. Automatica
35, 1215{1241.
Francis, B.A. (1987). In: A Course in H1 Control Theory. Vol. 88 of Lecture Notes
in Control and Information Sciences. Springer-Verlag. London, UK.
Gangsaas, D., K.R. Bruce, J.D. Blight and U.-L. Ly (1986). Application of modern
synthesis to aircraft control: Three case studies. IEEE Trans. on Automatic
Control AC-31, 995{1104.
Georgiou, T.T. and M.C. Smith (1990). Optimal robustness in the gap metric. IEEE
Trans. on Automatic Control AC-35, 673{686.
Gevers, M. (1991). Connecting identication and robust control: a new challenge. In:
Proc. of the IFAC/IFORS Symposium on Identication and Parameter Estima-
tion. Budapest, Hungary. pp. 1{10.
Gevers, M. (1993). Towards a joint design of identication and control? In: Essays
on Control: Perspectives in the Theory and its Applications (H.L. Trentelman
and J.C. Willems, Eds.). pp. 111{151. Birkhauser. New York, USA.
Gevers, M. (1995). Identication for control. In: Proc. of the 5th IFAC Symposium
on Adaptive Control and Signal Processing. Budapest, Hungary. pp. 1{12.
Gevers, M. and L. Ljung (1986). Optimal experiment designs with respect to the
intended model application. Automatica 22, 543{554.
Gevers, M., B. Codrons and F. De Bruyne (1999a). Model validation in closed loop.
In: Proc. of the 1999 American Control Conference. San Diego, California, USA.
pp. 326{330.
Gevers, M., B.D.O. Anderson and B. Codrons (1998). Issues in modeling for control.
In: Proc. of the 1998 American Control Conference. Philadelphia, Pennsylvania,
USA. pp. 1615{1619.
Gevers, M., R.R. Bitmead and V. Blondel (1997). Unstable ones in understood alge-
braic problems of modelling for control design. Mathematical Modelling of Sys-
tems 3(1), 59{76.
Gevers, M., X. Bombois, B. Codrons, F. De Bruyne and G. Scorletti (1999b). The
role of experimental conditions in model validation for control. In: Robustness in
Identication and Control (A. Garulli, A. Tesi and A. Vicino, Eds.). Vol. 245 of
Lecture Notes in Control and Information Sciences. pp. 72{86. Springer-Verlag.
London, UK.
Gevers, M., X. Bombois, B. Codrons, F. De Bruyne and G. Scorletti (2000a). Model
validation for robust control and controller validation in a prediction error frame-
work. In: CD-ROM Proc. of the IFAC System Identication Symposium (SYSID
2000). Santa Barbara, California, USA. Paper WeAM1-1.
Gevers, M., X. Bombois, B. Codrons, G. Scorletti and B.D.O. Anderson (2000b).
Model validation for control and controller validation: a prediction error identi-
cation approach. Submitted to Automatica.
Glover, K. (1984). All optimal Hankel-norm approximations of linear multivariable
systems and their L1 error bounds.
Glover, K. and J.C. Doyle (1988). State-space formulae for all stabilizing controllers
that satisfy an H1 norm bound and relations to risk sensitivity. Systems and
Control Letters 11, 167{172.
214 Bibliography

Goddard, P.J. and K. Glover (1993). Controller reduction: Weights for stability and
performance preservation. In: Proc. of the 32nd IEEE Conference on Decision
and Control. San Antonio, Texas, USA. pp. 2903{2908.
Goddard, P.J. and K. Glover (1994). Performance-preserving frequency weighted con-
troller approximation: a coprime factorisation approach. In: Proc. of the 33rd
IEEE Conference on Decision and Control. Lake Buena Vista, Florida, USA.
pp. 2720{2725.
Gu, G. (1995). Model reduction with relative/multiplicative error bounds and relation
to controller reduction. IEEE Trans. on Automatic Control AC-40, 1478{1485.
Hansen, F.R. (1989). A Fractional Representation Approach to Closed Loop System
Identication and Experiment Design. PhD thesis. Stanford University. Stanford,
California, USA.
Hjalmarsson, H., M. Gevers and F. De Bruyne (1996). For model-based control design,
closed-loop identication gives better performance. Automatica 32, 1659{1673.
Hjalmarsson, H., M. Gevers, F. De Bruyne and J. Leblond (1994). Identication for
control: Closing the loop gives more accurate controllers. In: Proc. of the 33rd
IEEE Conference on Decision and Control. Orlando, Florida, USA. pp. 4150{
4155.
Hjalmarsson, H., S. Gunnarsson and M. Gevers (1995). Model-free tuning of a ro-
bust regulator for a exible transmission system. European Journal of Control
1(2), 148{156.
Ingason, H.T. and G.R. Jonsson (1998). Control of the silicon ratio in ferrosilicon
production. Control Engineering Practice 6, 1015{1020.
Kim, S.W., B.D.O. Anderson and A.G. Madievski (1995). Error bounds for transfer
function order reduction using frequency weighted balanced truncation. Systems
and Control Letters 24, 183{192.
Landau, I.D., A. Karimi, A. Voda and D. Rey (1995a). Robust digital control of exi-
ble transmissions using the pole placement/sensitivity function shaping method.
European Journal of Control 1(2), 122{133.
Landau, I.D., D. Rey, A. Karimi, A. Voda and A. Franco (1995b). A exible trans-
mission system as a benchmark for robust digital control. European Journal of
Control 1(2), 77{96.
Lee, W.S., B.D.O. Anderson, R.L. Kosut and I.M.Y. Mareels (1993). A new approach
to adaptive robust control. Int. Journal of Adaptive Control and Signal Process-
ing 7, 183{211.
Liu, K. and R.E. Skelton (1990). Closed-loop identication and iterative controller de-
sign. In: Proc. of the 29th IEEE Conference on Decision and Control. Honolulu,
Hawaii. pp. 482{487.
Liu, Y., B.D.O. Anderson and U. Ly (1990). Coprime factorization controller reduc-
tion with Bezout identity induced frequency weighting. Automatica 26(2), 233{
249.
Ljung, L. (1978). Convergence analysis of parametric identication methods. IEEE
Trans. on Automatic Control AC-23, 770{783.
Ljung, L. (1985). Asymptotic variance expressions for identied black-box transfer
function models. IEEE Trans. on Automatic Control AC-30(9), 834{844.
Ljung, L. (1995). System Identication Toolbox { User's Guide. The MathWorks, Inc.
Cochituate Place, Natick, Massachussetts, USA.
Ljung, L. (1997). Identication, model validation and control. Plenary lecture, 36th
IEEE Conference on Decision and Control, San Diego, California, USA.
 Bibliography 215

Ljung, L. (1998). Identication for control { what is there to learn? In: Workshop on
Learning, Control and Hybrid Systems. Bangalore, India.
Ljung, L. (1999). System Identication: Theory for the User. 2nd ed. Prentice-Hall.
Upper Saddle River, New Jersey, USA.
Ljung, L. and L. Guo (1997). The role of model validation for assessing the size of
the unmodelled dynamics. IEEE Trans. on Automatic Control AC-42(9), 1230{
1239.
Meyer, D.G. (1988). A fractional approach to model reduction. In: Proc. of the 1988
American Control Conference. pp. 1041{1047.
Meyer, D.G. (1990). Fractional balanced reduction: Model reduction by fractional
representation. IEEE Trans. on Automatic Control AC-35, 1341{1345.
Moore, B.C. (1981). Principal component analysis in linear systems: Controllability,
observability, and model reduction. IEEE Trans. on Automatic Control AC-
26(1), 17{32.
Ninness, B., H. Hjalmarsson and F. Gustafsson (1999). The fundemental role of gen-
eral orthonormal bases in system identication. IEEE Trans. on Automatic Con-
trol AC-44(7), 1384{1406.
Nordin, M. and P.O. Gutman (1995). Digital QFT design for the benchmark problem.
European Journal of Control 1(2), 97{103.
Obinata, G. and B.D.O. Anderson (2000). Model Reduction for Control System De-
sign. Book to be published.
Schrama, R.J.P. (1992a). Accurate identication for control: the necessity of an iter-
ative scheme. IEEE Trans. on Automatic Control AC-37, 991{994.
Schrama, R.J.P. (1992b). Approximate Identication and Control Design. PhD thesis.
Delft University of Technology. Delft, The Netherlands.
Skelton, R.E. (1989). Model error concepts in control design. Int. Journal of Control
49(5), 1725{1753.
Soderstrom, T. and P. Stoica (1989). System Identication. Prentice-Hall Interna-
tional. Hemel Hempstead, Hertfordshire, UK.
Tjarnstrom, F. and L. Ljung (1999). L2 model reduction and variance reduction. Tech-
nical report LiTH-ISY-R-2158, Department of Electrical Engineering, Linkoping
University, Linkoping, Sweden.
Van den Hof, P.M.J. and R.J.P. Schrama (1995). Identication and control { closed-
loop issues. Automatica 31, 1751{1770.
van Donkelaar, E.T. and P.M.J. Van den Hof (1997). Analysis of closed-loop iden-
tication with a tailor-made parametrization. In: CD-ROM Proc. of the 4th
European Control Conference. Brussels, Belgium. Paper TH-A-F-4 (853).
Varga, A. (1998). Computation of normalized coprime factorizations of rational ma-
trices. Systems and Control Letters 33, 37{45.
Vidyasagar, M. (1984). The graph metric for unstable plants and robustness estimates
for feedback stability. IEEE Trans. on Automatic Control AC-29, 403{417.
Vidyasagar, M. (1985). Control System Synthesis: A Factorization Approach. MIT
Press. Cambridge, Massachusetts, USA.
Vidyasagar, M. (1988). Normalized coprime factorizations for non-strictly proper sys-
tems. IEEE Trans. on Automatic Control AC-33(3), 300{301.
Vinnicombe, G. (1993a). Frequency domain uncertainty and the graph topology. IEEE
Trans. on Automatic Control AC-38, 1371{1383.
Vinnicombe, G. (1993b). Measuring the Robustness of Feedback Systems. PhD thesis.
Cambridge University. Cambridge, UK.
216 Bibliography

Vinnicombe, G. (1999). Uncertainty and Feedback (H1 loop-shaping and the  -gap
metric). World Scientic Publishing Co.
Wortelboer, P.M.R. (1994). Frequency-Weighted Balanced Reduction of Closed-Loop
Mechanical Servo-Systems: Theory and Tools. PhD thesis. Technische Univer-
siteit Delft. Delft, The Netherlands.
Wortelboer, P.M.R. and O.H. Bosgra (1994). Frequency weighted closed-loop order
reduction in the control design conguration.
Zang, Z., R.R. Bitmead and M. Gevers (1991). H2 iterative model renement and
controller enhancement. In: Proc. of the 30th IEEE Conference on Decision and
Control. Brighton, UK. pp. 279{284.
Zang, Z., R.R. Bitmead and M. Gevers (1995). Iterative weighted least squares iden-
tication and weighted LQG control design. Automatica 31, 1577{1594.
Zhou, K. (1995). A comparative study of H1 controller reduction methods. In: Proc.
of the American Control Conference. Seattle, Washington, USA. pp. 4015{4019.
Zhou, K. and J. Chen (1995). Performance bounds for coprime factor controller re-
duction. Systems and Control Letters 26, 119{127.
Zhou, K. and J. Doyle (1998). Essentials of Robust Control. Prentice-Hall. Upper
Saddle River, New Jersey, USA.
Zhu, Y.C. (1989). Black-box identication of MIMO transfer functions: Asymptotic
properties of prediction error models. Int. Journal of Adaptive Control and Signal
Processing 3, 357{373.
 Index
ARMAX model, 12 optimal, 1
ARX model, 12 reduction, 4, 38, 145
asymptotic variance, 15 closed-loop, 157
coprime-factor, 156, 157
balanced realisation, 19 criterion, 157
balanced truncation, 9, 16, 20, 147 open-loop, 156
closed-loop, 5, 160 to-be-designed, 37
frequency-weighted, 5, 21 two-degree-of-freedom, 29
bias, 2, 15, 38, 41, 146 unstable, 4, 51, 63, 65
Box-Jenkins model, 12 validation, 112, 123
for performance, 126
certainty equivalence principle, 1 for stability, 125
chordal distance, 26, 34, 123 coprime factorisation, 10, 30, 68
worst-case, 123 normalised, 30, 31, 147
closed loop, 24 Cramer-Rao bound, 196
balanced truncation, 5, 160 crossover frequency, 42, 45, 197
identication, 2, 5, 38
coprime-factor approach, 4, 54, 58, directed gap, 36
63
direct approach, 4, 54, 61, 63 EDF, 5, 162

Electricit
e de France, 5, 162
Hansen scheme, 54, 63, 68
indirect approach, 4, 53, 55, 63 experimental conditions, 38, 40
tailor-made parametrisation, 54, experiment design, 62
65 ferrosilicon, 134
residuals, 116 rst principle, 9
stability, 2, 4, 24, 26, 62 FIR model, 12
transfer matrix, 25 exible transmission system, 71, 126
validation, 4, 112, 115
direct approach, 119 gap metric, 35, 147
indirect approach, 115 Gauss-Newton, 14
consistency, 14 generalised predictive control, 135
controllability, 17 generalised stability margin, 9, 26
Gramian, 17 generic prediction error uncertainty set,
controller 122
adjustment, 43 GPC, 135
low-order, 5, 38 Gramian, 19
nonminimum phase, 4, 51, 63, 65 controllability, 17
one-degree-of-freedom, 29 observability, 18
217
218 Index

Hankel singular value, 19 set, 9, 10

Hansen scheme, 54, 63, 68 validated, 44, 115
validation, 3, 115, 196, 199
identication, 1, 9, 10, 115 closed-loop, 4, 115
and control design, iterative, 43 for control, 4, 111, 123
closed-loop, 2, 5, 38 open-loop, 112
coprime-factor approach, 4, 54, 58, modelling for control, 37
63 model error, 196
direct approach, 4, 54, 61, 63 model error model, 111, 121, 199
Hansen scheme, 54, 63, 68 multi-input multi-output, 10
indirect approach, 4, 53, 55, 63
tailor-made parametrisation, 54, nonminimum phaseness, 53
65
for control, 2, 37 observability, 18
instability, 53 Gramian, 18
internal stability, 26 order selection, 195
iterative output error model, 12
design, 3, 38, 43 overmodelling, 195, 196
scheme, 3
parameter
Kalman lter, 165 estimate, 12, 15
vector, 11
LFT, 9, 28 Youla, 68
linear fractional transformation, 9, 28 performance, 3
linear quadratic Gaussian, 164 criterion, 126, 149
linear time-invariant, 10 robustness, 40
loop shaping lter, 73, 165 worst-case, 126
LQG, 164 pole, 52, 196
LTI, 10 blocking, 53
unstable, 4, 53, 54
MIMO, 10 zero cancellation, 197
minimality, 52 prediction, 11
model one-step ahead, 11
ARMAX, 12 prediction error, 9, 11, 113
ARX, 12 identication, 3, 9, 10, 115
Box-Jenkins, 12 uncertainty set, 122
error, 3, 46, 114 pressurised water reactor, 5, 162
FIR, 12 PWR, 5, 162
high-order, 4, 9, 16, 38
low-order, 9, 195 reduction
order selection, 195 L2, 202
output error, 12 controller, 4, 38, 145
parametrised, 10, 41 closed-loop, 157
reduction, 4, 9, 38 coprime-factor, 156, 157
L2, 202 open-loop, 156
closed-loop, 149 criterion, 149, 157
control-oriented, 145 model, 9, 38
coprime-factor, 147, 149 closed-loop, 149
criterion, 149 control-oriented, 4, 145
open-loop, 147 coprime-factor, 147, 149
 Index 219

open-loop, 147 winding number, 34, 35

residuals, 3, 46, 113 windsurfer approach, 43
closed-loop, 116
Riemann sphere, 35 Youla parametrisation, 63, 68
robustness zero, 53, 196
performance, 40 blocking, 53
stability, 40 nonminimum phase, 4, 53, 54
robust control design, 2, 63 transmission, 53
robust stabilisation
necessary condition, 125  -gap, 10, 34, 64, 123, 147
suÆcient conditions, 124 worst-case, 111, 123
sensitivity function, 41, 197
single-input single-output, 12
SISO, 12
stability
closed-loop, 2, 4, 24, 26, 62
generalised stability margin, 9, 26
internal, 26
nominal, 54, 62
radius, 126
robustness, 40
system
closed-loop, 24
LTI, 10
MIMO, 10
nonminimum phase, 53
true, 10
unstable, 53
tailor-made parametrisation, 54, 65
uncertainty, 2, 38
region, 46, 65, 111, 117, 120, 122,
199
generic, 121
in parameter space, 114
in transfer function space, 114, 196
undermodelling, 11
validation, 46, 115
closed-loop, 4, 112
controller, 112, 123, 125, 126
for control, 4, 123
model, 3, 4, 111, 123, 196, 199
closed-loop, 4, 115
open-loop, 112
open-loop, 112
variance, 2, 15, 38, 44, 115, 196