Link oping Studies in Science and Technology.

Dissertations
No. 716
Residual Generation for Fault
Diagnosis
Erik Frisk
Department of Electrical Engineering
Link oping University, SE581 83 Link oping, Sweden
Link oping 2001
Residual Generation for Fault Diagnosis
c _ 2001 Erik Frisk
[email protected]
http://www.fs.isy.liu.se
Department of Electrical Engineering,
Link oping University,
SE581 83 Link oping,
Sweden.
ISBN 91-7373-115-3 ISSN 0345-7524
Printed by Linus & Linnea AB, Linkoping, Sweden 2001
i
Abstract
The objective when supervising technical processes is to alarm an operator
when a fault is detected and also identify one, or possibly a set of components,
that may have been the cause of the alarm. Diagnosis is an expansive subject,
partly due to the fact that nowadays, more applications have more embedded
computing power and more available sensors than before.
A fundamental part of many model-based diagnosis algorithms are so called
residuals. A residual is a signal that reacts to a carefully chosen subset of
the considered faults and by generating a suitable set of such residuals, fault
detection and isolation can be achieved.
A common thread is the development of systematic design and analysis
methods for residual generators based on a number of dierent model classes,
namely deterministic and stochastic linear models on state-space, descriptor,
or transfer function form, and non-linear polynomial systems. In addition, it
is considered important that there exist readily available computer tools for all
design algorithms.
A key result is the minimal polynomial basis algorithm that is used to pa-
rameterize all possible residual generators for linear model descriptions. It also,
explicitly, nds those solutions of minimal order. The design process and its
numerical properties are shown to be sound. The algorithms and its principles
are extended to descriptor systems, stochastic systems, nonlinear polynomial
systems, and uncertain linear systems. New results from these extensions in-
clude: increased robustness by introduction of a reference model, a new type
of whitening lters for residual generation for stochastic systems both on state-
space form and descriptor form, and means to handle algorithmic complexity
for the non-linear design problem.
In conclusion, for the four classes of models studied, new methods have
been developed. The methods fullls requirements generation of all possible
solutions, availability of computational tools, and numerical soundness. The
methods also provide the diagnosis system designer with a set of tools with well
specied and intuitive design freedom.
ii
iii
Acknowledgments
This work was performed at the department of Electrical Engineering, division
of Vehicular Systems, Link oping University. I would like to thank my professor
and supervisor, Lars Nielsen for letting me join his group and for our research
discussions. I would also like to thank the sta at Vehicular Systems for creating
a positive atmosphere.
Foremost, I would like to thank my colleague Mattias Nyberg for all our
discussions, collaborative work, and for proofreading the manuscript. Dr. Didier
Henrion is gratefully acknowledged for patiently answering my questions and
writing some of the Matlab code used in this work. Marcus Klein, Mattias
Krysander, and Niten Olofsson are also acknowledged for proofreading early
versions of the manuscript.
This work has been supported by the Swedish Foundation for Strategic Re-
search through the graduate school ECSEL (Excellence Center in Computer
Science and Systems Engineering in Link oping).
Link oping, October 2001
Erik Frisk
iv
Contents
1 Introduction 1
1.1 Outline and contributions of the thesis . . . . . . . . . . . . . . . 3
1.2 Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2 Model Based Diagnosis 7
2.1 Introduction to model based diagnosis . . . . . . . . . . . . . . . 7
2.1.1 Fault modeling . . . . . . . . . . . . . . . . . . . . . . . . 8
2.1.2 Residuals and residual generators . . . . . . . . . . . . . . 10
2.1.3 Fault isolation . . . . . . . . . . . . . . . . . . . . . . . . 11
2.1.4 Model based diagnosis using residuals . . . . . . . . . . . 13
2.2 Consistency relations and residual generator implementation . . 14
2.3 Problem motivation and discussion . . . . . . . . . . . . . . . . . 16
2.3.1 The linear problem . . . . . . . . . . . . . . . . . . . . . . 17
2.3.2 The non-linear problem . . . . . . . . . . . . . . . . . . . 19
3 Residual Generation Based on Linear Models 21
3.1 The minimal polynomial basis approach . . . . . . . . . . . . . . 22
3.1.1 A general problem formulation . . . . . . . . . . . . . . . 22
3.1.2 Derivation of design methodology . . . . . . . . . . . . . . 23
3.2 Methods to nd a minimal polynomial basis for N
L
(M(s)) . . . . 25
3.2.1 Frequency domain solution . . . . . . . . . . . . . . . . . 25
3.2.2 State-space solution . . . . . . . . . . . . . . . . . . . . . 25
3.2.3 No disturbance case . . . . . . . . . . . . . . . . . . . . . 28
v
vi
3.2.4 Finding a minimal polynomial basis for the null-space of
a general polynomial matrix . . . . . . . . . . . . . . . . . 29
3.3 Matlab sessions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.4 Bounds on maximum and minimum row-degree of the basis . . . 30
3.4.1 Upper bound for the maximum row-degree of the basis . . 31
3.4.2 Bounds for the minimal row-degree of the basis . . . . . . 34
3.5 Relation to other residual generator design methods . . . . . . . 36
3.5.1 The parity-space approach . . . . . . . . . . . . . . . . . . 36
3.5.2 Frequency domain approaches . . . . . . . . . . . . . . . . 37
3.6 Design examples with nominal models . . . . . . . . . . . . . . . 38
3.6.1 Design example 1: Aircraft dynamics . . . . . . . . . . . . 38
3.6.2 Design example 2: Turbo-jet engine . . . . . . . . . . . . 44
3.7 Descriptor systems . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.7.1 Computing N
L
(M(s)) for descriptor systems . . . . . . . 48
3.7.2 Design example . . . . . . . . . . . . . . . . . . . . . . . . 50
3.7.3 Non-zero initial states . . . . . . . . . . . . . . . . . . . . 53
3.7.4 Links to observer design . . . . . . . . . . . . . . . . . . . 55
3.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.A Standard notions from linear systems theory . . . . . . . . . . . . 57
3.B Standard results from linear systems theory . . . . . . . . . . . . 59
3.C State-space matrices for descriptor example . . . . . . . . . . . . 63
4 Residual Generation Based on Non-Linear Models 65
4.1 Problem formulation . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.1.1 Elementary functions as polynomials . . . . . . . . . . . . 66
4.2 Basic elimination theory . . . . . . . . . . . . . . . . . . . . . . . 67
4.3 Design using elimination theory . . . . . . . . . . . . . . . . . . . 69
4.3.1 Algorithm outline . . . . . . . . . . . . . . . . . . . . . . 70
4.4 Realizable residual generator . . . . . . . . . . . . . . . . . . . . 72
4.5 Isolability analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.6 Complexity management . . . . . . . . . . . . . . . . . . . . . . . 76
4.7 Simulation example: Coupled water tanks . . . . . . . . . . . . . 77
4.7.1 Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4.7.2 Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.7.3 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5 Residual Generation Based on Stochastic Linear Models 87
5.1 Problem formulation . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.2 Spectral factorization theory . . . . . . . . . . . . . . . . . . . . 92
5.2.1 Note on the singular case . . . . . . . . . . . . . . . . . . 93
5.3 Introductory examples . . . . . . . . . . . . . . . . . . . . . . . . 94
5.3.1 Example 1: Successful design . . . . . . . . . . . . . . . . 94
5.3.2 Example 2: Zeros on the imaginary axis . . . . . . . . . . 95
5.3.3 Example 3: Innite zeros . . . . . . . . . . . . . . . . . . 95
vii
5.4 Design algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
5.4.1 Design of innovation lters . . . . . . . . . . . . . . . . . 96
5.4.2 Design of whitening residual generators . . . . . . . . . . 99
5.5 Design examples . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
5.5.1 Design example: Aircraft dynamics . . . . . . . . . . . . . 106
5.5.2 Example with purely imaginary zeros . . . . . . . . . . . 112
5.6 White residuals . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
5.6.1 Whiteness tests . . . . . . . . . . . . . . . . . . . . . . . . 113
5.6.2 Simulations and comparisons . . . . . . . . . . . . . . . . 114
5.6.3 Simulation conclusions . . . . . . . . . . . . . . . . . . . . 115
5.7 Time-discrete systems . . . . . . . . . . . . . . . . . . . . . . . . 120
5.8 The singular case . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
5.8.1 Singular complications . . . . . . . . . . . . . . . . . . . . 121
5.8.2 When does the non-singular case occur? . . . . . . . . . . 123
5.9 Stochastic descriptor example . . . . . . . . . . . . . . . . . . . . 124
5.10 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
5.A innovationlter.m . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
6 Residual Generation Based on Uncertain Linear Models 129
6.1 Robust residual generation . . . . . . . . . . . . . . . . . . . . . . 130
6.2 Reference model . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
6.3 Computation of a robust residual generator . . . . . . . . . . . . 132
6.3.1 Robustness criterion . . . . . . . . . . . . . . . . . . . . . 132
6.3.2 Computational framework . . . . . . . . . . . . . . . . . . 133
6.4 Background example on reference model design . . . . . . . . . . 138
6.5 Forming the reference model . . . . . . . . . . . . . . . . . . . . . 140
6.5.1 Nominal design . . . . . . . . . . . . . . . . . . . . . . . . 141
6.5.2 Discussion of design choices . . . . . . . . . . . . . . . . . 143
6.6 Summary of design method . . . . . . . . . . . . . . . . . . . . . 144
6.7 Illustrative dynamic example . . . . . . . . . . . . . . . . . . . . 145
6.7.1 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
6.7.2 Residual generator specications . . . . . . . . . . . . . . 146
6.7.3 Robustness comparison . . . . . . . . . . . . . . . . . . . 146
6.8 Robust fault detection . . . . . . . . . . . . . . . . . . . . . . . . 147
6.9 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
7 Conclusions 151
Notation 155
References 157
Author Index 165
Index 169
1
Introduction
Modern processes use more and more embedded computers and sensors to,
among other things, increase performance and introduce new functionality. At
the same time, the sensors combined with on-line computing power provide
means for on-line supervision of the process itself. In such more autonomous
operation, it is of important to detect faults before the fault seriously aects
system performance. Faults in a control loop are particularly important since
feedback from a faulty sensor very quickly can result in instability causing a
complete failure of the plant. Such faults might need to be detected within
a few samples (Blanke et al., 1997). Therefore it is important that faults are
detected during normal operation of the plant, without the need to perform
certain tests to perform the diagnosis.
Diagnosis
System
observations
Known variables
Process
(measurements,
Diagnosis
controller outputs,...)
Figure 1.1: Diagnosis application
Here, the word diagnosis means detection and location (isolation) of a faulty
component. A general structure of a technical diagnosis application is shown
in Figure 1.1, where the diagnosis system is fed all available knowledge (also
called observations) of the process. Such knowledge include measured variables,
1
2 Chapter 1. Introduction
controller outputs and any other information that is available to the diagnosis
system. The diagnosis system processes the observations and produce a diagno-
sis, which is a list of possible faults aecting the process. Often the process is
regulated by a controller and the known variables consist of controller outputs
and sensor data. Such a situation is depicted in Figure 1.2 which also illustrates
a fundamental complication the diagnosis system designer faces. Disturbances,
also called unknown inputs, not considered faults also inuence the process. The
diagnosis system must thereby separate the inuence caused by these unknown
inputs and the faults.
Control
System
Control Signals
-
Process
?
Faults
-
Disturbances
6
Measurements

-
Diagnosis
System
?
Diagnosis
Figure 1.2: Control oriented diagnosis application.
To detect and isolate faulty components, some sort of redundancy is needed.
The redundancy is used to make consistency checks between related variables.
In applications with very high security demands such as aircraft control-systems,
redundancy can be supplied by extra hardware, hardware redundancy. A criti-
cal component, for example a sensor, is then duplicated or triplicated and voting
schemes can be used to monitor signal levels and trends to detect and locate
faulty sensors. Hardware redundancy has the advantage of being reliable and
gives high performance, but the approach has drawbacks, e.g. extra hardware
costs, space and weight consideration, and some components can not be dupli-
cated.
Instead of using hardware redundancy, analytical redundancy can be used
where the redundancy is supplied by a process model instead of extra hardware.
Then the process can be validated by comparing a measured variable with an
1.1. Outline and contributions of the thesis 3
estimate, produced using the process model, of the same variable. Diagnosis
systems based on analytical redundancy are also called model based diagnosis
systems which is further described in Chapter 2.
1.1 Outline and contributions of the thesis
Often, a fundamental part of model based diagnosis systems is a residual gen-
erator. A residual is a computable quantity that is used to alarm if a fault is
present in the supervised process or not. They can also, if designed properly,
provide means for isolation of the faults. How to design residual generators for
dierent model descriptions is the topic of this thesis.
Chapter 2 gives an overview of the model-based diagnosis problem. First,
the diagnosis problem is dened and it is showed how residual generators t
into a complete supervision system, performing both fault detection and fault
isolation. Then, residual generator design based on consistency relations
1
is de-
scribed for linear and non-linear systems. Finally, the problems studied further
in the chapters to follow are indicated and motivated.
In Chapter 3, residual generation for deterministic linear systems is pre-
sented. A key contribution is the minimal polynomial basis approach to resid-
ual generation. A main property of the algorithm is that it can, in a straight-
forward and numerically sound way, utilize models on transfer function form,
state-space form, and also applies to a more general class of linear systems de-
scribed by dierential-algebraic equations, descriptor systems. The algorithm is
thoroughly exemplied on a linearized aircraft model to show basic properties
of the algorithm. A large, 24 state model of a jet engine is also included to show
numerical properties of the approach.The theoretical parts are are mainly based
on (Frisk and Nyberg, 2001) and the examples on (Frisk and Nyberg, 1999).
Chapter 4 continues with a derivation of a design procedure for non-linear
systems, mainly based on (Frisk, 2000b). The algorithm has considerable sim-
ilarities with the linear design procedure, and free design variables in the non-
linear case has direct counterparts in the linear case. The algorithm has strong
computational support in modern computer algebra systems like Mathematica
and Maple. A major concern is the computational complexity of the design
algorithm and it is shown how structural analysis of the model equations can
be used to manage the complexity.
The linear design problem is revisited in Chapter 5, where stochastic linear
model descriptions is considered. To systematically select the free design vari-
ables available after a deterministic design, additional modeling and additional
constraints on the residual generators are needed which reduces the available
design freedom. The design algorithm from Chapter 3, with all its merits on
simplicity and numerical stability, is extended to the stochastic design problem.
Finally, the design procedure is exemplied on both state-space and descriptor
systems. This chapter is mainly based on (Frisk, 2001).
1
Other commonly used words are parity or analytical redundancy relations
4 Chapter 1. Introduction
Chapter 6 provides an approach to make the residual generator as robust as
possible to parametric uncertainties in the model description. An optimization
procedure, based on H

-ltering theory, is used and a main contribution is the

systematic procedure to form the the optimization criterion. A key observation
is how a, at rst sight natural, criterion can result in unnecessary poor perfor-
mance of the residual generator. A systematic procedure, based on the nominal
design problem, to form a feasible optimization criterion to synthesize residual
generators is then outlined. The main objective with the procedure is to utilize
the design freedom as much as possible to make the residual optimally robust
against parametric uncertainties. The algorithm is mainly based on (Frisk and
Nielsen, 1999).
Finally, Chapter 7 provides the conclusions.
1.2 Publications
In the research work, leading to this thesis, the author has published the fol-
lowing conference and journal papers (in reversed chronological order):
E. Frisk (2001). Residual generation in linear stochastic systems - a poly-
nomial approach. To appear in proc. IEEE Conf. on Decision and Con-
trol, Orlando, USA.
E. Frisk and M. Nyberg (2001). A minimal polynomial basis solution
to residual generation for fault diagnosis in linear systems. Automatica
37(9), September, pp. 14171424.
I. Andersson and E. Frisk (2001). Diagnosis of evaporative leaks and
sensor faults in a vehicle fuel system. In proc. IFAC Workshop: advances
in automotive control, Karlsruhe, Germany.
E. Frisk (2000a). Order of residual generators - bounds and algorithms.
In proc. IFAC Safeprocess, Budapest, Hungary, pp. 599604.
E. Frisk (2000b). Residual generator design for non-linear, polynomial
systems a Grobner basis approach. In proc. IFAC Safeprocess, Bu-
dapest, Hungary, pp. 979984.
E. Frisk and M. Nyberg (1999). Using minimal polynomial bases for fault
diagnosis. In proc. European Control Conference, Karlsruhe, Germany.
2
E. Frisk and L. Nielsen (1999). Robust residual generation for diagnosis
including a reference model for residual behavior. In proc. IFAC World
Congress, Beijing, P.R. China, Vol. P, pp. 5560.
2
Awarded the Polyx prize for best paper on polynomial methods in 1999.
1.2. Publications 5
M. Nyberg and E. Frisk (1999). A minimal polynomial basis solution to
residual generation for fault diagnosis in linear systems. In proc. IFAC
World Congress, Beijing, P.R. China, Vol. P, pp. 6166.
E. Frisk, M. Nyberg, and L. Nielsen (1997). FDI with adaptive residual
generation applied to a DC-servo. In proc. IFAC Safeprocess, Hull, United
Kingdom, pp. 438442.
6 Chapter 1. Introduction
2
Model Based Diagnosis
The aim of this chapter is twofold, rst to give a brief introduction to the eld
of model based diagnosis and second to introduce problem formulations and
notation. It is intended to form a, both notational and conceptual, basis for the
chapters to follow, not to give a complete view of the eld.
Section 2.1 provides an introduction to model based diagnosis so that the
subsections of Section 2.1 give brief presentations of the concepts and subsys-
tems of residual-based diagnosis systems. This is done by rst introducing an
important concept, analytical redundancy. Then, in Subsection 2.1.1 fault mod-
els are described. The introductory presentation proceeds by discussing topics
of central importance in this work, residuals and residual generators, in Sub-
section 2.1.2. The basics of fault isolation is described in Subsection 2.1.3 and
nally, Subsection 2.1.4 describes how residual generators t into a complete
model-based diagnosis system. The introduction so far does not mention how
to design and implement the residual generators. This topic is approached in
Section 2.2 by exploring consistency relations, a concept that is central in the
chapters to follow. Finally, in Section 2.3 problems studied in the remaining
parts of the dissertation is discussed and motivated.
2.1 Introduction to model based diagnosis
Model based diagnosis methods has been developed for many model domains,
e.g. models from the AI-eld which are often logic based (Hamscher et al.,
1992), or Discrete Event Dynamic Systems for which automata descriptions are
common (Larsson, 1999; Sampath et al., 1995, 1996). A third model domain
7
8 Chapter 2. Model Based Diagnosis
that is commonly considered are models typically found in the eld of signals
and systems, i.e. models involving continuous variables in continuous or discrete
time. Typical model formulations are dierential/dierence equations, transfer
functions, and/or static relations. From now on in this work, only models from
this domain are used.
In Chapter 1 it was discussed how model-based diagnosis is used when re-
dundancy is supplied by a model instead of additional hardware. Redundancy
supplied by a model is called analytical redundancy and can be dened more
formally as:
Denition 2.1 (Analytical Redundancy). There exists analytical redun-
dancy if there exists two or more dierent ways to determine a variable x
by only using the observations z(t), i.e. x = f
1
(z(t)) and x = f
2
(z(t)), and
f
1
(z(t)) , f
2
(z(t)).
Thus, the existence of analytical redundancy makes it possible to check the
validity of the assumptions made to ensure that f
1
(z(t)) = f
2
(z(t)).
Example 2.1
Assume two sensors measure the variable x according to
y
1
=

x y
2
= x
The integrity of the two sensors can then be validated by ensuring that the
relation, represented by the equation y
2
1
y
2
= 0, holds.
In Example 2.1 it was easy to see that a malfunction in any of the two sensors
would invalidate the relation and a fault could be detected. In more general
cases, and to facilitate not only fault detection but also fault isolation, there is
a need to describe fault inuence on the process more formally, i.e. fault models
of some sort is needed.
2.1.1 Fault modeling
A fault model is a formal representation of the knowledge of possible faults and
how they inuence the process. More specic, the term fault means that com-
ponent behavior has deviated from its normal behavior. It does not mean that
the component has stopped working altogether. The situation where a com-
ponent has stopped working is, in the diagnosis community, called a failure.
So, one goal is to detect faults before they cause a failure. In general, utilizing
better fault models (assuming good and valid fault models) implies better diag-
nosis performance, i.e. smaller faults can be detected and more dierent types
of faults can be isolated. Here fault modeling is illustrated using an example.
For more elaborate discussions on fault modeling the reader is referred to e.g.
(Nyberg, 1999b; Gertler, 1998; Chen and Patton, 1999).
A common fault model is to model faults as deviations of constant parame-
ters from their nominal value. Typical faults that are modeled in this way are
2.1. Introduction to model based diagnosis 9
gain-errors and biases in sensors, process faults modeled as a deviation of
a physical parameter. In cases with constant parameter fault models, methods
and theory from parameter estimation have shown useful also for fault diag-
nosis, see for example (Isermann, 1993). However, other more elaborate fault
models exists e.g. fault models that utilizes the change-time characteristic of
the process(Basseville and Nikiforov, 1993).
The fault models used in the chapters to follow are typically time-varying
fault signals or constant parameter changes. An advantage with using fault-
signals when modeling faults is the simplicity and relatively few assumptions
made in modeling. A disadvantage with such a general fault model is that
fault isolability may be lost compared to more detailed fault models. A small
example is now included to illustrate fault modeling principles.
Example 2.2
A nonlinear state-space description including fault models can be written
x = g(x, u, f)
y = h(x, u, f)
where x, u and y are the state, control signals, and measurements respectively.
The signal f represents the fault, which in the fault-free case is f 0 and
non-zero in a faulty case. The signal f here represents an arbitrary fault that
can for example be a fault in an actuator or a sensor fault.
To illustrate fault modeling more concretely, consider a small idealized rst
principles model of the arm of an industrial robot. Linearized dynamics around
one axis can be described by equations looking something like the following
equations:
J
m

m
= F
v,m

m
+k
T
u +
spring
(2.1a)

spring
= k(
a

m
) +c(
a

m
) (2.1b)
J
a

a
=
spring
(2.1c)
y =
m
(2.1d)
where the model variables are:
Symbol Description
J
m
moment of inertia: motor
J
a
moment of inertia: arm

m
motor position

a
arm position
F
v,m
viscous friction, motor
k stiness coecient, gear box
c damping coecient, gear box
k
T
torque constant, is 1 when torque control-loop is working
u torque reference value, fed to the torque controller
10 Chapter 2. Model Based Diagnosis
Now, fault models are illustrated by modeling the following faults
1. A faulty torque-controller
2. Faulty arm position sensor, resulting in increased signal to noise ratio in
the sensor signal
3. The robot has a load attached to the tip of the robot arm which is dropped
4. Collision of the robot arm with the environment
Associate a fault-variable f
1
to f
4
with the faults above. Introducing fault
models in (2.1) gives for example
J
m

m
= F
v,m

m
+ (k
T
+f
1
(t))u +
spring

spring
= k(
a

m
) +c(
a

m
)
(J
a
+f
3
)
a
=
spring
+f
4
(t)
y =
m
+(f
2
)
(f
2
) =
_
N(0,
2
1
) f
2
= 0, fault-free case
N(0,
2
1
+f
2
2
) f
2
,= 0, faulty sensor

f
2
= 0

f
3
= 0
Here it is seen that faults f
2
and f
3
are assumed constant. Faults f
1
and f
4
is
however not assumed constant. Such an assumption for f
4
would of course lead
to a highly unrealistic fault model since f
4
is the torque exercised on the robot
arm by the environment which naturally would not be constant in a collision
situation. The time variability assumption of f
1
and f
4
is emphasized in the
model by adding explicit time dependence.
2.1.2 Residuals and residual generators
The second step in the introductory presentation of model based diagnosis is
a presentation of residuals and residual generators. Residuals is often a funda-
mental component in a diagnosis system. A residual is an, often time-varying,
signal that is used as a fault detector. Normally, the residual is designed to
be zero (or small in a realistic case where the process is subjected to noise and
the model is uncertain) in the fault-free case and deviate signicantly from zero
when a fault occurs. Note however that other cases exist. In case of a likelihood-
function based residual generator where the residual indicates how likely it is
that the observed data is generated by a fault-free process, the residual is large
in the fault-free case and small in a faulty case. But for the remainder of this
text it is assumed, without loss of generality, that a residual is 0 in the fault-free
case.
A residual generator is a lter that lters known signals to produce the
residual. A linear residual generator is thus a proper MISO (Multiple Input
2.1. Introduction to model based diagnosis 11
Single Output) lter Q(s), ltering known signals y and u (measurements and
control signals) producing an output r
r = Q(s)
_
y
u
_
Introduction to linear residual generator design is given in Section 2.2.
A more general non-linear residual generator on state-space form is given by
two non-linear functions g and h and the lter
z = g(z, y, u)
r = h(z, y, u)
A main diculty when designing residual generators is to achieve the distur-
bance decoupling, i.e. to ensure that the residual r is not inuenced by unknown
inputs that is not considered faults. This is was illustrated by Figure 1.2.
The main topic of the chapters to follow is procedures to design and analyze
residual generators, i.e. the transfer function Q(s) in the linear case and the
non-linear functions g and h in the non-linear case.
2.1.3 Fault isolation
Before it is described how residuals and residual generators t into a diagnosis
system in Subsection 2.1.4, basic fault isolation strategies is described. Since
fault isolation is not the topic of this thesis, this section illustrates fault isolation
mainly by example.
To achieve isolation, several principles exists. For methods originating from
the area of automatic control, at least three dierent approaches can be distin-
guished: xed direction residuals, structured residuals, and structured hypothesis
tests.
The idea of xed direction residuals (Beard, 1971) is to design a residual
vector such that the residual responds in dierent directions depending on what
fault that acts on the system. Fault isolation is then achieved by studying and
classifying the direction of the residual. This approach has not been so much
used in the literature, probably because the problems associated with designing
a residual vector with desired properties.
The idea of structured residuals (Gertler, 1991) is to have a set of residuals,
in which each individual residual is sensitive to a subset of faults. By study-
ing which residuals that respond, fault isolation can be achieved. Structured
residuals have been widely used in the literature, in both theoretical and prac-
tical studies. The basic idea is quite simple and many methods for constructing
suitable residuals have been presented both for linear and non-linear systems.
As a generalization of structured residuals, structured hypothesis tests has
been proposed where the isolation method is formally dened. This formal
denition makes it possible to use any possible fault models (Nyberg, 1999a),
12 Chapter 2. Model Based Diagnosis
I f
1
f
2
f
3
r
1
1 1 0
r
2
X 0 1
r
3
1 1 1
II f
1
f
2
f
3
r
1
1 1 0
r
2
1 0 1
r
3
1 1 1
III f
1
f
2
f
3
r
1
0 X X
r
2
X 0 X
r
3
X X 0
IV f
1
f
2
f
3
r
1
1 0 0
r
2
0 1 0
r
3
0 0 1
Figure 2.1: Examples of inuence structures.
and perform deeper and further analysis of isolation properties of fault diagno-
sis systems. However, isolation issues are addressed very briey here and the
isolation procedure is mainly illustrated by example.
Residuals, as described in Section 2.1.2, can not only be used for fault detec-
tion, they can also be used for fault isolation in a structured residual/hypothesis-
test isolation framework. The following example illustrates briey how the iso-
lation procedure works.
Example 2.3
To achieve isolation, in addition to fault detection, a set of residuals need to
be designed where dierent residuals are sensitive to dierent subsets of faults.
Which residuals that are sensitive to what faults, can be described by the inu-
ence structure
1
. Four examples of inuence structures are shown in Figure 2.1.
A number 1 in the i:th row and the j:th column represents the fact that resid-
ual r
i
is sensitive to fault j. A number 0 in the i:th row and the j:th column
represents the fact that residual r
i
is not sensitive to fault j. An X in the i:th
row and the j:th column represents the fact that residual r
i
is sometimes sen-
sitive to fault j. For example in structure I, it can be seen that residual r
2
is
sometimes sensitive to fault f
1
, not sensitive to fault f
2
, and always sensitive
to fault f
3
. The isolation can ideally be performed by matching fault columns
to the actual values of the residuals. Consider for example inuence structure
II in Figure 2.1, and assume that residuals r
1
and r
3
have signaled, but not r
2
.
The conclusion is then that fault f
2
has occurred.
In light of this illustration, it is convenient to introduce some notation. Consider
residual r
2
in inuence structure I which is completely insensitive to fault f
2
and
sensitive to faults f
1
, f
3
, i.e. two sets of faults are considered in each residual
generator design. The faults that the residual should be sensitive to are called
monitored faults and the faults the residual should be insensitive to are called
1
Also the method structured residuals uses inuence structures but under dierent names.
Names that have been used are for example incidence matrix(Gertler and Singer, 1990),resid-
ual structure(Gertler, 1998), and coding set(Gertler, 1991).
2.1. Introduction to model based diagnosis 13
non-monitored faults. The non-monitored faults are said to be decoupled in
the residual. Thus, the residual generator design problem is, in a structured
residuals/hypothesis-test framework, essentially a decoupling problem.
It is worth noting that in general, the more faults that are decoupled in each
residual, the greater is the possibility to isolate multiple faults. It is for example
easy to see that inuence structure IV facilitates isolation of 3 multiple faults
while inuence structure III only can handle single faults. The price to pay
for this increased isolation performance is that more sensors are needed and the
residual generators become more complex and model-dependent. These issues,
among others, are explored in detail in the chapters to follow.
2.1.4 Model based diagnosis using residuals
This section describes how residuals is used in a structured residuals based
diagnosis system. To be able to perform the fault isolation task, the residuals
must react to faults according to an isolating inuence structure. Thus, a design
procedure would follow a procedure looking something like
1. Select a desired isolating inuence structure. See (Gertler, 1998) for de-
tails on how, for example desired isolability properties restricts possible
inuence structures.
2. For each residual, collect unknown inputs and non-monitored faults, i.e.
faults corresponding to zeros in the current row of the inuence structure,
in a vector d. The rest of the faults, the monitored faults, are collected in
a vector f.
3. Design a residual that decouples d and verify what faults the residual is
sensitive to. Ideally it is sensitive to all monitored faults, but it is possible
that when decoupling d in the residual, also some of the monitored faults
are decoupled.
4. If, when all residuals are designed, the resulting inuence structure does
not comply with design specications, return to step 1 and re-design the
desired inuence structure.
It is clear from the procedure above that assuming a fault to be non-monitored
is equal to introducing a zero at the corresponding location in the inuence
structure. Thus, by moving faults between monitored and non-monitored faults,
the inuence structure becomes a design choice made by the designer. Note
that, for example the number of sensors and structural properties of the model
both restricts the available design freedom when forming the inuence structure.
Thus, the inuence structure is not entirely free for the designer to choose.
14 Chapter 2. Model Based Diagnosis
2.2 Consistency relations and residual genera-
tor implementation
A consistency relation is any relation between known or measured variables
that, in the fault free case always holds. This section is intended to provide a
background on consistency relations, how they can be used to form a residual
generator and also indicate fundamental dierences between how linear and
non-linear consistency relations can be used for implementation. Consistency
relations is not the only term used in fault diagnosis literature. Previously the
words parity relations and parity equations was most common, but lately other
words have appeared e.g. analytical redundancy relations (ARR) (Staroswiecki
and Comtet-Varga, 2001). Here, the word consistency relation is used.
A consistency relation is an analytical relation between known signals z (and
higher order derivatives) that is satised in the fault-free case. In case that the
known signals are measurements and control signals then the known signals is
the vector z = [y u]. Thus, g is a consistency relation if the following holds for
all z that satisfy the original system equations (the model) when f 0:
g(z, z, z, . . . ) = 0 (2.2)
In case of a time-discrete system, the time derivatives are substituted for time-
delays. For time-continuous linear systems, a consistency relation can always
(in the frequency domain) be written as
F(s)z = 0
where F(s) is a polynomial vector (or matrix if multidimensional consistency
relations are considered) in s. Such linear consistency relations are studied in
detail in Chapter 3. Note that this holds only if all initial states are zero. Details
on consistency relations for the case of non-zero initial states are discussed in
Section 3.7.3.
Now, clearly these consistency relations are interesting for fault diagnosis
since they describe a relation that is satised in the fault-free case and (possibly)
not satised in case of a fault. If all variables included in the consistency relation
(2.2) are known, a residual could be generated by letting
r = g(z, z, z, . . . )
For dynamic systems, the relation g in general contains time-dierentiated mea-
surements and control signals, i.e. u and y. Since these are not normally known,
it is usually not possible to use the consistency relation directly in an implemen-
tation of a residual generator. In the linear case, this implementation problem
is easily circumvented which is illustrated by the next example and described
in detail in Chapter 3.
2.2. Consistency relations and residual generator implementation 15
Example 2.4
Consider the linear model
y =
1
s
2
+as +b
u +f
The time domain interpretation of the transfer function is (with zero initial
conditions):
y +a y +by u

f a

f bf = 0 (2.3)
Equation (2.3) directly gives us a consistency relation, by examining the fault
free case, i.e. by setting f 0 (f =

f =

f = 0):
y +a y +by u = 0
and an equivalent frequency domain description of the relation:
(s
2
+as +b)y u = 0
It is clear that if y and y were known, we could calculate r = y + a y + by u
which would be 0 in the fault free case and deviate from 0 when f , 0. However,
the higher order derivatives are usually not known and one way to circumvent
this complication is to add, e.g. low-pass dynamics to the consistency relation.
That is, instead of computing the residual like r = y + a y + by u, compute
the residual according to the dierential equation
r +c
1
r +c
2
r = y +a y +by u (2.4)
where constants c
1
and c
2
has been chosen to ensure a stable residual generator.
In the frequency domain the residual generator transforms to
r =
s
2
+as +b
s
2
+c
1
s +c
2
y
1
s
2
+c
1
s +c
2
u
which can be realized on explicit state-space form, i.e. higher order derivatives
of y and u need not be used. The lter still has the property that r = 0 in the
fault free case.
Consistency relations are frequently used for linear systems, but are equally ap-
plicable in the nonlinear case. However, in the example above it was straightfor-
ward to add dynamics to form a residual generator based on the linear consis-
tency relation. The main property used was the linearity property. The follow-
ing small example illustrates non-linear consistency relations and the problem
that arise when using non-linear consistency relations to form a residual gener-
ator.
16 Chapter 2. Model Based Diagnosis
Example 2.5
A non-linear consistency relation is best represented in the time domain since
no straightforward frequency domain description is possible. Consider the non-
linear system described by state-space equations (inspired by ow equations in
water-tank systems):
x =

x +ku
y =

x
A consistency relation for the model above can be derived by using the mea-
surement equation which gives that
y

x = 0
Dierentiating both sides, another equation is obtained
y
1
2

x
(

x +ku) = 0
Using these two equations, the state-variable x can easily be eliminated and
2y y +y ku = 0 (2.5)
is obtained which is a consistency relation for the example model.
Unfortunately, it is not as easy as in the linear case to use the derived consistency
relation to form a realizable residual generator. Adding linear dynamics like in
(2.4) is in general not sucient to be able to state the residual generator on
state-space form. Further discussions on this topic is found Chapter 4.
The example also illustrates close links with elimination theory when deriving
consistency relations. To obtain the consistency relation, unknown variables
such as the state x and possibly other unknown inputs have to be eliminated
from a set of equations derived from the original model equations. A well known
method for linear residual generator design is the Chow-Willsky scheme, rst
described in (Chow and Willsky, 1984) and later extended to include unknown
inputs in (Frank, 1990). This method is very similar to the non-linear example
above where the model equations are dierentiated a number of times until a set
of equations is obtained where unknown variables can be eliminated
2
. A non-
linear extension of this approach is investigated in Chapter 4 with an approach
that is also closely related to obtaining input-output descriptions of a system
described on state-space form (Jirstrand, 1998).
2.3 Problem motivation and discussion
Having now introduced a few basic principles of model based diagnosis, some
background to the problems studied further in the chapters to follow are now
illustrated and motivated.
2
See Section 3.5.1 for more details on this approach.
2.3. Problem motivation and discussion 17
2.3.1 The linear problem
For linear models on state-space form (or proper transfer functions), any con-
sistency relation based design can be performed in an observer framework and
vice versa. See for example (Ding et al., 1998, 1999a) for a recent description
of these connections and (Patton, 1994) for a more historic view. However, this
does not mean that the design algorithms are equivalent or have equal proper-
ties. To illustrate consistency based residual generator design, consider a small
example system with two sensors, one actuator, and one fault, given by the
block-diagram:
y
2

1
s+b
y
1
d
?
f

1
s+a

u
On analytical matrix form, the model description consists of the following linear
equations:
_
y
1
y
2
_
=
_
1
(s+a)(s+b)
1
(s+a)
_
u +
_
1
0
_
f (2.6)
This system consists of two linear equations, i.e. we can expect to nd two
linearly independent consistency relations. Two linearly independent consis-
tency relations are directly given, in the frequency domain, by the two model
equations as:
(s +a)(s +b)y
1
u = 0 (2.7a)
(s +a)y
2
u = 0 (2.7b)
which are both satised in the fault-free case. Any consistency relation for the
system can now be written as a linear combination of these two. However, the
block diagram gives that a rst order relationship exists between y
1
and y
2
since they are only separated with rst order dynamics. This gives that also
the following two consistency relations spans all consistency relations for the
system,
(s +b)y
1
y
2
= 0 (2.8a)
(s +a)y
2
u = 0 (2.8b)
These two equations clearly captures the most local relationships in the model
and reects the structure of the process. Thus, a consistency relation based de-
sign algorithm should parameterize all solutions in these two relations. The de-
sign variables, free for the designer to choose, are then which linear-combination,
with rational coecients, of the two relations that should form the residual-
generator. Since the consistency relations are closely related to the process
model, this gives a natural interpretation of the design variables.
18 Chapter 2. Model Based Diagnosis
It is desirable to nd a unied design/analysis procedure that is applicable
to all linear model descriptions and all design problems. Of course, such an
algorithm need good numerical performance to be able to cope with large or
ill-conditioned model descriptions. In Chapter 3, a design algorithm for the
decoupling problem is developed based on these considerations for systems de-
scribed by proper transfer functions (or linear state-space descriptions) and in
Section 3.7 it is shown how the algorithm also covers descriptor systems. The
algorithm nds minimal order relations that span all possible consistency rela-
tions like (2.8). It is worth noting that a design method not considering the
order easily results in a residual generator of the same order as the process
model, and the dierence can be signicant. For example, with the 26:th order
jet-engine model studied in Section 3.6.2, it was possible to design a residual
generator based on a 4:th order relationship with fault sensitivity according to
design specication.
Robustness
Low order relationships can also imply robustness against model uncertainty.
Consider the following two residual generators for detecting f in (2.6) where
the rst is based on relation (2.7a) and the second on (2.8a):
r
1
= y
1

1
(s +a)(s +b)
u r
2
= y
1

1
s +b
y
2
Examining the expressions gives that both r
1
and r
2
has the same fault-response
but r
1
relies on the accuracy of both model parameters a and b while r
2
only
on parameter b. Thus, the lower order residual generator r
2
is less dependent
on the model accuracy compared to r
1
. This is not a general result; model
dependency does not always decrease with the order. However, if the model
has such a property, systematic utilization of low-order residual generators is
desirable.
Uncertainty models
The last step in a residual generator design is to select the free design vari-
ables. To guide the selection, or at least restrict the design freedom, additional
modeling/requirements on the residual generator is needed. A common way to
introduce such extra requirements is to consider uncertain models. Two natu-
ral ways to model this uncertainty are parametric uncertainties in the model or
subjecting the model to stochastic noise and investigate what available design
freedom that is available with these extended models.
For stochastic linear systems, i.e. noise aected linear systems, there is not
much work published. A common approach for such systems is to use Kalman-
lters as residual generators which then produces residuals that are zero-mean
and white with known covariance. However, for fault diagnosis, faults must be
2.3. Problem motivation and discussion 19
decoupled in the residuals to facilitate fault isolation which means that the diag-
nosis decision should not be based on any residuals that are dependent on these
unknown signals, i.e. they should be decoupled in the residual. Unknown input
decoupling is not handled directly using basic, straightforward Kalman ltering
theory. The nominal design algorithm from Chapter 3 handles decoupling and
Chapter 5 extends the nominal design algorithm to also address disturbance
decoupling in stochastic linear systems.
When the model is subjected to parametric uncertainties, it is a common ap-
proach to rst state an optimization criterion reecting diagnosis performance.
Synthesis of residual generators is then performed by minimizing
3
the inuence
from worst-case uncertainties. For the optimization to produce a useful result,
the criterion must be stated such that inuence from both control signals and
disturbances are attenuated while fault sensitivity is kept. A main diculty is
how to state the desired fault sensitivity without violating structural properties
of the model. An algorithm, based on the nominal design algorithm, to form
the optimization criterion is developed in Chapter 6.
2.3.2 The non-linear problem
When approaching the full non-linear problem, it quite naturally gets more dif-
cult to derive complete solutions similar to what is available in the linear case.
In Chapter 4, systematic methods, with strong computational support, to derive
consistency relations for non-linear systems is pursued. Deriving consistency re-
lations is closely related to variable elimination, which in a general non-linear
system of equations is dicult. Therefore, only a class of non-linear systems is
considered, namely models consisting of a set of polynomial dierential-algebraic
equations. For this class of systems, a design algorithm that nds polynomial
consistency relations is derived. The algorithm then produces non-linear ver-
sions of relations like (2.7). The available design freedom is then similar to the
design freedom in the linear case, i.e. which combination of the relations that
should form the residual-generator. Here, a fundamental dierence between
the linear and non-linear case appears in the diculty of using a non-linear
consistency relation for residual generation which was discussed in Section 2.2.
The computational support provided by symbolic computer algebra packages
such as Mathematica and Maple enables highly automated design procedures.
However, the high computational complexity of these algorithms forces us to,
if anything but small sized systems are considered, take additional measures to
handle the complexity. In Chapter 4, structural analysis of the model equations
is used to handle such complexity problems.
3
assuming a small value of the criterion indicates good performance
20 Chapter 2. Model Based Diagnosis
3
Residual Generation Based on
Linear Models
In this chapter, design and analysis tools for residual generators based on deter-
ministic linear dynamic models are developed. For this linear design problem,
a plethora of design methods for designing linear residual generators have been
proposed in literature, see for example (Chen and Patton, 1999; W unnenberg,
1990; Massoumnia et al., 1989; Nikoukhah, 1994; Chow and Willsky, 1984; Ny-
berg and Nielsen, 2000). However there still exists important topics that have
not been resolved. Based on the discussion in Chapter 2, focus of the approach
described here is a number of natural questions. For example
Does the method nd all possible residual generators?
Does the method nd residual generators of minimal order?
What types of model descriptions can the method cope with? Due to
the simple nature of linear systems, a design method for linear residual
generators should be able to cope with any linear description, i.e. transfer
functions, state-space descriptions or descriptor models.
What are the numerical properties of the design algorithm?
Based on these fundamental questions, a design methodology is developed. Al-
though the results are quite straightforward, the details proofs requires theory
for polynomial matrices, rational vector spaces, and polynomial bases for these
spaces (Kailath, 1980; Forney, 1975; Chen, 1984). Basic denitions and the-
orems used are, for the sake of convenience, collected in Appendix 3.A and
3.B.
21
22 Chapter 3. Residual Generation Based on Linear Models
3.1 The minimal polynomial basis approach
This section introduces the minimal polynomial basis approach to the design of
linear residual generators. All derivations are performed in the time-continuous
case but the same results for the time-discrete case can be obtained by substi-
tuting s by z and improper by non-causal. In the stochastic case, additional
dierences exist between the discrete and continuous-time case. This topic is
further investigated in Chapter 5.
3.1.1 A general problem formulation
First, a general problem formulation is presented which has been used in many
papers, e.g. (Gertler, 1991).
The class systems studied are assumed to be modeled by
y = G
u
(s)u +G
d
(s)d +G
f
(s)f (3.1)
where y R
m
is the measurements, u R
k
u
the known inputs to the system,
d R
k
d
the disturbances including non-monitored faults, and f R
k
f
the mon-
itored faults. The transfer functions G
u
(s), G
d
(s), and G
f
(s) are all assumed
to be proper and of suitable dimensions.
Since we are considering linear systems also linear residual generators are
considered, i.e. the residual is produced by linear ltering of measurements and
control signals. For a system (3.1), linear residual generators can be dened as
follows:
Denition 3.1 (residual generator for deterministic systems). A stable
and proper linear lter Q(s) is a residual generator for (3.1) if and only if when
f 0 it holds that
r = Q(s)
_
y
u
_
0 (3.2)
for all u, d.
From now on, all initial conditions is assumed 0. Since only strictly stable
residual generators are considered, inuence from these unknown initial states
will vanish exponentially. If the transfer functions is non-proper, this is not
generally true. This will be discussed further in Section 3.7 for so called de-
scriptor systems. Note that for the residual to be useful for fault detection it
must also hold that the transfer function from faults to the residual is non-zero.
Sometimes this requirement is also included in the denition of the residual gen-
erator. Also, the requirement that the residual should be zero in the fault-free
case is too strict in the general case. This since perfect decoupling is not always
possible even in the deterministic case and we have to resort to approximate
decoupling of disturbances. Such issues is further explored in Chapter 6. From
now on, without loss of generality, r is assumed to be a scalar signal.
3.1. The minimal polynomial basis approach 23
3.1.2 Derivation of design methodology
Inserting (3.1) into (3.2) gives
r = Q(s)
_
G
u
(s) G
d
(s)
I
k
u
0
k
u
k
d
_ _
u
d
_
+Q(s)
_
G
f
(s)
0
k
u
k
f
_
f (3.3)
To make r = 0 when f = 0, it is required that disturbances and the control
signal are decoupled, i.e. for Q(s) to be a residual generator, it must hold that
Q(s)
_
G
u
(s) G
d
(s)
I
k
u
0
k
u
k
d
_
= 0
This implies that Q(s) must belong to the left null-space of
M(s) =
_
G
u
(s) G
d
(s)
I
k
u
0
k
u
k
d
_
(3.4)
This null-space is denoted N
L
(M(s)). The matrix Q(s) need to fulll two
requirements to form a good residual generator: belong to the left null-space of
M(s) and provide good fault sensitivity properties in the residual. This lter
Q(s) can be, and has been, designed by observer methodology or by numerous
other methods. Here however, the design method directly considers the null-
space of M(s) which will show to lead to appealing analysis possibilities and a
straightforward and numerically good design algorithm. If, in the rst step of
the design, all Q(s) that fullls the rst requirement is found and parameterized,
then in a second step a single Q(s) with good fault sensitivity properties can
be selected. Thus, in a rst step of the design, f or G
f
(s) do not need to be
considered. The problem is then to nd and parameterize all rational Q(s)
N
L
(M(s)). Of special interest are residual generators of low order for reasons
discussed in Chapter 2.
Finding all Q(s) N
L
(M(s)) can be done by nding a minimal polynomial
basis for the rational vector-space N
L
(M(s)). Algorithms for computing such a
basis for N
L
(M(s)) will be described in Section 3.2. For now, assume that such
a basis has been found and is formed by the rows of a matrix denoted N
M
(s).
By inspection of (3.4), it can be realized that the dimension of N
L
(M(s)) (i.e.
the number of rows of N
M
(s)) is
dim N
L
(M(s)) = mrank G
d
(s)

= mk
d
(3.5)
The last equality, marked

=, holds only if rank G
d
(s) = k
d
, but this should be
the normal, or generic, case. More formal thoughts on genericity can be found
in (Wonham, 1979).
Forming the residual generator
The second and nal design-step is to use the polynomial basis N
M
(s) to form
the residual generator. A decoupling polynomial vector is a polynomial row-
vector F(s) for which it holds that F(s) N
L
(M(s)). It is immediate that such
24 Chapter 3. Residual Generation Based on Linear Models
a row-vector corresponds to a consistency relation
F(s)
_
y
u
_
= 0
The minimal polynomial basis N
M
(s) is irreducible according to Theorem 3.B.1,
and then according to Theorem 3.B.2, all decoupling polynomial vectors F(s)
can be parameterized as
F(s) = (s)N
M
(s) (3.6)
where (s) is a polynomial vector of suitable dimension. Thus, all consistency
relations can be parameterized by a polynomial row-vector (s). Since N
M
(s) is
a basis, the parameterization vector (s) have minimal number of elements, i.e.
N
M
(s) gives a minimal parameterization of all decoupling polynomial vectors,
not only minimal order.
It is also straightforward to show that since N
M
(s) is a minimal polynomial
basis, one of the rows corresponds to a decoupling polynomial vector of minimal
row-degree, see proof of Lemma 3.2 on page 31. Consistency relations was
discussed thoroughly in Section 2.2 where also Example 2.4 indicated how, in
the linear case, such a consistency relation could be used to design a residual
generator. A general formulation of that example gives that a realizable rational
transfer function Q(s), i.e. the residual generator, can be found as
Q(s) = c
1
(s)F(s) (3.7)
where the scalar polynomial c(s) has greater or equal degree compared to the
row-degree of F(s). The degree constraint is the only constraint on c(s) apart
from a stability constraint. This means that the dynamics, i.e. poles, of the
residual generator Q(s) can be chosen freely as long as the roots of c(s) lies in
the open left half-plane. Therefore, (s) and c(s) includes all design freedom
that can be used to shape the fault-to-residual response. This also means that
the minimal order of a realization of a residual generator is determined by the
row-degree of the polynomial vector F(s).
This design freedom can be used in many ways, e.g. can the poles of the
residual generator be selected to impose a low-pass characteristic of the resid-
ual generator to lter out noise or high frequency uncertainties. However, if
the residual generator problem is stated as in Denition 3.1 and the model is
given by (3.1), any choice of (s) and c(s) are theoretically equally good. In
practice this is of course not true, but to be able to form a systematic de-
sign procedure where the parameterization matrices are determined, additional
modeling/requirements need to be introduced. Two such natural extensions
are explored in subsequent chapters; the rst is introduction of stochastic noise
in the model and the second is introduction of parametric uncertainties in the
model equations.
3.2. Methods to nd a minimal polynomial basis for N
L
(M(s)) 25
3.2 Methods to nd a minimal polynomial basis
for N
L
(M(s))
The problem of nding a minimal polynomial basis to the left null-space of
the rational matrix M(s) can be solved by a transformation to the problem of
nding a polynomial basis for the null-space of a polynomial matrix. This latter
problem is then a standard problem in linear systems theory where standard
algorithms can be applied (Kailath, 1980).
The transformation from a rational problem to a polynomial problem can be
done in dierent ways. In this section, two methods are demonstrated, where
one is used if the model is given on transfer function form and the other if the
model is given in state-space form. If there are no disturbances d, the problem
of nding a basis to the left null-space of M(s) is simplied and a method
applicable in this case will also be described. Altogether, the results in this
section will give us a computationally simple, ecient, and numerically stable
method to nd a polynomial basis for the left null-space of M(s).
3.2.1 Frequency domain solution
When the system model is given on transfer function form (3.1), the trans-
formation from the rational problem to a polynomial problem can be done by
performing a right MFD (Kailath, 1980) of M(s), i.e.
M(s) =

M
1
(s)

D
1
(s) (3.8)
By nding a polynomial basis for the left null-space of the polynomial matrix

M
1
(s), a basis is found also for the left null-space of M(s) since

D(s) is full
rank. Thus the problem of nding a minimal polynomial basis for N
L
(M(s))
has been transformed into nding a minimal polynomial basis for N
L
(

M
1
(s)).
3.2.2 State-space solution
When the system model is available in state-space form, it is here shown how
the system matrix in state-space form can be used to nd the left null-space of
M(s). The system matrix has been used before in the context of fault diagnosis,
see for example (Nikoukhah, 1994; Magni and Mouyon, 1994).
Assume that the fault-free system is described in state-space form by,
x = Ax +B
u
u +B
d
d (3.9a)
y = Cx +D
u
u +D
d
d (3.9b)
To be able to obtain a basis that is irreducible, it is required that the state x is
controllable from [u
T
d
T
]
T
.
26 Chapter 3. Residual Generation Based on Linear Models
Denote the system matrix M
s
(s), describing the system with disturbances
as inputs:
M
s
(s) =
_
C D
d
sI
n
+A B
d
_
(3.10)
Dene a matrix P as
P =
_
I
m
D
u
0
nm
B
u
_
(3.11)
The rationale of these denitions is that the Laplace transform of the model
equations (3.9) can then be written as
M
s
(s)
_
x
d
_
= P
_
y
u
_
Then the following theorem gives a direct method on how to nd a minimal
polynomial basis to N
L
(M(s)) via the system matrix.
Theorem 3.1. If the pair A, [B
u
B
d
] is controllable and the rows of a poly-
nomial matrix V (s) form a minimal polynomial basis for N
L
(M
s
(s)), then the
rows of W(s) = V (s)P form a minimal polynomial basis for N
L
(M(s)).
Before this theorem can be proven, a lemma is needed:
Lemma 3.1.
dimN
L
(M(s)) = dimN
L
(M
s
(s))
Proof. In this proof, controllability of (3.9) is assumed. See (Nyberg, 1999b)
for the general proof.
The dimension of N
L
(M(s)) can immediately be seen as
dimN
L
(M(s)) = m+k
u
rank M(s) = mrank G
d
(s) (3.12)
since rank M(s) = k
u
+ rank G
d
(s). By utilizing generalized Bezout-identities
like in (Kailath, 1980, Sec. 6.4.2), it is seen that
M
s
(s)
s

_
I
n
0
0 N(s)
_
where N(s) is the numerator in a right MFD G
d
(s) = N(s)D
1
(s) and
s

represents Smith-form similarity. This gives that

rank M
s
(s) = n + rank N(s) = n + rank G
d
(s)
Then, dimN
L
(M
s
(s)) can be written as
dimN
L
(M
s
(s)) = n +mrank M
s
(s) = mrank G
d
(s)
which is equal to (3.12) which gives the theorem.
3.2. Methods to nd a minimal polynomial basis for N
L
(M(s)) 27
Now, return to the proof of Theorem 3.1:
Proof. In the fault free case, i.e. f = 0, consider the following relation between
the matrices M(s) and M
s
(s):
P
_
y
u
_
= PM(s)
_
u
d
_
= M
s
(s)
_
x
d
_
If V (s)M
s
(s) = 0, then the above expression is zero for all x and d, which
implies that W(s)M(s) = V (s)PM(s) = 0, i.e. W(s) N
L
(M(s)). It is also
immediate that if V (s) is polynomial, W(s) = V (s)P is also polynomial. Also,
Lemma 3.1 gives that that dimN
L
(M
s
(s)) = dimN
L
(M(s)). Then since both
V (s) and W(s) has the same number of rows, the rows of W(s) must span the
whole null-space N
L
(M(s)), i.e. W(s) must be a basis for N
L
(M(s)).
To now show that W(s) is a minimal polynomial basis, it is according to
Theorem 3.B.1 sucient to prove that W(s) is irreducible and row-reduced. It
is clear that the following relation must hold:
V (s)[P M
s
(s)] = V (s)
_
I D
u
C D
d
0 B
u
(sI A) B
d
_
= [W(s) 0] (3.13)
Since the state x is controllable from u and d, the PBH test (Theorem 3.B.3)
implies that the lower part of the matrix [P M
s
(s)] has full rank for all s, i.e.
it is irreducible. Now assume that W(s) is not irreducible. This means that
there exists a s
0
and a ,= 0 such that V (s
0
)[P M
s
(s
0
)] = [W(s
0
) 0] = 0.
Since [P M
s
(s
0
)] has full row-rank it must hold that V (s
0
) = 0. However, this
contradicts the fact that V (s) is a minimal polynomial basis. This contradiction
implies that W(s) must be irreducible.
Now, partition V (s) = [V
1
(s) V
2
(s)] according to the partition of M
s
(s).
Since V (s) N
L
(M
s
(s)), it holds that
V
1
(s)C = V
2
(s)(sI A) = sV
2
(s) V
2
(s)A
Also, since each row-degree of sV
2
(s) is strictly greater than the corresponding
row-degree of V
2
(s)A, it holds that for each row i
row-deg
i
sV
2
(s) = row-deg
i
V
2
(s) + 1 = row-deg
i
V
1
(s)C
The above equation can be rearranged into the inequalities
row-deg
i
V
2
(s) < row-deg
i
V
1
(s)C row-deg
i
V
1
(s)
This implies that V
hr
= [V
1,hr
0] where V
hr
and V
1,hr
are the highest-row-degree
coecient matrices of V (s) and V
1
(s) respectively. Since V (s) is a minimal
polynomial basis V
hr
has full row rank from which it follows that V
1,hr
has full
row rank.
From the denition of P it follows that
[W
1
(s) W
2
(s)] = [V
1
(s) (V
1
(s)D
u
V
2
(s)B
u
)] (3.14)
28 Chapter 3. Residual Generation Based on Linear Models
From (3.14) it follows that the highest-row-degree coecient matrix of W(s)
looks like W
hr
= [V
1,hr
] where is any constant matrix. Since V
1,hr
has full
row-rank so has W
hr
, i.e. W(s) is row reduced.
What happens if the controllability assumption in Theorem 3.1 is dropped
is directly given by the following corollary.
Corollary 3.1. Let the polynomial matrix V (s) form a minimal polynomial
basis for N
L
(M
s
(s)), then the rows of W(s) = V (s)P form a polynomial basis,
not necessarily irreducible, for N
L
(M(s)).
Proof. Following the proof of Theorem 3.1 it is seen that if the realization is
not controllable from
_
u
T
d
T

T
, then the matrix [P M
s
(s)] in (3.13) does not
have full row-rank for all s. Thus, W(s) = V (s)P is a basis but not necessarily
irreducible. This has the implication that all decoupling polynomial vectors
F(s) can not be parameterized as in (3.6).
3.2.3 No disturbance case
If there are no disturbances, i.e. G
d
(s) = 0, the matrix M(s) becomes M
nd
(s) =
[G
T
u
(s) I]
T
. A minimal basis is then given directly by the following theorem:
Theorem 3.2. If G(s) is a proper transfer matrix,

D
G
(s),

N
G
(s) form an
irreducible left MFD of G(s), and

D
G
(s) is row-reduced, then
N
M
nd
(s) = [

D
G
(s)

N
G
(s)] (3.15)
forms a minimal polynomial basis for the left null-space of the matrix M
nd
(s).
Proof. It is immediate to evaluate
[

D
G
(s)

N
G
(s)]
_
G
u
(s)
I
_
= 0
Also, the dimension of the left null-space of M
nd
(s) has dimension m, i.e. the
number of measurements, which equals the number of rows in N
M
(s). Since

D
G
(s) and

N
G
(s) is co-prime, N
M
(s) will be irreducible. Let

D
G
(s) = S
D
(s)D
hr
+L
D
(s)

N
G
(s) = S
N
(s)N
hr
+L
N
(s)
where D
hr
and N
hr
be the highest row-degree coecient matrices for

D
G
(s)
and

N
G
(s) where D
hr
will be of full rank since

D
G
(s) is row-reduced.
Since the transfer function G(s) is proper, i.e. the row degrees of

N
G
(s) is less
or equal to the row degrees of

D
G
(s). A high-degree coecient decomposition
of the basis N
M
(s) will look like
[

D
G
(s)

N
M
(s)] = S
D
(s)[D
hr
] +

L(s)
where is any matrix. Since D
hr
is full rank, so is [D
hr
], which gives that
the basis is row-reduced which ends the proof.
3.3. Matlab sessions 29
Note that not just any irreducible MFD will suce, the row-reducedness
property is also needed and an algorithm that provides such an MFD is found
in (Strijbos, 1996) and is implemented in (Polynomial Toolbox 2.5 for Matlab
5, 2001). The row-degrees of the minimal polynomial basis for N
L
(M
nd
(s)) are
closely related to the observability indices according to the following theorem:
Theorem 3.3. The set of observability indices of a proper transfer function
G(s) is equal to the set of row degrees of

D
G
(s) and also (3.15) in any row-
reduced irreducible left MFD G(s) =

D
1
G
(s)

N
G
(s).
Proof. A proof of the dual problem, controllability indices, can be found in
(Chen, 1984, p. 284).
Thus, a minimal polynomial basis for the left null-space of matrix M
nd
(s)
is given by a left MFD of G(s) and the order of the basis is the sum of the
observability indices of G(s).
Remark 1: Note that, in the general case, the observability indices of the pair
A, C do not give the row-degrees of a minimal polynomial basis for N
L
(M(s)).
However, as will be shown in Theorem 3.6, the minimal observability index of
A, C does give a lower bound on the minimal row-degree of the basis.
Remark 2: The result (3.15) implies that nding the left null-space of the
rational transfer matrix (3.4), in the general case with disturbances included,
can be reduced to nding the left null-space of the rational matrix

M
2
(s) =

D
G
(s)H(s) (3.16)
In other words, this is an alternative to the use of the matrix

M
1
(s) in (3.8).
This view closely connects with the so called frequency domain methods, which
are further examined in Section 3.5.
3.2.4 Finding a minimal polynomial basis for the null-
space of a general polynomial matrix
For the general case, with disturbances included, the only remaining problem
is how to nd a minimal polynomial basis to for the left null-space of a general
polynomial matrix. This is a well-known problem in the general literature on
linear systems. When numerical performance is considered, a specic algorithm
based on the polynomial echelon form (Kailath, 1980) has been proven to be
both fast and numerically stable. Such an algorithm is implemented in the
command null in Polynomial Toolbox 2.5 for Matlab 5 (2001).
3.3 Matlab sessions
To illustrate the simplicity of the design algorithm, a complete Matlab-session
(requires control and polynomial toolbox) for design of residual generators is
30 Chapter 3. Residual Generation Based on Linear Models
included for both cases when the model is given on transfer function form and
state-space form. In both cases, (s) in (3.6) is set to (s) = [1 0] and c(s) in
(3.7) to c(s) = (s + 1)
3
. First the state-space case:
1 Ms = [C Dd;-(s*eye(nx)-A) Bd];
2 P = [eye(m) -Du;zeros(nx,m) -Bu];
3 Nms = null(Ms.).;
4 fi = [1 0]; c = (s+1)^3;
5 [Qa,Qb,Qc,Qd] = lmf2ss(fi*Nms*P,c);
6 Q = ss(Qa,Qb,Qc,Qd);
A Matlab-session for the case when the model is given on transfer function form
is similar:
1 M = [Gu Gd;eye(ku) zeros(ku,kd)
2 [M1,D] = ss2rmf(M.a,M.b,M.c,M.d);
3 Nm = null(M1.).;
4 fi=[1 0]; c = (s+1)^3;
5 [Qa,Qb,Qc,Qd] = lmf2ss(fi*Nm,c);
6 Q = ss(Qa,Qb,Qc,Qd);
As shown above, no diagnosis specic code need to be developed and the design
procedure solely relies on high performance numerical routines in established
Matlab toolboxes. The two main operations are null which computes a basis
for the null-space of a polynomial matrix and lmf2ss which derives a state-space
realization of a left MFD description. The simplicity of the design algorithm is
possible because of the abstraction made to consider polynomial matrices. The
numerical performance in diagnosis applications of the above code is illustrated
further in Section 3.6 and Section 3.7.
3.4 Bounds on maximum and minimum row-
degree of the basis
In Section 2.3, inuence of the order of the residual generator on e.g. robustness
properties was discussed. This section will provide a deeper analysis on the
residual generator order.
In Section 3.1 it was shown how the row-degrees of a minimal polynomial
basis for the matrix M(s) were connected with the order of the residual gener-
ator. Now follows an analysis of these row-degrees where upper limits on the
maximum and minimum row-degree of a matrix is derived. The notation n is
used to denote the number of states in a given state-space representation and
n
x
will be used to denote the number of states controllable from
_
u
T
d
T

T
.
3.4. Bounds on maximum and minimum row-degree of the basis 31
3.4.1 Upper bound for the maximum row-degree of the
basis
Theorem 3.4 (Nyberg,1999b). A matrix whose rows form a minimal poly-
nomial basis for N
L
(M(s)) has all row-degrees n
x
.
Before Theorem 3.4 can be proven, a few lemmas are needed.
Lemma 3.2. Let the rows of F(s) form a minimal polynomial basis for a ratio-
nal vector space T. Denote the row-degrees of F(s) with
1

. Then it
holds that
i
m
i
, i = 1, . . . , where m
i
is the row-degrees of any polynomial
basis for T.
Proof. Let P(s) be a polynomial basis for T with row-degrees m
i
. Let the rows
in P(s) be ordered such that m
1
m

.
The theorem is proved by contradiction. Assume that
1
m
1
, . . . ,
i1

m
i1
but
i
> m
i
. Since F(s) is an irreducible basis, it holds that
p
j
(s) =

l=1
f
l
(s)q
l
(s) j = 1, . . . , i (3.17)
where q
l
(s) is polynomials.
If i = 1, then deg p
1
(s) <
j
j = 1, . . . , , i.e. according to Theorem 3.B.5,
p
1
(s) can not be a linear combination of the rows in the row reduced matrix
F(s). However, this contradicts (3.17).
If i > 1, according to the assumption, the following relations hold:
deg p
j
(s) m
i
<
i
j = 1, . . . , i
According to the predictable degree property it must hold that in (3.17), q
l
(s)
0, l = i, . . . , . Thus, the upper summation limit can at maximum be i 1, i.e.
equation (3.17) can be rewritten as:
p
j
(s) =
i1

l=1
f
l
(s)q
l
(s) j = 1, . . . , i
This contradicts the linear independence of the p
1
(s), . . . , p
i
(s) polynomial row
vectors since they are spanned by f
1
(s), . . . , f
i1
(s) ending the proof.
Lemma 3.3. Let P(s) be a matrix with maximum row-degree 1. Then the
maximum row-degree of a minimal polynomial basis for N
L
(P(s)) is less or
equal to rank P(s).
Proof. Since P(s) is a matrix pencil it can be transferred to Kronecker Ca-
nonical Form (Theorem 3.B.4) by pre- and post-multiplication with constant
non-singular matrices U and V , i.e. P
KCF
(s) = UP(s)V . Also, since all ma-
trices besides

L

, in the block-diagonal Canonical Form has full row-rank the

32 Chapter 3. Residual Generation Based on Linear Models
left null-space structure of P
KCF
(s) is fully characterized by the left Kronecker
indices
1
, . . . ,

.
It is also easy to check that the left null-space of

L

is given by
v

(s) = [1 s s

]
i.e, the degree of the left null-space vectors is directly given by the left Kronecker
indices. Thus, a basis for the left null-space of P
KCF
(s) is given by a matrix on
the form
N
P
KCF
(s) =
_

_
0 v

1
(s) 0 0
0 0 v

2
(s) 0 0 0
0 0
.
.
.
0 0
0 0 0 v

(s) 0
_

_
i.e. the maximum row-degree of N
P
KCF
(s) is max
i

i
. A basis for the left null-
space of P(s) is given by N
P
KCF
(s)U where U is constant and non-singular
which gives that the maximum row-degree its maximum row-degree is max
i

i
.
It holds that rank

L

= which gives that rank P(s)

i=1

i
, i.e. the maxi-
mum row-degree of a polynomial basis (and thereby also a minimal polynomial
basis according to Lemma 3.2) is less than rank P(s).
Lemma 3.4. The row-degrees of a minimal polynomial basis for N
L
(M(s)) is
equal to the row-degrees of a minimal polynomial basis for N
L
(M
s
(s)), where
M
s
(s) is a system matrix with the pair A, [B
u
B
d
] controllable.
Proof. Let V (s) be a minimal polynomial basis for N
L
(M
s
(s)) and partition
V (s) = [V
1
(s) V
2
(s)] according to the partition of M
s
(s). Then, since we know
that V (s) N
L
(M
s
(s)), it holds that
V
1
(s)C = V
2
(s)(sI A) = sV
2
(s) V
2
(s)A
Also, since each row degree of sV
2
(s) is strictly greater than the corresponding
row-degree of V
2
(s)A, it holds that for each row i
row-deg
i
sV
2
(s) = 1 + row-deg
i
V
2
(s) = row-deg
i
V
1
(s)C
The above equation can be rearranged to
row-deg
i
V
2
(s) < row-deg
i
V
1
(s)C row-deg
i
V
1
(s) (3.18)
i.e. row-deg
i
V (s) = row-deg
i
V
1
(s). From the denition of P in (3.11) it follows
that
W(s) = [W
1
(s) W
2
(s)] = V (s)P = [V
1
(s) (V
1
(s)D
u
V
2
(s)B
u
)] (3.19)
Equations (3.18) and (3.19) directly give
row-deg
i
W(s) = row-deg
i
V
1
(s) = row-deg
i
V (s),
i.e. the row degrees of W(s) and V (s) are equal. According to Theorem 3.1,
W(s) and V (s) are minimal polynomial bases for N
L
(M(s)) and N
L
(M
s
(s))
respectively and the lemma follows immediately.
3.4. Bounds on maximum and minimum row-degree of the basis 33
Remark:
Lemma 3.4 implies that the row-degrees of N
M
(s) equals the left Kronecker
indices of the matrix pencil M
s
(s). There exist a lot of literature and algorithms
for computing the Kronecker indices of a general matrix pencil, e.g. (Misra et al.,
1994; Wang et al., 1975; Aling and Schumacher, 1984; Kagstrom, 1986).
Now return to the proof of Theorem 3.4:
Proof. Let n
x
be the order of a minimal state-space realization of (3.9), con-
trollable from
_
u
T
d
T

T
. Let M
s
(s) be the corresponding system matrix, i.e.
M
s
(s) =
_
C D
d
(sI A) B
d
_
and let the rows of N
DB
be a basis for the left null-space of [D
T
d
B
T
d
]
T
. Then
we have that
N
DB
M
s
(s) =
_
N
DB
_
C
(sI A)
_
, 0

(3.20)
The left part of the matrix (3.20) has rank n
x
. From Lemma 3.3 we know
that a minimal polynomial basis for (3.20) has row degrees less or equal to n
x
.
Let the rows of a matrix Q(s) form such a basis.
The matrix Q(s)N
DB
forms a polynomial basis for N
L
(M
s
(s)) and since
Q(s) has row degrees less or equal to n
x
, the row degrees of the basis Q(s)N
DB
is also less or equal to n
x
. Thus, according to Lemma 3.2, a minimal polynomial
basis for N
L
(M
s
(s)) has lower or equal row-degrees than the polynomial basis
Q(s)N
DB
.
Since a minimal polynomial basis for ^
L
(M
s
(s)) has maximum row-degree
n
x
, Lemma 3.4 implies that also a minimal polynomial basis for ^
L
(M(s))
has maximum row-degree n
x
, ending the proof.
The result of Theorem 3.4 is important for several reasons, the residual gen-
erators obtained directly from the vectors of the minimal basis, are in one sense
the only ones needed. All other are ltered versions (i.e. linear combinations)
of these residual generators. With this argument, Theorem 3.4 shows that we
do not need to consider residual generators of orders greater than n
x
.
Remark
Related problems have been investigated in (Chow and Willsky, 1984) and
(Gertler et al., 1990). In (Chow and Willsky, 1984), it was shown that, in
the no-disturbance case, there exist a parity equation of order n. In (Gertler
et al., 1990), it was shown that for a restricted class of disturbances, there exist
a parity function of order n. However the result of Theorem 3.4 is stronger
since it includes arbitrary disturbances and shows that there exist a basis in
which the maximum row-degree is n
x
.
34 Chapter 3. Residual Generation Based on Linear Models
3.4.2 Bounds for the minimal row-degree of the basis
Theorem 3.5 (Frisk,2000a). An upper bound for the minimal row-degree
min
of a basis for N
L
(M(s)) is given by

min

n
x
+n
d
mn
d
|
where
n
d
= Rank
_
B
d
D
d
_
is the number of linearly independent disturbances.
Before Theorem 3.5 can be proven, some more results are needed. If n
d
< k
d
,
i.e. there exists linear dependencies between disturbances, rewrite the system
description with a new set of n
d
linearly independent disturbances. That is,
nd

B
d
and

D
d
with dimensions n
x
n
d
and mn
d
respectively such that
Im
_
B
d
D
d
_
= Im
_

B
d

D
d
_
Then, denote

=
. .
( + 2)(n
x
+n
d
)
_

_
Q R
Q R
.
.
.
Q R
_

_
_

_
( + 1)(m+n
x
)
where M
s
(s) = Q+sR. Then, the following lemma can be stated:
Lemma 3.5. The space N
L
(M
s
(s)) contains a -degree polynomial vector if
and only if

M

does not have full row rank.

Proof. Let F(s) be a -degree polynomial matrix in N
L
(M
s
(s)). Then it holds
that
0 = F(s)M
s
(s) = (F
0
+F
1
s + +s

)M
s
(s) =
= [F
0
F
1
F

]
_

_
M
s
(s)
sM
s
(s)
.
.
.
s

M
s
(s)
_

_
=

F

_
I
sI
.
.
.
s

I
_

_
From the equation above it is clear that a -degree polynomial F(s) is in
N
L
(M
s
(s)) if and only if

F

= 0. The lemma follows directly because such a

F can only exist if

M

does not have full row-rank.

3.4. Bounds on maximum and minimum row-degree of the basis 35
A similar result can also be found in (Karcanias and Kalogeropoulos, 1988).
Now, return to the proof of Theorem 3.5.
Proof. Using Lemma 3.4 and 3.5 it is clear that a -degree polynomial vector
is in N
L
(M(s)) if and only if

M

does not have full row rank. A sucient

condition for

M

not to have full row-rank is that the number of rows is larger

than the number of columns, i.e.
( + 1)(m+n
x
) > ( + 2)(n
x
+n
d
)
Straightforward manipulations of the inequality results in
>
n
x
+n
d
mn
d
1
Note that the inequality m > n
d
always holds if a residual generator exists which
can be seen directly in (3.5). Therefore, the smallest integer that fullls the
inequality is
n
x
+n
d
mn
d
| which completes the proof.
Remark: A similar result without disturbance decoupling, i.e. when n
d
= 0,
can be found in (Mironovskii, 1980).
Theorem 3.2 and 3.3 states that in the no-disturbance case, the observability
indices of the pair A, C give exactly the row-degrees of a minimal polynomial
basis for N
L
(M(s)). In the more general case, including disturbances, this is
not true anymore. However, the minimal observability index of A, C can still
be used to obtain a lower bound according to the following theorem
Theorem 3.6. A lower bound for the minimal row-degree
min
of a basis for
N
L
(M(s)) is the minimal observability index
min
of A, C, i.e.
min

min
with equality when n
d
= 0.
Proof. Denote the system matrix without disturbances with M
(nd)
s
(s), i.e.
M
(nd)
s
(s) =
_
C
sI A
_
A direct consequence of Theorem 3.3 is that the row-degrees of a minimal poly-
nomial basis for the left null-space of M
(nd)
s
(s) is equal to the observability
indices of the pair (A, C). Let c
min
be the minimum observability index of
(A, C). Then, according to Lemma 3.5, c
min
is the lowest such that

M
(nd)

does not have full row-rank. Let

Q = [Q
1
Q
2
] =
_
C D
d
A B
d
_
R = [R
1
R
2
] =
_
0 0
I 0
_
36 Chapter 3. Residual Generation Based on Linear Models
Then, by a trivial column reordering,

M

can be written on the form

=
_

_
Q
1
R
1
Q
2
R
2
Q
1
R
1
Q
2
R
2
.
.
.
.
.
.
Q
1
R
1
Q
2
R
2
_

_
L
= [

M
(nd)

]L
where L is a square, full rank pivoting matrix and is a matrix of suitable
dimensions who is not of further interest here. From the equation above, it is
clear that if

M
(nd)

has full row-rank, then also

M

has full row-rank. Also, for

all < c
min
,

M
(nd)

and thereby also

M

, will have full row-rank. The theorem

then follows directly from Lemma 3.5, i.e. there exists no -degree polynomial
in N
L
(M
s
(s)) where < c
min
.
A similar result can also be found in (Ding et al., 1999b) in a parity-space
setting and in (Karcanias and Kalogeropoulos, 1988) in a geometrical setting.
3.5 Relation to other residual generator design
methods
This section discusses the relation between the minimal polynomial basis ap-
proach and two other design methods for linear residual generation, the parity-
space approach and the so called frequency domain approach. Relations to these
two approaches are included because they both exhibit interesting relations to
the minimal polynomial basis approach.
3.5.1 The parity-space approach
The parity space approach to the design of linear residual generators was pre-
sented in (Chow and Willsky, 1984; Frank, 1990). Its basic version is based on
a state-space description (3.9). By dierentiating the measurement equation
times, and substituting in (3.9a) a new formulation of the system equations is
obtained. For notational simplicity, let

(s) =
_
I
m
sI
m
. . . s

I
m

T
The fault-free system description is then, in the frequency domain,

(s)y(t) = O

x(t) +Q

(s)u(t) +H

d(t)
where O

is the observation matrix up to order and Q

,H

is lower trian-
gular Toeplitz matrices describing fault/disturbance inuence on the process.
The exact appearance of these matrices is not of importance here, see any of
3.5. Relation to other residual generator design methods 37
the referred works cited above for details. A consistency relation can then be
obtained as
w[

(s) Q

(s)]
_
y(t)
u(t)
_
= 0
where w N
L
([O

]). A relation to the minimal polynomial basis approach

is straightforward to verify; w[

(s) Q

(s)] N
L
(M(s)) which should
be compared to (3.6). Also, Theorem 3.4 gives that with n, this is an
alternative way to parameterize all consistency relations, i.e. all polynomial
row-vectors in N
L
(M(s)). This basic approach does not however provide a
basis since an over-parameterized solutions is obtained. This issue is clearly
illustrated in a high-order design example in Section 3.6.2.
Also, minimal order consistency relations are not explicitly found with this
basic version of the algorithm. But, by starting with = 0 and iteratively
increase until a consistency relation is found, a minimal order consistency
relation is found. However, a basis can not be obtained in such a way and
several modications to the original algorithm is necessary to obtain a modied
scheme which will produce a minimal polynomial basis for the left null-space of
M(s). The modied scheme and the equivalence with the minimal polynomial
basis approach is thoroughly described in (Nyberg, 1999b).
3.5.2 Frequency domain approaches
A number of design methods described in literature are called frequency domain
methods where the residual generators are designed with the help of dierent
transfer matrix factorization techniques. Examples are (Frank and Ding, 1994)
for the general case with disturbances and (Ding and Frank, 1990; Viswanadham
et al., 1987) in the non-disturbance case. The methods can be summarized as
methods where the residual generator is parameterized as
r = R(s)[

D(s)

N(s)]
_
y
u
_
(3.21)
where

D(s) and

N(s) form a left co-prime factorization of G(s) over H

. Note
the close relationship with Equation (3.15) where the factorization is performed
over polynomial matrices instead of over H

.
This solution however does not generally generate a residual generator with
minimal order. In (Ding and Frank, 1990) and (Frank and Ding, 1994), the
co-prime factorization is performed via a minimal state-space realization of the
complete system, including the disturbances as in Equation (3.9). This results
in

D(s) and

N(s) of a degree that, in the general case, is larger than the lowest
possible order of a disturbance decoupling residual generator. Thus, to nd a
lowest order residual generator extra care is required since excess states need
to be canceled out.
38 Chapter 3. Residual Generation Based on Linear Models
3.6 Design examples with nominal models
This section includes two design examples, each included to illustrate dierent
properties of the algorithm described. The rst example, a model of aircraft
dynamics, illustrates design methodology, the available design freedom and ad-
vantages of the minimality property. It also shows in principle how a set of
residual generators can be designed to t in a diagnosis system based on struc-
tured hypothesis tests/structured residuals. The second example is included
to illustrate some numerical properties of the algorithm and also show conse-
quences of the minimality property of the algorithm.
3.6.1 Design example 1: Aircraft dynamics
This model, taken from (Maciejowski, 1989), represents a linearized model of
vertical-plane dynamics of an aircraft. The inputs and outputs of the model are
Inputs Outputs
u
1
: spoiler angle [tenth of a degree] y
1
: relative altitude [m]
u
2
: forward acceleration [ms
2
] y
2
: forward speed [ms
1
]
u
3
: elevator angle [degrees] y
3
: Pitch angle [degrees]
The model has state-space matrices:
A =
_

_
0 0 1.132 0 1
0 0.0538 0.1712 0 0.0705
0 0 0 1 0
0 0.0485 0 0.8556 1.013
0 0.2909 0 1.0532 0.6859
_

_
B =
_

_
0 0 0
0.12 1 0
0 0 0
4.419 0 1.665
1.575 0 0.0732
_

_
C = [I
3
0] D = 0
33
Suppose the faults of interest are sensor-faults (denoted f
1
, f
2
, and f
3
), and
actuator-faults (denoted f
4
, f
5
, and f
6
). Also, assume that the faults are mod-
eled with additive fault models. The total model, including fault models then
becomes:
_
_
y
1
y
2
y
3
_
_
= G(s)
_
_
_
_
u
1
u
2
u
3
_
_
+
_
_
f
4
f
5
f
6
_
_
_
_
+
_
_
f
1
f
2
f
3
_
_
where
G(s) =
_
A B
C D
_
Thus, the transfer function from fault vector f to measurement vector y be-
comes, G
yf
(s) = [I
3
G(s)].
3.6. Design examples with nominal models 39
Decoupling of faults in the elevator angle actuator
The rst design example is intended to illustrate the design procedure and also
illustrate how available design freedom can be utilized, e.g. when selecting the
inuence structure from Section 2.1.3, or selecting the dynamics of the residual
generator.
The design objective here is to design a residual generator Q
1
(s) that de-
couples faults in the elevator angle actuator, f
6
. Then, matrix G
d
(s) from
Equation 3.1 correspond to all signals that are to be decoupled, i.e. consid-
ered disturbances. In this case then, G
d
(s) becomes the column in G
yf
(s)
corresponding to f
6
. Matrix G
f
(s) corresponds to the monitored faults and
therefore G
f
(s) becomes the other columns.
Without actually deriving the minimal polynomial basis N
M
(s) for N
L
(M(s)),
some properties of the basis N
M
(s) can be obtained by using results from Sec-
tion 3.4. According to formula (3.5), the dimension of the null-space N
L
(M(s))
is 2, i.e. there exists exactly two linearly independent polynomial row-vectors
that decouples f
6
. The maximum row-degree of N
M
(s) will, according to The-
orem 3.4 be n = 5. Theorem 3.5 and Theorem 3.6 give an upper and lower
bound on the minimum row-degree
min
of N
M
(s). The minimal observability
index is
min
= 1, and thus 1
min

5+1
31
| = 3. This implies that the
minimum degree residual generator will have an order of 1, 2, or 3.
Since the model is given in state-space form, Theorem 3.1 is used to extract
N
M
(s). Calculations in Matlab give
N
M
(s) =
_
0.0705s s + 0.0538
22.7459s
2
+ 14.5884s 6.6653
0.091394 0.12 1 0
s
2
0.93678s 16.5141 31.4058 0 0
_
An additive actuator fault is decoupled if and only if the actuator is not used in
the calculation of the residual. This is veried by the two 0 in the last column in
the basis. This basis also gives that the dimension of the null-space N
L
(M(s)) is
2, i.e. there exists exactly two linearly independent numerators that decouples
f
6
which was the expected result according to (3.5). The row-degrees of the
basis is 1 and 2, i.e. it is a basis of order 3. From this it is clear that the lter
of least degree decoupling f
6
is a rst order lter corresponding to row 1 in the
basis. By setting in (3.6) to = [1 0] and c(s) in (3.7) to c(s) = 1 + s the
lter is made realizable and results in the following 1:st order lter
Q
1
(s) =
1
1 +s
_
0.0705s s + 0.0538 0.091394 0.12 1 0

(3.22)
The polynomial, here scalar, c(s) need to have degree 1 to make the lter
realizable since the row-degree of the rst row in N
M
(s) is 1. In this example,
a rst order polynomial is chosen to get minimal order. The polynomial c(s) is
chosen to 1 +s to detect faults with energy in frequency ranges up to 1 rad/s.
40 Chapter 3. Residual Generation Based on Linear Models
Let G
rd
(s) = Q(s)
_
G
u
(s) G
d
(s)
I 0
_
, which should be zero if innite precision
arithmetics were used. Computing the size of G
rd
(s) using the innity norm
gives |G
rd
(s)|

220 dB which is close to machine precision. Figure 3.1

shows how the monitored faults inuence the residual. Maximum gain from
fault to residual is 0 dB, i.e. control signals and the decoupled fault has
no signicant inuence compared to monitored faults. The leftmost plot in
10
5
10
0
10
5
150
100
50
0
f1
|
G
r
f
(
s
)
|

[
d
B
]
10
5
10
0
10
5
150
100
50
0
f2
10
5
10
0
10
5
150
100
50
0
f3
w [rad/s]
10
5
10
0
10
5
150
100
50
0
f4
10
5
10
0
10
5
150
100
50
0
f5
Figure 3.1: Magnitude bode plots for the monitored faults to the residual.
Figure 3.1 also shows that DC-gain from fault 1 to the residual is 0 and f
1
is
therefore dicult to detect since the eect in the residual of a constant fault
f
1
disappears. If a fault has zero DC-gain for any residual generator, the fault
is said to be weakly detectable. Since the algorithm from Chapter 3 nds all
residual generators, then to investigate if any fault is weakly detectable, evaluate
N
M
(s)
_
L(s)
0
_

s=0
If any column is identically zero, then the corresponding fault is weakly de-
tectable. A thorough investigation on fault detectability properties can be found
in (Nyberg, 2000).
The design freedom can also be used to avoid non minimum-phase behavior
of the fault response in the residual. Such behavior is undesirable since the
response to a step fault would be something like Figure 3.2. This is clearly
undesirable since the residual would rst indicate a fault, then pass below the
threshold again, cancelling the alarm, before settling above the threshold. Non
minimum-phase behavior can easily be detected by computing zeros of the poly-
nomial elements in the basis N
M
(s).
Decoupling several faults
As noted in the example above, the dimension of the null-space when decoupling
f
6
was 2. This was expected according to (3.5) and therefore additional freedom
exists. This freedom can for example be used to decouple more than one fault
in each residual to shape the inuence structure, e.g. for multiple fault isolation.
3.6. Design examples with nominal models 41
0 1 2 3 4 5 6 7
0.5
0
0.5
1
t [s]
r
Figure 3.2: Non minimum-phase fault response in the residual
For example, designing a lter that decouples both f
1
and f
4
, i.e. faults in the
rst sensor and the rst actuator results in a null-space of dimension 1 and a
minimal basis of order 3. The polynomial c(s) must be at least a third order
polynomial to make the lter realizable since the row-degree of the basis vector
is 3. Selecting c(s) to be c(s) = (1 +s)
3
, the residual generator becomes
Q
2
(s) =
1
(1 +s)
3
_

_
0
36.825s
2
+ 13.8953s + 1.1157
s
3
+ 0.61619s
2
+ 4.4322s + 2.048
0
36.825s 11.9626
1.665s 0.28273
_

_
T
where the decoupling of faults in sensor 1 is evident since the transfer function
from y
1
(the rst column in Q
2
(s)) is 0.
Designing a complete diagnosis system
As was noted above, there exists design-freedom to insert one or two zeros
in the inuence structure which means that there exists some freedom when
choosing the inuence structure. One example is to decouple one fault in each
residual as illustrated in the inuence structure
f
1
f
2
f
3
f
4
f
5
f
6
r
1
0 X X X X X
r
2
X 0 X X X X
r
3
X X 0 X X X
r
4
X X X 0 X X
r
5
X X X X 0 X
r
6
X X X X X 0
42 Chapter 3. Residual Generation Based on Linear Models
With this structure is it possible to detect and isolate all 6 dierent faults,
assuming only single faults. Calculations in Matlab shows that to achieve
this, the minimum order lter that generates all residuals is an 11:th order lter
(2:nd order for r
1
to r
5
and a 1:st order lter for r
6
).
Another example of an inuence structure that also isolates the 6 dierent
faults is
f
1
f
2
f
3
f
4
f
5
f
6
r
1
0 0 X X X X
r
2
X 0 0 X X X
r
3
X X 0 0 X X
r
4
X X X 0 0 X
where the possibility to introduce more than one 0 in each row is utilized.
Because of this additional freedom, only 4 residuals are required here to form
an isolating inuence structure. Continuing in the same fashion, decoupling
two faults in each residual it is possible to design residuals with an inuence
structure that is able to isolate two simultaneous faults. The price for this
increased isolation performance is that the minimal lter generating residuals
according to the second inuence structure above is a 15:th order lter where
calculations in Matlab gives the (minimal) orders of residuals r
1
to r
4
to be
4, 5, 4, and 2 respectively.
To illustrate how the nal diagnosis system could work, residual generators
r
1
to r
6
have been designed according to the rst inuence structure (the one
with 0 in the diagonal). Simulations are shown in Figure 3.3 and 3.4. These
simulations only show principle operation of the system, e.g. no noise attenua-
tion aspects has been considered. In the simulations, step faults of amplitude
1 are introduced at t = 1. It is evident that, nominally, the f
3
response in r
3
and the f
4
response in r
4
is 0. The isolation procedure is then simply to match
observed residual pattern to columns in the inuence structure.
3.6. Design examples with nominal models 43
0 2 4 6 8 10
0.2
0.1
0
0.1
0.2
r
1
0 2 4 6 8 10
1
0
1
r
2
0 2 4 6 8 10
1
0.5
0
0.5
1
r
3
0 2 4 6 8 10
2
1
0
1
2
r
4
0 2 4 6 8 10
0.5
0
0.5
t [s]
r
5
0 2 4 6 8 10
0.1
0.05
0
0.05
0.1
t [s]
r
6
Figure 3.3: Fault simulation with step fault in f
3
at t = 1. Especially, note
that fault inuence on residual 3 is zero in correspondence with column 3 in the
inuence structure.
0 2 4 6 8 10
0.04
0.02
0
0.02
0.04
r
1
0 2 4 6 8 10
5
0
5
r
2
0 2 4 6 8 10
4
2
0
2
4
r
3
0 2 4 6 8 10
1
0.5
0
0.5
1
r
4
0 2 4 6 8 10
2
1
0
1
2
t [s]
r
5
0 2 4 6 8 10
0.2
0.1
0
0.1
0.2
t [s]
r
6
Figure 3.4: Fault simulation with step fault in f
4
at t = 1. Note that fault
inuence on residual 4 is zero in correspondence with column 4 in the inuence
structure.
44 Chapter 3. Residual Generation Based on Linear Models
3.6.2 Design example 2: Turbo-jet engine
The turbo-jet engine used in this example is developed by Volvo Aero Corpo-
ration (VAC) and used in ghter aircrafts of the Swedish air force. Figure 3.5
shows a schematic picture of the engine. VAC uses a high-order non-linear model
Figure 3.5: The turbo-jet engine
of the engine for analysis and control design. This model can also be used for
diagnosis purposes. In this example, the model has been linearized in an oper-
ating point and the resulting model, after non-controllable and non-observable
modes are eliminated, is a 26:th order model. The model used includes 8 sen-
sors, indicated by the acronyms at the bottom half of the gure, and 5 actuators
indicated by the acronyms on the upper half of the gure. In the operating point
where the model is linearized, only 4 actuators are used. The meanings of the
acronyms are given in the following table.
Inputs Outputs
FVG: Variable fan geometry T1: Inlet temperature
CVG: Variable compressor T25: Compressor temperature
geometry
WFM: Main fuel injector T5: Low-pressure turbine
temperature
A8: Variable Nozzle Area P1: Inlet pressure
WFR: Afterburner fuel injector PS3: Compressor pressure
P5: Low-pressure turbine pressure
NL: Low-pressure turbine speed
NH: High-pressure turbine speed
3.6. Design examples with nominal models 45
The model is numerically sti since it models fast dynamics, such as ther-
modynamics in small control volumes, and slow dynamics such as heating phe-
nomena of metal. The largest time-constant in the model is about 10
5
times
larger than the smallest time constant. This, in connection with the high-order
model makes the model numerically sensitive which demands good numerical
performance of the design algorithm.
Suppose actuator and sensor faults are studied, modeled in the same manner
as in the aircraft dynamics example with additive fault models. Due to the large
amount of sensors in the process, there exists a large amount of freedom shaping
the inuence structure. If no faults are decoupled, the dimension of N
M
(s) is
8, according to Theorem 3.2. This means that up to 7 independent faults can
be decoupled in the residual, i.e. it is possible to introduce up to 7 zeros in a
row of the inuence structure. This issue will however not be explored further
here.
A design with the minimal polynomial basis approach is performed where
all 4 actuator faults are decoupled, i.e. only relations between measurements
are desired. Due to the number of states, and sensors, solving for N
M
(s) in
this problem is computationally (and numerically) quite demanding. The null-
space N
L
(M(s)) is of dimension 4 and the basis has row-degrees 4, 5, 5, and
6. Suppose faults with frequency up to 1 rad/s is of interest, then choosing the
rst row in N
M
(s) as a numerator and selecting c(s) = (s +1)
4
results in a 4:th
10
0
10
5
300
200
100
0
100
f1
d
B
10
0
10
5
300
200
100
0
100
f2
10
0
10
5
300
200
100
0
100
f3
10
0
10
5
300
200
100
0
100
f4
10
0
10
5
300
200
100
0
100
f5
d
B
10
0
10
5
300
200
100
0
100
f6
10
0
10
5
300
200
100
0
100
f7
10
0
10
5
300
200
100
0
100
f8
10
0
10
5
300
200
100
0
100
f9
[rad/s]
d
B
10
0
10
5
300
200
100
0
100
f10
[rad/s]
10
0
10
5
300
200
100
0
100
f11
[rad/s]
10
0
10
5
300
200
100
0
100
f12
[rad/s]
Figure 3.6: Fault response in residual. Residual generator designed with meth-
ods from Chapter 3. Note how the gains from the decoupled faults f
9
, . . . , f
12
is signicantly lower than the gains from the other faults.
46 Chapter 3. Residual Generation Based on Linear Models
order lter.
Figure 3.6 shows the absolute values of the transfer functions from the dif-
ferent faults to the residual. Faults 1 to 8 is the sensor faults and 9 to 12 the
actuator faults. It is clear that f
9
to f
12
(the actuator faults) are attenuated
according to design specications.
To illustrate the numerical diculties in this example, a design is also per-
formed with a well-known method, the Chow-Willsky scheme. Performing the
same design with the basic Chow-Willsky design method on a balanced realiza-
tion of the model gives Figure 3.7 which should be compared to Figure 3.6. It
10
5
10
0
10
5
300
200
100
0
f1
d
B
10
5
10
0
10
5
200
150
100
50
0
f2
10
5
10
0
10
5
150
100
50
0
f3
10
5
10
0
10
5
300
200
100
0
f4
10
5
10
0
10
5
300
200
100
0
f5
d
B
10
5
10
0
10
5
300
200
100
0
f6
10
5
10
0
10
5
300
200
100
0
f7
10
5
10
0
10
5
400
300
200
100
0
f8
10
5
10
0
10
5
100
50
0
50
100
f9
[rad/s]
d
B
10
5
10
0
10
5
100
50
0
50
f10
[rad/s]
10
5
10
0
10
5
150
100
50
0
f11
[rad/s]
10
5
10
0
10
5
100
50
0
50
f12
[rad/s]
Figure 3.7: Fault response in residual. Residual generator designed with the
Chow-Willsky scheme. The decoupling has failed since the gains fromf
9
, . . . , f
12
is not lower than the gains from f
1
, . . . , f
8
.
is clear that the decoupling of f
9
to f
12
here has failed and this is because of
severe numerical problems during design.
In the minimal polynomial basis approach there were some freedom in se-
lecting the numerator, since there were 4 to choose from. This kind of freedom
also exists in the Chow-Willsky approach, however the solution is in the ba-
sic form over-parameterized so there are dependent numerators. In this rather
large example, where in the case of the minimal polynomial basis approach there
were 4 numerators, the standard Chow-Willsky solution gives 211 numerators.
Because of the large over-parameterization not all numerators were evaluated
but when generating Figure 3.7 a representative numerator was chosen. Note
that the Chow-Willsky solution gives all numerators, thus also those of mini-
mal order. The minimal solution can be obtained by selecting a clever linear
3.7. Descriptor systems 47
combination of all 211 numerators. However, this operation is non-trivial and
is of the same complexity order as solving the original design problem. Readers
interested in details regarding the relationship between the minimal polyno-
mial basis approach and Chow-Willsky schemes are referred to (Nyberg, 1999b)
where this is investigated in great detail.
3.7 Descriptor systems
This section shows how the design and analysis algorithms developed in this
chapter easily can be extended to cover also descriptor systems, or generalized
state-space systems. Pros and cons of descriptor formulations is discussed in
(M uller, 2000). Linear descriptor systems are described on state-space form as
E x = Ax +B
u
u +B
d
d (3.23a)
y = Cx +D
u
u +D
d
d (3.23b)
where the only dierence from a standard state-space description is the matrix
E where E can be square, non-square, singular, or non-singular. A model with a
square, non-singular E can of course be transferred to a state-space description.
However, due to numerical reasons this may not be a suitable operation. In
(Sincovec et al., 1981) it is argued that the descriptor formulation in general has
numerical advantages compared to an ordinary state-space descriptions and in
addition, it is more closely related to physical variables and the physical system
structure.
A descriptor system is solvable, i.e. a unique solution exists given input
signals and initial conditions, if and only if the matrix pencil sE A is regular
(or equivalently matrix sE A is square and full rank) (Yip and Sincovec,
1981). A solvability assumption of (3.23), sometimes referred to as a regularity
assumption, is often made when analyzing descriptor systems for fault diagnosis
(Kratz et al., 1995; Shields, 1994). However, in general, for example for under-
modeled or dierential-algebraic systems, matrix E need not even be square
and it is shown in e.g. (M uller and Hou, 1993; Hou, 2000) that a solvability
assumption is not necessary in observer design or residual generator design.
Therefore, no solvability assumption is made here either, i.e. the matrix pencil
sE A can be square or non-square, singular or non-singular.
The design methodology for descriptor systems is exactly as described for
non-descriptor systems in Section 3.1 and all residual generators can be param-
eterized as in (3.6) and (3.7). A slight modication of the design is needed
though, since for non-solvable descriptor systems, it is not even possible to
write the system on a transfer function form like in (3.1). This minor dierence
is dealt with in the next section where a set T of all polynomial consistency
relations is dened instead of the left null-space of matrix M(s). With this
notational dierence, the only dierence from the non-descriptor case is a slight
modication on how to compute a basis for T/the left null-space of M(s) which
48 Chapter 3. Residual Generation Based on Linear Models
will be discussed next. In Chapter 5, the design algorithm is extended to con-
sider stochastic linear models where it will also be shown how the algorithm
handles stochastic descriptor systems.
3.7.1 Computing N
L
(M(s)) for descriptor systems
If the descriptor system (3.23) is solvable, it can be written on a, possibly non-
proper, transfer-function form. Thus, just as for non-descriptor systems, the
system description can be on transfer function form (3.1) or on a generalized
state-space form (3.23). In case of a transfer function model, the frequency
domain solution from Section 3.2 is directly applicable. The only dierence
from the non-descriptor case is that M(s) might be non-proper, which does not
inuence the computation of N
M
(s).
The state-space solution from Section 3.2 of course need a slight modica-
tion since a new state-space matrix E has been introduced. Here however, a
fundamental dierence from the non-descriptor case appears. The singularity
of matrix E has implications on minimality properties and this will be inves-
tigated further in this section. The system matrix M
s
(s) from (3.10) is in the
descriptor case replaced by
M
s
(s) =
_
C D
d
(sE A) B
d
_
(3.24)
When the system was written on transfer function form we saw that for all poly-
nomial row-vectors F(s) N
L
(M(s)), a consistency relation could be formed
as
F(s)
_
y
u
_
= 0
However, as indicated previously, in the general case it may not even be possible
to write the system on transfer function form, a replacement for the space
N
L
(M(s)) is needed that characterizes all consistency relations. From now on,
let the set T, over rational functions, be dened by
T = F(s) R
1(m+k
u
)
(s). < y, u > satises (3.23) F(s)
_
y
u
_
= 0
A version of Theorem 3.1, the theorem used for design based on a state-space
description of the process, for descriptor systems is then given by
Theorem 3.7. Assume that the model equations (3.23) is linearly indepen-
dent and let the rows of a polynomial matrix V (s) form a polynomial basis for
N
L
(M
s
(s)). Then the rows of W(s) = V (s)P form a polynomial basis for T
where P is given by (3.11).
Proof. The proof idea is to rst prove that W(s) spans T, and then show that
the rows are linearly independent.
3.7. Descriptor systems 49
First, note that for the descriptor system (3.23) it holds that
P
_
y
u
_
= M
s
(s)
_
x
d
_
(3.25)
and that matrix F(s) T if and only if
x, d; M
s
(s)
_
x
d
_
= P
_
y
u
_
F(s)
_
y
u
_
= 0
The spanning property of W(s) is proven by observing that:
Measurements satisfy model equations x, d; M
s
(s)
_
x
d
_
= P
_
y
u
_

P
_
y
u
_
Im M
s
(s) N
M
s
(s)P
_
y
u
_
= 0 (3.26)
Thus, any F(s) T can be written as F(s) = (s)W(s), i.e. W(s) spans T.
The last equivalence in (3.26) is due to the fact that a vector x is in the column
space of matrix A if and only if x is orthogonal to the complementary space of
A, i.e. x Im A N
A
x = 0 where N
A
is a basis for the left null-space of A.
The linear independence of the rows in W(s) is easily seen by observing the
following matrix:
V (s)[P M
s
(s)] = V (s)
_
I D
u
C D
d
0 B
u
(sE A) B
d
_
= [W(s) 0] (3.27)
Since V (s) is a polynomial basis for N
L
(M
s
(s)), it has full row-rank. Also, the
assumption on the linear independence of the descriptor model (3.23) gives that
the matrix [P M
s
(s)] has full row rank and multiplication of two full row-rank
matrices gives a full row-rank matrix. Thereby is linear independence of the
rows of W(s) proved, i.e. W(s) has been proven full-rank and to span T thus
ending the proof.
Remark: If the assumption on linear independence of the descriptor equa-
tions (3.23) is not fullled, it is evident from the proof of Theorem 3.7 that W(s)
may not be a basis even though it spans T. So, linear independence among the
model equations might lead to an over-parameterized solution.
Also, it is important to note that the W(s) in the theorem is not a minimal
polynomial basis, it is merely a basis for T. The two properties needed for
W(s) to be a minimal polynomial basis is according to Theorem 3.B.1, row-
reducedness and irreducibility. The matrix W(s) can always be made row-
reduced and irreducible by a full-rank transformation matrix T(s)
W
min
(s) = T
1
(s)W(s)
50 Chapter 3. Residual Generation Based on Linear Models
where T(s) is e.g. the greatest left divisor of W(s). However, if T(s) has non-
stable zeros, feasible residual generators can not (unless cancellation of the
non-stable zeros occurs) be parameterized as
Q(s) = c
1
(s)(s)W
min
(s) (3.28)
This is because of unknown initial conditions, thus it is not always desirable to
make W(s) an irreducible basis. This is further discussed in Section 3.7.3 where
inuence from non-zero initial states are considered. If W(s) is irreducible but
not row-reduced, a minimal polynomial basis can always safely be obtained by
a unimodular transformation T(s) (which of course is stable). This transfor-
mation can be done in several ways, see (Kailath, 1980, p. 386) for principles
and e.g. (Polynomial Toolbox 2.5 for Matlab 5, 2001) for numerically stable
algorithms.
For state-space systems, the irreducibility property was ensured by assum-
ing that the pair A, [B
u
B
d
] was controllable. A similar condition for the
descriptor case is given by the following corollary
Corollary 3.2. Matrix W(s) in Theorem 3.7 is irreducible if E, A, [B
u
B
d
]
is R- or C-controllable.
Proof. If E, A, [B
u
B
d
] is R-controllable, then the bottom block-row of matrix
(3.27) is irreducible according to Theorem 3.B.8 which gives that [P M
s
(s)] is
irreducible. The product of two full row-rank, irreducible matrices results in a
irreducible matrix and since V (s) is irreducible by denition, W(s) is irreducible.
Theorem 3.B.8 and Theorem 3.B.9 directly gives that C-controllability implies
R-controllability and therefore is also C-controllabilty a sucient condition on
irreducibility of W(s).
Another important topic is fault detectability. Nyberg (2000) provides a
thorough treatment on criterions for fault detectability for non-descriptor sys-
tems based on the minimal polynomial basis approach. It is trivial to show that
the criterions is identical for descriptor systems, e.g. a fault is detectable if and
only if
Im
_
D
f
B
f
_
, Im
_
C D
d
(sE A) B
d
_
which can be controlled with a simple rank test.
3.7.2 Design example
The design example is taken from (Hou, 2000), where a descriptor model of
a three-link planar manipulator is used, see Figure 3.8. The process works
by moving the end eector repeatedly from point A to point B, e.g. cleaning
a facade. The manipulator is equipped with three actuators that can apply
torques at all three joints. Four sensors is used measuring the height of the end
eector, the contact force in the x direction, and tracking signals. The fault-free
3.7. Descriptor systems 51
B
A
y
x
Figure 3.8: The three-link manipulator
model is stated on descriptor form in (Hou and M uller, 1996). The model has
11 states: Cartesian coordinates of the end eector (3 states), derivatives of
the Cartesian coordinates (3 states), two Lagrangian multipliers (2 states), and
controller states (3 states).
The model is given on the form:
E x = Ax +B
u
u +B
f
1
f
1
+B
f
2
f
2
y = Cx
The process is subjected to two faults, f
1
and f
2
corresponding to malfunctions
in an actuator and in the tracking reference signal. Numerical values for the
state-space matrices are taken from (Hou and M uller, 1996) and (Hou, 2000)
and are also included in Appendix 3.C. The descriptor system is solvable, i.e.
E is square and the pencil sE A has full rank.
The design goal is to design two residual generators Q
1
(s) and Q
2
(s), where
fault f
2
is decoupled in Q
1
(s) and fault f
1
is decoupled in Q
2
(s). Performing the
design using Theorem 3.7 followed by a transformation to a minimal polynomial
basis gives
N
M
s
(s) =
_
_
0.38s 1.6 1.3 0.99s 49
0.12s + 0.52 0.42 0.027s + 1.1
42s
2
2.4 10
2
s 5.3 10
2
3.7 10
2
0.27s 64
0.098s + 4.5 0 0.99 0.098
0.77s 38 0 0.027 0.77
0.65s + 64 0 0.27 0.65
_
_
52 Chapter 3. Residual Generation Based on Linear Models
Since N
M
s
(s) has three rows, N
L
(M
s
(s)) has dimension 3 and that N
M
s
(s)
has row-degrees 1, 1, 2. This means that there exists exactly three linearly
independent residual generators and a residual generator of minimal order has
1 state. The rst residual generator in this example is formed in accordance
with (3.7) as:
Q
1
(s) =
1
1 +s/3
[1 1 0]N
M
s
(s) (3.29)
The parameterization (s) from (3.6) is here chosen, ad-hoc, to [1 1 0]. For the
second design, the dimension of N
L
(M
s
(s)) is also three and the row-degrees
of N
M
s
(s) is also 1, 1, 2 and a similar design as (3.29) is performed to form
Q
2
(s). Thus, two rst order residual generators was designed, adding low-pass
dynamics c(s) = 1 +s/3.
In the simulations, control variables u
2
and u
3
is chosen to be zeros while u
1
is chosen according to Figure 3.9. The time response of the residual generators
0 5 10 15 20 25 30
0.5
0.4
0.3
0.2
0.1
0
0.1
0.2
0.3
0.4
0.5
t [s]
Figure 3.9: Control signal u
1
during simulations.
is shown in Figure 3.10 and it is clear that the residual generators meet the
specications. The work (Hou, 2000) is a chapter in the book (Patton et al.,
2000) which is an edited book, collecting state of the art methods and algorithms
as of year 2000. It is then interesting to make a few comparisons with the design
made there. The design made by Hou resulted in two 4:th order observers while
with the minimal polynomial basis approach, it was straightforward to nd two
rst order residual generators meeting the design specications. Also, since
this is a quite large example (11 states), some numerical performance of the
design algorithm can be evaluated. Figure 3.10 shows that the residual has no
inuence from control signals and the decoupled faults. Computing the norm
3.7. Descriptor systems 53
0 5 10 15 20 25 30
1
0
1
2
3
4
5
x 10
3
t [s]
(a) The solid line is residual r
1
, the
dashed fault f
1
and the dash-dotted
fault f
2
. Residual r
1
is zero in fault-
free operation and is sensitive to f
1
and
insensitive to f
2
according to specica-
tions.
0 5 10 15 20 25 30
1
0
1
2
3
4
5
x 10
3
t [s]
(b) The solid line is residual r
2
, the
dashed fault f
1
and the dash-dotted
fault f
2
. Residual r
2
is zero in fault-
free operation and is sensitive to f
2
and
insensitive to f
1
according to specica-
tions.
Figure 3.10: Time-responses of the residual generators.
of the transfer functions from u and d to r, which ideally should be 0, gives
that |Q(s)M(s)| 200 dB which is close to machine precision indicating
a successful design. In the design made by Hou, severe numerical problems is
evident since the residual uctuates notably in fault free operation and inuence
from u
1
is clearly visible in the residuals. This indicates that the design problem
studied is a numerically demanding problem for which the algorithm presented
here produces a small and feasible design.
3.7.3 Non-zero initial states
Previously, considering state-space descriptions, initial conditions has been ne-
glected in the design and this works under the, very reasonable, assumption
that only stable residual generators are considered. In such cases, inuence
from the unknown initial conditions will disappear exponentially. However, for
descriptor systems this is not always so which is shown by the following small
example.
Example 3.1
Consider an equation stating that the derivative of an input equals the derivative
of the output.
y = u
54 Chapter 3. Residual Generation Based on Linear Models
A descriptor formulation of this model equation is given by
_
1 1
0 0
_
x =
_
0 0
0 1
_
x +
_
0
1
_
u
y = x
1
Computing the basis W(s) according to Theorem 3.7 gives the expected result
N
M
s
(s)P = [s s]
Here it is immediate that the basis [s s] is not irreducible since it loses rank
for s = 0. It would be tempting to obtain a lower-degree basis by making it
irreducible by factoring out the polynomial s and then form a stable residual
generator
Q(s) =
1
s + 1
[1 1]
But with initial conditions y(0) = y
0
, u(0) = u
0
, and a Laplace limit theorem
we would get
lim
t
r(t) = lim
s0
s
s + 1
(y u) = lim
s0
1
s + 1
(sy su) =
= lim
s0
1
s + 1
(y
0
u
0
) = y
0
y
0
Thus, if y(0) ,= u(0), the inuence from initial conditions will not vanish.
For a general treatment considering arbitrary initial conditions x(0) = x
0
, per-
form a single-sided Laplace transform of (3.23). Description (3.25) then trans-
forms into
M
s
(s)
_
x
d
_
= P
_
y
u
_

_
0
E
_
x
0
Now, assume that N
M
s
(s)P is made irreducible with a full-rank invertible trans-
formation T(s), and then a residual generator is formed according to (3.28). The
internal form of the residual generator is then in fault-free operation
r = c
1
(s)(s)W
min
(s)
_
y
u
_
= c
1
(s)(s)T
1
(s)N
M
s
(s)P
_
y
u
_
=
= (s)T
1
(s)N
M
s
(s)
_
0
E
_
x
0
Now, it is clear that if T(s) has zeros in the closed right half-plane, the inuence
from the initial state x
0
will not vanish (or even grow exponentially). This
could not happen in the state-space case if, according to Theorem 3.1, the pair
A, [B
u
B
d
] was controllable since N
M
s
(s) then would be irreducible to begin
with. For non R-controllable descriptor formulations, we can not guarantee
3.8. Conclusions 55
irreducibility of N
M
s
(s)P. These issues are dealt with in detail in (Frisk and
Nyberg, 2002) for a more general class of linear system descriptions; systems
described by dierential-algebraic linear equations. The general rule is briey
that if the greatest left divisor of N
M
s
(s)P is strictly stable, it is safe to make the
residual generator irreducible, otherwise not. Also, a non row-reduced residual
generator is always safe to make row-reduced since this can be made by a
unimodular transformation T(s) (which by denition is stable since it has no
nite zeros).
3.7.4 Links to observer design
As was discussed in Chapter 2, in the non-descriptor case close relations to
observers exists. To briey point out a close link between observers and residual
generators also in the descriptor case, consider Equation (3.25) which form
a straightforward link for polynomial design of state observers for descriptor
systems. Assume that no unknown inputs are present and that (3.23) is R-
observable. Then, M
s
(s) is given by [C
T
(sE
T
A
T
)]
T
which has full
row-rank for all s. According to Theorem 3.B.11, matrices C and sE +A are
right co-prime which together with Theorem 3.B.10 gives that there exists a
matrix H(s) such that
H(s)
_
C
(sE A)
_
= I
Then, equation (3.25) directly gives that a state-observer is given by
x = c
1
(s)H(s)P
_
y
u
_
where c
1
(s) determines the dynamics, ensures stability, and properness of the
observer. In case there are unknown input signals you may have to settle for
estimating a set of (less than number of states) linear combinations of state-
variables. This is e.g. achieved by left multiplication of (3.25) with a basis for
the left null-space of [D
T
d
B
T
d
]
T
and treating the result as the no-disturbance
case.
3.8 Conclusions
Residual generation for systems described by deterministic models on trans-
fer function, state-space, or descriptor form has been considered. The de-
sign/analysis algorithms described in this chapter is formulated by using poly-
nomial matrices as a fundamental mathematical object, rather than constant
matrices commonly used. This, together with the existence of high performance
computational tools for polynomial matrices, enables the use of polynomial ma-
trix theory not only as a theoretical tool, but also as a design tool. Further, this
makes it possible to state a simple and straightforward design algorithm. The
56 Chapter 3. Residual Generation Based on Linear Models
computational procedure is based on standard operations on polynomial matri-
ces and no diagnosis specic code is necessary. The design procedure consists
of three steps,
1. Compute a basis for the left null-space of a matrix which is directly given
by the model equations.
2. Select the free design variables, i.e. dynamics of the residual generator and
which consistency relations to utilize
3. Realize the residual generator on state-space form for implementation
which can be written in about two lines of Matlab-code.
An important property of the algorithm is that it, with a minimum num-
ber of parameters, parameterizes all possible consistency relations and residual
generators. Also, the polynomial framework proves benecial when performing
e.g. order analysis of residual generators/consistency relations. Upper and lower
bounds for orders of consistency relations is given based on number of states,
sensors, disturbances, and observability indices of the model.
Also, the use of polynomial matrices and the simple design specication
makes it possible to extend the algorithm to directly apply to not only standard
state-space models but also transfer function and descriptor models. Later, in
Chapter 5, the algorithm is extended to include also stochastic state-space and
stochastic descriptor models.
The examples included showed important properties of the approach. Me-
thodology of design was presented in a rst example, where also interpretations
of available design freedom were shown, e.g. in the residual structure choice,
detectability analysis, and minimality issues. Implications of minimal order
residual generators were particularly clear in the second turbo jet-engine exam-
ple where a 4:th order lter was found with the polynomial approach, instead
of a 26:th order lter that would have been found with a method neglecting the
minimality issue. Finally, the third example, taken from (Hou, 2000), shows
how the algorithm is directly applicable also to descriptor models. Numerical
performance of the algorithm is clearly demonstrated in both the turbo jet-
engine example and in the descriptor example, where the algorithm produced a
feasible design where the algorithm in (Hou, 2000) clearly experiences numerical
diculties.
3.A. Standard notions from linear systems theory 57
Appendix
This appendix is intended to serve as a compilation of denitions, theorems,
and basic properties of linear systems, polynomial matrices, and polynomial
bases for rational vector spaces used in this work. Sources describing these
matters in detail are e.g. (Forney, 1975; Kailath, 1980; Chen, 1984) for con-
trol oriented views, and (Lancaster and Tismenetsky, 1985) for a mathematical
presentation.
3.A Standard notions from linear systems the-
ory
Denition 3.A.2 (normal rank). The normal rank of a polynomial matrix
A(s) R
mn
[s] is dened as e.g.
1
1. the number
rank A(s) = max
sC
rank A(s)
2. the number of linearly independent columns (or rows) of A(s)
3. the number of invariant polynomials of the Smith form of A(s)
4. The dimension of the space spanned by the rows/columns
The normal rank is sometimes also referred to as generic rank.
Denition 3.A.3 (rank of a rational matrix). The rank of a matrix M(s)
R
mn
(s) is dened as e.g.
1. number of linearly independent rows/columns with rational coeecients
2. the (normal) rank of N(s) where N(s) is the numerator in a left or right
MFD of M(s)
3. The dimension of the space spanned by the rows/columns
Denition 3.A.4 (matrix pencil). A polynomial matrix M(s) of degree 1 is
called a matrix pencil. Often, a matrix pencil is written as
M(s) = sE +F
where E and F are constant matrices.
Denition 3.A.5 (regular matrix pencil). A matrix pencil M(s) is said to
be regular if M(s) is square and full (normal) rank.
1
There exists several other, equivalent, denitions in the literature, but these are the ones
most suitable for this presentation.
58 Chapter 3. Residual Generation Based on Linear Models
Denition 3.A.6 (irreducible polynomial matrix). A polynomial matrix
M(s) R
mn
[s] is irreducible if it has full rank for all s C.
Denition 3.A.7 (unimodular polynomial matrix). A square, irreducible,
polynomial matrix is called unimodular. A dening property of unimodular ma-
trices is that they have polynomial inverses.
Denition 3.A.8 (row-degree of a polynomial vector). A polynomial
vector v(s) R
1n
[s] has row-degree p if
p = max
j
deg v
i
(s)
Denition 3.A.9 (polynomial basis). A polynomial basis for a rational vec-
tors pace T is represented by a matrix M(s) where the rows spans T and the
order of the basis is dened as the sum of row-degrees of the rows of matrix
M(s). A minimal polynomial basis is thus a polynomial basis for T that mini-
mizes this order.
Denition 3.A.10 (leading row coecient-matrix). A matrix M(s)
R
mn
[s] with row-degrees
1
, . . . ,
m
can always be written
M(s) = S(s)M
hr
+L(s)
where
S(s) = diag s

j
, j = 1, . . . , m
M
hr
= the leading row coecient matrix
The matrix L(s) denotes the remaining terms and is of row-degrees strictly less
than those of M(s).
Denition 3.A.11 (row-reduced matrix). A matrix M(s) is said to be row-
reduced if its leading row coecient matrix is full rank.
Denition 3.A.12 (relatively prime/coprime matrices). Two polynomial
matrices with the same number of columns(rows) are relatively right(left) co-
prime if all their greatest common right (left) divisors are unimodular.
Denition 3.A.13 (R and C controllable descriptor systems). A de-
scriptor system
E x = Ax +Bu
is completely controllable (C-controllable) if one can reach any state from any
initial state. The system is controllable within the set of reachable states (R-
controllable) if one can reach any state in the set of reachable states from any
admissible initial state.
3.B. Standard results from linear systems theory 59
Denition 3.A.14 (Sylvester matrix). Let W(s) be a polynomial matrix of
degree d, i.e.
W(s) =
d

i=0
W
i
s
i
with W
d
,= 0. Then, the q:th order Sylvester matrix for W(s) is dened as
sylv(W(s), q)
_

_
W
0
W
1
W
d
0 0 0
0 W
0
W
1
W
d
0 0
0 0
.
.
.
.
.
.
.
.
.
0 0 W
0
W
1
W
d
_

_
3.B Standard results from linear systems theory
Theorem 3.B.1 (Kailath,1980; Theorem 6.5-10). The rows of a matrix
F(s) form a minimal polynomial basis for the rational vector space they generate,
if and only if F(s) is irreducible and row-reduced.
Theorem 3.B.2. If the rows of F(s) form an irreducible polynomial basis for a
rational vector space F, then all polynomial row vectors x(s) F can be written
x(s) = (s)F(s) where (s) is a polynomial row vector.
Proof. Assume x(s) F, then there exists co-prime polynomials n(s) and d(s)
such that
d(s)x(s) = n(s)F(s)
Now, assume that d(s) has a zeros at s = s
0
. Then it holds that
0 = n(s
0
)F(s
0
)
But since n(s) and d(s) were co-prime, n(s) ,= 0 which in turn means that n(s
0
)
lies in the left null-space of F(s
0
). However, this is a contradiction since F(s)
was assumed irreducible. This means that d(s) can not have any zeros, i.e. d(s)
is proven to be a constant which ends the proof.
Remark: Even though the basis vectors are polynomial, the vector-space
they generate is not a polynomial vector space. This is because there can not
exist such a thing as a polynomial vector space. A standard, formal denition
on a vector space from (Lancaster and Tismenetsky, 1985) read as follows: Let
be a set on which a closed binary operation (+) (like vector addition) is
dened. Let F be a eld and let a binary operation (like scalar multiplication)
be dened from F to . If for there exists an additive zero, additive
inverse, and the operations obey distributive and associative laws, then o is
a vector-space. However, since polynomials have no polynomial multiplicative
inverse, the set of polynomials can not be a eld, only a ring. However, with a
suitable basis, we can still characterize all polynomial elements in the rational
vector-space.
60 Chapter 3. Residual Generation Based on Linear Models
To illustrate the concept of rational vector-spaces and polynomial bases, the
following example has been included.
Example 3.2
Let the rows of the matrix F(s) be a basis for the rational vector-space F.
F(s) =
_
_
s 0 1
1 1 0
0 s 2
_
_
It is clear that F(s) is a basis since det(F(s)) = s , 0, i.e. the matrix has full
rank and therefore, the rows are linearly independent. Any polynomial vector
of dimension 3 will of course belong to F. Consider for example the vector
b
1
(s) =
_
s 0 0

F
This vector can be written as a linear combination of the columns as follows:
b
1
(s) =
_
2 s 1

_
_
s 0 1
1 1 0
0 s 2
_
_
= x(s)F(s)
Here, x(s) happens to be a polynomial vector. In general however, rational
vectors are needed. Consider for example the vector
b
2
(s) =
_
1 0 0

=
_
2
s
1
1
s

_
_
s 0 1
1 1 0
0 s 2
_
_
= x(s)F(s)
In this case, x(s) is rational and there exists no polynomial x(s) such that
b
2
(s) = x(s)F(s).
If the polynomial basis is irreducible, then according to Theorem 3.B.2, only
polynomial x(s):s are needed. An irreducible, but not row-reduced, basis for
the same vector-space F is for example
F

(s) =
_
_
2 1 0
1 1 0
s 0 1
_
_
= T
1
(s)F(s)
which is achieved by extraction of the greatest left divisor T(s) of F(s). Now
b
2
(s) can be written
b
2
(s) = [1 1 0]F

(s)
Theorem 3.B.3 (Kailath,1980;p. 366). The PBH Rank Tests
1. A pair A, B will be controllable if and only if the matrix
[sI A B] has full-rank n for all s C
3.B. Standard results from linear systems theory 61
2. A pair C, A will be observable if and only if the matrix
_
C
sI A
_
has full-rank n for all s C
Theorem 3.B.4 (Kronecker Canonical Form of a matrix pencil).
For any linear matrix pencil A sB, it is possible to nd constant, square,
and nonsingular matrices U and V such that
U(AsB)V = block-diagL

1
, . . . , L

1
, . . . ,

, sJ I, sI F
where
1. F is in Jordan form
2. J is a nilpotent Jordan matrix
3.

L

is a ( + 1) matrix of the form

. .

_
s
1 s
1
.
.
.
s
1
_

_
_

_
+ 1
4. L

=

L
T

The
i
and
i
are called left and right Kronecker indices.
Theorem 3.B.5 (Kailath,1980;Theorem 6.3-13). Let D(s) be a polynomial
matrix of full row-rank, and for any polynomial vector p(s), let
q(s) = p(s)D(s)
Then, D(s) is row-reduced if and only if
deg q(s) = max
i:p
i
(s)0
[deg p
i
(s) +
i
]
where p
i
(s) is the i:th entry of p(s) and
i
is the degree of the i:th row of D(s).
This result is also called the predictable-degree property of row-reduced matrices.
Theorem 3.B.6 (Kailath,1980; Theorem 6.3-12). Let N(s) R
nm
and
D(s) R
nn
[s] be full-rank and row-reduced. Then the solution H(s) to the
matrix equation
H(s)D(s) = N(s)
is proper if and only if each row of N(s) has row-degrees less than or equal to
corresponding row in D(s).
62 Chapter 3. Residual Generation Based on Linear Models
Theorem 3.B.7 (Division Theorem for Polynomial Matrices; Theo-
rem 6.3-15 in Kailath,1980.). Let D(s) be an mm nonsingular polynomial
matrix. Then, for any pm polynomial matrix N(s), there exists unique polyno-
mial matrices W(s), R(s) such that N(s) = D(s)W(s)+R(s) and D
1
(s)R(s)
strictly proper. Matrices W(s) and R(s) is called the polynomial matrix quotient
and remainder of D
1
(s)N(s).
Theorem 3.B.8 (R-controllability;Yip and Sincovec,1981). A descriptor
model
E x = Ax +Bu
is R-controllable if and only if the matrix [(sE A) B] has full row-rank for all
s C.
Theorem 3.B.9 (C-controllability;Yip and Sincovec,1981). A descriptor
model
E x = Ax +Bu
is C-controllable if and only if the matrix [(sE A) B] has full row-rank for all
s C and matrix [E B] has full row-rank.
Theorem 3.B.10 (Bezout identity). Polynomial matrices D(s) and N(s)
will be right co-prime if and only if there exists polynomial matrices X(s) and
Y (s) such that
X(s)D(s) +Y (s)N(s) = I
Theorem 3.B.11 (Rank criterion for relative primeness). Polynomial
matrices N(s) and D(s) will be right coprime if and only if matrix
_
D(s)
N(s)
_
has full rank for all s C. Similarly, N(s) and D(s) will be left coprime if and
only if matrix
_
D(s) N(s)

has full rank for all s C.

3.C. State-space matrices for descriptor example 63
3.C State-space matrices for descriptor example
E =
_

_
1 0 0 0 0 0 0 0 0 0 0
0 1 0 0 0 0 0 0 0 0 0
0 0 1 0 0 0 0 0 0 0 0
0 0 0 18.75 7.95 7.95 0 0 0 0 0
0 0 0 7.95 31.82 26.82 0 0 0 0 0
0 0 0 7.95 26.82 26.82 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 1 0 0
0 0 0 0 0 0 0 0 0 1 0
0 0 0 0 0 0 0 0 0 0 1
_

_
A =
_

_
0 0 0 1 0 0
0 0 0 0 1 0
0 0 0 0 0 1
68.7 98.96 77.74 601.16 43.92 107.77
45.23 402.43 337.54 906.97 177.27 179.24
4.48 339.82 219.17 697.11 149.56 360.37
1 0 0 0 0 0
0 0 1 0 0 0
0 1 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
1 0 68.83 34.83 6.41
0 0 280.46 58.29 24.22
0 1 236.89 48.76 69.93
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
1 0 0 0 0
0 1 0 0 0
_

_
B
u
=
_
_
0 0 0 0 0 0 0 0 1 0 0
0 0 0 0 0 0 0 0 0 1 0
0 0 0 0 0 0 0 0 0 0 1
_
_
T
C =
_

_
0 1 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 1 0 0
0 0 0 0 0 0 0 0 0 1 0
0 0 0 0 0 0 0 0 0 0 1
_

_
B
1
=
_
0 0 0 36.334 76.914 76.914 0 0 0 0 0

T
B
2
=
_
0 0 0 0 0 0 0 0 1 0 0

T
64 Chapter 3. Residual Generation Based on Linear Models
4
Residual Generation Based on
Non-Linear Models
The perhaps most common approach to residual generation for non-linear mod-
els is to use state-observers based on a state-space description of the process
dynamics. However, as was discussed in Section 2.2, consistency relations is
an equally valid notion for non-linear systems and have been used in a num-
ber of works for supervising non-linear systems e.g. (Krishnaswami et al., 1994;
Krishnaswami and Rizzoni, 1994; Guernez et al., 1997; Zhirabok, 1999; Zhang
et al., 1998). In the linear case, consistency relations proved to be benecial
in the analysis and design of residual generators. Therefore, in this chapter, a
systematic design procedure for non-linear system is outlined, exhibiting strong
similarities with the linear design method described in Chapter 3 and with
similar design freedom. The goal, to state a constructive design algorithm for
non-linear systems with computer support is rather ambitious and not yet re-
alistic. To still reach for this goal, the class of non-linearities is restricted and
only a class of nonlinear systems is considered, systems described by polynomial
dierential-algebraic equations.
Section 4.1 describes the non-linear design problem and denes the class
of models that is considered, polynomial dierential-algebraic systems. Then,
basic notation from standard elimination theory, needed for the design proce-
dure, is introduced in Section 4.2 and the computational tool is exemplied.
Section 4.3 describes a systematic design procedure for non-linear consistency
relations and Section 4.4 addresses the problem on how to compute a residual
based on the consistency relation. Section 4.5 proceeds to describe how isolabil-
ity analysis can be performed on models in this class of systems. One major issue
that, even for moderately sized models, becomes a problem is the computational
65
66 Chapter 4. Residual Generation Based on Non-Linear Models
complexity of the design procedure. Section 4.6 shows how structural analysis
can be used to manage this complexity. Finally in Section 4.7, the proposed
design procedure is applied to a non-linear model in a noisy environment.
4.1 Problem formulation
The model is a set of polynomial dierential-algebraic equations:
g
i
( y, u,

d,

f, x) = 0 i = 1, . . . , p (4.1)
where u is the control vector, y the measurement vector, d disturbance vector,
f the fault vector, and x a vector of unknown internal states. The notation y
denotes y, y, y, . . . and u,

d,

f, x correspondingly for u, d, f, and x. The type
of non-linearities considered is polynomials, i.e. the g
i
in (4.1) are polynomials
in y, u,

d,

f, and x.
The design problem is, based on the model description (4.1), to nd a consis-
tency relation that satises the design specications and then form a realizable
residual generator based on that consistency relation. As before, all faults that
are to be decoupled and any modeled disturbances are collected in a vector d,
and the faults we wish to detect are collected in a vector f. Thus, given a set
of model equations (4.1), nd a computable quantity r that is a function of y
and u only, such that when

f = 0 it holds that
r( y, u) = 0

d, x
4.1.1 Elementary functions as polynomials
The restriction to polynomial non-linearities is not as restrictive as it may seem,
many non-polynomial non-linearities can be rewritten in polynomial form, e.g.
y = sin u y
2
u
2
(1 y
2
) = 0
These two forms are not really strictly equivalent, there exists a number of solu-
tions that satises the right hand side but not the left hand side. For example,
all solutions on the form y(t) = sin(u(t) +c) where c is any constant, is a solu-
tion to the dierential equation on the right hand side. In (Lindskog, 1996) a
translation table is included, where a polynomial description is provided for all
elementary functions. This is done by describing the function as the solution of
a polynomial dierential equation. More formally stated; the elementary func-
tions are dierentially algebraic and since many of the commonly used functions
e.g. trigonometric, inverse trigonometric, exponential, logarithmic functions are
elementary functions, quite general systems can be handled within the frame-
work of polynomial systems. Also, any smooth non-linearity can be described
by polynomials to an arbitrary accuracy by Taylor expansion.
4.2. Basic elimination theory 67
4.2 Basic elimination theory
This section will introduce some notation and a theorem from basic elimination
theory that is needed for the design procedure in Section 4.3. It will not describe
elimination theory other than with a small example. For more details, see e.g.
(Cox et al., 1991).
Let k[x
1
, . . . , x
n
] denote the set of polynomials in variables x
1
, . . . , x
n
with
coecients in k, e.g.
x
1
x
2
+x
3
3
2x
2
1
x
2
x
3
R[x
1
, x
2
, x
3
]
An important concept, ideal, is now dened:
Denition 4.1 (Ideal). Let g
1
, . . . , g
s
be polynomials in k[x
1
, . . . , x
n
]. Then
denote
I =< g
1
, . . . , g
s
>=
_
s

i=1
h
i
g
i
: h
i
k[x
1
, . . . , x
n
]
_
I is called the ideal generated by the polynomials g
1
, . . . , g
s
.
This means that I is the set of all linear combinations of the polynomials g
i
with polynomial coecients h
i
. It is helpful to think of ideals using an analogy
with subspaces from linear algebra. Both ideals and subspaces are closed under
addition and multiplication, except that for subspaces we multiply with scalars
but for ideals we multiply with polynomials. However, the analogy should not
be taken too far. For example, in linear algebra a basis always consists of
linearly independent vectors whereas for an ideal a basis is only concerned with
spanning, independence is not mentioned. This is natural since it is in fact
easy to prove that for any two polynomials g
i
and g
j
, zero can be written as
a linear combination with polynomial coecients. Thus, if f is an element in
< g
1
, . . . , g
s
> with coecients h
i
, i.e.
f =
s

k=1
h
i
g
i
(4.2)
the coecients h
i
is not unique.
Connections to the linear case
Expression (4.2) indicates relations with polynomial elements in rational vector
spaces which appeared when deriving linear consistency relations in Chapter 3.
If the rows of F(s) span a rational vector-space T, a vector f(s) T if and only
if there exists co-prime polynomials n(s), d(s) such that d(s)f(s) = n(s)F(s).
However, if F(s) is irreducible, Theorem 3.B.2 gives that f(s) T if and only
if there exists a polynomial n(s) such that
f(s) = n(s)F(s)
68 Chapter 4. Residual Generation Based on Non-Linear Models
This irreducibility property was important in equation (3.6) where all linear
consistency relations were parameterized by the polynomial row-vector (s).
This unique polynomial parameterization was only possible due to the fact that
N
M
(s) was irreducible. The polynomials g
i
in Denition 4.1 therefore share
similar traits with irreducible polynomial bases from Chapter 3.
Main elimination theorem
Now the main theorem used in the design, the well known elimination theorem,
is stated.
Theorem 4.1 (Elimination Theorem).
Let I k[x
1
, . . . , x
n
] be an ideal and let G be a Gr obner basis of I with respect
to lex order x
1
~ x
2
~ ~ x
n
. Then, for every 0 m n, the set
G
m
= G k[x
m+1
, . . . , x
n
]
is a Gr obner basis of the m:th elimination ideal I
m
I
m
= I k[x
m+1
, . . . , x
n
]
This means that all polynomials, where variables x
1
, . . . , x
m
have been elim-
inated, can be written as in Denition 4.1 where g
1
, . . . , g
s
are the polynomials
in G
m
. Now follows a small example showing how the elimination theorem can
be used in practice.
Example 4.1
Consider the following set of equations
x
2
+y +z = 1
x +y
2
+z = 1
x +y +z
2
= 1
Suppose we wish to eliminate all variables but z, then according to the elimi-
nation theorem, rst compute a Gr obner basis for the ideal
I =< x
2
+y +z 1, x +y
2
+z 1, x +y +z
2
1 >
with lex order x ~ y ~ z. In a Mathematica session, this is done by:
In[1]:= F={-1+x^2+y+z, -1+x+y^2+z, -1+x+y+z^2};
GroebnerBasis[F,{x,y,z}]
Out[2]= {-1 + x + y + z^2, y^2 + z-z^2-y,
2yz^2 + z^4-z^2, z^2 -4z^3 + 4z^4-z^6}
4.3. Design using elimination theory 69
Thus, according to the elimination theorem I k[z] =< z
2
4z
3
+ 4z
4
z
6
>.
Note the triangular structure of the Gr obner basis, the last polynomial is a
function of z only, the second and third depends on y and z while the rst
depends on all three x, y, z.
Another useful property of the Gr obner basis is given by the extension theorem
(Cox et al., 1991), which gives a procedure to determine all solutions to the
polynomial equation system. First, solve for z in the fourth equation, then
substitute the roots into the second and third polynomial to obtain possible
solutions y. Finally, solve for possible x using the rst equation by substituting
possible y and z.
4.3 Design using elimination theory
The model is a set of polynomial equations on the form:
g
i
( y, u,

d,

f, x) = 0 i = 1, . . . , p (4.3)
The basic step in the design algorithm is to manipulate the model equations (4.3)
such that a consistency relation is obtained where all disturbances

d (including
faults that are to be decoupled) and internal states x have been eliminated. This
relation can then be used to form a residual generator where the computational
form is a function of y and u only and the internal form is a function of

f
only. Calculating a Gr obner basis and using Theorem 4.1 provides a systematic
procedure to perform such manipulations. All derivations in this section are
made in the time-continuous case. However, corresponding results for the time-
discrete case is immediate by exchanging the time dierentiation operator with
the time shift operator. A dierence between the time-discrete and the time-
continuous case occurs when the consistency relation is used to realize a residual
generator. More comments on this topic are included in Section 4.4.
Gr obner bases is a non-dierential tool, i.e. the dierentiation operator is
not considered and when applying Gr obner basis methodology to systems like
(4.1), x and x are seen as two completely unrelated variables. Of course, the
dierentiation operator is essential when analyzing dynamic systems. To be able
to use non-dierential tools like Gr obner bases, the dierentiation has to be done
by-hand. Thus, the model equations (4.3) is dierentiated a number of times
and a new, larger set of equations is obtained. When dierentiating the model
equations, new model equations are obtained. It is however dicult to know
when to stop dierentiating. For an n:th order linear system, n dierentiations
is enough to extract all information in the model equations. For non-linear
systems, no such limit exists. You may even have to dierentiate innitely
many times to extract all information. In e.g. (Jirstrand, 1998), the input-
70 Chapter 4. Residual Generation Based on Non-Linear Models
output ideal of the rst order system
x
1
= x
2
x
1
x
2
= 0
y = x
1
is proven to be non-nitely generated
1
. Therefore, when using the method out-
lined here, residual generators and consistency relations up to a certain order
are considered, equal to the number of times the model equations were dier-
entiated.
Another approach is to use dierential tools like Ritts characteristic sets
(Ritt, 1950) to perform the elimination. However, the dierential theory is
currently not as developed as the non-dierential, and dierential bases is an
active research topic.
4.3.1 Algorithm outline
The rst step in the design procedure is to obtain a set of equations consisting
of the original equations and dierentiated model equations. The next step is
to compute the (reduced) Gr obner basis for the elimination ideal where

d, x has
been eliminated. Denote this basis with
GB =< b
1
, . . . , b
r
>
where b
i
are polynomials in all variables but

d and x. This Gr obner basis
GB means that any polynomial, analytical relation, inferred from the model
equations (4.3) without using the dierentiation operation, where the unknown
signals are eliminated can be written as
r

i=1
h
i
b
i
(4.4)
for a set of polynomials h
i
. Each of the r polynomials in GB, or any combination
as in (4.4), can be used to form a consistency relation where the fault-free
relation is
r

i=1
h
i
b
i

f
i
=0
= 0
The h
i
polynomials are design variables available to the designer and this choice
should be seen as a non-linear equivalent to the design matrix (s) in (3.6) for
the linear design case. These can be used e.g. to shape the fault response in the
residual or select the residual structure, i.e. to get sensitivity in the residual to
a desired set of faults.
1
In the example, no input signal was included. The input-output ideal then consists of
relations including y, y, . . . , only.
4.3. Design using elimination theory 71
As usual, the consistency relations includes higher order derivatives of y and
u which are not known. Using the consistency relation as a residual genera-
tor is, in this non-linear case, non-trivial in the lack of a general realization
theory. This issue is further discussed in Section 4.4. If however, these higher
order derivatives were known, the computational and internal form of a residual
generator could be formed as:
r
comp
( y, u) =
r

i=1
h
i
b
i

f
i
=0
(4.5)
r
int
(

f, y, u) =
r

i=1
h
i
b
i
r
comp
(4.6)
For examples on computational and internal forms from the upcoming design
example, see page 80.
A design variable, apart from choosing h
i
polynomials in (4.5), is the variable
ordering when calculating the Gr obner basis. Dierent variable orderings can
produce very dierent bases with highly dierent complexity and computing
time. The issue of variable ordering is not pursued further here, see e.g. (Boege
et al., 1986) for examples of heuristic optimal variable orderings. Here, a
natural way to choose variable ordering can be stated as:
1. Order the variables to be decoupled as the highest ranked variables, i.e.
the faults to be decoupled, the disturbances, and the unknown internal
states should be ordered rst when using Theorem 4.1.
2. Next, order the known variables and their derivatives. To get a consis-
tency relation of as low order as possible, order high order time-derivatives
higher than low-order time-derivatives.
3. Finally, order the faults not to be decoupled as the lowest ranked variables
to get as good fault sensitivity properties in the consistency relations as
possible.
Related approaches
Related work has been published by (Zhang et al., 1998) where consistency
relations are derived using Ritts algorithm (although, they could just as well
have used Gr obner bases as described above). However, in the Zhang et al.
(1998) paper, faults are not decoupled in the residual, only the state variables.
The reason for this is that additional decoupling of fault signals results in higher
order consistency relations, rendering a more dicult realization problem. Is
also implies that fault isolation is not performed using structured residuals,
instead local statistical properties of the consistency relations are utilized and
fault isolation is performed using the so called local approach (Basseville et al.,
1987; Basseville and Nikiforov, 1993). The gain using the local approach rather
72 Chapter 4. Residual Generation Based on Non-Linear Models
than structured residuals for fault isolation is, as was noted above, an easier
realization problem due to lower order derivatives. The prize to pay is fault
isolation performance. Zhang et al. concludes for large parameter changes,
FD still works in most practical situations (and even becomes easier), but FI is
likely to become incorrect
2
.
4.4 Realizable residual generator
Note that y, u etc. normally appear in r
comp
( y, u). Usually, these derivatives
are not known. This problem can be handled by realization theory or approxi-
mations. Realization is exemplied below, but rst some simple approximation
methods are recalled.
A simple method to compute the residual is to approximate dierentiated
variables with

y =
s
sT
d
+ 1
y
with a suitably chosen T
d
. This simple method may not be a satisfactory so-
lution in many cases. Another, less noise sensitive, solution may be to rst
t suitable analytical functions, e.g. cubic splines, to possibly low-pass ltered
data. The derivatives can then be estimated by extracting the derivatives ana-
lytically from tted analytical expressions.
Other possibilities than estimating the derivatives, not explored further
here other than the brief discussions below, are realization theory and time-
discretization of the time-continuous model.
Realization theory
In the linear case, it was straightforward to add stable linear dynamics to form
a realizable residual generator based on a consistency relation where the added
dynamics is free to choose as long as it is stable. Thus, it would be desirable
to approach the non-linear problem in the same way, e.g. add stable (possibly
linear) low-pass dynamics
r +
1
r +
2
r = r
comp
and nd an explicit state-space realization of the residual generator. Unfortu-
nately, realization theory with inputs is dicult, even when only polynomial
systems are considered (Forsman, 1991). Below, a small example where the
realization step is immediate demonstrates the idea on how to use realization
theory.
2
Zhang et al. uses the abbreviations FD for fault detection and FI for fault isolation.
4.4. Realizable residual generator 73
Example 4.2
Consider a system described by the dierential equation
x = sin
3
(x)(u +f)
2
y = x + (u +f)
where f is an actuator fault that has to be supervised. A consistency relation
for the system above can easily be derived by dierentiating the measurement
equation and eliminating the state-variable x.
y + sin
3
(y u)u
2
u = h(y, u, f) (4.7)
where
h(y, u, f) =

f sin
3
(y u f)(u +f)
2
+ sin
3
(y u)
2
u
2
Thus, if y and u were known, the left hand side of (4.7) could be used to compute
a residual that could be used to detect the actuator fault.
y + sin
3
(y u)u
2
u =
_
0 f 0
c(t) , 0 f , 0
3
Here, the time derivatives are assumed to be unknown. In the light of the
previous discussion, add stable rst-order linear dynamics to the left hand side
of (4.7), i.e.
r + r = y + sin
3
(y u)u
2
u (4.8)
with > 0 and try to nd an explicit state-space representation of (4.8) with
y and u as inputs and the residual r as output. The choice of corresponds
to c(s) = 1 + s in equation (3.7). In this particular case this is easy and a
realization is e.g.
z =
1

z
1

(y u) + sin
3
(y u)u
2
r =
1

z +
1

(y u)
The internal form for this lter is
r + r = h(y, u, f)
which will be 0 in the fault-free case and non-zero when a fault occurs.
As noted above, general non-linear realization of input-output descriptions is a
dicult task. However, there exist an important dierence between the basic
realization problem (Sadegh, 2001; Isidori, 1995; Sontag and Wang, 1990; der
3
This is not entirely true, of course there exists particular solutions f(t) to the dierential
equation 0 = h(y, u, f) and for these solutions, the residual will be 0.
74 Chapter 4. Residual Generation Based on Non-Linear Models
Schaft, 1987) and using realization theory to form a computable residual gener-
ator. The dynamics of the residual generator is free for the designer to choose
(as long as it is stable). This means that quite a bit of freedom exists that is
normally not present when studying the realization problem. In the example,
linear dynamics r + r was added. For the general problem, this can be any
dynamics f(r, r, . . . ) such that r = 0 is a globally stable operating point of the
dierential equation
f(r, r, . . . ) = 0
Realization of residual generators based on non-linear consistency relations is a
promising topic where further study is needed.
Time-discretization
A second approach is to transform the original, time-continuous model, to a
time-discrete model. Then, an analogous design can be made using the time-
discrete polynomial model, which results in time-discrete consistency relations.
These time-discrete consistency relations can be directly used as residual gener-
ators (or arbitrary low-pass dynamics can be added) since no time-dierentiated
signals occur, only time-delayed known signals.
4.5 Isolability analysis
An interesting and important topic in diagnosis is diagnosability analysis, e.g.
deciding if it is theoretically possible to separate the modeled faults. This
section serves as an introduction the isolability analysis problem and also to
indicate how theory from related areas of research can be applicable.
For example, if all faults are modeled as constant parameters, isolability
is very closely connected to identiability, and results from the identication
community can be directly applied. An example for polynomial systems is
(Ljung and Glad, 1994), where Ritts algorithm and characteristic sets are used
for identiability analysis of non-linear dynamical systems on polynomial form.
Briey stated, Ritts algorithm can be used to compute a characteristic set for
the considered dierential ideal. In (Ljung and Glad, 1994) it is then shown
that a necessary and sucient condition for global identiability (and here also
for isolability) is that all fault parameters appears linearly in the characteristic
set. Thus, the polynomials in the characteristic sets should appear like
P( y, u)f
i
+Q( y, u) = 0, i = 1, . . . , k
f
for all polynomials including the fault signals/parameters f
i
. When doing isola-
bility analysis, it is important to utilize all restrictions on control signals and
model parameters available to get as precise answers as possible. In particular it
is desirable to only consider real solutions and also to incorporate any inequal-
ity constraints available. The above result from the identication community
4.5. Isolability analysis 75
does not consider only real solutions. However, quantier elimination is such
an interesting tool where it is possible to do real algebra and incorporating in-
equality constraints. See e.g. (Jirstrand, 1997, 1998) for theory and applications
of quantier elimination for controller design/analysis. Here, isolability anal-
ysis by quantier elimination is illustrated only briey, mainly to illustrate a
fruitful view of the isolability analysis problem. But rst, a small general exam-
ple on quantier elimination is included to illustrate the principles of quantier
elimination.
Example 4.3
Consider the problem of deciding for which real a that the equation x
2
+ax +
1 = 0 has real solutions. This problem can be posed and solved by quantier
elimination
x R. x
2
+ax + 1 = 0 [a[ 2
The above simplication is called quantier elimination since variables, bound
by quantiers, is eliminated resulting in an equivalent expression, free of quan-
tied variables.
For the above operation, constructive computational tools exist and are nding
their way into commercial computational algebra packages
4
. Unfortunately, the
complexity of quantier elimination is huge, and at the moment only toy-sized
examples can be handled. However, this view of the isolability analysis problem
is appealing which will be demonstrated next on a small, static, scalar model.
Example 4.4
Consider the static model:
y(t) = (1 +f
1
(t))u(t) +f
2
(t)
It is immediate by observing the model equation that for any f
1
(t) there exists
a f
2
(t) that can reproduce the observed behavior. The converse is not true since
when u(t) = 0, no f
1
(t) can reproduce the output of any f
2
(t) ,= 0. Thus, fault
f
2
is isolable from f
1
but f
1
is not isolable from f
2
. However, if we know that
u(t) > 0, then the faults are not isolable from each other. Now, lets see how
quantier elimination can answer these questions. First, verify that f
1
is not
isolable from f
2
by verifying that for all control signals and all f
1
, there exists
y and f
2
that are consistent with the model equation in both fault modes.
uf
1
f
2
y. y (1 +f
1
)u = 0 y u f
2
= 0
Eliminating the quantiers in Mathematica yields True, i.e. f
1
is not isolable
from f
2
. Exchanging f
1
and f
2
gives
uf
2
f
1
y. y (1 +f
1
)u = 0 y u f
2
= 0
4
The command resolve, which is used in the examples here, is included as experimental
with version 4 of Mathematica.
76 Chapter 4. Residual Generation Based on Non-Linear Models
which resolves to False which means that f
2
is isolable from f
1
. If we want to
know the exact condition on u, when a fault f
1
can explain the behavior caused
by f
2
, simplify the formula
f
2
f
1
y. y (1 +f
1
)u = 0 y u f
2
= 0
which is equivalent to u ,= 0 as expected. Finally, if we know that u(t) > 0, then
it is true that for any f
2
there exists an f
1
that produces the same observables.
This is veried by the formula
uf
2
f
1
y. (u > 0) y (1 +f
1
)u = 0 y u f
2
= 0
which resolves to True when the quantiers are eliminated.
4.6 Complexity management
Although strong computational support exists, the computation of a Gr obner
basis is a time and memory consuming operation that, even for moderately sized
problems, quickly becomes computationally intractable. The quantier elimina-
tion procedure illustrated in the previous section is even more computationally
demanding. Thus, some means of handling complexity is desired.
A promising approach, for the elimination problem, is to perform structural
analysis (Staroswiecki and Comtet-Varga, 2001) on the model equations, iden-
tifying a subset of the model equations from which it is possible to derive a
consistency relation. Since only a subset of the equations is used, a computa-
tionally smaller problem is identied. An example, with equations borrowed
from (Persis and Isidori, 2001), is used to illustrate such a procedure.
Example 4.5
The example is a 4 state model with 3 measurements given by the following 7
equations:
x
1
= x
2
x
2
1
x
2
= x
3
1
x
2
4

1
+
2
x
2
1
u
1
+dx
2
1
x
3
= x
4
x
1
x
4
= 2x
2
x
4
+
2
(u
2
+f)
y
1
= x
1
y
2
= x
3
y
3
= x
4
To derive a consistency relation, the equations need to be dierentiated at least
once. Dierentiating all the equations and feeding the resulting 14 equations
into Mathematica gives an intractable problem (the Gr obner basis computation
4.7. Simulation example: Coupled water tanks 77
did not nish within 30 minutes on a standard PC). With a structural analysis
of the model equations it is, under certain technical conditions, possible to ex-
tract subsets of the model equations where it is possible to derive a consistency
relation. Krysander and Nyberg (2002) call these subsets of equations struc-
turally singular. A structurally singular set of equations is basically m equations
with, at most, m 1 unknown variables. A structurally singular set is said to
be minimal if no proper subset is also structurally singular. In the table below,
a structural description of the model equations is given where an x indicates
that the corresponding variable appears in the equation.
d x
1
x
2
x
3
x
4
y
1
y
2
y
3
u
1
u
2
f
(1) x x
(2) x x x x x
(3) x x
(4) x x x x x
(5) x x
(6) x x
(7) x x
From this table, it is clear that equations 1,4,5,7 (marked with ) form a
structurally-singular set. This is because these four equations only include 3
unknown variables (x
1
, x
2
, x
4
). It is also a minimally structurally-singular set
since any proper subset of these equations is not structurally singular. Thus,
these 4 equations can be used instead of the complete model. Now, using these
4 equations, Mathematica delivers (in 0.02 seconds!) the rst order relation:
y
1
y
3
+ 2 y
1
y
3

2
u
2
=
2
f
which clearly can be used to detect the fault assuming that the realization
problem can be solved. For more details on structural analysis and algorithms
to nd structurally singular sets, the interested reader is referred to (Krysander
and Nyberg, 2002) and the references therein.
4.7 Simulation example: Coupled water tanks
The model used to illustrate the approach is two coupled water tanks, shown in
Figure 4.1. The process is equipped with four sensors, two sensors measuring
the water level in each tank and two sensors measuring the outow of water
from each tank. The process is controlled by a pump.
78 Chapter 4. Residual Generation Based on Non-Linear Models
FS
y
1
FS
y
4
y
2
y
3
u
clogging
Figure 4.1: The simulation process: Coupled water tanks equipped with ow
and water-level sensors.
4.7.1 Modeling
A rst-principles model of the fault-free process, utilizing Bernoullis law for
the ows, is given by:

h
1
= a
1
u a
2
_
h
1

h
2
= a
3
_
h
1
a
4
_
h
2
y
1
= h
1
y
2
= h
2
y
3
= a
5
_
h
1
y
4
= a
6
_
h
2
(4.9)
where a
i
are the model parameters, y
i
the measurements, u the control signal
to the pump, and h
i
the height of water in each tank. The water level in the
tanks can be between 0 and 10, i.e. 0 h
1
, h
2
10.
The faults considered are faults in the actuator, sensors, and clogging in
the pipe between the two water tanks at the point indicated by the arrow in
Figure 4.1. Let f
1
denote an unknown additive fault on the actuator signal and
f
2
, f
3
, f
4
, f
5
additive faults on the four sensors. The clogging fault is modeled by
f
6
where f
6
= 1 represents a completely clogged pipe and 0 < f
6
< 1 represents
partial clogging. Extending model (4.9) with fault models and introducing
4.7. Simulation example: Coupled water tanks 79
auxiliary variables z
i
to get a polynomial description gives:

h
1
= a
1
(u +f
1
) a
2
(1 f
6
)z
1

h
2
= a
3
(1 f
6
)z
1
a
4
z
2

f
i
= 0, i = 1, . . . , 6
z
2
1
= h
1
z
2
2
= h
2
y
1
= h
1
+f
2
y
2
= h
2
+f
3
y
3
= a
5
(1 f
6
)z
1
+f
4
y
4
= a
6
z
2
+f
5
(4.10)
The non-polynomial model (4.9) has been transferred into a polynomial (de-
scriptor) model by adding the auxiliary variables z
i
. Model (4.10) is not equiv-
alent to (4.9) since both negative and positive z
i
can satisfy (4.10), but

h
i
0.
Thus, the model specied by (4.10) covers (4.9) in fault-free mode. For the mod-
els to really be equivalent, conditions z
i
0 has to be added, but since such
inequalities can not be handled by the computational framework used here this
issue is not pursued further. Note that this is not a problem in the design stage,
if however we wish to answer e.g. detectability and isolability questions, these
kind of issues need to be resolved.
Here, only constant faults are considered, i.e.

f
i
= 0. Note that this assump-
tion is not required by the approach, it is only made here to limit the size of
the example.
Finally, the model equations can not be used on the form (4.10) since the
mathematical tool used is non-dierential. Thus, the static equations (last 6
equations in (4.10)) is dierentiated. Therefore, dierentiate the static equa-
tions k times, substitute for

f
i
and

h
i
using the dynamic equations, and collect
the results. The higher order residual generators that is considered, the more
of the dynamic model can be utilized. However, higher orders means a more
dicult implementation problem according to the discussion in Section 4.4. In
this design example, only rst order residual generators are considered. Since
there are 6 static equations in (4.10), the result is 12 equations on form (4.3).
4.7.2 Design
The object of the design is to nd a set of residuals that form a fault isolating
residual structure. To form a isolating inuence structure, 6 elimination ide-
als (and the corresponding Gr obner bases), are calculated where one fault is
eliminated in each ideal. Residual generators are then selected among the basis
polynomials of the calculated Gr obner bases.
The design is performed as described in Section 4.3. The following variable
80 Chapter 4. Residual Generation Based on Non-Linear Models
ordering is used when eliminating f
1
:
f
1
~ h
i
~ z
i
~ z
i
~ y
i
~ y
i
~ u ~ f
2
~ f
3
~ f
4
~ f
5
~ f
6
and corresponding ordering when eliminating the other variables. The variables
y
i
are given a high ordering since it is desirable to eliminate those variables to
get simple computational forms of the residual generators.
Then, 6 consistency relations that creates a fault isolating structure is se-
lected to form the residual generators. The computational form of the residual
generators are:
r
comp
1
= a
2
6
y
2
+y
2
3
y
2
4
a
2
5
y
1
r
comp
2
= a
2
6
y
2
+a
2
a
5
y
3
+ 2y
3
y
3
y
2
4
a
1
a
2
5
u
r
comp
3
= y
2
3
a
2
5
y
1
r
comp
4
= a
2
6
y
2
y
2
4
r
comp
5
= y
2
3
+ 2a
3
y
3
y
2
+a
5
(a
2
4
y
2
y
2
2
) a
2
5
y
1
a
2
3
a
5
y
1
r
comp
6
= a
2
6
y
2
+a
1
a
5
uy
3
+a
6
(a
3
y
3
a
5
y
2
) 2a
5
y
1
y
3
y
2
4
a
4
a
5
y
4
a
2
y
2
3
and the corresponding internal forms are:
r
int
1
=f
2
4
+a
2
5
(f
2
1 + f
2
6
(2 + f
6
)f
6
y
1
) 2f
4
y
3
+
+ 2f
5
y
4
f
2
5
a
2
6
f
3
r
int
2
=2a
2
a
5
f
4
f
6
2a
2
a
5
f
6
y
3
+a
2
a
5
f
2
6
y
3
+ 2f
5
y
4
2f
4
y
3

a
1
a
2
5
(f
1
1 + f
2
6
+ (2 + f
6
)f
6
u)
a
2
a
5
f
4
f
2
6
f
2
5
a
2
a
5
f
4
a
2
6
f
3
r
int
3
=a
2
5
(f
2
1 + f
2
6
(2 + f
6
)f
6
y
1
) + f
4
(f
4
2y
3
)
r
int
4
=2f
5
y
4
f
2
5
a
2
6
f
3
r
int
5
=a
2
3
a
5
(f
2
1 + f
2
6
(2 + f
6
)f
6
y
1
)+
+a
2
5
(f
2
1 + f
2
6
(f
6
2)f
6
y
1
) + f
4
(f
4
2y
3
) 2a
3
f
4
y
2
a
2
4
a
5
f
3
r
int
6
=a
4
a
5
f
5
+ 2a
2
f
4
y
3
+ 2f
5
y
4
+ 2a
5
f
2
y
3
a
1
a
5
(f
4
u + f
1
(f
4
y
3
))
f
2
5
a
2
f
2
4
a
3
a
6
f
4
a
2
6
f
3
By inspection of the internal forms, the inuence structure can be concluded
to be as in Table 4.1. All instances of fault variables in the internal forms has
been marked by shaded boxes. A 1 in column i and row j of the table means
that fault i ideally inuences residual j. The rst aim was to design residual
generator r
i
to be sensitive to all faults but f
i
. This is possible, but at the cost
of more complex residual generators. The design made here is made as simple
as possible while keeping single-fault isolability, i.e. uniqueness of all columns
4.7. Simulation example: Coupled water tanks 81
f
1
f
2
f
3
f
4
f
5
f
6
r
1
0 1 1 1 1 1
r
2
1 0 1 1 1 1
r
3
0 1 0 1 0 1
r
4
0 0 1 0 1 0
r
5
0 1 1 1 0 1
r
6
1 1 1 1 1 0
Table 4.1: Inuence structure
in Table 4.1. Detectability properties of the residual generators can be seen
directly in the internal forms of the residual generators. For example, fault f
2
will be dicult to detect with residual r
6
since in case of a single fault f
2
r
int
6
= 2a
5
f
2
y
3
Thus, r
6
will only deviate from zero if y
3
is non stationary. In the simulation
study that follows below, the process is regulated by a controller making water
levels stationary making detection of f
2
using r
6
unreliable. This does not mean
that f
2
can not be isolated, what it indicates is that the corresponding position
in the decision structure need to be an X.
The computational forms of the residual generators can not be used directly
since the time dierentiated variables are not directly available and need to be
estimated. In the simulations the process is subjected to noise and the time
derivatives are estimated using the spline procedure outlined in Section 4.4. A
three step procedure is used:
1. Low-pass lter the measurements and control signals.
2. Estimate cubic spline polynomials and extract

y
i
from the estimated poly-
nomials.
3. Compute the residuals according to r
comp
i
and low-pass lter the residual
again.
All residuals are also scaled such that a threshold 1 is used for all residuals.
4.7.3 Simulations
In the simulations, a simple proportional controller is used to control the water
level in the upper tank to follow a square reference signal. All 4 sensors are
subjected to rather high intensity measurement noise. Figure 4.2 shows the
water levels in both tanks in a fault-free, but noisy, simulation. Noise-free
simulations gives, as expected, ideal performance of the residual generators.
Figure 4.3 shows the residuals in the fault-free case. All residuals are below the
dotted thresholds. Here, only two fault scenarios are shown, a constant fault in
82 Chapter 4. Residual Generation Based on Non-Linear Models
0 10 20 30 40 50 60 70
1
2
3
4
5
6
7
8
t [s]
Figure 4.2: Water level in the upper tank, y
1
, and the lower tank y
2
, during
fault-free simulations.
0 20 40 60
1.5
1
0.5
0
0.5
1
1.5
r
1
0 20 40 60
1.5
1
0.5
0
0.5
1
1.5
r
2
0 20 40 60
1.5
1
0.5
0
0.5
1
1.5
r
3
0 20 40 60
1.5
1
0.5
0
0.5
1
1.5
r
4
t [s]
0 20 40 60
1.5
1
0.5
0
0.5
1
1.5
r
5
t [s]
0 20 40 60
1.5
1
0.5
0
0.5
1
1.5
r
6
t [s]
Figure 4.3: Residuals in the fault-free case. All residuals are below the dotted
thresholds, i.e. no false alarms during the simulation.
4.7. Simulation example: Coupled water tanks 83
the actuator and a fault in sensor y
2
, measuring the water level in the second
tank. Figure 4.4 shows the residuals when the actuator fault f
1
= 0.1 is induced
at time t = 40 sec. It is clear that the residuals respond as expected by the
inuence structure in Table 4.1, i.e. residuals r
2
and r
6
respond to the fault
while r
1
, r
3
, r
4
, and r
5
does not. The fault is correctly isolated. Figure 4.5
0 20 40 60
1.5
1
0.5
0
0.5
1
1.5
r
1
0 20 40 60
6
4
2
0
2
4
6
r
2
0 20 40 60
1.5
1
0.5
0
0.5
1
1.5
r
3
0 20 40 60
1.5
1
0.5
0
0.5
1
1.5
r
4
t [s]
0 20 40 60
1.5
1
0.5
0
0.5
1
1.5
r
5
t [s]
0 20 40 60
3
2
1
0
1
2
3
r
6
t [s]
Figure 4.4: Residuals when fault f
1
= 0.1 is induced at time t = 40 sec. Resid-
uals r
2
and r
6
respond to the fault while r
1
, r
3
, r
4
, and r
5
does not, i.e. f
1
is
correctly isolated according to Table 4.1.
shows the residuals when fault is sensor y
2
appears abruptly. The fault is also
here induced at time t = 40 sec. Also here, the residuals respond according
to the inuence structure in Table 4.1, i.e. all residuals but r
3
respond to the
fault. The fault is correctly isolated. Simulating the other faults gives similar
results corresponding to the inuence structure. Thus, for this simulation of
a non-linear process, subjected to measurement noise the approach produced
a feasible solution. The design was highly automated in Mathematica and the
design choices were similar as in the linear case, i.e. mainly
Choice of desired inuence structure.
Choice of consistency relations to realize the inuence structure.
Choice of low-pass dynamics to make residual generators realizable.
84 Chapter 4. Residual Generation Based on Non-Linear Models
0 20 40 60
5
0
5
r
1
0 20 40 60
6
4
2
0
2
4
6
r
2
0 20 40 60
1.5
1
0.5
0
0.5
1
1.5
r
3
0 20 40 60
6
4
2
0
2
4
6
r
4
t [s]
0 20 40 60
3
2
1
0
1
2
3
r
5
t [s]
0 20 40 60
1.5
1
0.5
0
0.5
1
1.5
r
6
t [s]
Figure 4.5: Residuals when fault f
3
= 0.6 is induced at time t = 40 sec. All
residuals but r
3
respond to the fault which means that f
3
was correctly isolated.
4.8. Conclusions 85
4.8 Conclusions
A systematic and constructive design procedure for non-linear consistency re-
lations has been developed with strong computational support in standard
computer-algebra packages. Finding consistency relations is closely linked with
variable elimination, and a suitable class of systems to consider is models de-
scribed by polynomial, dierential-algebraic equations. This is a suitable class
since, for these types of non-linearities, elimination theory exists and practically
any analytical expression can be restated in polynomial form.
A systematic approach for design and analysis of disturbance decoupling
consistency relations is presented. The basic design step in the design procedure
is to compute a Grobner basis for an elimination ideal where all disturbances
have been eliminated. A nice property of the approach is that the available
design freedom is closely connected to the design freedom available in the linear
case for which the design freedom is well understood. In the linear case, it
was straightforward to use a consistency relation to form a realizable residual
generator. The nonlinear case is more dicult and dierent aspects of this
problem is discussed and approximate solutions are suggested in Section 4.4.
The strong theoretical support for polynomial systems also introduces some
interesting possibilities for the future; isolability analysis is such a possibility.
It is shown by example how advanced tools from real algebra, quantier elim-
ination, can be used for automatic isolability analysis. It is also noted how
identiability analysis by characteristic sets from dierential algebra can be
used for isolability analysis when only constant faults are considered.
Two major limitations of the approach exists, the computational complexity
of computing Gr obner bases and secondly how to utilize the non-linear consis-
tency relations to form a realizable residual generator. Computing Gr obner
bases may become computationally intractable even for moderately sized prob-
lems. Section 4.6 exemplies how structural analysis can be of great assistance
in reducing the size of the problem. In an example, an intractable problem that
could not be solved on a standard PC was reduced to a problem that could be
solved in 0.02 seconds.
The approach is nally demonstrated on a small, but non-trivial, example
model consisting of two coupled water tanks. The example shows how the
design freedom can be used and how fault isolation properties of the model can
be analyzed from the calculated Gr obner bases. The example also shows how
both constant and time-varying faults are handled equally in the design process.
86 Chapter 4. Residual Generation Based on Non-Linear Models
5
Residual Generation Based on
Stochastic Linear Models
This chapter investigates residual generation in linear, stochastic systems. In
Chapter 3, a design algorithm was developed for deterministic linear models
based on polynomial methods. That algorithm will now be extended to cover
also linear stochastic model descriptions.
A fundamental contribution to this problem is given by Nikoukhah (1994)
where a class of residual generators, innovation lters, for models stated on
state-space form were considered. The basic stochastic design requirement on
an innovation lter is that the residuals should be zero mean and white in the
fault-free case and that the whiteness property can be achieved without loosing
any design freedom.
Here, the aim is to extend the polynomial methods that proved benecial
in the deterministic case to the stochastic case and address problems posed in
(Nikoukhah, 1994) and also extend the problem formulation and solve a more
general problem. In the more general problem formulation, the requirement on
design-freedom is dropped which proves benecial.
The use of polynomial theory facilitates, just as in the deterministic case,
the development of an algorithm that covers not only state-space models, but
also general descriptor models. The main algorithmic tool is J-spectral co-
factorization which is shown to handle the stochastic problem. Algorithms for
spectral factorization of polynomial matrices has recently received much at-
tention since it plays a fundamental role in the solution of polynomial H

-
(Green et al., 1990) and H
2
-(Kwakernaak, 2000b) standard problems. There-
fore, feasible and numerically appealing algorithms and implementations has
been proposed (Kwakernaak and

Sebek, 1994; Kwakernaak, 2000a).
87
88 Chapter 5. Residual Generation Based on Stochastic Linear Models
The design algorithm is mainly described for the continuous-time case. Ad-
ditional considerations exists for the time-discrete case which are discussed in
Section 5.7. Note that in the nominal design problem, continuous and discrete
time systems could be handled analogously, but it will be shown that in the
stochastic case, small but important dierences exists.
5.1 Problem formulation
The system under consideration in this stochastic investigation is similar to
what was considered in Chapter 3 with the dierence that a stochastic term is
added to (3.1), i.e. the following class of models is studied:
y = G
u
(s)u +G
d
(s)d +G
f
(s)f +G
n
(s)n (5.1)
where y R
m
is the measurement vector, u R
k
u
control signals, d R
k
d
unknown disturbances, f R
k
f
faults, n R
k
n
noise, and G
u
(s), G
d
(s), G
f
(s),
and G
n
(s) are proper transfer matrices of suitable dimensions. The dierence
between the disturbances d and the noise n is that the disturbances are assumed
to have no stochastic description and must be decoupled while the noise is
modeled as a white stationary stochastic process with unit covariance. The
noise is not decoupled but is handled otherwise.
For deterministic models, residual generators were dened in Denition 3.1.
This denition is the basis also for stochastic residual generators and also here it
is assumed that perfect decoupling of disturbances d is possible. In case perfect
decoupling of d is not possible, new sensors may be needed or signals may
have to be transferred from d to n and the model augmented with stochastic
descriptions of these signals. Now, the residual generator denition needs an
extension where stochastic properties of the residual generator are specied.
For linear models with no unknown inputs, the innovation process associated
with the Kalman lter is often used as a residual because of its zero-mean and
whiteness properties in the fault-free case. Once the innovations is generated,
the fault detection problem reduces to a whiteness test of the residual. Also,
other more elaborate decision algorithms can be used based on more deep uti-
lization of stochastic properties of the residual (Basseville and Nikiforov, 1993).
Trying to achieve the same properties but also including unknown disturbances
in the system leads to the following extension to Denition 3.1:
Denition 5.1 (Whitening residual generator). A stable and proper linear
lter Q(s) is a residual generator for (5.1) if and only if when f 0 it holds
that
r = Q(s)
_
y
u
_
is zero mean and white for all u and d.
Note that here, for the sake of convenience, the scalar assumption on r is
dropped. This is mainly to be able to keep the presentation close to Nikoukhah
5.1. Problem formulation 89
(1994). Of course, for the residual generator to be useful for fault detection,
when f ,= 0 the zero mean whiteness property need to be violated.
Finally, a restricted class of residual generators dened by Nikoukhah (1994),
where the whiteness property of the residual is achieved without restricting the
number of linearly independent residuals. This has the consequence that the
whiteness property of the residual is achieved without sacricing any design
freedom or fault detectability. This is a smaller class of residual generators
than those dened in Denition 5.1.
Denition 5.2 (Innovation lter). A nite-dimensional linear time-invariant
system Q(s) is called an innovation lter for system (5.1) if it is stable with the
least number of outputs such that, in the absence of failure,
1. its output
r = Q(s)
_
y
u
_
is zero-mean, white and decoupled from u and d,
2. if Q

(s) is any nite-dimensional linear time-invariant system such that

r

= Q

(s)
_
y
u
_
is decoupled from u and d, then there exists a linear system L(s) such that
Q

(s) = L(s)Q(s).
Note that for an innovation lter, the residual is no longer a scalar but equals
the dimension of N
L
(M(s)).
Assumption: From now on it is assumed that perfect decoupling of both the
noise n and disturbances d is not possible. A brief discussion on the case that
arises when the noise is perfectly decoupled is presented in Section 5.8.
This chapter now describes an extension, under the assumption above, of the
algorithms presented in Chapter 3 to synthesize whitening residual generators
and innovation lters. Characterization of residual generators and innovation
lters will be derived and presented in Theorem 5.1 and Theorem 5.2. Let the
fault-free system (5.1) be described by a state-space realization on the form
x = Ax +B
u
u +B
d
d +B
n
n (5.2a)
y = Cx +D
u
u +D
d
d +D
n
n (5.2b)
As shown in Chapter 3, any deterministic residual generator Q(s) can be written
as Q(s) = (s)N
M
s
(s)P
x
where
M
s
(s) =
_
C D
d
(sI
n
x
A) B
d
_
P
x
=
_
I
k
m
D
u
0
n
x
k
m
B
u
_
(5.3)
where n
x
is the number of states i.e. the size of x and N
M
s
(s) is a minimal
polynomial basis for the left null-space of M
s
(s). In Chapter 3, Theorem 3.1, it
90 Chapter 5. Residual Generation Based on Stochastic Linear Models
was also required that the pair A, [B
u
B
d
] was controllable to get a minimal
polynomial basis. This requirement is relaxed here because, due to clarity of
presentation, minimality issues is neglected here. The consequence of this re-
laxation is that it possible that matrix N
M
s
(s)P
x
is not irreducible according
to Corollary 3.1. However, according to the proof of Theorem 3.1, it is still
guaranteed to be row-reduced regardless if (5.2) is controllable or not. In this
chapter, row-reducedness will show to be important while irreducibility is not.
For nominal designs in Chapter 3, all solutions were parameterized by a
single, rational row-vector (s) as in (3.6) and (3.7)
(s) = c
1
(s)(s)
There, no additional modeling was available to guide the selection of the param-
eterization matrix. Here, additional constraints on the residual is imposed, i.e.
the whiteness requirement. So here, parts of the available design freedom will
be used to fulll the whiteness requirements and other constraints according to
Denitions 5.1 and 5.2. Now follows a characterization of the parameterization
matrix (s) to fulll the added requirements:
Theorem 5.1. A transfer matrix Q(s) is a whitening residual generator for
(5.2) if and only if there exists a (s) such that
Q(s) = (s)N
M
s
(s)P
x
is proper, stable and it holds that
s.H(s)H
T
(s) =
where H(s) = (s)N
M
s
(s)
_
D
n
B
n
_
and is a constant matrix.
Proof. All disturbance decoupling residual generators can be written as
Q(s) = (s)N
M
s
(s)P
x
Insertion of (5.2) into r = Q(s) (
y
u
) gives, after some straightforward calcula-
tions,
r = (s)N
M
s
(s)P
x
_
y
u
_
= (s)N
M
s
(s)
_
C D
d
D
n
(sI A) B
d
B
n
_
_
_
x
u
n
_
_
=
= (s)N
M
s
(s)
_
D
n
B
n
_
n (5.4)
Whiteness of r is equivalent to
r
(j) constant for all which, since
r
(s) is
rational, is equivalent to
r
(s) is constant for all s. The spectrum
r
(s) can be
5.1. Problem formulation 91
written as

r
(s) = (s)N
M
s
(s)
_
D
n
B
n
_

n
(s)
_
D
n
B
n
_
T
N
T
M
s
(s)
T
(s) =
= (s)H(s)H
T
(s)
T
(s)
and the theorem follows immediately.
Theorem 5.2. A transfer matrix Q(s) is an innovation lter for system (5.2)
if and only if there exists a matrix (s) such that
Q(s) = (s)N
M
s
(s)P
x
is proper, stable and it holds that
s.H(s)H
T
(s) =
where H(s) = (s)N
M
s
(s)
_
D
n
B
n
_
, R
rr
is a constant full-rank matrix, and
r = dim N
L
(M(s)).
Proof. Following the same lines of proof as in Theorem 5.1, it is seen that Q(s)
satises the rst requirement in Denition 5.2 if and only if there exists a (s)
such that the conditions given in the theorem are satised.
For the second requirement from the denition, consider a disturbance de-
coupling lter Q

(s). From Chapter 3 we know that Q

(s) can be parameterized

by
Q

(s) =

(s)N
M
s
(s)P
x
The second requirement is fullled if and only if Q

(s) = L(s)Q(s) have a

solution for any

(s). For this to be true, the rank of (s) must equal

dimN
L
(M(s)). This gives that Q(s) fullls both requirements if and only if
there exists a (s) with full column-rank that satises the conditions in the
theorem.
Since the least number of rows of an innovation lter is given by r =
dimN
L
(M(s)), Q(s) has the least number of outputs if and only if (s) is
square and therefore has dimension r r. Also, is full-rank under the main
assumption that it was not possible to perfectly decouple the noise n.
Now, the parameterization matrix (s) has been characterized for both
whitening residual generators and innovation lters. However, the results so
far does not indicate a design algorithm how to nd (s). Before such an algo-
rithm can be described, some theory on polynomial J-spectral factorization is
needed.
92 Chapter 5. Residual Generation Based on Stochastic Linear Models
5.2 Spectral factorization theory
The material for this section comes mainly from (Kwakernaak and

Sebek, 1994;

Sebek, 1990), but also from (Jezek and Kucera, 1985; Callier, 1985). The nu-
merical implementation used in this work is described in (Kwakernaak, 2000a).
A corresponding discrete time version of this theory is also available and the
time-discrete case is discussed in Section 5.7.
A polynomial matrix Z(s) is said to be para-hermitian if Z
T
(s) = Z(s).
Para-hermitian is sometimes abbreviated as p.h. From now on, only para-
hermitian matrices Z(s) with real coecients are considered in this work. A
factorization
Z(s) = P
T
(s)JP(s)
is called a J-spectral factorization if J is a signature matrix and P(s) a square
matrix with real coecients such that det P(s) is Hurwitz. This is equivalent
to all zeros of the invariant polynomials lying in the closed left half plane. For
short, P(s) is said to be strictly stable. Sometimes the J is omitted and the
factorization is called a spectral factorization. The signature matrix J has the
following form
J =
_
_
I
1
0 0
0 I
2
0
0 0 0
_
_
Figure 5.1 show how zeros of a para-hermitian matrix and the zeros of its
spectral factor is related. A factorization on the form Z(s) = P(s)JP
T
(s) is
x
x
x
x
x x
x
x
(2)
(2)
Re
Im
Figure 5.1: Distribution of zeros of a para-hermitian matrix Z(s) and its spectral
factor P(s). The X marks the zeros of the invariant polynomials of Z(s) and
the dotted line marks the zeros of a spectral factor P(s). Only half of the zeros
on the imaginary axis is a zero of P(s).
5.2. Spectral factorization theory 93
called a J-spectral co-factorization. A specic class of spectral factorizations is
of particular importance in this work.
Denition 5.3 (Canonical spectral factorization). Let P(s) be a spectral
factor of the para-hermitian matrix Z(s). The spectral factorization is said to
be canonical if P(s) is column-reduced and the column-degrees equals the half
diagonal degrees of Z(s).
No necessary and sucient existence conditions are known for J-spectral
factorization (Kwakernaak and

Sebek, 1994). However, the following necessary
condition due to (Jakubovic, 1970) (referred in (Kwakernaak and

Sebek, 1994))
gives a necessary condition.
Theorem 5.3 (Existence of J-Spectral Factorization). Suppose that the
multiplicity of the zeros on the imaginary axis of each of the invariant polyno-
mials of the para-hermitian polynomial matrix Z(s) is even, then Z(s) has a
spectral factorization Z(s) = P
T
(s)JP(s).
A related issue is uniqueness of the factorization
Theorem 5.4 (Non-uniqueness of J-Spectral Factorization). Let the
polynomial Matrix P(s) be a spectral factor of the full-rank para-hermitian ma-
trix Z(s) with corresponding signature matrix J.
1. All other spectral factors of Z(s) are of the form U(s)P(s) with U(s)
unimodular such that
U
T
(s)JU(s) = J
Matrix U(s) is said to be a J-unitary unimodular matrix.
2. If the factorization is canonical, i.e. P(s) is column reduced, any other
spectral factor is on the form UP(s) with U constant J-unitary.
Proof. See (Kwakernaak and

Sebek, 1994)
Theorem 5.5. Let Z(s) be positive denite on the imaginary axis, then the
J-spectral factorization of Z(s) is canonical.
Proof. See (Kwakernaak and

Sebek, 1994)
5.2.1 Note on the singular case
A brief description of spectral (co-)factorization of singular para-hermitian ma-
trices is now given. This presentation follows (

Sebek, 1990), and the character-

ization of spectral factors of singular para-hermitian matrices will be used in
Section 5.8 where the design of residual generators where the assumption from
Section 5.1 is violated, i.e. it is possible to perfectly decouple the stochastic
noise n.
94 Chapter 5. Residual Generation Based on Stochastic Linear Models
Let Z(s) be a para-hermitian, n n matrix of rank m < n. First, nd the
greatest right divisor of Z(s), i.e. nd Z
R
(s) and a unimodular U(s) such that
Z(s) = U(s)
_
Z
R
(s)
0
_
By symmetrical extraction of U(s) we get
Z(s) = U(s)
_

Z(s) 0
0 0
_
U
T
(s)
where

Z(s) is a square mm non-singular para-hermitian matrix. Let a

P(s)
and

J be a spectral co-factor and the signature of

Z(s). Then a spectral co-
factorization of Z(s) is given by
Z(s) = U(s)
_

P(s) 0
0 I
_ _

J 0
0 0
_ _

P
T
(s) 0
0 I
_
U
T
(s)
That is, a spectral co-factor and the signature of Z(s) can always be written as
P(s) = U(s)
_

P(s) 0
0 I
_
J =
_

J 0
0 0
_
where

P(s) and

J is a spectral co-factor and signature of a non-singular para-
hermitian matrix.
Also, a spectral factorization of a singular para-hermitian matrix is not
necessarily canonical and therefore the row-reducedness and degree property of
Theorem 5.5 does not hold.
5.3 Introductory examples
Before going into details, describing a design algorithm and existence condi-
tions, three small illustrative examples are presented that illustrates various
issues a design algorithm must consider. The rst describes a successful design
and the last two illustrates the two cases when whitening residual generators
do not exist. All these introductory examples are scalar in the sense that only
one, linearly independent, disturbance decoupling residual generator exists and
thereby is, in these cases, a whitening residual generator equivalent to an inno-
vation lter.
5.3.1 Example 1: Successful design
Consider a system described by
y =
_
1
s+1
1
s(s+1)
_
u +
_
0
1
_
d +
_
1
0
_
f +
_
n
1
n
2
_
(5.5)
5.3. Introductory examples 95
The only residual generator that decouples d is parameterized by the free vari-
able (s) as
r = (s)[s + 1 0 1]
_
y
u
_
(5.6)
Inserting (5.5) into (5.6), the internal form, in the fault-free case becomes
r = (s)(s + 1)n
1
It is clear that by letting (s) =
1
s+1
we get a white residual in the fault-free
case by the stable and proper residual generator
Q(s) = [1 0
1
s + 1
]
5.3.2 Example 2: Zeros on the imaginary axis
Consider the same example as above, but switch the positions of f and d, i.e.
y =
_
1
s+1
1
s(s+1)
_
u +
_
1
0
_
d +
_
0
1
_
f +
_
n
1
n
2
_
In the same way as before it is clear that all disturbance decoupling residual
generators can be parameterized as
r = (s)[0 s(s + 1) 1]
_
y
u
_
for which the fault-free internal form is given by
r = (s)s(s + 1)n
2
Here it is clear that no strictly stable (s) exists making r white, all because of
the nite zero on the imaginary axis in the transfer function from n to r.
This also shows a link to non strongly-detectable faults (Chen and Patton,
1994; Nyberg, 2000). A zero at s = 0 will appear in the transfer function from
n to r if n enters the system in the same way as a non strongly detectable fault
f which was the case in the example above.
5.3.3 Example 3: Innite zeros
Consider the scalar system
y =
1
s + 1
u +f +
1
(s + 2)
2
n
All residual generators can be written
r = (s)[s + 1 1]
_
y
u
_
96 Chapter 5. Residual Generation Based on Stochastic Linear Models
for which the internal form is
r = (s)
s + 1
(s + 2)
2
n (5.7)
It is clear that for r to be white (s) =
(s+2)
2
s+1
which gives an improper, and
thus non-realizable, residual generator
r = [(s + 2)
2

(s + 2)
2
s + 1
]
_
y
u
_
And this was caused by the innite zero of the transfer function
s+1
(s+2)
2
in (5.7).
Now, with these three examples in mind, a design algorithm is described in
the next section.
5.4 Design algorithm
The main step in designing both whitening residual generators and innovation
lters is to rst compute N
M
s
(s) in (5.3) and then nd (s) such that Q(s) =
(s)N
M
s
(s)P
x
is stable, proper, and the spectrum of r is constant for all s. For
the innovation lter, additional requirements on (s) is needed.
Now, existence conditions and design procedures for these lters will be de-
rived. First, results for innovation lters is derived in Section 5.4.1 and then,
the more involved case of whitening residual generators is addressed in Sec-
tion 5.4.2. The proofs are constructive, outlining design algorithms that nds
all possible whitening residual generators and innovation lters.
First, for sake of notational convenience, let Z(s) R
mm
[s] denote
Z(s) = N
M
s
(s)
_
D
n
B
n
_ _
D
n
B
n
_
T
N
T
M
s
(s) (5.8)
for the remaining part of this chapter. Then, from the proof of Theorem 5.1,
the spectrum of r can be written

r
(s) = (s)Z(s)
T
(s)
This also implies that the assumption made in Section 5.1, that it is not possible
to perfectly decouple the stochastic noise n, is equivalent to Z(s) being full-rank.
If Z(s) would be rank decient, there would exist a (s) such that the spectrum
of r would be 0, i.e. the noise would be perfectly decoupled. Therefore, in this
section it is assumed, unless otherwise noted, that Z(s) is full-rank. Further
discussions on the case when Z(s) is not full-rank is found in Section 5.8.
5.4.1 Design of innovation lters
Before the main results can be stated, a lemma characterizing the parameteri-
zation matrix (s) in Theorem 5.1 and 5.2 is needed:
5.4. Design algorithm 97
Lemma 5.1. Assume Z(s) full-rank. Then there exists a (s) such that the
linear time-invariant lter Q(s) = (s)N
M
s
(s)P
x
produces white residuals if
and only if (s) can be written
(s) = (s)P
1
(s)
where P(s) is a spectral co-factor of Z(s) and (s)
T
(s) = for some constant
matrix .
Proof. According to the proof of Theorem 5.1, the spectrum of r can be written

r
(s) = (s)Z(s)
T
(s) (5.9)
Note that Z(s) is a para-hermitian polynomial matrix. Now, let P(s) be a
spectral co-factor and J a signature of Z(s), i.e.
Z(s) = P(s)JP
T
(s) (5.10)
Since Z(s) is assumed positive denite it has signature J = I
m
. Insertion of
(5.10) into (5.9) and denoting (s) = (s)P(s) gives

r
(s) = (s)P(s)JP
T
(s)
T
(s) = (s)
T
(s)
Thus,
r
(s) is constant for all s if and only if (s)
T
(s) = for some constant
. The parameterization matrix (s) is found by solving for (s) in the equation
(s) = (s)P(s) (5.11)
which has only one unique solution (s) = (s)P
1
(s).
Remark: Any constant (s) yields a constant (s)
T
(s) for all s. Note however
that also non-constant (s) exists. One such example is (s) = [
1
s+1
s
s+1
].
Now, we are ready to present the main theorem on design of innovation
lters.
Theorem 5.6. If Z(s) is full rank, an innovation lter exists if and only if
i.row-deg
i
N
M
s
(s)
_
D
n
B
n
_
= row-deg
i
N
M
s
(s)
and Z(s) has no roots on the imaginary axis. Furthermore, if an innovation
lter exist, all innovation lters can be parameterized as
Q(s) = (s)P
1
(s)N
M
s
(s)P
x
where P(s) is a spectral co-factor of Z(s) and (s) is any strictly stable, full-rank
matrix, such that (s)
T
(s) is constant.
98 Chapter 5. Residual Generation Based on Stochastic Linear Models
Proof. According to Theorem 5.2 and Lemma 5.1, an innovation lter exists if
and only if there exists an (s) such that
Q(s) = (s)P
1
(s)N
M
s
(s)P
x
is stable, proper, (s)
T
(s) is constant and full-rank of dimension r r with
r = dim N
L
(M
s
(s)).
First it will be shown, by contradiction, that Z(s) have no roots on the
imaginary axis is a necessary condition for the existence of an innovation lter.
For this, assume Q(s) is an innovation lter and that Z(s) has a zero at s
0
= j
0
.
Since Q(s) is strictly stable, lim
sj
0
Q(s) exists. But, Z(s) has a zero at s
0
implies that P(s
0
) is rank decient. Since, according to assumption, Q(s
0
)
exists, it must hold that (s) looses rank at s
0
since N
M
s
(s)P
x
is irreducible.
However, this contradicts = (s)
T
(s) being full-rank which gives that
full-rank of Z(s) on the imaginary axis is a necessary condition for Q(s) to be
stable.
Next it will be shown, also by contradiction, that the row-degree condition
in the theorem is also a necessary condition for the existence of an innovation
lter. Assume Q(s) is an innovation lter and that there exists an i such that
row-deg
i
N
M
s
(s)
_
D
n
B
n
_
< row-deg
i
N
M
s
(s) (5.12)
Partition N
M
s
(s) = [V
1
(s) V
2
(s)] according to the block-structure of (5.3).
Then, from the proof of Theorem 3.1, we know that V
1
(s) is row-reduced,
row-deg
i
N
M
s
(s) = row-deg
i
V
1
(s), and that V
2
(s) = V
1
(s)C(sI A)
1
. Since
V
1
(s) is row-reduced, we can rewrite (5.12) as
row-deg
i
S
V
1
(s)V
1,hr
D
n
+

V
1
(s)D
n
+V
2
(s)B
n
< row-deg
i
V
1
(s)
where the decomposition of V
1
(s) is done according to Denition 3.A.10. Since
the row-degrees of

V
1
(s) and V
2
(s) is strictly less than the row-degrees of S
V
1
(s),
the inequality can only be fullled if V
1,hr
D
n
does not have full row-rank. This
also gives that
lim
s
V
1
(s)D
n
= lim
s
S
V
1
(s)V
1,hr
D
n
(5.13)
does not have full row-rank. Now, since Q(s) is an innovation lter, there exist
an (s) such that
Q(s) = (s)P
1
(s)N
M
s
(s)P
x
and H(s)H
T
(s) is square, full-rank, and constant where
H(s) = (s)P
1
(s)N
M
s
(s)
_
D
n
B
n
_
But, when s goes to innity, it holds that
lim
s
H(s) = lim
s
(s)P
1
(s)V
1
(s)
_
C(sI A)
1
B
n
+D
n
_
=
= lim
s
(s)P
1
(s)V
1
(s)D
n
5.4. Design algorithm 99
which does not have full row-rank due to (5.13) and the fact that (s) and
P(s) is square and full-rank. Thus, lim
s
H(s) does not have full rank which
contradicts that H(s)H
T
(s) is constant and full-rank.
Now, suciency. Since Z(s) does not have zeros on the imaginary axis, a
spectral co-factor P(s) will be strictly stable and, according to Theorem 5.5 and
Denition 5.3, row-reduced with row-degrees satisfying
row-deg
i
P(s) = row-deg
i
N
M
s
(s)
_
D
n
B
n
_
Thus, Theorem 3.B.6 gives that
Q(s) = P
1
(s)N
M
s
(s)P
x
will be proper, strictly stable and fulll all requirements in Denition 5.2, i.e.
Q(s) is an innovation lter.
Finally, if Q(s) is an innovation lter, it is immediate that Q

(s) is an inno-
vation lter if and only if
Q

(s) = (s)Q(s)
where (s) is a square, full-rank, all-pass link i.e. (s)
T
(s) is constant and
full rank.
Summary of design procedure
1. Form M
s
(s) according to (5.3) and compute N
M
s
(s).
2. Form Z(s) as in (5.8). If Z(s) is full-rank, an innovation lter exists if
and only if P(s) is strictly stable and
i.row-deg
i
N
M
s
(s)
_
D
n
B
n
_
= row-deg
i
N
M
s
(s)
3. All innovation lters, if any exists, are then given by:
Q(s) = (s)P
1
(s)N
M
s
P
x
where P(s) is a spectral co-factor of Z(s) and (s) is any invertible matrix
such that (s)
T
(s) is constant for all s.
5.4.2 Design of whitening residual generators
The design procedure for whitening residual generators is a bit more complex
due to the increased design freedom, resulting in a possibly more involved design
procedure. It will be shown that the design procedure is not necessarily more
complex, only in special cases when Z(s) has zeros on the imaginary axis and/or
if no row of P
1
(s)N
M
s
(s)P
x
is proper.
100 Chapter 5. Residual Generation Based on Stochastic Linear Models
A whitening residual generator is, according to Denition 5.1, a stable and
proper lter that produces white residuals in the fault-free case. The whiteness
property has already been characterized in Lemma 5.1. Now, more details
on the properness and stability properties of the residual generator is needed
before the main result can be stated. We begin with considering the properness
condition.
First, a result is needed characterizing all rational (s) such that (s)
T
(s)
is constant which is fundamental for the whiteness property but also inuences
properness/stability results later on.
Lemma 5.2. Let (s) be a 1 m row-vector of transfer functions such that
(s)
T
(s) = with ,= 0. Then it holds that (s) is proper and at least one
element of (s) is not strictly proper.
Proof. Since (s)
T
(s) = , it holds that
= (j)
T
(j) =
m

k=1
[
k
(j)[
2
Assume that element k of (s) is improper, then the limit lim

[
k
(j)[
2
would not exist. Since for all l, [
l
(j)[
2
0, the improper assumption has lead
to a contradiction, i.e. (s) is proper.
Left to prove is that at least one element of (s) is not strictly proper.
Denote the constant coecient matrix of (s) with H
0
, i.e. H
0
= lim
s
(s).
The limit exists since (s) is proper. Then,
(s) = H
0
+ (s)
where it holds that (s) is strictly proper. Now, assume H
0
= 0, i.e.
= (j)
T
(j) = (j)
T
(j)
Since all elements of (s) is strictly proper, lim

[
k
(j)[
2
= 0 which is a
contradiction, i.e. H
0
,= 0 which means that at least one element of (s) is not
strictly proper.
Now we can almost characterize proper, whitening residual generators. But
rst some additional notation is needed. From now on, let matrix W(s) denote
W(s) = quotient [P
1
(s)N
M
s
(s)] (5.14)
See Theorem 3.B.7 for denitions on matrix quotient and remainder. Let
W(s) R
mn
[s] have degree d, then a matrix that will prove useful is the
Sylvester matrix of W(s) which is dened as
sylv(W(s), q)
_

_
W
0
W
1
W
d
0 0 0
0 W
0
W
1
W
d
0 0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0 0 W
0
W
1
W
d
_

_
5.4. Design algorithm 101
with q + 1 block-rows. Finally, for 0 q d, introduce the notation
J
q

_

_
W
q
W
q+1
. . . W
d
0 0
.
.
.
.
.
.
W
d
0
W
1
W
2
. . . . . . . . W
d1
W
d
_

_
(5.15)
Now, we are ready to state the properness result in the following lemma.
Lemma 5.3. Let W(s)R
mn
[s], dened in (5.14), have degree d. Then there
exists a rational (s) such that
Q(s) = (s)P
1
(s)N
M
s
(s)P
x
is proper (s)
T
(s) is constant
if and only if N
L
(J
d
) ,= . Then (s) = d
1
(s)n(s) where n(s) is parameterized
by the scalar polynomial
1
(s) and the polynomial row-vector
0
(s) such that
n(s) =
1
(s)[l
0
l
q
]
_

_
I
.
.
.
s
q
I
_

_
+
0
(s)
where row-deg
0
(s) < deg
1
(s) = r and [l
0
l
q
] N
L
(J
q
) with q d. A
denominator d(s) always exists that ensures (s)
T
(s) constant.
Proof. Let H(s) = P
1
(s)N
M
s
(s)P
x
and let W(s) and E(s) be the polynomial
matrix quotient and remainder of H(s) respectively, i.e.
Q(s) = (s)H(s) = (s)W(s) +(s)P
1
(s)E(s)
First note that (s) is proper according to Lemma 5.2. Also, since P
1
(s)E(s)
is strictly proper it holds that Q(s) is proper if and only if (s)W(s) is proper.
Now, write (s) as
(s) = d
1
(s)n(s)
where d(s) is a scalar polynomial and n(s) is a polynomial row-vector of the
same dimensions as (s). Since (s) is proper, but not strictly proper it must
hold that
deg d(s) = row-deg n(s) (5.16)
Now, for (s)W(s) to be proper it must hold that deg d(s) row-deg n(s)W(s).
This together with (5.16) gives that Q(s) is proper if and only if there exists an
n(s) such that
row-deg n(s) row-deg n(s)W(s) (5.17)
Let W(s) R
mn
[s] be a matrix of degree d and let n(s) =

q
i=0
n
i
s
i
with
q d and n
q
,= 0. Assume that no n(s) of degree less than q exists such that
102 Chapter 5. Residual Generation Based on Stochastic Linear Models
(5.17) is satised. Then,
n(s)W(s) = [n
0
n
q
]sylv(W(s), q)
_

_
I
sI
.
.
.
s
q+d
I
_

_
= [n
0
n
q
][ [ J
q
]
_

_
I
.
.
.
s
q
I
. . . . .
s
q+1
I
.
.
.
s
q+d
I
_

_
where denotes a matrix that is of no importance here. From this, (5.17) is
satised if and only if
[n
0
n
q
] N
L
(J
q
)
Also, for q > d, it is straightforward to see that (5.17) is satised if and only if
[n
0
n
q
] N
L
(
_
0
J
d
_
)
Since, it was assumed that no n(s) of degree less than q satised the lemma, J
d
has full row-rank and therefore n
i
= 0, i = q d, . . . , q, i.e. n(s) has row-degree
less than q. Thus, there exists a n(s) such that Q(s) is proper if and only if for
any q = 1, . . . , d
N
L
(J
q
) ,=
And due to the triangular structure of J
q
this is equivalent to N
L
(J
d
) ,= .
Also, if n(s) satises (5.17), then so does
1
(s)n(s) +
0
(s) when
i
(s) satises
the conditions in the theorem.
To form (a stable) (s) that satises the lemma, a d(s) always exists such
that (s)
T
(s) is constant. One such d(s) is e.g. found by a spectral factor-
ization of n(s)n
T
(s) = d(s)d
T
(s).
In the case where W(s) is row-reduced, which is the normal case since W(s)
is the quotient between two row-reduced matrices, the above result can be stated
little simpler
Lemma 5.4. Let W(s) R
mn
[s] be row-reduced, then there exists an n(s)
R
1m
[s] such that
row-deg n(s) row-deg n(s)W(s)
if and only if N
L
(J
0
) is non-empty.
Proof. Theorem 3.B.5 gives that
row-deg n(s)W(s) = max
i.n
i
(s)0
[
i
+ deg n
i
(s)]
5.4. Design algorithm 103
where
i
is the row-degrees of W(s). The condition in the theorem can thus be
rewritten
max
i.n
i
(s)0
[deg n
i
(s)] max
i.n
i
(s)0
[
i
+ deg n
i
(s)] (5.18)
which can be satised if and only if at least one
i
= 0.
Let W
i
be the coecient matrices of matrix W(s), i.e.
W(s) =
d

i=0
W
i
s
i
Then, since W(s) is row-reduced matrix [W
0
W
d
] has full-row rank and
[W
1
W
d
] looses rank if and only if any row-degree of W(s) is zero. Thus
i.
i
= 0 N
L
(J
0
) ,=
which together with (5.18) end the proof.
Now that properness of Q(s) is analyzed, we proceed to the stability prop-
erty, i.e. when does a rational (s) exist such that
Q(s) = (s)P
1
(s)N
M
s
(s)P
x
is stable (s)
T
(s) is constant
The word stability is used slightly sloppy here, meaning no poles in the closed
right half plane. Therefore, a non-proper Q(s) can here be said to be stable al-
though it is clearly not BIBO-stable. However, since properness is also required
later on, having noted this (mis-)use of the word here, stability is used in this
sense from now on.
The cause of non strictly-stable lters was, as was demonstrated in the small
scalar example in Section 5.3.2, the existence of zeros on the imaginary axis of
the transfer function from noise to residual. To be able to proceed with the
stability analysis, we need some additional notation. Assume Z(s) has purely
imaginary zeros j
i
, i = 1, . . . , r with multiplicities 2m
i
, i = 1, . . . , r. Then,
denote a polynomial
p
imag
(s) =
r

i=1
(s j
i
)
m
i
(5.19)
These are the poles we need to cancel/nullify to be able to form a strictly stable
residual generator. Also denote
p
stab
(s) =
q

i=1
(s
i
)
n
i
(5.20)
where
i
, i = 1, . . . , q are the strictly stable zeros of Z(s) with multiplicities n
i
.
This implies that for a spectral co-factor P(s) of Z(s) it holds that det(P(s)) =
kp
stab
(s)p
imag
(s) where k is a real constant. Also, the rational (s) can always
be rewritten as
(s) = d
1
(s)n(s)
104 Chapter 5. Residual Generation Based on Stochastic Linear Models
where d(s) and n(s) are relatively prime. Now, with this notation, Q(s) can be
rewritten as
Q(s) =
1
kp
stab
(s)p
imag
(s)
d
1
(s)n(s) adj P(s)N
M
s
(s)P
x
and it is immediate that Q(s) is stable if and only if p
imag
(s) divides
n(s)adj P(s)N
M
s
(s)P
x
and d(s) is strictly stable. A way to compute n(s), if
any exists, is to compute the row-Hermite form T(s) of adj P(s)N
M
s
(s)P
x
, i.e.
nd a unimodular matrix U(s) such that
T(s) = U(s)adj P(s)N
M
s
(s)P
x
Then, Q(s) can be written
Q(s) =
1
kp
stab
(s)p
imag
(s)
d
1
(s)n(s) U
1
(s)T(s)
Let
n(s) = [
1
(s)
m
(s)]U(s) (5.21)
Due to the upper triangular structure and properties of the Hermite form,
p
imag
(s) divides n(s) adj P(s)N
M
s
(s)P
x
, if and only if p
imag
(s) divides
i
(s)T
i
(s),
i = 1, . . . , m where T
i
(s) is the i:th row of T(s). This gives a direct method of
selecting the
i
(s) polynomials, and thereby n(s):

i
(s) = c
i
(s)p
imag
(s)/gld [p
imag
(s) T
i
(s)] (5.22)
where c
i
(s) is an arbitrary polynomial (possibly 0) and gld is the greatest left
divisor. Finally, a strictly stable denominator d(s) of (s) exists if and only if the
polynomials c
i
(s) is chosen such that n(s) has no zeros on the imaginary axis.
All steps in the above discussion goes both ways, i.e. a stable solution exists
if and only if a rational (s) can be found by rule (5.22), and for referential
convenience this is also summarized in a lemma.
Lemma 5.5. There exist a rational (s) such that
Q(s) = (s)P
1
(s)N
M
s
(s)P
x
is stable (s)
T
(s) is constant
if and only if (s) = d
1
(s)n(s) where n(s) is given by (5.21) and (5.22) and
d(s) is strictly stable.
Now, a simple and sucient existence condition that is easy to evaluate is
immediate which does not require the computation of Hermite forms.
Corollary 5.1. If Z(s) is full rank with no zeros on the imaginary axis, a
whitening residual generator exists if
i.row-deg
i
N
M
s
(s)
_
D
n
B
n
_
= row-deg
i
N
M
s
(s)
5.4. Design algorithm 105
Proof. Since, according to assumption in the theorem, Z(s) has no zeros on
the imaginary axis, strict stability of the residual generator is assured. By
Theorem 5.5 and Denition 5.3, P(s) is row-reduced and the row-degrees of
P(s) equals the row-degrees of N
M
s
(s)
_
D
n
B
n
_
. Thus, at least one row of
P
1
(s)N
M
s
(s)P
x
is also proper according to Theorem 3.B.6, and therefore fullls the requirements
of a whitening residual generator.
Summary of design procedure
Even though the proofs are constructive, the design steps might be a little
concealed by the details in the proofs. Therefore is the design procedure sum-
marized here.
1. Form M
s
(s) according to (5.3) and compute N
M
s
(s).
2. Form Z(s) as in (5.8) and compute a spectral co-factor P(s). If Z(s) is
not full-rank, it is possible to perfectly decouple n in the residual and the
initial assumption made is violated. If Z(s) has purely imaginary zeros,
form the polynomials p
imag
(s) and p
stab
(s) according to (5.19) and (5.20).
3. Now, four cases can occur, regarding zeros on the imaginary axis of Z(s)
and properness of H(s) = P
1
(s)N
M
s
(s)P
x
.
i) No zeros on the imaginary axis and H(s) proper.
This is the easiest case where any Q(s) = (s)H(s) with (s)
T
(s)
constant is a whitening residual generator.
ii) No zeros on the imaginary axis but H(s) not proper.
Here, caution has to be taken to achieve a proper residual generator.
First, compute the quotient W(s) and matrix J
d
according to (5.14)
and (5.15). If N
L
(J
d
) ,= , any (s) formed as in Lemma 5.3 gives a
whitening residual generator Q(s) = (s)H(s). If N
L
(J
d
) = , then
no whitening residual generator exists.
iii) Zeros on the imaginary axis and H(s) proper.
Here on the other hand, properness is guaranteed but stability has
to be ensured. Compute the row-Hermite form of adj P(s)N
M
s
(s)P
x
,
i.e. nd a unimodular matrix U(s) such that the Hermite form T(s)
can be written
T(s) = U(s)adj P(s)N
M
s
(s)P
x
Then a rational (s) giving a stable residual generator exists if and
only if it can be written as (5.21) using rule (5.22). The denomi-
nator d(s) of (s) is most easily found by spectral factorization of
106 Chapter 5. Residual Generation Based on Stochastic Linear Models
n(s)n
T
(s). A suitable procedure to nd a state-space realization is
to rst compute n(s), then perform the polynomial matrix division
1
p
imag
(s)
adj P(s)N
M
s
(s)P
x
= V (s) +
1
p
imag
(s)
R(s)
and then realize
Q(s) =
1
kd(s)p
stab
(s)
V (s)
which will be strictly stable. The remainder R(s) should be 0 (or
close to zero due to nite precision arithmetics).
iv) Zeros on the imaginary axis but H(s) not proper.
This case is not well covered. Even though all stable and all proper
solutions is characterized, it is not obvious how to combine the re-
sults. A feasible heuristic is to solve e.g. the stability constraint in
a rst step, and then try to nd a proper solution among the stable
solutions.
5.5 Design examples
This section includes 4 design examples that illustrates dierent aspects of the
design problem and the proposed design algorithm. The rst three examples are
based around the same linearized airplane model that was used in Section 3.6
to demonstrate the deterministic design problem.
In the rst example, a complete design of an innovation lter and a whitening
residual generator is shown. In the second example, only the noise environment
is changed, leaving the rest of the model the same. For this second model setup
it is shown that no innovation lter or whitening residual generator exists. In yet
a third example, using a third noise setup, it is shown that an innovation lter
does not exists but a whitening residual generator exists that has acceptable
fault sensitivity. The case with purely imaginary zeros is demonstrated in a
nal fourth example.
All calculations are done in Matlab using Polynomial Toolbox 2.5 for Mat-
lab 5 (2001). All functions used are included in the toolbox
1
. Included in
Appendix 5.A is a full Matlab implementations of innovation lter design.
5.5.1 Design example: Aircraft dynamics
The model from Section 3.6 is here extended with noise models. The same set
of faults is considered, i.e. additive actuator and sensor faults. Therefore, the
1
The spectral factorization procedure used is (Kwakernaak, 2000a) which probably will be
included in future versions of the Toolbox. The spectral factorization command shipped with
the toolbox can not handle zeros on the imaginary axis.
5.5. Design examples 107
total model including fault models and noise descriptions becomes:
x = Ax +B
u
u +B
f
f +B
n
n
y = Cx +D
u
u +D
f
f +D
n
n
Details on fault models and numerical values for the state-space matrices is pro-
vided in Section 3.6. The noise is assumed white with unit covariance. The three
dierent examples is based on this simulation model, each case with dierent
noise assumptions, i.e. dierent B
n
and D
n
matrices.
The design goal in all the three examples based on this model is a residual
generator Q(s) that decouples faults in the elevator angle actuator, and produces
a white residual in the fault-free case.
Process and measurement noise
In this rst example, both measurement noise and process noise is considered
and state-space matrices B
n
and D
n
is set to
B
n
= [I
5
0
53
] D
n
= [0
35
I
3
]
First, an innovation lter design is performed. Calculations in Matlab give
N
M
(s) =
_
0.0496s 0.703s + 0.0378 . . .
0.421s
2
+ 0.27s 0.123 . . .
. . . 0.0643 0.0844 0.703 0
. . . 0.0185s
2
0.0174s 0.306 0.582 0 0
_
(5.23)
Thus, the dimension of the null-space N
L
(M(s)) is 2, i.e. there exists exactly
two linearly independent numerators that decouples f
6
.
Step 2 from the summary in Section 5.4.1 was to compute matrix Z(s) and
checking full-rank condition. Matrix Z(s) is shown to be
Z(s) =
_
0.5s
2
+ 0.5 0.021s
3
0.012s
2
0.11s 0.011
0.021s
3
0.012s
2
+ 0.11s 0.011 0.18s
4
0.26s
2
+ 0.57
_
which has full rank. Performing a J-spectral co-factorization gives:
P(s) =
_
0.59s 0.61 0.39s 0.37
0.26s
2
0.57s 0.38 0.33s
2
+ 0.75s + 0.66
_
J = I
2
The spectral factor P(s) is strictly stable which can be seen by computing the
zeros of the invariant polynomials. Computing the zeros of P(s) in Matlab gives
s = 1.0196 and s = 1.1124 j0.7305.
Checking for existence of innovation lter according to Theorem 5.6 gives
row-deg N
M
s
(s) = 1, 2
row-deg N
M
s
(s)
_
D
n
B
n
_
= 1, 2
108 Chapter 5. Residual Generation Based on Stochastic Linear Models
i.e. an innovation lter exists and can be formed as Q(s) = P
1
(s)N
M
s
(s)P
x
.
Thus, the parameterization matrix (s) in Step 3 of the design summary in
Section 5.4.1 is chosen to be the identity matrix.
Next, a scalar whitening residual generator is to be designed. The rst 2
steps in the design summary in Section 5.4.2 has already been performed, left
is to nd an (s) in step 3 achieving unit variance in the fault-free residual. In
this case it is easy since case i from step 3 applies and one choice of (s) that
satises the design requirements is
(s) =
1

2
[1 1]
and the whitening residual generator can be formed as
Q(s) = (s)P
1
(s)N
M
s
(s)P
x
which is a 3:rd order realizable and strictly stable residual generator. The order
of the residual generator is, due to the choice of (s), equal to the sum of
row-degrees of P(s). Figure 5.2 shows how the faults inuence the residual
and Figure 5.3 shows the fault-free spectrum
r
(j) which is 1 for all as
expected. Especially note that the desired decoupling of fault f
6
has succeeded
while keeping the spectrum of r constant for all .
10
5
10
0
10
5
300
250
200
150
100
50
0
f
1
|
G
r
f
(
s
)
|

[
d
B
]
10
5
10
0
10
5
300
250
200
150
100
50
0
f
2
10
5
10
0
10
5
300
250
200
150
100
50
0
f
3
10
5
10
0
10
5
300
250
200
150
100
50
0
f
4
|
G
r
f
(
s
)
|

[
d
B
]
10
5
10
0
10
5
300
250
200
150
100
50
0
f
5
[rad/s]
10
5
10
0
10
5
300
250
200
150
100
50
0
f
6
Figure 5.2: Magnitude bode plots for the gain from faults to the residual.
5.5. Design examples 109
10
5
10
0
10
5
10
8
6
4
2
0
2
4
6
8
10
[rad/s]

r
(
j

)

[
d
B
]
Figure 5.3: Spectrum
r
(j).
Only process noise
In this second example, only process noise is considered and state-space matrices
B
n
and D
n
is set to:
B
n
= I
5
D
n
= 0
35
The null-space basis N
M
s
(s) is identical to the rst example (5.23). The row-
degrees of N
M
s
(s) is 1, 2 and the row-degrees of N
M
s
(s)
_
D
n
B
n
_
is 0, 1, i.e. no
innovation lter exists according to Theorem 5.6. Next it will be shown that no
whitening residual generator exists either. Performing the spectral factorization
gives a spectral co-factor:
P(s) =
_
1 0
0 0.9975s + 1.611
_
Since the roots of Z(s) is 1.6152, stability is ensured and only properness is
left. The matrix quotient W(s) of P
1
(s)N
M
s
(s)P
x
from (5.14) becomes
W(s) =
_
0.0703s 0.998s 0.0537 0.0912 . . .
0.997s 0.971 0.0703s + 0.0646 0.044s 0.119
. . . 0.12 0.998 0
. . . 0.00843 0.0703 0
_
110 Chapter 5. Residual Generation Based on Stochastic Linear Models
and it is easily checked that W(s) is row-reduced. Also, computing J
0
according
to (5.15) gives
J
0
=
_
0.0703 0.998 0 0 0 0
0.997 0.0703 0.044 0 0 0
_
which has full row-rank. Then, according to Lemma 5.4, no proper whitening
residual generator exists.
Noise on all states and sensor 3
In this case the process is subjected to noise on all states and on sensor 3, i.e.
the matrices B
n
and D
n
are given by
B
n
= [I
5
0] D
n
=
_
_
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 1
_
_
Now, computing N
M
s
(s) and Z(s) as before gives that Z(s) is strictly stable
and that the row-degrees of N
M
s
(s) is (as before) 1, 2 and row-degrees of
N
M
s
(s)
_
D
n
B
n
_
is 0, 2. This gives that no innovation lter exists according to
Theorem 5.6. But because Z(s) has no imaginary zeros and the row-degree
of the second row of N
M
s
(s) does not decrease when multiplied by the noise
distribution matrices, Corollary 5.1 proves existence of a whitening residual
generator. Computing W(s) gives that
W(s) =
_
0.14s + 0.4 2s + 0.12 0.2 0.24 2 0
0.014s 46 0.2s 1.5 2 0.024 0.2 0
_
Matrix W(s) has degree 1, and the existence of a whitening residual generator
is proven by computing J
1
for which a, non-empty, left null-space basis is
computed. Using Lemma 5.3, a (here constant) (s) is found
(s) =
_
0.0994 0.995

Then, a whitening residual generator can be formed by

Q(s) = P
1
(s)N
M
s
(s)P
x
Figures 5.4 and 5.5 shows the fault inuence on the residual and the spectrum
of the fault-free residual. Here it is clear that the residual still is able to detect
all faults, besides from fault f
6
which should be decoupled according to design
specications. Thus, even though an innovation lter didnt exist, a whiten-
ing residual generator with satisfactory fault detectability properties existed.
This residual generator could not have been designed algorithm proposed in
(Nikoukhah, 1994).
5.5. Design examples 111
10
5
10
0
10
5
300
250
200
150
100
50
0
f
1
|
G
r
f
(
s
)
|

[
d
B
]
10
5
10
0
10
5
300
250
200
150
100
50
0
f
2
10
5
10
0
10
5
300
250
200
150
100
50
0
f
3
10
5
10
0
10
5
300
250
200
150
100
50
0
f
4
|
G
r
f
(
s
)
|

[
d
B
]
10
5
10
0
10
5
300
250
200
150
100
50
0
f
5
[rad/s]
10
5
10
0
10
5
300
250
200
150
100
50
0
f
6
Figure 5.4: Magnitude bode plots for the fain from faults to the residual.
10
5
10
0
10
5
10
8
6
4
2
0
2
4
6
8
10
[rad/s]

r
(
j

)

[
d
B
]
Figure 5.5: Spectrum
r
(j).
112 Chapter 5. Residual Generation Based on Stochastic Linear Models
It is straightforward to realize that innovation lters preserve any fault de-
tectability properties since (s) is invertible, i.e. the number of outputs of an
innovation lter equals the dimension of N
L
(M(s)). However, if an innovation
lter does not exist, there may very well exist a whitening residual generator
with desirable fault detectability properties which was the case in this example.
5.5.2 Example with purely imaginary zeros
To study how purely imaginary zeros of Z(s) inuences the design procedure,
consider the following system:
y =
_
1
s+1
1
s(s+1)
_
(u +f) +
_
1 2
3 4
_
n
Simple calculations by hand gives that no innovation lter exists and all whiten-
ing residual generators can be written as
Q(s) = (s)
1

5
(y
1

1
s + 1
u) (5.24)
where (s) is any scalar all-pass link. Now, lets see how the design algorithm
outlined in Section 5.4.2 arrives at the same conclusion.
Forming M
s
(s), computing N
M
s
(s) and Z(s) gives
Z(s) =
_
25s
2
+ 5 11s
2
+ 16s 5
11s
2
16s 5 5s
2
+ 5
_
which has zeros 0, 0, 1. This gives that Z(s) has zeros on the imaginary axis
and p
imag
(s) = s.
T(s) =
_
s
2
+s 0 s
0 s
2
+s 1
_
U(s) =
_
0.089975 1.1144
0.24782 2.4877
_
Since the rst, but not the second row, is divided by p
1
imag
(s), the numerator of
(s) can be chosen using (5.21) and rule (5.22) as e.g.
n(s) = [1 0]U(s)
The scalar denominator d(s) is easily obtained by scalar spectral factorization
(here it is actually even more simple since n(s) is constant). Performing the
left division p
1
imag
(s) adj P(s)N
M
s
(s)P
x
gives a zero remainder as expected and
a quotient N(s) = [(s + 1) 0 1]. Realization of the residual generator, where
the imaginary zeros has been cancelled, is then
Q(s) = [0.4472 0
0.4472
s + 1
]
which, since 1/

5 0.4472 is identical to (5.24).

5.6. White residuals 113
5.6 White residuals
Why is white residuals desirable and are there advantages with white residuals
compared to low-pass ltered residuals to be thresholded?
One reason for white residuals is that basic change detection algorithms, e.g.
cusum (Basseville and Nikiforov, 1993), is directly applicable. Although these
algorithms are optimal, the algorithm relies on deep knowledge of amplitude
distribution of the residual and also the type of change (step/ramp/etc.). Often,
such knowledge is not available and a perhaps more important advantage with
white residuals is that the simple thresholding test for testing if the residual is
small can be replaced by a whiteness test of the residual.
The next section describes a basic whiteness test and Section 5.6.2 shows a
few simulation results where a whiteness test based detection is compared with
a low-pass and threshold approach. The simulation study does not provide any
hard theoretical results on the merits of either method, but still indicates some
properties.
5.6.1 Whiteness tests
A whiteness tests is a test to decide between the hypotheses
H
0
: n(t) is white H
1
: n(t) is coloured (5.25)
where n(t) is the signal being analyzed. This section is mainly based on
(S oderstr om and Stoica, 1989) and (Mehra and Peschon, 1971).
Typically, a test quantity to separate H
0
from H
1
is based on an estimate
of the covariance function c
k
. A basic property of white noise processes is that
c
k
=
2
if k = 0 and c
k
= 0 for k ,= 0. The covariance function of n(t) is dened
(for zero-mean stationary stochastic processes),
c
k
= En(t)n(t k) k = 0, 1, . . .
and an estimate of c
k
can be obtained from data as
c
k
=
1
N
N

t=1
n(t)n(t +k) k = 0, 1, . . . (5.26)
Let c denote the vector
c =
_
_
_
_
_
c
1
c
2
.
.
.
c
n
k
_
_
_
_
_
,
then it holds that (asymptotically)
N c
T
c
c
0

2
(n
k
) (5.27)
114 Chapter 5. Residual Generation Based on Stochastic Linear Models
Thus, a test quantity for hypotheses (5.25) can be formed by estimating c
0
to
c
n
k
from M data points and calculate
T(x) =
M c
T
c
c
0
(5.28)
The null hypothesis is rejected when T(x) > J. The thresholds are selected
such that the false alarm rate is lower than a specied level . The threshold
selection is done by assuming asymptotic properties (5.27) of T(x), i.e. assume
M large enough.
One nice property of the test T(x), assuming asymptotic properties, is that
it is invariant to distribution and input noise power.
5.6.2 Simulations and comparisons
In this section, a few simulations is done to compare the whiteness test from the
previous section with a simple LP-and-threshold test. The thresholds in both
tests is set to achieve a false-alarm rate of = 0.01.
In the simulations, the system under consideration is given by
y = G
u
(q)u +n +f
i.e. sensor noise and a sensor fault is considered. The raw residual is given by
r
raw
= y G
u
(q)u = n +f
In the simulations, n is white Gaussian noise and y, u is collected during 100 sec-
onds, sampled with sampling frequency 10 Hz. The residual is then evaluated,
either with a white-noise test or a LP-and-threshold test.
The parameters in the whiteness is set to (without any excessive tuning)
M = 80 and n
k
= 20. In the LP-and-threshold test, a rst order LP-link with
unit DC-gain is used, i.e.
H
lp
(z) =
1 a
z a
The parameter in the test quantity is therefore the placement of the pole in the
lter. The threshold is set by (correctly) assuming Gaussian distribution and
unit variance. In the simulations, four dierent pole placements with dierent
cut-o frequencies are used, a 0.9, 0.95, 0.97, 0.99.
Three kinds of faults are simulated
1. Step fault. The fault signal is
f(t) =
_
0 t < 50
0.85 t 50
The simulation results is shown in Figures 5.6, 5.7, and 5.8.
5.6. White residuals 115
2. Sinus fault. The fault signal is
f(t) =
_
0 t < 50
1.5 sin(t) t 50
The simulation results is shown in Figures 5.9, 5.10, and 5.11.
3. Ramp fault. The fault signal is
f(t) =
_
0 t < 50
1.5
50
(t 50) t 50
The simulation results is shown in Figures 5.12, 5.13, and 5.14.
0 10 20 30 40 50 60 70 80 90 100
3
2
1
0
1
2
3
4
t [s]
R
e
s
id
u
a
l
(a) Residual
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
t [s]
F
a
u
lt

s
ig
n
a
l
(b) Fault signal
Figure 5.6: The raw residual and fault signal in the step-fault simulation.
5.6.3 Simulation conclusions
From this simulation study it is not possible to draw any hard theoretical con-
clusions regarding the eectiveness of a whitening test compared to a low-pass
and threshold test. The simulation study does however indicate that, in the case
studied, the whiteness test is generally better to handle a wide variety of fault
signal characteristics than the lp-and-threshold approach. This is achieved at
the expense of relying on additional stochastic assumptions of signal properties.
116 Chapter 5. Residual Generation Based on Stochastic Linear Models
0 10 20 30 40 50 60 70 80 90 100
0
2
4
6
8
10
12
t [s]
T
w
h
ite
(a) Test quantity
0 10 20 30 40 50 60 70 80 90
0
0.2
0.4
0.6
0.8
1
t [s]
D
e
c
is
io
n
(b) Decision
Figure 5.7: Test quantity and the decision, i.e. thresholded test quantity, for
the whiteness based test in the step-fault simulation.
0 20 40 60 80 100
2
0
2
4
z
L
P
=
0
.
9
0
0
LP residuals
0 20 40 60 80
0
0.5
1
Decisions
0 20 40 60 80 100
2
0
2
4
z
L
P
=
0
.
9
5
0
0 20 40 60 80
0
0.5
1
0 20 40 60 80 100
2
0
2
4
z
L
P
=
0
.
9
7
0
0 20 40 60 80
0
0.5
1
0 20 40 60 80 100
5
0
5
10
z
L
P
=
0
.
9
9
0
t [s]
0 20 40 60 80
0
0.5
1
t [s]
Figure 5.8: Residuals and decisions for the LP-and-threshold test in the step-
fault simulation.
5.6. White residuals 117
0 10 20 30 40 50 60 70 80 90 100
5
4
3
2
1
0
1
2
3
4
t [s]
R
e
s
id
u
a
l
(a) Residual
0 10 20 30 40 50 60 70 80 90 100
1.5
1
0.5
0
0.5
1
1.5
t [s]
F
a
u
lt

s
ig
n
a
l
(b) Fault signal
Figure 5.9: The raw residual and fault signal in the sinus-fault simulation.
0 10 20 30 40 50 60 70 80 90 100
0
1
2
3
4
5
6
7
8
t [s]
T
w
h
ite
(a) Test quantity
0 10 20 30 40 50 60 70 80 90
0
0.2
0.4
0.6
0.8
1
t [s]
D
e
c
is
io
n
(b) Decision
Figure 5.10: Test quantity and the decision, i.e. thresholded test quantity, for
the whiteness based test in the sinus-fault simulation.
118 Chapter 5. Residual Generation Based on Stochastic Linear Models
0 20 40 60 80 100
2
0
2
z
L
P
=
0
.
9
0
0
LP residuals
0 20 40 60 80
0
0.5
1
Decisions
0 20 40 60 80 100
2
1
0
1
z
L
P
=
0
.
9
5
0
0 20 40 60 80
0
0.5
1
0 20 40 60 80 100
2
1
0
1
z
L
P
=
0
.
9
7
0
0 20 40 60 80
0
0.5
1
0 20 40 60 80 100
2
1
0
1
z
L
P
=
0
.
9
9
0
t [s]
0 20 40 60 80
0
0.5
1
t [s]
Figure 5.11: Residuals and decisions for the LP-and-threshold test in the sinus-
fault simulation.
0 10 20 30 40 50 60 70 80 90 100
4
3
2
1
0
1
2
3
4
5
t [s]
R
e
s
id
u
a
l
(a) Residual
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
1.5
t [s]
F
a
u
lt

s
ig
n
a
l
(b) Fault signal
Figure 5.12: The raw residual and fault signal in the ramp-fault simulation.
5.6. White residuals 119
0 10 20 30 40 50 60 70 80 90 100
0
5
10
15
t [s]
T
w
h
ite
(a) Test quantity
0 10 20 30 40 50 60 70 80 90
0
0.2
0.4
0.6
0.8
1
t [s]
D
e
c
is
io
n
(b) Decision
Figure 5.13: Test quantity and the decision, i.e. thresholded test quantity, for
the whiteness based test in the ramp-fault simulation.
0 20 40 60 80 100
2
0
2
4
z
L
P
=
0
.
9
0
0
LP residuals
0 20 40 60 80
0
0.5
1
Decisions
0 20 40 60 80 100
5
0
5
z
L
P
=
0
.
9
5
0
0 20 40 60 80
0
0.5
1
0 20 40 60 80 100
5
0
5
10
z
L
P
=
0
.
9
7
0
0 20 40 60 80
0
0.5
1
0 20 40 60 80 100
5
0
5
10
z
L
P
=
0
.
9
9
0
t [s]
0 20 40 60 80
0
0.5
1
t [s]
Figure 5.14: Residuals and decisions for the LP-and-threshold test in the ramp-
fault simulation.
120 Chapter 5. Residual Generation Based on Stochastic Linear Models
5.7 Time-discrete systems
In the deterministic case, time-discrete systems and time-continuous systems
could be handled identically by just replacing s with z and proper with causal
as shown in Chapter 3. In the stochastic case, small but important dierences
exists which are briey discussed below.
First, a few words on the discrete time spectral factorization problem. Let
r = H(z)n with n unit covariance white noise, then the spectrum of r is given
by

r
() = H(e
jT
)H
T
(e
jT
)
Thus, a time-discrete version of (5.8) becomes:
Z(z) = N
M
s
(z)
_
D
n
B
n
_ _
D
n
B
n
_
T
N
T
M
s
(z
1
)
In discrete time, a matrix A(z) is para-hermitian if A(z) = A
T
(z
1
). A J-
spectral co-factorization of Z(z) then becomes
Z(z) = P(z)JP
T
(z
1
)
where P(z) have all zeros inside and on the unit circle. However, here only
factorizations with signature J = I and Z(z) full rank on the unit-circle is
considered. This because, to the authors knowledge, no numerically reliable
factorization algorithm exists for the general case. In Polynomial Toolbox 2.5
for Matlab 5 (2001), an algorithm described in (Jezek and Kucera, 1985) is
used.
The main dierence between the time-continuous and time-discrete cases is
that properness of the residual generator can always be achieved. This is imme-
diate since a non-proper (non-causal) lter can always be made proper (casual)
by inserting a number of time-delays and since z
1
is an all-pass link, the white-
ness property is not violated. Thus, it is immediate to prove that the existence
conditions for full-rank innovation lters and whitening residual generators is
identical to the time-continuous case where the properness condition has been
removed. For example, the existence of discrete time innovation lters is given
by.
Theorem 5.7. If Z(z) is full rank, an innovation lter exists if and only if
Z(z) has no roots on the imaginary axis.
Therefore, for the innovation lter design, the algorithmic restriction to Z(z)
with no zeros on the imaginary axis is no restriction. However in the whitening
residual generator case this is a restriction as illustrated by the following small
example which shows how the continuous algorithm directly transfers to the
time-discrete case.
y =
_
1
za
1
(z1)(za)
_
(u +f) +n
5.8. The singular case 121
with [a[ < 1. It is clear that a residual generator r = y
1

1
za
u produces
white residuals in the fault-free case, an innovation lter does however not
exist. Straightforward calculations give that for any residual generator it holds
that the transfer function from noise to residual is given by
r = (z)
_
z a 0
0 1 z
_
n (5.29)
This gives that
r
(z) is given by

r
(z) = (z)Z(z)
T
(1/z) = (z)P(z)P
T
(1/z)
T
(1/z)
with
P(z) =
_
z a 0
0 z 1
_
Therefore, Q(z) = (z)P
1
(z)N
M
s
(z)P
x
is a whitening residual generator if
and only if there exists a (z) such that
Q(z) stable (z)(1/z) = 1
One such choice is (z) = [1 0] which gives the residual generator
r = y
1

1
z a
u
i.e. same as (5.29). However, to the authors knowledge, no computational tools
seems to be available that can handle general cases with zeros on the unit circle.
5.8 The singular case
This section deals with the case when it is possible to perfectly decouple the
noise. This is called the singular case since perfect decoupling of the noise is
possible if and only if the matrix Z(s) is singular.
5.8.1 Singular complications
Now follows a version of Lemma 5.1 where the non-singularity assumption is
removed. After that, a discussion on the diculties of formulating singular
versions design methodologies for innovation lters and whitening residual gen-
erators using this polynomial approach.
Lemma 5.6. There exists a (s) such that the linear time-invariant lter
Q(s) = (s)N
M
s
(s)P
x
produces white residuals if and only if (s) can be written
(s) = [(s) (s)]P
1
(s)
where P(s) is a spectral co-factor of Z(s), (s)
T
(s) = , and [(s) (s)] is
partitioned according to the signature of the spectral co-factorization.
122 Chapter 5. Residual Generation Based on Stochastic Linear Models
Proof. According to the proof of Theorem 5.1, the spectrum of r can be written

r
(s) = (s)Z(s)
T
(s) (5.30)
Note that Z(s) is a para-hermitian polynomial matrix. Now, let P(s) be a
spectral co-factor and J a signature of Z(s), i.e.
Z(s) = P(s)JP
T
(s) (5.31)
Since Z(s) is positive semidenite on the imaginary axis it is positive semidef-
inite in the whole complex plane. This implies that the signature of Z(s) has
the form
J =
_
I
r
0
mr
_
where r = rank Z(s). Let P
1
(s) be the columns of P(s) that corresponds to
the non-zero part of the signature. Insertion of (5.31) into (5.30) and denoting
(s) = (s)P
1
(s) gives

r
(s) = (s)P(s)JP
T
(s)
T
(s) = (s)
T
(s)
which also gives that
r
(s) = according to the assumptions in the theorem.
The parameterization matrix (s) is found by solving for (s) in the equation
(s) = (s)P
1
(s) (5.32)
If Z(s) is full-rank, P
1
(s) = P(s) will be a square full-rank matrix and (5.32) has
only one unique solution (s) = (s)P
1
(s). If Z(s) is rank decient, several
solutions exists, parameterized the matrix (s) as
(s) = (s)P
L
1
(s) +(s)N
P
1
(s)
where P
L
1
(s) is any left-inverse of matrix P
1
(s) and N
P
1
(s) is a basis for the
left null-space of P
1
(s). The existence of such a left-inverse is ensured by the
full column-rank property of P
1
(s). A stable inverse can be found by the inverse
of P(s). Let
P
1
(s) =
_
P
1i
(s)
P
2i
(s)
_
then it holds that P
1i
(s) is a stable left inverse of P
1
(s) and P
2i
is a basis for
the null-space of P
1
(s). Thus, in the rank decient case, (s) satises (5.32) if
and only if it can be written
(s) = [(s) (s)]P
1
(s)

5.8. The singular case 123

Remark 1:An example of an inverse P
L
1
(s) from the proof is the Moore-Penrose
inverse. However, even though det P(s) is Hurwitz, the Moore-Penrose inverse
of P
1
(s) is not necessarily stable.
Remark 2: One might argue that the parameterization from the theorem intro-
duces unnecessary poles in (s) since whole P(s) is inverted, not only P
1
(s) as
needed. However, as shown in Section 5.2.1, the spectral factor P(s) can always
be written on the form
P(s) = U(s)
_

P(s) 0
0 I
_
where P(s) is partitioned according to the signature J, matrix U(s) is unimod-
ular, and

P(s) is a spectral factor of the non-singular part. Thus, all poles in
P
1
(s) origins from the non-singular part of the spectral factorization.
Now, existence conditions of innovation lters and whitening residual gen-
erators would be performed like in Section 5.4. Such conditions that are easily
computable is dicult to state mainly because
1. The spectral factorization of a singular para-hermitian matrix is not cano-
nical, i.e. we can not assume generic row-degrees of a spectral factor as in
Theorem 5.5. To the authors knowledge, no canonical spectral factoriza-
tion exists for the singular case and no characterization of the row-degrees
of non-canonical spectral factors is available. This problem is also stated,
in another problem setting, in the conclusions of (

Sebek, 1990). There-

fore, easily computable results like Corollary 5.1 and Theorem 5.6 is not
available.
2. The freedom included in the design with (s) from Lemma 5.6 to ensure
properness and stability of the residual generator is not yet fully under-
stood.
5.8.2 When does the non-singular case occur?
Now, a result giving necessary condition for the problem to be non-singular
where the results from Section 5.4 is directly applicable.
Theorem 5.8. A sucient condition for Z(s) to be non-singular is
rank
_
D
n
B
n
_
mrank
_
D
d
B
d
_
Proof. If Z(s) is non-singular case it is possible to achieve full decoupling of
the noise n. This is evident by following the proof of Lemma 5.6 and letting
(s) = 0 and (s) ,= 0 which will give a residual generator (always possible to
get proper and stable) such that n is decoupled in the residual.
Following the proof of Theorem 5.1 it is seen that this is achievable if and
only if the left null-space of
M(s) =
_
C D
d
D
n
(sI A) B
d
B
n
_
124 Chapter 5. Residual Generation Based on Stochastic Linear Models
is non-empty. A dimensionality analysis of M(s) immediately gives the theorem.

The following corollary follows directly:

Corollary 5.2. Let all sensors in process (5.1) be subjected to measurement
noise, then Z(s) is non-singular.
Proof. If all sensors are subjected to noise, rank D
n
= m and Theorem 5.8 is
trivially fullled.
5.9 Stochastic descriptor example
In Section 3.7 it was shown how the minimal polynomial basis approach was
applicable to also nominal descriptor systems. It is straightforward to realize
that the extension described in this chapter for stochastic systems is equally
applicable in the stochastic descriptor case.
The same calculations as in the proof of Theorem 5.1 gives that Equa-
tion (5.4) also holds for descriptor systems, i.e. the design algorithm is exactly
as for the non-descriptor case with M
s
(s) given by (3.24).
Again, consider the descriptor example from Section 3.7 with sensor noise
added, i.e. the stochastic descriptor model is given by
E x = Ax +B
u
u +B
f
1
f
1
+B
f
2
f
2
y = Cx +n
The design goal was to design residual generators Q
1
(s) and Q
2
(s), decoupling
f
2
and f
1
respectively. In both cases, innovation lters exists and computing
the spectral co-factor for Z(s) in the rst design gives
P(s) =
_
_
42s
2
2.6 10
2
s 6.5 10
2
0.013s
2
+ 0.99s 42
0.12s + 0.58 0.23s 12
0.38s 1.8 0.9s 44
0.0019s
2
0.13s 84
0.73s + 36
0.42s 20
_
_
The free parameter (s) is chosen to be (s) = 1/

3[1 1 1]. The procedure

is repeated for the second residual generator Q
2
(s) and Figure 5.15 shows the
magnitude of the transfer functions from fault f
1
and f
2
to residual 1 and 2
respectively. The gure also reveals that fault f
1
will be more dicult to detect
than fault f
2
since both residuals have the same fault-free behavior and the
gain from f
1
to r
1
is signicantly lower than the f
2
to r
2
gain. Figure 5.16
demonstrates that the whiteness property of the residuals in the fault-free case
holds since the spectra are constant for all .
5.9. Stochastic descriptor example 125
10
2
10
0
10
2
300
250
200
150
100
50
0
[rad/s]
|
G
r
f (
s
)
|

[
d
B
]
10
2
10
0
10
2
300
250
200
150
100
50
0
[rad/s]
(a) Fault sensitivity in residual 1
10
2
10
0
10
2
300
250
200
150
100
50
0
[rad/s]
|
G
r
f (
s
)
|

[
d
B
]
10
2
10
0
10
2
300
250
200
150
100
50
0
[rad/s]
(b) Fault sensitivity in residual 2
Figure 5.15: Transfer functions from both faults to residual 1 and 2 respectively.
The design specications are met since fault f
2
is decoupled in residual 1, fault
f
1
in residual 2 and the residuals are sensitive to the remaining faults.
10
0
10
1
10
2
10
3
10
8
6
4
2
0
2
4
6
8
10
[rad/s]
M
a
g
n
it
u
d
e

(
d
B
)
(a)
r
1
(j)
10
1
10
0
10
1
10
2
10
3
10
8
6
4
2
0
2
4
6
8
10
[rad/s]
M
a
g
n
it
u
d
e

(
d
B
)
(b)
r
2
(j)
Figure 5.16: Spectrum of residuals r
1
and r
2
in the fault-free case.
126 Chapter 5. Residual Generation Based on Stochastic Linear Models
5.10 Conclusions
Linear residual generator design for stochastic models on state-space, descriptor,
and transfer-function form has been considered, both in continuous-time and
discrete time. The basic design formulation is based on innovation lters as
formulated by Nikoukhah (1994) where the basic stochastic design requirement
is that the residuals should be zero mean and white in the fault-free case. The
problem formulation is also further developed to whitening residual generators,
a more general class of residual generators.
The main extension introduced by considering whitening residual generators
instead of innovation lters is to consider the existence of any scalar residual
generator producing white residuals. Previously with innovation lters, only the
case where whiteness of the residual can be achieved without loosing any design
freedom is considered. In such cases, the stochastic requirements can never
reduce fault detectability in the residual. Here however, all residual generators
that satises the whiteness requirement is parameterized. If there among these
exist a residual generator with acceptable fault sensitivity, a feasible design has
been found.
The algorithm is an extension of the deterministic design algorithm Chap-
ter 3, where in a rst step all residual generators for the deterministic problem is
derived. Then, in a second step, all whitening residual generators are found by
post ltering. Since the algorithm from Chapter 3 handled descriptor systems,
so do this stochastic extension. In contrast to the deterministic design problem,
the continuous and discrete-time problems is not analogous and a small, but
important, dierence exists. In the continuous-time case, an innovation lter or
whitening residual generator may not exists because no proper whitening resid-
ual generator exists. For discrete-time, this is never an issue since a necessary
number of time-delays always can be introduced and a time-delay is an all-pass
link, i.e. the whiteness property is not violated. Computationally, the two main
steps in the design algorithm is extraction of a polynomial basis for the left
null-space of a polynomial matrix (as in the deterministic case) followed by a
J-spectral co-factorization of a para-hermitian polynomial matrix. For both
these operations there exists good numerical tools. To show the simplicity of
the design methods, full Matlab implementation of innovation lter design is
included in Appendix 5.A.
The design algorithm is successfully demonstrated on a number of non-
trivial examples that in detail illustrates dierent properties of the algorithm
and the design problems. One key observation is shown by example in which an
innovation lter, as proposed by (Nikoukhah, 1994), does not exist. However,
due to the extended class of residual generators considered, a residual generator
do exist that fullls all design specications.
5.A. innovationlter.m 127
Appendix
5.A innovationlter.m
function Q=innovationfilter(A,Bu,Bd,Bn,C,Du,Dd,Dn)
% INNOVATIONFILTER - Innovation filter design
%
% Given a fault-free system desciption
% x= Ax + Bu u + Bd d + Bn n
% y = Cx + Du u + Dd d + Dn n
%
% Design (if one exists) an innovation filter Q(s).
% Any innovation filter W(s) can then be parameterized by
% W(s) = L(s)Q(s)
% where L(s) is invertible and satisfies L(s)L(s)=I
%
% Syntax: Q=innovationfilter(A,Bu,Bd, Bn,C,Du,Dd,Dn)
tol=1e-5; % Tolerance for detecting purely imaginary zeros
nx = length(A); % Number of states
nmeas = size(C,1); % Number of measurements
ndist = size(Bd,2); % Number of disturbances
Mx = [C Dd;-(s*eye(nx)-A) Bd];
Px = [eye(nmeas) -Du;zeros(nx,nmeas) -Bu];
Nmx = null(Mx.).;
Z = Nmx*[Dn;Bn]*[Dn;Bn]*Nmx;
if rank(Z)<size(Z,1)
error(Singular cases is not covered by this function);
end
if (max(deg(Nmx,row)-deg(Nmx*[Dn;Bn],row))>0) | ...
(sum(abs(real(roots(Z)))<tol)>0)
error(No innovation filter exists);
end
[P,J] = spf(Z.); P = P.; %% Spectral factorization
[Qa,Qb,Qc,Qd] = lmf2ss(Nmx*Px,P);
Q = minreal(ss(Qa,Qb,Qc,Qd));
128 Chapter 5. Residual Generation Based on Stochastic Linear Models
6
Residual Generation Based on
Uncertain Linear Models
This chapter addresses the problem of synthesizing and analyzing robust resid-
ual generators in the presence of parametric uncertainties and deterministic
additive disturbances that inuences the process. This problem is well studied
and there exists many papers with dierent design focuses on the subject, e.g.
(Frank and Ding, 1994; Mangoubi et al., 1995; Edelmayer et al., 1994; Man-
goubi et al., 1994, 1992; Eich and Oehler, 1997; Sauter et al., 1997). A common
trait of these works is that an optimization problem is stated to synthesize an
optimally robust residual generator. Among these optimization criterions two
common classes of optimization criterions stand out, one can be stated on the
form:
min
Q(s)
|G
rd
(s)|
|G
rf
(s)|
where Q(s) is the residual generator. Dierent choices of norm | | have been
proposed. With such a criterion, model uncertainties are modeled as additive
input signals. However, model uncertainty most often appears as parametric
uncertainty where physical parameters can only be determined within a range.
Theory and methodology for design of robust controllers for such systems is
available in -synthesis of H

/H
2
controllers (Zhou et al., 1995). This theory
is also applicable to the ltering problem, and is therefore easily adopted to
the residual generation problem. This leads to a second class of optimization
criterions that is in common use:
min
Q(s)
|r f|
|d|
129
130 Chapter 6. Residual Generation Based on Uncertain Linear Models
i.e. the residual should be a fault estimator in the presence of uncertainties d.
Since focus in this chapter is on parametric uncertainties, this type of optimiza-
tion criterion is used. A main observation is that using the residual as a fault
estimator is not always a good idea; a more rened criterion is often necessary.
For this purpose, it is advantageous to introduce a reference model that de-
scribes the desired behavior of the residuals with respect to faults. A theory
is developed where the reference model idea and a new optimization criterion
are key elements. It is shown how this reference model is an intuitive design
parameter in the synthesis problem, but that it has to be chosen with care.
6.1 Robust residual generation
The system under consideration is again assumed to be on the form
y = G

u
(s)u +G

d
(s)d +G

f
(s)f (6.1)
like in (3.1) with a small dierence, the superscript . Matrices G

u
(s), G

d
(s),
and G

f
(s) are all rational transfer matrices and the superscripts indicate
that the model is subject to bounded parametric uncertainties. The class of
uncertainties studied here is a quite general class where it is assumed that the
uncertain system can be described by a linear fractional transformation (LFT).
Section 6.3.2 further claries which types of systems that can be represented in
this way. More details on linear fractional transformations can be found in e.g.
(Zhou et al., 1995, Chapter 10).
The residual generator is, as in previous chapters, a nite dimensional linear
lter Q(s) that uses available known signals, i.e. y and u, to form a residual, r,
that can be used to detect and isolate the dierent faults f.
r = Q(s)
_
y
u
_
(6.2)
The basic requirement on Q(s), besides being H

(i.e. stable and proper), is

that the residual should be insensitive to control actions, u, and disturbances,
d, but it should also be sensitive to faults f.
The principle of fault isolation considered here is again structured residuals
as described in Section 2.1.3, where a subset of faults are decoupled in each
residual. Note that in Chapter 3, non-monitored faults were included in the
d-vector. This re-organization of d and f vectors is, for the sake of convenience,
not performed in this chapter. The fault-vector f will in this chapter include
all modeled faults and the vector d all modeled disturbances.no
By generating a set of residuals where dierent subsets of faults are decou-
pled in each residual, fault isolation is possible. Inserting (6.1) into (6.2) gives
r = Q(s)
_
G

u
(s) G

d
(s)
I 0
_ _
u
d
_
+Q(s)
_
G

f
(s)
0
_
f (6.3)
6.2. Reference model 131
Nominally, to achieve decoupling of u and d, the rst term of (6.3) must be 0
while the second term must be ,= 0 which was discussed in detail in Chapter 3.
This issue will be re-addressed in Section 6.5 where the freedom available in the
robust design problem is explored. However, with uncertain models it is in most
cases impossible to get the rst term = 0 for all possible , i.e. for all possible
instances of uncertainties, without loosing some or all of the desired fault sen-
sitivity. Note that it is not always so; in some cases parametric uncertainty can
be transformed into unknown input signals that can be decoupled with methods
based on nominal models, see e.g. (Patton and Hou, 1998). However, generally
some tradeo between sensitivity to faults and disturbance/uncertainty atten-
uation is required.
6.2 Reference model
When synthesizing a robust residual generator, it is desired that the design
freedom available should be used to achieve both robust fault-free behavior
and robust detection performance. The question arises how this performance
should be formulated. A natural choice is to introduce a reference model, R(s),
to describe the desired behavior of the residual vector r, with respect to faults,
f. Dene desired residual behavior, r
0
(s), of the residual, via the reference
model as
r
0
= R(s)f(s)
which then describes both fault-free behavior and the fault response in the
residual. The matrix R(s) is an arbitrary H

transfer matrix of appropriate

dimensions. The idea of a reference model has successfully been used to describe
signal behavior in other elds like controller design (

Astrom and Wittenmark,

1984) and adaptive control (

Astrom and Wittenmark, 1989).

It is of course necessary that the reference model, R(s), contains the nec-
essary structure for Q(s) to be a residual generator corresponding to a desired
inuence structure.
Example 6.1
Suppose there are three modeled faults and it is desired to design a residual
that:
1. Decouples the rst fault
2. Responds to high frequency components ( >
h
) of the second fault
3. Responds to low frequency components ( <
l
) of the third fault
i.e. the residual should correspond to a row [0 1 1] in the residual structure. An
example of R(s) could then be
R(s) =
_
0
s
s+
h
1
s/
l
+1
_
where
h
and
l
was chosen to reect the frequency ranges of interest.
132 Chapter 6. Residual Generation Based on Uncertain Linear Models
Later on in this chapter a design algorithm is outlined that tries to achieve the
fault response given by the reference model while being as robust as possible
against variation of the unknown parameters. However, it will later show in
Section 6.4 that it is important to put some thought into the reference model
selection since an ill specied reference model will lead to an optimization prob-
lem where the algorithm will spend unnecessary freedom on achieving the badly
posed reference model. In the next section, the design algorithm and how to
compute the solution is outlined and the reference model selection is revisited
in Section 6.5.
6.3 Computation of a robust residual generator
This section will provide an algorithm, that is based on standard H

-optimi-
zation, that computes an optimally robust residual generator. First in Sec-
tion 6.3.1, an optimization criterion is posed that incorporates the specied
desired residual behavior, the reference model. Then in Section 6.3.2, the re-
sults needed to compute the optimally robust residual generator is presented.
6.3.1 Robustness criterion
The optimization criterion used here is formulated as a robust H

-ltering
problem (Zhou et al., 1995), with an intuitive and appealing interpretation
which is given after Eq. (6.6). The criterion is
J = sup
vL
2
|r
0
(t) r(t)|
2
|v(t)|
2
(6.4)
where v = [u
T
f
T
d
T
]
T
. The norm | |
2
is dened as
|u|
2
2
=
_

0
[u(t)[
2
dt
The optimization criterion J is thus the worst case distance between the residual
r and the idealized residual r
0
, dened by transfer matrix R(s), normed by
the size of the inputs. The optimal residual generator Q(s) is the lter that
minimizes J for all such that || for some scalar > 0.
The optimization criterion J can be rewritten as
J = sup
vL
2
|r
0
r|
2
|v|
2
= sup
vL
2
|T
zv
(s)v|
2
|v|
2
= |T
zv
(s)|

(6.5)
where z(t) = r
0
(t) r(t), and
T
zv
(s) =
_
G

ru
(s)
_
R(s) G

rf
(s)
_
G

rd
(s)
_
(6.6)
6.3. Computation of a robust residual generator 133
is the transfer matrix from v(t) to z(t). The transfer matrices from u to r,
G

ru
(s), from d to r, G

rd
(s), and from f to r, G

rf
(s) all depend on the residual
generator Q(s). Minimizing J, i.e. minimizing the -norm of expression (6.6),
has a simple interpretation. Keeping |T
zv
(s)|

small ensures that the rst and

third element are small which makes sure that the inuence from u and d on
the residual are attenuated. The middle term keeps fault sensitivity, and also
shapes the fault to residual transfer function G

rf
(s) by minimizing the distance
to the reference model R(s).
The optimization/performance index minimizes the absolute dierence be-
tween R(s) and G

rf
(s). A reasonable assumption is that it is the relative dif-
ference that need to be minimized, otherwise in e.g. high-gain models even very
small relative errors will dominate the loss function and therefore move away
optimization focus from robustness to fault sensitivity in an unwanted manner.
Also, dierent signal levels between e.g. u and d have similar eects. Therefore
is it important to normalize and/or weight the model appropriately to avoid
such eects.
Remark
It is obvious that the optimal value of J also delivers constraints on the size
(norm) of r in the fault-free and faulty case. Thus, the optimal value of the
optimization criterion can be used to guide robust threshold selection. This
issue will be briey revisited in Section 6.8.
6.3.2 Computational framework
The above stated problem can be naturally formulated in an LFT (Linear Frac-
tional Transformation)-framework, and then be solved by conventional methods
from robust and optimal control, i.e. H

-ltering. Now follows a brief descrip-

tion on the formulation and computation of an optimal solution to (6.5). First,
a standard denition:
Denition 6.1 (LFT). Let M be a complex matrix partitioned as
M =
_
M
11
M
12
M
21
M
22
_
C
(p
1
+p
2
)(q
1
+q
2
)
and let
l
C
q
2
p
2
and
u
C
q
1
p
1
be two other complex matrices. Then,
provided that all inverses exists, the lower LFT of M with respect to
l
is
dened as
T
l
(M,
l
) M
11
+M
12

l
(I M
22

l
)
1
M
21
and the upper LFT of M with respect to
u
is dened as
T
u
(M,
u
) M
22
+M
21

u
(I M
11

u
)
1
M
12
A graphical representation of these two transformations can be seen as:
134 Chapter 6. Residual Generation Based on Uncertain Linear Models

-
M

z
a)
M

v

-
b)
where T
l
(M,
l
) and T
u
(M,
u
) are the transfer functions from v to z in Figure
a and b respectively.
Now, to be able to use standard optimization theory, an LFT formulation of
the optimization problem (6.6) stated in the Section 6.3.1 is needed. Figure 6.1
shows such a representation, where all parametric uncertainty is lumped into
the block-diagonal matrix , matrix P(s) is an augmented plant description
and Q(s) is the residual generator. The LFT framework is quite general and
P(s)

v =
_
_
u
f
d
_
_

-
Q(s)

z = r
0
r
-
_
y
u
_

r
Figure 6.1: LFT formulation of residual generation problem
for example, all polynomial, rational functions can be represented by an LFT
(Zhou et al., 1995). Thus, the class of uncertainties is that are covered by the
LFT representation is quite large.
Now follows a small, static example illustrating an LFT formulation of a
system with parametric uncertainties.
Example 6.2
Consider a rational, static transfer function with two uncertain parameters
1
and
2
given by
z =
a +b
2
1
1 +c
1

2
w = Gw
Even though the inuence from uncertainty is non-linear, it is straightforward
6.3. Computation of a robust residual generator 135
to realize that G = T
u
(M, ) with
M =
_

_
0 0 0 0 1
1 0 0 0 0
0 b 0 c a
0 0 1 0 0
0 b 0 c a
_

_
=
_

1
I
3
0
0
2
_
When the optimization problem is formulated according to Figure 6.1, there
exist algorithms minimizing J with respect to Q(s) by e.g. -synthesis. The
algorithm used in this work is basic DK-iterations which, although heuristic and
with no convergence guarantees, have in practice shown reliable performance
(Balas et al., 1993). The parametric uncertainties are here assumed complex
to be able to use standard tools in Matlab. This gives a conservative solution
since the parametric uncertainties modeled are physical constants which are real.
A better solution would be to restrict the uncertainties corresponding to physical
parameters to real values. This would lead to a mixed problem, with both
real and complex uncertainties, which is addressed in e.g. (Helmersson, 1995).
However, such topics are outside the scope of this thesis and does not alter the
main ideas. The DK-iteration procedure nds a solution by iteratively solving
for Q(s) (the K-step) and iteratively solving for some scaling matrices (the D-
step). Details on this procedure can be found in e.g. in (Zhou et al., 1995).
Some details on the K-step is given here because it gives some insight in how
to, in the residual generation problem, form the augmented plant description
P(s).
The standard H

-ltering problem
The K-step in the design of the residual generation is to solve a standard
H

-ltering problem (Zhou et al., 1995).

Suppose a dynamic system is described by
x = Ax +B
1
v (6.7a)
r
0
= C
1
x +D
11
v (6.7b)
y

= C
2
x +D
21
v (6.7c)
where y

is the measurables/known variables and r

0
the signal that is to be
estimated. To simplify expressions, the following standard notation is used
_
A B
C D
_
= C(sI A)
1
B +D
The ltering problem is then to nd a linear estimator r = Q(s)y

that mini-
mizes
sup
vL
2
|z|
2
|v|
2
136 Chapter 6. Residual Generation Based on Uncertain Linear Models
_
_
A B
1
C
1
D
11
C
2
D
21
_
_

r
0
Q(s)
f

r ?

r
0
r
Figure 6.2: Block diagram of the H

-ltering problem.
where z = r
0
r. Figure 6.2 shows a block representation of the ltering
problem. The ltering problem in Figure 6.2 can then be reformulated in an
LFT framework as is shown in Figure 6.3. The residual generation problem
P(s)

v
Q(s)

r
0
r
-
y

r
P(s) =
_
_
A B
1
0
C
1
D
11
I
C
2
D
21
0
_
_
Figure 6.3: LFT formulation of the H

-ltering problem.
can be tted into the problem formulation (6.7) by setting v
T
= [u
T
f
T
d
T
],
y
T

= [y
T
u
T
], and letting r
0
= R(s)f. Therefore the augmented matrix P(s)
is formed via (6.7) and Figure 6.3.
When P(s) is formed, the problem is ready to be solved and a solution to
the ltering problem is characterized by the following theorems, more or less,
directly from (Zhou et al., 1995). The solution given by the theorems does
not nd the optimal solution because it is generally hard to obtain. However,
a suboptimal solution where |T
l
(P, Q)|

< is more easy to come by. A

bisection algorithm can then be used to iterate to a solution suciently close
to optimum.
The rst theorem gives an existence test of a solution for a given which
can be used to iterate to a solution near optimum. The second theorem pa-
rameterizes all solutions that fullls |T
l
(P, Q)|

< given that one exists.

These results involve the solution of Ricatti equations and before the results are
presented, some notation is needed. When solving Ricatti equations
A

X +XA+XRX +Q = 0
it is useful to dene an associated Hamiltonian matrix which has the form
H =
_
A R
Q
T
A
_
6.3. Computation of a robust residual generator 137
where A, Q, and R are real nn matrices with Q and R symmetric. In particu-
lar, we are interested in Ricatti equations where there exists a unique, stabilizing
solution X. The set of all Hamiltonian matrices such that the associated Ri-
catti equation has such a unique, stabilizing solution is denoted dom(Ric) and
the solution with X = Ric(H). Conditions for when H dom(Ric) and other
properties of Ricatti equations can be found in for example (Zhou et al., 1995).
Now, we are ready to state the results.
Theorem 6.1 (Existence). Suppose (C
2
, A) is detectable and
_
AjI B
1
C
2
D
21
_
has full row rank for all . Let D
21
be normalized and D
11
partitioned con-
formably as
_
D
11
D
21
_
=
_
D
111
D
112
0 I
_
.
Then there exists a causal lter Q(s) H

such that J < if and only if

(D
111
) < and J

dom(Ric) with Y

= Ric(J

) 0 where

R =
_
D
11
D
21
_ _
D
11
D
21
_

_

2
0
0 0
_
J

=
_
A

0
B
1
B

1
A
_

_
C

1
C

2
B
1
D

11
B
1
D

21
_

R
1
_
D
11
B

1
C
1
D
21
B

1
C
2
_
.
Proof. See (Zhou et al., 1995).
Theorem 6.2 (Parameterization). If the conditions in Theorem 6.1 are
satised, then all rational causal lters Q(s) = T
l
(M

, F) satisfying J =
|T
zv
(s)|

< are parameterized by

M

(s)

F(s)

r
-

Q(s)
where F(s) H

, |F(s)|

< and
M

(s) =
_
_
A+L
2
C
2
+L
1
D
112
C
2
L
1
D
112
L
2
L
1

D
12
C
1
D
112
C
2
D
112

D
12
C
2
I 0
_
_
138 Chapter 6. Residual Generation Based on Uncertain Linear Models
where

D
12
is any real matrix satisfying

D
12

D

12
= I
2
D
111
D

111
and
_
L
1
L
2

=
_
B
1
D

11
+Y

1
B
1
D

21
+Y

R
1
.
Proof. Start with a general parameterization of H

controllers, e.g. in (Zhou

et al., 1995, p. 451) and utilize that B
2
= 0 and D
12
= I. Straightforward
but lengthy calculations give the parameterization above.
6.4 Background example on reference model de-
sign
The idea of a reference model for residual behavior has been introduced, and was
in the previous section used in the criterion for robust design. Before going into
details of the design of the reference model, a simple stylized static example
will illustrate the necessity of using a well formulated reference model. The
discussion forms a background and leads to the detailed design in Section 6.5.
Example 6.3
Consider the static system
y
1
= (1 +
1
)x + 3f
1
y
2
= (2 +
2
)x + 4f
2
where x is an unknown disturbance and [[ <
1
2
. Suppose it is desired that the
disturbance x does not inuence the residual while f
1
and f
2
has signicant
inuence, which is the same as saying that r should respond to a row [1 1] in a
residual structure.
The residual generator is formed as
r = q
1
y
1
+q
2
y
2
First attempt
A naive choice of R(s) without any further thought might be to aim for unit
gain from both faults, i.e. R(s) = [1 1]. A design is performed as described in
the previous section. The resulting residual generator is denoted q
a
= [q
T
1
q
T
2
]
T
.
Robustness is evaluated by determining the worst-case gain from the disturbance
to the residual. This corresponds to measuring the size of the third term in
(6.6). Since there is no control signal in this example, this measure indicates
the performance of the lter in a no-fault situation.
max
||<
1
2
|k
rx
| = max
||<
1
2
|[(1 +
1
) (2 +
2
)]q
a
| = 1.683
That is, the worst-case gain from x to r is k
rx
= 1.683.
6.4. Background example on reference model design 139
Second attempt
A little more thought on the problem leads to a second attempt based on a
nominal design. If no uncertainty is present, then = 0 and a natural choice of
q would satisfy
q
2
=
1
2
q
1
to completely decouple the disturbance x. The relation between the gain from
f
1
to r, k
rf
1
, and f
2
to r , k
rf
2
, is then related by
k
rf
1
k
rf
2
=
3q
1
4q
2
=
2 3
1 4
=
3
2
Using this nominal information will signicantly improve robustness properties
of the design. Let the elements in R(s) be proportional to the nominal gains
from the faults to the residual k
rf
1
and k
rf
2
,
R(s) = [
3
2
1] (6.8)
Note that this choice of R(s) also corresponds to a [1 1] residual structure. A
new design with this R(s) results in a q
b
, normed to be of the same size as
q
a
, in the sense that [q
a
[ = [q
b
[. This is important to enable a fair robustness
comparison. The second design has robustness properties
max
||<
1
2
|[(1 +
1
) (2 +
2
)]q
b
| = 0.7906
Comparison
Thus, a signicant improvement on robustness properties is achieved in the
second design compared to the rst design. This is because in the rst design,
R(s) reected unrealistic performance specications that inicted unnecessary
poor robustness properties on the residual generator. This issue will be further
explored below and in the concluding example in Section 6.7.
It is worth noting that even in a no uncertainty case with
1
=
2
= 0,
the rst design attempt would not have found a solution where the unknown
disturbance x is completely decoupled, even though such a solution exists. The
reason for this, which is the same reason that made the second solution more
robust, is that during optimization, freedom is spent to make G
rf
(s) R(s)
as small as possible. If an unrealistic R(s) is used, freedom is spent on an im-
possible task that often is of no importance to the FDI problem. A suitable
choice of R(s) helps focusing on the robustness properties while keeping fault
sensitivity and conforming to the specied performance. Note that it is straight-
forward to introduce weighting matrices for f, u, d, and respectively to help
the optimization procedure to get a proper trade-o between robustness- and
140 Chapter 6. Residual Generation Based on Uncertain Linear Models
P(s)

W
v

v =
_
_
u
f
d
_
_

-
Q(s)

W
z

z
-
_
y
u
_

r
Figure 6.4: Weighting matrices
performance-focus in the optimization. This is an important non-trivial step
in the design process. Examples of weighting matrices are shown in Figure 6.4
where W
v
(s) = block-diag[W
u
(s), W
f
(s), W
d
(s)]. The weighted counterpart of
(6.6) is then
T
zv
(s) = W
z
(s)
_
G

ru
(s)W
u
(s) [R(s) G

rf
(s)]W
f
(s) G

rd
(s)W
d
(s)

The issue illustrated above needs to be considered also when only one fault at
a time is supposed to inuence the residual, e.g. in a diagonal residual structure
or when the notion of dont care (Nyberg, 1999b) is used. For example, consider
a fault that is only weakly detectable (Nyberg and Nielsen, 1997), i.e. G
rf
(0) = 0
for all residual generators. Figure 6.5 shows typical magnitude bode plot of
the transfer function to a residual from a weakly detectable fault. Then, by
specifying R(s) = 1, for low frequencies the second component in (6.6) will
be large regardless of Q(s), and the robustness properties will therefore be de-
emphasized in the criterion. A more suitable reference model would then be
something similar to
R(s) =
s
s +
c
6.5 Forming the reference model
From the example above it is clear that a poorly chosen reference model, i.e. a
reference model with unrealistic performance properties, can result in a residual
generator with unnecessary poor robustness properties. To form a methodology
on how to select the reference model, attainable reference model is dened.
Denition 6.2 (Attainable Reference Model). A reference model is said
to be attainable if, with no model uncertainty, there exists a residual generator
with the specied fault response.
6.5. Forming the reference model 141
10
1
10
0
10
1
20
18
16
14
12
10
8
6
4
2
0

|
G
r
f
(
s
)
|

[
d
B
]
[rad/s]
Figure 6.5: Typical transfer function (amplitude plot) from a weakly detectable
fault to the residual
In other words, a reference model is attainable if, for some xed values of
the uncertain parameters, there exists a residual generator that produces the
fault response specied by the reference model. In the examples used here, the
xed values of the uncertain parameters are set to their nominal values.
The main idea is thus to use a nominal design of the residual generator
to shape the reference model when synthesizing the robust residual generator,
thus assuring attainable reference models. This is to avoid specifying an unre-
alistic performance criterion and thereby inicting unnecessary poor robustness
properties on the residual generator.
The formation of the criterion for the robust design is straightforward, given
that a nominal residual generator, i.e. a Q
nom
(s), has been derived that nomi-
nally fullls all demands. The reference model R(s) is then selected as
R(s) = Q
nom
(s)
_
G
f
(s)
0
_
(6.9)
since this is the nominal fault to residual transfer function, compare with
Eq. (6.3).
Of course, if no design based on a nominal model is available that meets
the requirements of the application, then no feasible design with an uncertain
model is available either.
6.5.1 Nominal design
The residual generator design problem based on nominal linear models can be
solved by any of a number of methods. The design method used here to shape
the reference model, is the algorithm from Chapter 3. The reason for this is that
142 Chapter 6. Residual Generation Based on Uncertain Linear Models
the design freedom available in the nominal design, and therefore in the design
of the reference model, is clear, explicit and condensed into the two polynomial
matrices (s) and c(s) from (3.6) and (3.7).
Remark
It is possible that, even with nominal models, no disturbance decoupling resid-
ual generator can be achieved that conforms to the desired residual structure,
i.e. dim N
L
(M(s)) = 0. Then a residual generator can be synthesized where
disturbances and faults (according to the residual structure) are approximately
decoupled. This is often stated as an optimization problem and solved in dif-
ferent ways by many dierent methods, e.g. by H

methods (Frank and Ding,

1994; Qiu and Gertler, 1994), singular value truncation (Lou et al., 1986) and
other methods (W unnenberg, 1990). If the resulting residual generator, which
doesnt perfectly decouple disturbances, still complies with the requirements
on the diagnosis system, the nominal residual generator can be used to form a
reference model according to (6.9) and robust synthesis following this chapter
can be performed.
Example of reference model design
Below follows an example on how to, given a system description, form the
reference model.
Example 6.4
Assume a 1-input, 2-output, second order process with sensor and actuator
faults:
y =
_

s+

s(s+)
_
(u +f
3
) +
_
f
1
f
2
_
= G(s)u + [I G(s)]
_
_
f
1
f
2
f
3
_
_
= G(s)u +L(s)f
A (minimal) polynomial basis for N
L
(M(s)) can, in this no-disturbance case,
be obtained by a row-reduced and irreducible left MFD of G
u
(s) = D
1
(s)N(s)
according to Theorem 3.2. It is straightforward to verify that
G(s) =
_

s+

s(s+)
_
=
_
s + 0
1 s
_
1
_

0
_
= D(s)
1
N(s)
Then it holds that
N
M
(s) = [D(s) N(s)] =
_
s + 0
1 s 0
_
Now it is possible to form a matrix K(s) that characterizes all attainable fault
to residual transfer functions, i.e. it characterizes all possible residual structures.
The matrix K(s) becomes
K(s) = N
M
(s)
_
L(s)
0
_
=
_
s + 0
1 s 0
_
6.5. Forming the reference model 143
Here K(s) has full row-rank and it is therefore clear that residuals corresponding
to rows in the residual structure with maximum 1 zero is possible, i.e. it is
possible to perfectly decouple maximum 1 fault. If a residual with structure
[1 0 1] is desired, then Q
nom
(s) and R(s) become e.g.
Q
nom
(s) =

s +
[s + 0 ] (6.10)
R(s) = Q
nom
(s)
_
L(s)
0
_
=

s +
[s + 0 ] (6.11)
This is achieved by choosing a (s) = [1 0] that selects the rst row in N
M
(s)
and adding arbitrary rst order dynamics D
F
(s) =
1

s+1, i.e. is free to shape

the fault response in the residual which species desired response time.
6.5.2 Discussion of design choices
When choosing Q
nom
(s), and thereby the reference model R(s), two degrees
of freedom have to be chosen by the designer (here it is assumed that in the
rst-step nominal design, perfectly decoupling residual generators exist):
The numerator may need to be chosen, especially since the space of decou-
pling numerators can have dimensionality greater than one, which means
that there are several numerators to choose from.
The denominator need to be chosen. The dynamics (poles) of the residual
generator is completely free, conditioned that the degree is greater or equal
to the numerator degree.
These choices are then held x during optimization. Instead of specifying the
poles individually, they could of course be obtained almost automatically using
a band-width requirement together with a Butterworth structure (

Astrom and
Wittenmark, 1984).
It might be argued that by xating these two degrees of freedom limits
the freedom in the optimization to achieve robustness against parametric vari-
ations and that these two variables should be optimized over. However, any
xed set of uncertain parameters would only inuence the numerator of the
residual generator. This is clear from Chapter 3, since when xating the
problem is a nominal problem as in Chapter 3 where the poles, i.e. matrix
D
F
(s) was completely free to choose, except for degree constraints. Therefore,
it is believed that xating the poles does not severely inuence the robustness
properties of the optimal residual generator. If dim N
L
(M(s)) > 1, dierent
base vectors (numerators) might generate dierent robustness properties of the
resulting residual generator. Therefore is it desired to optimize over all linear
combinations (s) with some constraint on (s) such that fault sensitivity is
kept. If no such constraint is enforced, (s) = 0 (R(s) = 0) and Q(s) = 0
144 Chapter 6. Residual Generation Based on Uncertain Linear Models
would be optimal, which obviously is a useless design. One possible heuristic
way, of course with no convergence guarantees, of performing this optimization
is to iteratively solve for Q(s) and (s), much like the use of D-K iterations
when performing -synthesis. The search for (s) could be constrained to e.g.
all constant vectors with [[ = 1.
6.6 Summary of design method
An advantage with the problem formulation used here is that it is possible to
incorporate desired fault to residual performance in the optimization criterion,
i.e. the reference model. This is not included in any other optimization criterion
used for synthesizing robust residual generators.
Recall the optimization criterion
J = sup
vL
2
|r
0
r|
2
|v|
2
= |T
zv
(s)|

=
= |
_
G

ru
(s) [R(s) G

rf
(s)] G

rd
(s)

where G

ru
(s) is the transfer function from u to r, G

rf
(s) is the transfer function
from f to r, and G

rd
(s) is the transfer function from d to r. These are given
directly by Eq. (6.3) and they are all dependent on the residual generator Q(s).
The optimization goal is then to nd the Q(s) that minimizes J under bounded
parametric uncertainties . The rst and third components of J then decouples
u and d and the second shapes the residual response of the faults.
The synthesis procedure is as follows
1. Perform a nominal design Q
nom
(s) as in Chapter 3.
2. Select the reference model as R(s) = Q
nom
(s)
_
G
f
(s)
0
_
3. Introduce weighting matrices to focus on important frequency ranges
4. Optimize (minimize) J under structured uncertainties with -synthesis.
5. Apply model reduction techniques on the resulting residual generator and
re-evaluate robust performance via -analysis.
The main designer interaction is in step 1 where the nominal design, and thereby
the reference model, is selected and step 3 where knowledge of the process or
demands on the diagnosis system is used to shape the optimization criteria.
Design choices during step 1 were discussed in Section 6.5.2.
A special case of Eq. (6.4) is treated in e.g. (Niemann and Stoustrup, 1997),
where integrated residual generator and controller design is presented. There,
the residual generator design is performed by letting r
0
(t) = f(t) in (6.4), i.e.
strive for a diagonal residual structure with unit gain from fault i to residual i,
6.7. Illustrative dynamic example 145
the optimization criteria becomes the same. This may be a good design objective
if it is possible. However, as shown previously in this chapter, this is not always
possible and might lead to a poor design. A simple, generic, condition for
nominal models on the existence of a residual generator with diagonal inuence
structure is (assuming detectability of faults)
m rank [G
d
(s) G
f
(s)]
where m is the number of measurements.
6.7 Illustrative dynamic example
In this section, the concepts introduced in this chapter are applied to a small
dynamic example with parametric uncertainties. Three designs are made, one
nominal, one with an ad-hoc reference model and one design with a reference
model based on the nominal design. Robustness properties of the three designs
are then compared.
6.7.1 Model
Reconsider Example 6.4, inspired from a second order DC-servo model. Assume
uncertainties in moment of inertia, modeled by
1
, and in the viscous friction,
modeled by
2
. The model can then be written on a state-space form as
(1 +
1
) x
1
= ( +
2
)x
1
+(u +f
3
)
x
2
= x
1
y
1
= x
1
+f
1
y
2
= x
2
+f
2
Block representation of the fault-free system where the input and output
signal to/from
i
is denoted
i
and
i
respectively:
u -

-
+
f -
_ y
1
-
_
-
y
2

f
6

2
6

2

2

1
?

1

1
Straightforward calculations give the transfer-functions from
i
and u to
i
and
146 Chapter 6. Residual Generation Based on Uncertain Linear Models
y.

1
=
s
s +
(u
1

2
)

2
=
1
s +
(u
1

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

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

2
)
An upper LFT formulation of the uncertain model (without faults) is then
s
s+
s
s+
s
s+
1
s+
1
s+

s+
1
s+
1
s+

s+
1
s(s+)
1
s(s+)

s(s+)

y
1

y
2
-

2
-

1

1

u
6.7.2 Residual generator specications
Suppose a residual generator is to be synthesized to conform to a [1 0 1] struc-
ture, i.e. decouple sensor fault 2 while keeping sensitivity to the other two faults.
The parameters are = 1 and = 4. Three dierent designs are considered:
The nominal design is made via polynomial methods as in Example 6.4.
The nominal residual generator Q
nom
(s) was given by (6.10).
The robust residual generator with a non-attainable R(s) = [1 0 1] is syn-
thesized using the method in Section 6.6. The resulting residual generator
is denoted Q
na
(s). This choice of R(s) is related to the rst naive attempt
in Example 6.3.
A robust design with an attainable R(s), based on the nominal design
Q
nom
(s), is performed using the method in Section 6.6 and is denoted
Q
att
(s).
6.7.3 Robustness comparison
It is hard to make a fair comparison of robustness because of the many dierent
denitions of robust residual generators. Here, the designs have been normed
to have equal static fault-to-residual gains (the 2-norm of G
rf
(0) is equal for all
designs). Then, the robustness comparison measure used is how large G
ru
(s)
6.8. Robust fault detection 147
gets under bounded parametric uncertainties in a worst-case situation. This
is the same robustness evaluation criterion as in the static Example 6.3 and
corresponds to the operation of the residual generators in a fault-free situation.
Since, compared to Example 6.3, this is a dynamic example, we need some
theory to perform the robustness comparison. What is needed is a theorem
regarding robust performance from (Zhou et al., 1995) where the worst-case
size of G

ru
(s) can be computed. Consider the following model setup, where the
LFT describes the transfer function from u to r:
P(s)

r
(s)

-
Then, it holds that
Theorem 6.3 (Robust Performance). Let > 0. For all |(s)|

<
1

,
the loop shown above is well-posed, internally stable and |T
u
(P, )|

if
and only if
sup
R

P
(P(j))
where

P
(P(j)) is the structured singular value and

P
=
__
0
0
f
_
: ,
f
C
q
2
q
2
_
So the smaller , i.e. peak-value of , the smaller is the worst-case value of
|G
ru
(s)|

and thereby a more robust fault-free behavior of the residual gener-

ator is achieved.
Figure 6.6 shows the -plot,

P
(G
ru
(j)) for the three designs. The inter-
pretation of the plot is, according to the theorem above, when

P
(G
ru
(j)) <
then for all ||

< 1/ it holds that |G

ru
(s)| < . The plot gives that
Q
nom
(s),Q
na
(s), and Q
att
(s) gives = 0.6, 0.55, 0.5 respectively. Thus, Q
na
(s)
suers from up to 15% unnecessary loss of robustness compared to Q
att
(s). In
fact, Q
na
(s) is even worse than the nominal design for large frequencies. This
means that a robust design not considering a proper reference model can in fact
be less robust than a nominal design.
This example, and the example in Section 6.4, show that a signicant in-
crease in robustness was achieved by robust design with a reference model chosen
via a nominal design.
6.8 Robust fault detection
The performance index J has up to now only been used as an optimization
index, used to synthesize the residual generator. However, the optimal value of
148 Chapter 6. Residual Generation Based on Uncertain Linear Models
10
3
10
2
10
1
10
0
10
1
10
2
0.3
0.35
0.4
0.45
0.5
0.55
0.6
0.65

[rad/s]
Figure 6.6: Plot of

P
(G
ru
(j)) to evaluate robustness properties. The
dashed, dash-dotted, and the solid lines corresponds to Q
nom
(s), Q
na
(s) and
Q
att
(s) respectively.
the optimization function can also be used to evaluate absolute performance.
Below follows a theorem presenting sucient conditions on J such that robust
fault detection is possible. This means that it is possible to nd a threshold
which (some norm of) the residual always exceeds when a fault, of a certain
size, is present and never exceeds when no fault is present.
Theorem 6.4 (Robust Fault Detection). Let the problem be normalized so
that
|f
i
|
2
1 i = 1, . . . , q if f
i
,= 0
|d|
2
1
|u|
2
1
and denote
f
= min

(R(j)) and = sup

p
(T
zv
(j)).
Then, if
<

f

2 +

3
there exists a J
th
> 0 such that
|r|
2
< J
th
if f = 0
|r|
2
> J
th
if |f
i
| 1
for all ||

< 1/.
6.8. Robust fault detection 149
Proof. By denition:
v L
2
:
|r
0
r|
2
|v|
2
(6.12)
In the fault-free case r
0
= 0 and |v|
2
=
_
|d|
2
2
+|u|
2
2
. With (6.12), an upper
bound on |r|
2
can be derived as:
|r|
2

_
|d|
2
2
+|u|
2
2

2 (6.13)
thus |r|
2

2 in the fault-free case.

With fault(s), r
0
= R(s)f and v =
_
_
u(t)
f(t)
d(t)
_
_
. With (6.12) a lower bound on
|r|
2
can be derived as:
|r|
2
|r
0
|
2
|r
0
r|
2
|R(s)f|
2
|v|
2


f
|f|
2

_
|f|
2
2
+ 2 = J
1
(6.14)
If min
f
2
1
J
1
>

2 there exists a J
th
,

2 < J
th
< min
f
2
>1
J
1
such
that
|r|
2
< J
th
if f = 0
|r|
2
> J
th
if |f| 1
is fullled. From (6.14) it follows that
min
f
2
1
J
1
=
_

f

3 if
f
if >
f
(6.15)
The condition on ,
<

f

2 +

3
(6.16)
ensures that <
f
. Equation 6.16 is equivalent with

3 >

2
from which it follows, by inspection of (6.13) and (6.15), that
min
f
2
1
J
1
>

2.

Example 6.5
Consider a design where R(s) = I, i.e. r is an estimation of the fault-vector.
Then any fault, of size larger than 1, can be robustly detected if
<
1

2 +

3
150 Chapter 6. Residual Generation Based on Uncertain Linear Models
6.9 Conclusions
A theory for robust residual design has been developed where the key element
is the use of a reference model. The reference model represents desired per-
formance of the synthesized residual generator. It is a condensed formulation
including structural requirements, to make the synthesized residual generator
t in a larger diagnosis system based on structured residuals. It also includes
performance issues such as fault response in the residual. Without considering
structural constraints, it is possible to form unrealistic performance demands
and it is shown by examples how this can de-emphasize the robustness parts of
the optimization and lead to a design with unnecessary poor robustness prop-
erties. A methodology how to select realistic reference models is presented
where all design freedom available is explicit and intuitive. The optimization
algorithms used to synthesize the residual generator rely on established and
ecient methods. The designer of a diagnosis system is thus provided with a
tool where it is easy to specify desired behavior without violating structural re-
quirements. Finally it is shown how the optimization procedure, theoretically,
provides an absolute measure on the size of disturbances/model uncertainty
that is acceptable to be able to robustly detect a fault.
7
Conclusions
The objective when supervising technical processes is to alarm an operator when
a fault is detected and also identify one, or possibly a set of components, that
may have been the cause of the alarm. Diagnosis is an expansive subject, both
in the academic research community and in industry. This is partly due to the
fact that nowadays, more applications have more embedded computing power
and more available sensors than before. This provide means for, in addition
to more advanced control strategies, advanced diagnosis algorithms and on-line
supervision by the process itself.
A fundamental part of many model-based diagnosis algorithms are so called
residuals. A residual is a signal that reacts to a carefully chosen subset of
the considered faults and by generating a suitable set of such residuals, fault
detection and isolation can be achieved. Design of such residual generators has
been the topic of this work.
A common thread throughout this dissertation is the development of sys-
tematic design and analysis methods for residual generators based on a number
of dierent model classes. In particular, exploration of available design freedom
and simple, intuitive, parameterization of that freedom is pursued. In addition,
it is considered important that there exist readily available computer tools for
all design algorithms where as little as possible diagnosis specic code need to
be developed. Of course, the numerical performance of the algorithms is impor-
tant to be able to cope with large or ill-conditioned design problems and still
produce feasible solutions.
The model classes that were considered are:
deterministic linear systems on state-space, transfer-function, and descrip-
151
152 Chapter 7. Conclusions
tor form
non-linear systems described by polynomial dierential-algebraic equa-
tions
stochastic linear systems on state-space, transfer-function, and descriptor
form
linear systems on state-space and transfer-function form with bounded
parametric model uncertainties
The simplest case of these are deterministic linear models on state-space or
transfer-function form. It is benecial to study these fundamental systems to
thoroughly investigate the residual generation problem and explore fundamen-
tal properties of the design problem. It is shown in Chapter 2 how consistency
relations, representing the most local relations in the model, forms a suitable
basis for residual generator design. The main reason for this is good robustness
properties of the residual generator. Based on this, a design algorithm, the min-
imal polynomial basis approach, is formed by using theory and algorithms for
polynomial matrices. The design algorithm mainly consist of two standard op-
erations on polynomial matrices, computation of a basis for the left null-space
of a polynomial matrix and realization of an MFD on state-space form. All
design freedom is contained in the row-vector (s) from (3.6) and the scalar
polynomial c(s) in (3.7). A consequence of the formulation of the design algo-
rithm is that the extension of the algorithm to also cover descriptor systems is
immediate. The numerical performance is shown to be good on relatively large
state-space and descriptor systems.
To further help guide the selection of the available design freedom, additional
modeling is necessary. Two natural choices of model classes to consider are
then models with parametric uncertainties or models with noise descriptions
of uncertainty. When considering linear systems with uncertain parameters, a
design algorithm is formulated based on an optimization criterion and standard
H

-ltering techniques. A fundamental part of the optimization criterion is the

reference model that species desired performance of the residual generator. A
key observation is that a natural, and often used, reference model can impose
unnecessary poor robustness properties of the residual generator because the
criterion violates structural properties of the model. Based on this observation,
a methodology to select a well formed reference model is developed in Chapter 6
that is based on the minimal polynomial basis approach from Chapter 3.
For the stochastic linear systems systems, design of innovation lters is
studied and also an extension to a more general class of residual generators,
whitening residual generators. Also here, the nominal design procedure forms a
foundation for the design steps and only one additional operation, spectral fac-
torization, is needed for the design of residual generators. Because of this, also
stochastic descriptor systems can be handled without any modication of the
algorithm. In a series of non-trivial examples, both state-space and descriptor,
153
it is shown how the design works and what design freedom that is still available.
To show the merits of the extended problem formulation, a design example is
included where an innovation lter does not exist while a whitening residual
generator exist that meets all design specications.
The last model class considered is non-linear models. To achieve the rather
ambitious goals, the class of non-linear models is restricted to models described
by polynomial dierential-algebraic equations. For this, relatively large, class
of non-linearities, powerful computational tools exists and a design algorithm
with great similarities with the linear design algorithm is developed. The key
step is here, instead of computing a basis for the left null-space of a polynomial
matrix, to compute a Gr obner basis for a specied elimination ideal. A main
problem with the approach is the inherent algorithmic complexity which makes
it dicult to handle anything but small models. Structural analysis of the model
equations, which can be performed for huge models, is then shown to provide a
way to manage this complexity. To study this, it is shown how a relatively small
4 state model poses an intractable problem on a standard personal computer.
However, after a structural analysis is used to reduce the problem, a computer
algebra package provide a solution in only 0.02 seconds.
In conclusion, for the four classes of models studied, new methods have
been developed. The methods fullls requirements generation of all possible
solutions, availability of computational tools, and numerical soundness. The
methods also provide the diagnosis system designer with a set of tools with well
specied and intuitive design freedom.
154 Chapter 7. Conclusions
Notation
R,C The eld of real/complex numbers.
H

The Hardy space that consists of all complex-valued functions

that are analytical and bounded in the open left half-plane.
H

The real rational subspace of H

is denoted H

which then
consists of all proper and real rational stable transfer matrices.
rank A(s) Denotes normal rank of A(s) R
mn
[s].
R
mn
[s] M(s) R
mn
[s] if and only if M(s) is a matrix of dimension
m n with polynomial elements in variable s with real coe-
cients.
R
mn
(s) M(s) R
mn
(s) if and only if M(s) is a matrix of dimension
mn with rational elements in variable s with real coecients.
N
L
(M(s)) The rational left null-space of a matrix M(s) R
mn
[s] or
M(s) R
mn
(s).
N
M
(s) The rows of N
M
(s) forms a minimal polynomial basis for
N
M
s
(s).
u, y, f, d Signals representing control signals, measurements, faults, and
disturbances respectively. Signals are denoted in lower-case,
both in time-domain and frequency domain. When it is clear
from context if the signals is specied in time or frequency
domain, argument t or s dropped for notational convenience.
G
uv
(s) The transfer function from v to u.
_
A B
C D
_
The transfer function C(sI A)
1
B +D
155
156
References
H. Aling and J.M. Schumacher. A nine-fold canonical decomposition for linear
systems. Int. J. of Control, 39(4):779805, 1984.
K.

Astrom and B. Wittenmark. Computer Controlled Systems - Theory and
Design. Prentice Hall, 1984.
K.

Astrom and B. Wittenmark. Adaptive Control. Prentice Hall, 1989.
G.J. Balas, J.C. Doyle, K. Glover, A. Packard, and R. Smith. -Analysis and
Synthesis Toolbox. Math Works Inc., 1993.
M. Basseville, A. Benveniste, G. Moustakides, and A. Rougee. Detection and
diagnosis of changes in the eigenstructure of nonstationary multivariable sys-
tems. Automatica, 23(4):479489, 1987.
M. Basseville and I.V. Nikiforov. Detection of Abrupt Changes - Theory and
Application. Prentice Hall, 1993.
R.V. Beard. Failure Accommodation in Linear System Through Self Reorgani-
zation. PhD thesis, MIT, 1971.
M. Blanke, R. Izadi-Zamanabadi, S. Bgh, and C. Lunau. Fault-tolerant con-
trol systems - a holistic view. Control Eng. Practice, 5(5):693701, 1997.
W. Boege, R. Gebauer, and H. Kredel. Some examples for solving systems of
algebraic equations by calculating groebner bases. Journal of Symbolic Com-
puting, 2(1):8398, 1986.
157
158 References
F.M. Callier. On polynomial matrix spectral factorization by symmetric factor
extraction. IEEE Trans. Automatic Control, 30(5):453464, 1985.
C.T. Chen. Linear System Theory and Design. Holt, Rinehart and Winston,
New York, 1984.
J. Chen and R.J. Patton. A re-examination of fault detectability and isolability
in linear dynamic systems. In Proc. IFAC Safeprocess1994, pages 567573,
Espoo, Finland, 1994.
J. Chen and R.J. Patton. Robust Model-Based Fault Diagnosis for Dynamic
Systems. Kluwer, 1999.
E.Y. Chow and A.S. Willsky. Analytical redundancy and the design of robust
failure detection systems. IEEE Trans. on Automatic Control, 29(7):603614,
1984.
D. Cox, J. Little, and D. OShea. Ideals, Varieties, and Algorithms. Springer-
Verlag, 1991.
A.J.van der Schaft. On realization of nonlinear systems described by higher
order dierential equations. Math. Systems Theory, 19:239275, 1987.
S.X. Ding, E.L. Ding, and T. Jeinsch. A numerical approach to analysis and
design of FDI systems. IEEE Conf. on Decision and Control, 1998.
S.X. Ding, E.L. Ding, and T. Jeinsch. An approach to analysis and design of
observer and parity relation based FDI systems. volume P of Proc. of IFAC
1999 World Congress, pages 3742, Beijing, P.R. China, 1999a.
X. Ding and P.M. Frank. Fault detection via factorization approach. Systems
& control letters, 14(5):431436, 1990.
X. Ding, L. Guo, and T. Jeinsch. A characterization of parity space and its
application to robust fault detection. IEEE Trans. on Automatic Control, 44
(2):337342, 1999b.
A. Edelmayer, J. Bokor, and L. Keviczky. An H

ltering approach to robust

detection of failures in dynamic systems. In Proceedings Decision & Control,
pages 30373039, 1994.
J. Eich and R. Oehler. On the application of the generalized structured singular
value to robust fdi system design. In Proc. IFAC Safeprocess1997, pages 885
890, Hull, UK, 1997.
G.D. Forney. Minimal bases of rational vector spaces, with applications to
multivariable linear systems. SIAM J. Control, 13(3):493520, May 1975.
K. Forsman. Constructive Commutative Algebra in Nonlinear Control Theory.
PhD thesis, Link oping University, 1991.
References 159
P.M. Frank and X. Ding. Frequency domain approach to optimally robust
residual generation and evaluation for model-based fault diagnosis. Automat-
ica, 30(5):789804, 1994.
P.M. Frank. Fault diagnosis in dynamic systems using analytical and
knowledge-based redundancy - a survey and some new results. Automatica, 26
(3):459474, 1990.
E. Frisk and L. Nielsen. Robust residual generation for diagnosis including
a reference model for residual behavior. In Proc. IFAC World Congress, vol-
ume P, pages 5560, Beijing, China, 1999.
E. Frisk and M. Nyberg. Using minimal polynomial bases for fault diagnosis.
In Proc. European Control Conference, 1999.
E. Frisk and M. Nyberg. A minimal polynomial basis solution to residual
generation for fault diagnosis in linear systems. Automatica, 37(9):14171424,
2001.
E. Frisk and M. Nyberg. Residual generation for fault diagnosis of systems
described by general linear dierential algebraic equations. In submitted to
Proc. 15th IFAC World Congress, Barcelona, Spain, 2002.
E. Frisk. Order of residual generators bounds and algorithms. In Proc. IFAC
Safeprocess2000, pages 599604, Budapest, Hungary, 2000a.
E. Frisk. Residual generatior design for non-linear, polynomial systems a
gr obner basis approach. In Proc. IFAC Safeprocess2000, pages 979984, Bu-
dapest, Hungary, 2000b.
E. Frisk. Residual generatior in linear stochastic systems - a polynomial ap-
proach. In Proc. IEEE Conference on Decision and Control, Orlando, USA,
2001.
J. Gertler, X. Fang, and Q. Luo. Detection and diagnosis of plant failures:
the orthogonal parity equation approach. Control and Dynamic Systems, 37:
159216, 1990.
J. Gertler and D. Singer. A new structural framework for parity equation-based
failure detecation and isolation. Automatica, 26(2):381388, 1990.
J. Gertler. Analytical redundancy methods in fault detection and isolation;
survey and synthesis. In IFAC Fault Detection, Supervision and Safety for
Technical Processes, pages 921, Baden-Baden, Germany, 1991.
J. Gertler. Fault Detection and Diagnosis in Engineering Systems. Marcel
Dekker, 1998.
M. Green, K. Glover, D.J.N. Limebeer, and J. Doyle. A J-spectral factorization
approach to H

-control. SIAM J. Contr. Opt., 28:13501371, 1990.

160 References
C. Guernez, J.Ph. Casar, and M. Staroswiecki. Extension of parity space
to non-linear polynomial dynamic systems. In Proc. IFAC Safeprocess 1997,
pages 861866, Hull, UK, 1997.
W. Hamscher, L. Console, and J. de Kleer, editors. Readings in Model-Based
Diagnosis. Morgan Kaufmann Publishers, 1992. ISBN 1-55860-249-6.
A. Helmersson. Methods for Robust Gain Scheduling. PhD thesis, Link oping
University, 1995.
M. Hou and P.C. M uller. Tracking control for a class of descriptor systems. In
Proc. 13th IFAC World Congress, pages 109114, San Fransisco, USA, 1996.
M. Hou. Fault detection and isolation for descriptor systems, chapter 5. In ,
Patton et al. (2000), 2000.
R. Isermann. Fault diagnosis of machines via parameter estimation and knowl-
edge processing - tutorial paper. Automatica, 29(4):815835, 1993.
A. Isidori. Nonlinear control systems. Springer, 3rd edition, 1995.
V.A. Jakubovic. Factorization of symmetric matrix polynomials. Dokl. Akad.
Nauk SSSR, 194(3), 1970.
J. Jezek and V. Kucera. Ecient algorithm for matrix spectral factorization.
Automatica, 21(6):663669, 1985.
M. Jirstrand. Nonlinear control design by quantier elimination. Journal of
Symbolic Computation, 24(2):137152, August 1997.
M. Jirstrand. Constructive Methods for Inequality Constraints in Control. PhD
thesis, Link oping University, 1998.
B. Kagstr om. RGSVD - an algorithm for computing the kronecker structure
and reducing subspaces of singular AB pencils. SIAM J. Sci. Stat. Comput,
7(1):185211, 1986.
T. Kailath. Linear Systems. Prentice-Hall, 1980. ISBN 0-13-536961-4.
R. Karcanias and G. Kalogeropoulos. Right, left characteristic sequences and
column, row minimal indices of a singular pencil. Int. J. of Control, 47(4):
937946, 1988.
F. Kratz, S. Bousghiri, and W. Nuninger. A nite memory observer structure
of continuous descriptor systems. In Proc. American Control Conference, pages
39003904, Seattle, USA, 1995.
V. Krishnaswami, G.C. Luh, and G. Rizzoni. Fault detection in IC engines
using nonlinear parity equations. In Proceedings of the American Control Con-
ference, pages 20012005, Baltimore, Maryland, 1994.
References 161
V. Krishnaswami and G. Rizzoni. Nonlinear parity equation residual gener-
ation for fault detection and isolation. In IFAC Fault Detection, Supervision
and Safety for Technical Processes, pages 305310, Espoo, Finland, 1994.
M. Krysander and M. Nyberg. Diagnosis of large scale dierential-algebraic
systems utilizing structural properties. In Submitted to IFAC World Congress,
Barcelona, Spain, 2002.
H. Kwakernaak and M.

Sebek. Polynomial J-spectral factorization. IEEE
Tran. on Automatic Control, 39(2):315328, 1994.
H. Kwakernaak. A descriptor algorithm for the spectral factorization of poly-
nomial matrices. In Proc. Robust Control Design, Prague, Czech Republic,
2000a. IFAC.
H. Kwakernaak. H
2
-optimization theory and applications to robust control
design. In Proc. Robust Control Design, Prague, Czech Republic, 2000b. IFAC.
P. Lancaster and M. Tismenetsky. The theory of matrices. Academic Press,
second ed. edition, 1985.
M. Larsson. Behavioral and Structural Model Based Approaches to Discrete
Diagnosis. PhD thesis, Link oping University, Sweden, 1999.
P. Lindskog. Methods, Algorithms and Tools for System Identication Based
on Prior Knowledge. PhD thesis, Link oping University, 1996.
L. Ljung and T. Glad. On global identiability for arbitrary model
parametrizations. Automatica, 30(2):265276, 1994.
X.C. Lou, A.S. Willsky, and G.C. Verghese. Optimally robust redundancy
relations for failure detection in uncertain systems. Automatica, 22(3):333
344, 1986.
J.M. Maciejowski. Multivariable Feedback Design. Addison Wesley, 1989.
J.F. Magni and P. Mouyon. On residual generation by observer and parity
space approaches. IEEE Trans. on Automatic Control, 39(2):441447, 1994.
R. Mangoubi, B.D. Appleby, and J. Farrell. Robust estimation in fault detec-
tion. In Proceedings Decision & Control, pages 23172322, 1992.
R. Mangoubi, B.D. Appleby, G.C. Verghese, and W.E. VanderVelde. A robust
failure detection and isolation algorithm. In Proceedings Decision & Control,
pages 23772382, 1995.
R. Mangoubi, B.D. Appleby, and G.C. Verghese. Stochastic interpretation of
H

and robust estimation. In Proceedings Decision & Control, pages 3943

3948, 1994.
162 References
M.A. Massoumnia, G.C. Verghese, and A.S. Willsky. Failure detection and
identication. IEEE Trans. on Automatic Control, AC-34(3):316321, 1989.
R.K. Mehra and J. Peschon. An innovations approach to fault detection and
diagnosis in dynamic systems. Automatica, 7:637640, 1971.
L.A. Mironovskii. Functional diagnosis of linear dynamic systems. Automation
and Remote Control, pages 11981205, 1980.
P. Misra, P. Dooren, and A. Varga. Computation of structural invariants of
generalized state-space systems. Automatica, 30(12):19211936, 1994.
P.C. M uller and M. Hou. On the observer design for descriptor systems. IEEE
Trans. on Automatic Control, 38(11):16661671, 1993.
P.C. M uller. Descriptor systems: pros and cons of system modelling by
dierential-algebraic equations. Mathematics and computers in simulation,
53:273279, 2000.
H. Niemann and J. Stoustrup. Integration of control and fault detection:
nominal and robust design. In IFAC Fault Detection, Supervision and Safety
for Technical Processes, pages 341346, Hull, UK, 1997.
R. Nikoukhah. Innovations generation in the presence of unknown inputs:
Application to robust failure detection. Automatica, 30(12):18511867, 1994.
M. Nyberg and L. Nielsen. Parity functions as universal residual generators and
tool for fault detectability analysis. In IEEE Conf. on Decision and Control,
1997.
M. Nyberg and L. Nielsen. A universal Chow-Willsky scheme and detectability
criteria. IEEE Trans. on Automatic Control, 45(1):152156, 2000.
M. Nyberg. Model based diagnosis of both sensor-faults and leakage in the
air-intake system of an si-engine. In Proc. SAE World Congress, number 1999-
01-0860 in SAE, Detroit, USA, 1999a.
M. Nyberg. Model Based Fault Diagnosis: Methods, Theory, and Automotive
Engine Applications. PhD thesis, Link oping University, May 1999b.
M. Nyberg. Criterions for detectability and strong detectability of faults in
linear systems. In Proc. IFAC Safeprocess2000, 2000.
R.J. Patton, P.M. Frank, and R.N. Clark, editors. Issues of Fault Diagnosis
for Dynamic Systems. Springer, 2000.
R.J. Patton and M. Hou. Design of fault detection and isolation observers: A
matrix pencil approach. Automatica, 34(9), 1998.
References 163
R.J. Patton. Robust model-based fault diagnosis:the state of the art. In IFAC
Fault Detection, Supervision and Safety for Technical Processes, pages 124,
Espoo, Finland, 1994.
C. De Persis and A. Isidori. A geometric approach to nonlinear fault detection
and isolation. IEEE Trans. on Automatic Control, 46(6):853865, 2001.
The Polynomial Toolbox 2.5 for Matlab 5. Polyx, Czech Republic. URL:
http://www.polyx.com, 2001.
Z. Qiu and J. Gertler. Robust FDI systems and H

-optimization. In IFAC
Fault Detection, Supervision and Safety for Technical Processes, pages 247
252, Espoo, Finland, 1994.
J. Ritt. Dierential Algebra. Dover Publications, 1950.
N. Sadegh. Minimal realization of nonlinear systems described by input-output
dierence equations. IEEE Trans. on Automatic Control, 46(5):698710, 2001.
M. Sampath, R. Sengupta, S. Lafortune, K. Sinnamohideen, and D.C. Teneket-
zis. Diagnosability of discrete event systems. IEEE Trans. on Automatic Con-
trol, 40(9):15551575, 1995.
M. Sampath, R. Sengupta, S. Lafortune, K. Sinnamohideen, and D.C. Teneket-
zis. Failure diagnosis using discrete-event models. IEEE Trans. on Control
Systems Technology, 4(2):105124, 1996.
D. Sauter, F. Rambeaux, and F. Hamelin. Robust fault diagnosis in an H

set-
ting. In IFAC Fault Detection, Supervision and Safety for Technical Processes,
pages 879884, Hull, UK, 1997.
D.N. Shields. Robust fault detection for descriptor systems. In Proc. IASTED
International Conf., modelling, identication, and control., pages 301304,
Anaheim, USA, 1994.
R.F. Sincovec, A.N. Erisman, E.L. Yip, and M.A. Epton. Analysis of descriptor
systems using numerical algorithms. IEEE Trans. on Automatic Control, 26
(1):139147, 1981.
T. S oderstr om and P. Stoica. System Identication. Prentice Hall, 1989.
E.D. Sontag and Y. Wang. Input/output equations and realizability. In Proc.
Int. Symposium MTNS-89, volume 1, pages 125132, 1990.
M. Staroswiecki and G. Comtet-Varga. Analytical redundancy relations for
fault detection and isolation in algebraic dynamic systems. Automatica, 37(5):
687699, 2001.
164 References
R.C.W. Strijbos. Calculation of right matrix fraction descriptions; an algo-
rithm. In Proceedings of the 4th IEEE Mediterranean Symposium on New Di-
rections in Control and Automation,, pages 478482, Maleme, Krete, Greece,
1996.
M.

Sebek. An algorithm for spectral factorization of polynomial matrices with
any signature. Technical Report ISSN 0169-2690, Faculty of Applied Mathe-
matics, University of Twente, The Netherlands, 1990.
N. Viswanadham, J.H. Taylor, and E.C. Luce. A frequency-domain approach
to failure detection and isolation with application to GE-21 turbine engine
control systems. Control - Theory and advanced technology, 3(1):4572, March
1987.
S.H. Wang, P. Dorato, and E.J. Davison. On the computation of minimal
indices for linear multivariable systems. International Journal of Control, 21
(4):537543, 1975.
W.M. Wonham. Linear Multivariable Control: a Geometric Approach.
Springer-Verlag, 2nd edition, 1979.
J. W unnenberg. Observer-Based Fault Detection in Dynamic Systems. PhD
thesis, University of Duisburg, 1990.
E.L. Yip and R.F. Sincovec. Solvability, controllability, and observability of
continuous descriptor systems. IEEE Trans. on Automatic Control, 26(2):
702707, 1981.
Q. Zhang, M. Basseville, and A. Benveniste. Fault detection and isolation
in nonlinear dynamic systems: A combined input-output and local approach.
Automatica, 34(11):13591373, 1998.
A.N. Zhirabok. Parity space approach to fault diagnosis in nonlinear systems.
In European Control Conference, 1999.
K. Zhou, J.C. Doyle, and K. Glover. Robust and Optimal Control. Prentice
Hall, 1995.
Author Index
A
Aling, H . . . . . . . . . . . . . . . . . . . . . . . . . . 33
Appleby, B.D. . . . . . . . . . . . . . . . . . . . 129
B
Basseville, M. . . . . . . 9, 65, 71, 88, 113
Beard, R.V. . . . . . . . . . . . . . . . . . . . . . . . 11
Benveniste, A . . . . . . . . . . . . . . . . . 65, 71
Blanke, M. . . . . . . . . . . . . . . . . . . . . . . . . . 1
Boege, W. . . . . . . . . . . . . . . . . . . . . . . . . 71
Bgh, S . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Bokor, J . . . . . . . . . . . . . . . . . . . . . . . . . 129
Bousghiri, S. . . . . . . . . . . . . . . . . . . . . . . 47
C
Callier, F.M. . . . . . . . . . . . . . . . . . . . . . . 92
Casar, J.Ph. . . . . . . . . . . . . . . . . . . . . . . 65
Chen, C.T. . . . . . . . . . . . . . . . . 21, 29, 57
Chen, J. . . . . . . . . . . . . . . . . . . . . 8, 21, 95
Chow, E.Y. . . . . . . . . . . . . . . . . 16, 33, 36
Comtet-Varga, G. . . . . . . . . . . . . . 14, 76
Console, L . . . . . . . . . . . . . . . . . . . . . . . . . 7
Cox, D. . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
D
Ding, E.L. . . . . . . . . . . . . . . . . . . . . . . . . 17
Ding, S.X. . . . . . . . 17, 36, 37, 129, 142
Dooren, P . . . . . . . . . . . . . . . . . . . . . . . . 33
Doyle, J.C. . . . . . . . . . . . . . . . . . . . . . . 129
E
Edelmayer, A. . . . . . . . . . . . . . . . . . . . 129
Erisman, A.N. . . . . . . . . . . . . . . . . . . . . 47
F
Fang, X. . . . . . . . . . . . . . . . . . . . . . . . . . . 33
Forney, G.D. . . . . . . . . . . . . . . . . . . 21, 57
Forsman, K. . . . . . . . . . . . . . . . . . . . . . . 72
Frank, P.M. . . . . . . 16, 36, 37, 129, 142
Frisk, E . . . . . . . . . . . . . . . . . . 3, 4, 34, 55
G
Gebauer, R . . . . . . . . . . . . . . . . . . . . . . . 71
Gertler, J. . . . . . . . . . . . 8, 1113, 22, 33
Glad, T. . . . . . . . . . . . . . . . . . . . . . . . . . . 74
Glover, K . . . . . . . . . . . . . . . . . . . . 87, 129
Green, M . . . . . . . . . . . . . . . . . . . . . . . . . 87
Guernez, C . . . . . . . . . . . . . . . . . . . . . . . 65
Guo, L. . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
H
Hamscher, W . . . . . . . . . . . . . . . . . . . . . . 7
165
166 Author Index
Helmersson, A . . . . . . . . . . . . . . . . . . . 135
Hou, M. . . . . . . . . . . . . . . 47, 50, 51, 131
I
Isermann, R. . . . . . . . . . . . . . . . . . . . . . . . 9
Isidori, A. . . . . . . . . . . . . . . . . . . . . . 74, 76
Izadi-Zamanbadi, R . . . . . . . . . . . . . . . . 1
J
Jakubovic. . . . . . . . . . . . . . . . . . . . . . . . . 93
Jezek, J . . . . . . . . . . . . . . . . . . . . . . 92, 120
Jeinsch, T. . . . . . . . . . . . . . . . . . . . . 17, 36
Jirstrand, M. . . . . . . . . . . . . . . 16, 69, 75
K
Kailath, T. . . . . . 21, 25, 26, 29, 50, 57
Kalogeropoulos, G. . . . . . . . . . . . . 35, 36
Karcanias, R . . . . . . . . . . . . . . . . . . 35, 36
Keviczky, L. . . . . . . . . . . . . . . . . . . . . . 129
de Kleer, J . . . . . . . . . . . . . . . . . . . . . . . . . 7
Kratz, F . . . . . . . . . . . . . . . . . . . . . . . . . . 47
Kredel, H. . . . . . . . . . . . . . . . . . . . . . . . . 71
Krishnaswami, V. . . . . . . . . . . . . . . . . . 65
Krysander, M. . . . . . . . . . . . . . . . . . . . . 77
Kucera, V. . . . . . . . . . . . . . . . . . . . 92, 120
Kwakernaak, H. . . . . . . . . . . 87, 92, 106
Kagstrom, B. . . . . . . . . . . . . . . . . . . . . . 33
L
Lafortune, S . . . . . . . . . . . . . . . . . . . . . . . 7
Lancaster, P . . . . . . . . . . . . . . . . . . 57, 59
Larsson, M. . . . . . . . . . . . . . . . . . . . . . . . . 7
Limebeer, D.J.N. . . . . . . . . . . . . . . . . . . 87
Lindskog, P. . . . . . . . . . . . . . . . . . . . . . . 66
Little, J. . . . . . . . . . . . . . . . . . . . . . . . . . . 67
Ljung, L . . . . . . . . . . . . . . . . . . . . . . . . . . 74
Lou, X.C. . . . . . . . . . . . . . . . . . . . . . . . . 142
Luce, E.C. . . . . . . . . . . . . . . . . . . . . . . . . 37
Luh, G.C. . . . . . . . . . . . . . . . . . . . . . . . . . 65
Luo, Q. . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
M
Maciejowski, J.M. . . . . . . . . . . . . . . . . . 38
Magni, J.F. . . . . . . . . . . . . . . . . . . . . . . . 25
Mangoubi, R. . . . . . . . . . . . . . . . . . . . . 129
Massoumnia, M.A. . . . . . . . . . . . . . . . . 21
Mehra, R.K. . . . . . . . . . . . . . . . . . . . . . 113
Mironovskii, L.A. . . . . . . . . . . . . . . . . . 35
Misra, P . . . . . . . . . . . . . . . . . . . . . . . . . . 33
Moustakides, G . . . . . . . . . . . . . . . . . . . 71
Mouyon, P. . . . . . . . . . . . . . . . . . . . . . . . 25
M uller, P.C. . . . . . . . . . . . . . . . . . . . 47, 51
N
Nielsen, L. . . . . . . . . . . . . . . . . . . . . . . . . . 4
Niemann, H. . . . . . . . . . . . . . . . . . . . . . 144
Nikhoukhah, R . . . . . . . . . . . 87, 89, 110
Nikiforov, I.V. . . . . . . . . . 9, 71, 88, 113
Nikoukhah, R. . . . . . . . . . . . . . . . . . . . . 25
Nuninger, W. . . . . . . . . . . . . . . . . . . . . . 47
Nyberg, M3, 8, 11, 26, 31, 37, 40, 47,
50, 55, 95
O
OShea, D . . . . . . . . . . . . . . . . . . . . . . . . 67
P
Patton, R.J. . . . . . . . . 8, 21, 52, 95, 131
de Persis, C. . . . . . . . . . . . . . . . . . . . . . . 76
Peschon, J . . . . . . . . . . . . . . . . . . . . . . . 113
R
Ritt, J . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
Rizzoni, G . . . . . . . . . . . . . . . . . . . . . . . . 65
S
Sadegh, N . . . . . . . . . . . . . . . . . . . . . . . . 74
Sampath, M . . . . . . . . . . . . . . . . . . . . . . . 7
Sangupta, R . . . . . . . . . . . . . . . . . . . . . . . 7
van der Schaft, A.J. . . . . . . . . . . . . . . . 74
Schumacher, J.M. . . . . . . . . . . . . . . . . . 33

Sebek, M . . . . . . . . . . . . . 87, 92, 93, 123

Shields, D.N. . . . . . . . . . . . . . . . . . . . . . . 47
Sincovec, R.F. . . . . . . . . . . . . . . . . . . . . 47
Singer, D . . . . . . . . . . . . . . . . . . . . . . . . . 12
Sinnamohideen, K. . . . . . . . . . . . . . . . . . 7
Staroswiecki, M. . . . . . . . . . . . 14, 65, 76
Stoica, P. . . . . . . . . . . . . . . . . . . . . . . . . 113
Stoustrup, J . . . . . . . . . . . . . . . . . . . . . 144
Strijbos, R.C.W. . . . . . . . . . . . . . . . . . . 29
S oderstr om, T . . . . . . . . . . . . . . . . . . . 113
Author Index 167
T
Taylor, J.H. . . . . . . . . . . . . . . . . . . . . . . . 37
Teneketzis, D.C. . . . . . . . . . . . . . . . . . . . 7
Tismenetsky, M . . . . . . . . . . . . . . . 57, 59
V
Varga, A. . . . . . . . . . . . . . . . . . . . . . . . . . 33
Verghese, G.C. . . . . . . . . . . 21, 129, 142
Viswanadham, N. . . . . . . . . . . . . . . . . . 37
W
Willsky, A.S. . . . . . . 16, 21, 33, 36, 142
Wittenmark, B. . . . . . . . . . . . . . 131, 143
Wonham, W.M. . . . . . . . . . . . . . . . . . . . 23
W unnenberg, J. . . . . . . . . . . . . . . 21, 142
Y
Yip, E.L. . . . . . . . . . . . . . . . . . . . . . . . . . 47
Z
Zhang, Q. . . . . . . . . . . . . . . . . . . . . . 65, 71
Zhirabok, A.N. . . . . . . . . . . . . . . . . . . . . 65
Zhou, K . . . . . . . . . . . . . . . . . . . . . . . . . 129

Astrom, K . . . . . . . . . . . . . . . . . . 131, 143

168 Author Index
Index
Symbols
H

ltering . . . . . . . . . . . . . . . . . . . . . 135
synthesis. . . . . . . . . . . . . . . . . . . . . . . 135
A
AI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
approximate decoupling . . . . . . . . . . . 22
attainable reference model . . . . . . . 140
automata . . . . . . . . . . . . . . . . . . . . . . . . . . 7
B
Bezout identity . . . . . . . . . . . . . . . . . . . 62
C
C-controllable. . . . . . . . . . . . . . . . . 50, 58
characteristic sets . . . . . . . . . . . . . 70, 74
Chow-Willsky scheme . . . . . 16, 36, 46
coding set. . . . . . . . . . . . . . . . . . . . . . . . . 12
complexity management. . . . . . . . . . . 76
computational form. . . . . . . . . . . . . . . 71
consistency relations . . . . . . . . . . . . . . 14
controllable . . . . . . . . . . . . . . . . . . . . . . . 54
coprime matrices . . . . . . . . . . . . . . . . . . 58
D
decoupling polynomial . . . . . . . . . . . . 23
descriptor systems . . . . . . . . . . . . . . . . 47
stochastic . . . . . . . . . . . . . . . . . . . 124
design example
aircraft dynamics . . . . . . . . . . . . . 38
stochastic . . . . . . . . . . . . . . . . . 106
coupled water tanks . . . . . . . . . . 77
planar manipulator . . . . . . . . . . . 50
stochastic . . . . . . . . . . . . . . . . . 124
turbo-jet engine . . . . . . . . . . . . . . 44
uncertain system . . . . . . . . . . . . 145
dierential-algebraic
linear . . . . . see descriptor systems
polynomial . . . . . . . . . . . . . . . . . . . 66
dierentially algebraic. . . . . . . . . . . . . 66
Discrete Event Dynamic Systems . . . 7
discrete time . . . . . . . . . . . . . . . . . . . . . . 22
division theorem . . . . . . . . . . . . . . . . . . 62
DK iterations . . . . . . . . . . . . . . . . . . . . 135
E
elementary functions . . . . . . . . . . . . . . 66
elimination theorem. . . . . . . . . . . . . . . 68
F
failure. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
fault . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
169
170 Index
fault detectability . . . . . . . . . . . . . . . . . 50
fault isolation . . . . . . . . . . . . . . . . . . . . . 11
fault modeling . . . . . . . . . . . . . . . . . . . . . 8
FDI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
xed direction residuals . . . . . . . . . . . 11
G
Gr obner basis . . . . . . . . . . . . . . . . . . . . . 68
I
ideal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
identiability . . . . . . . . . . . . . . . . . . . . . 74
incidence matrix . . . . . . . . . . . . . . . . . . 12
inequality constraints . . . . . . . . . . . . . 74
inuence structure . . . . . . . . . . . . . . . . 12
innovation lter . . . . . . . . . . . . . . . . . . . 88
innovation process . . . . . . . . . . . . . . . . 88
input-output forms. . . . . . . . . . . . . . . . 16
internal form. . . . . . . . . . . . . . . . . . . . . . 71
irreducible polynomial matrix . . . . . 58
isolability analysis . . . . . . . . . . . . . . . . 74
K
Kalman lter . . . . . . . . . . . . . . . . . . . . . 88
KCF. . see Kronecker Canonical Form
Kronecker Canonical Form. . . . . . . . 61
Kronecker indices . . . . . . . . . . . . . . . . . 33
L
leading row coecient-matrix . . . . . 58
lex order . . . . . . . . . . . . . . . . . . . . . . . . . . 68
LFT . . . . . . . . . . . see Linear Fractional
Transformation
likelihood-function . . . . . . . . . . . . . . . . 10
Linear Fractional Transformation 133
local relations . . . . . . . . . . . . . . . . . . . . . 18
M
matrix pencil . . . . . . . . . . . . . . . . . 57, 61
regular . . . . . . . . . . . . . . . . . . . 47, 57
minimal polynomial basis . . . . . 23, 59
frequency domain . . . . . . . . . . . . 25
no disturbances. . . . . . . . . . . . . . . 28
state-space . . . . . . . . . . . . . . . . . . . 25
minimal polynomial basis approach22
minimum-phase . . . . . . . . . . . . . . . . . . . 40
model uncertainty. . . . . . . . . . . . . . . . . 18
monitored faults . . . . . . . . . . . . . . . . . . 12
Moore-Penrose inverse . . . . . . . . . . . 123
N
non-monitored faults . . . . . . . . . . 13, 22
non-zero initial states . . . . . . . . . . . . . 53
normal rank . . . . . . . . . . . . . . . . . . . . . . 57
numerical performance . . . . . . . . . . . . 45
O
observations. . . . . . . . . . . . . . . . . . . . . . . . 1
P
para-hermitian. . . . . . . . . . . . . . . . . . . . 92
parity-space. . . . . . . . . . . . . . . . . . . . . . . 36
PBH test . . . . . . . . . . . . . . . . . . . . . . . . . 60
perfect decoupling . . . . . . . . . . . . . . . . 22
polynomial basis . . . . . . . . . . . . . . . . . . 58
polynomial echelon form . . . . . . . . . . 29
Q
quantier elimination . . . . . . . . . . . . . 75
R
R-controllable . . . . . . . . . . . . . 50, 54, 58
rational vector space . . . . . . . . . . . . . . 59
real algebra . . . . . . . . . . . . . . . . . . . . . . . 74
realization theory . . . . . . . . . . . . . . . . . 72
redundancy . . . . . . . . . . . . . . . . . . . . . . . . 2
analytical . . . . . . . . . . . . . . . . . . . 2, 8
hardware . . . . . . . . . . . . . . . . . . . . . . 2
reference model . . . . . . . . . . . . . . . . . . 131
relatively prime. see coprime matrices
residual . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
residual generator . . . . . . . . . . . . . . . . . 10
deterministic systems . . . . . . . . . 22
stochastic systems . . . . . . . . . . . . 88
uncertain system . . . . . . . . . . . . 130
residual structure . . . . . . . . . . . . . . . . . 12
robust fault detection . . . . . . . . . . . . 148
robust performance . . . . . . . . . . . . . . 147
robust residual generation. . . . . . . . 130
robustness . . . . . . . . . . . . . . . . . . . . . . . . 18
robustness criterion . . . . . . . . . . . . . . 132
row-degree . . . . . . . . . . . . . . . . . . . . . . . . 58
Index 171
row-reduced polynomial matrix . . . 58
S
signature matrix . . . . . . . . . . . . . . . . . . 92
solvable descriptor system. . . . . . . . . 47
spectral co-factorization . see spectral
factorization
spectral factorization. . . . . . . . . . . . . . 92
canonical . . . . . . . . . . . . . . . . . . . . . 93
discrete time . . . . . . . . . . . . . . . . 120
singular . . . . . . . . . . . . . . . . . . . . . . 93
sti models . . . . . . . . . . . . . . . . . . . . . . . 45
stochastic linear systems . . . . . . . . . . 18
strongly detectable . . . . . . . . . . . . . . . . 95
structural analysis . . . . . . . . . . . . . . . . 76
structurally singular sets . . . . . . . . . . 77
minimal . . . . . . . . . . . . . . . . . . . . . . 77
structured hypothesis tests . . . . . . . . 11
structured residuals . . . . . . . . . . . . . . . 11
Sylvester matrix . . . . . . . . . . . . . . . . . . 59
system matrix . . . . . . . . . . . . . . . . . . . . 26
descriptor case . . . . . . . . . . . . . . . 48
U
unimodular polynomial matrix . . . . 58
V
variable ordering . . . . . . . . . . . . . . . . . . 80
optimal . . . . . . . . . . . . . . . . . . . . . . . 71
W
weakly detectable . . . . . . . . . . . . . . . . . 40
whiteness test. . . . . . . . . . . . . . . . . . . . 113
172 Notes
Notes 173
174 Notes