Ault Detection and Isolation For Linear Dynamic Systems: E Hang
Ault Detection and Isolation For Linear Dynamic Systems: E Hang
Ault Detection and Isolation For Linear Dynamic Systems: E Hang
DYNAMIC SYSTEMS
by
Z E Z HANG
A report submitted in fulfilment of requirements for the Doctor of Philosophy at the Department
of Electrical and Electronic Engineering, Imperial College London.
OCTOBER 2011
Abstract
As modern control systems and engineering processes become increasingly more complex and
integrated, the consequences of system failures and faults can be disastrous environmentally and
economically. This thesis considers the fault detection and isolation (FDI) problem for linear
time-invariant (LTI) systems subject to faults, disturbances and model uncertainties.
Firstly, a novel on-line approach to the robust FDI problem for linear discrete-time systems is
proposed by using input/output measurement analysis over a finite estimation horizon. Upper and
lower bounds on the fault signal are computed at each sampling instant so that a fault is detected
and isolated when its upper bound is smaller than zero or its lower bound is larger than zero.
Moreover, a subsequent-state-estimation technique, together with an estimation horizon update
procedure are given to allow the on-line FDI process to be repeated in a moving horizon scheme.
Secondly, an optimal solution to the H− /H∞ fault detection (FD) problem is given for linear time-
invariant systems subject to faults, disturbances and model uncertainties by using an observer-
based approach. A new performance index is developed to capture both fault detection and dis-
turbance rejection requirements which is particularly suitable for handling model uncertainties. A
class of optimal solutions to the problem is then given in the form of simple linear matrix inequal-
ities (LMI) with two degrees of freedom. By appropriately choosing these degrees of freedom,
fault isolation can also be achieved.
Thirdly, in order to improve the FD performance and remove restrictive rank assumptions, rou-
tinely made in the literature, observer-based FD problems are investigated at a single frequency
and over a finite frequency range, respectively. An optimal solution is derived such that, at a given
frequency, the static observer generates a residual signal which minimizes the sensitivity of the
residual to disturbances while maintaining a minimum level of sensitivity to faults. Then, an ini-
tial investigation is carried out for the FD problem over a finite frequency range. A solution is
derived in the form of an LMI optimization by using the generalized KYP lemma followed by a
linearization procedure. Conditions under which this solution is optimal are also derived.
Fully worked out numerical examples, mostly from the literature, are given to illustrate the effec-
tiveness of all the proposed schemes.
1
Contents
Acknowledgements 3
List of Publications 5
1 Introduction 7
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.1.1 Importance of fault detection . . . . . . . . . . . . . . . . . . . . . . . . 7
1.1.2 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2 A brief history of fault detection and diagnosis . . . . . . . . . . . . . . . . . . . 10
1.3 Outline of the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.4 Notation and preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3 Online fault detection and isolation for linear discrete-time uncertain systems 32
3.1 FDI problem setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2 Fault signal bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.3 Fault signal energy bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.4 Subsequent state estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.5 Numerical examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.5.1 Example 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.5.2 Example 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.6 Concluding remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
6 Conclusion 99
6.1 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
6.2 Future Research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
6.2.1 Flexible horizon length . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
6.2.2 Comparison between static and dynamic observers . . . . . . . . . . . . 101
6.2.3 Comparison between the approach in Chapter 3 and other FD approaches 102
6.2.4 FD over a finite frequency range . . . . . . . . . . . . . . . . . . . . . . 102
6.2.5 Uncertainty estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
Bibliography 114
3
Acknowledgements
A thousand words of my English can only convey a grey and weak gratitude to my PhD supervisor
Dr. Imad Jaimoukha. This fact makes me recall one story.
Well, some one thousand and three hundred years ago, a speechless Chinese poet’s deep appreci-
ation of his friend inspired him to write a simple but memorable line. At this moment, the very
last stage of my PhD, I, with the same speechless feeling, would like to borrow that poem line
and the poet’s one thousand years sincere appreciation and send them to my supervisor Dr. Imad
Jaimoukha.
(Translation: Peach Blossom Pool is a thousand feet deep, yet not so deep, as my thank for you
and your favour for me.)
I would like to express my genuine gratitude to Prof. Richard Vinter for his priceless help and
suggestions during my PhD study and job searching.
I am sincerely indebted to all my families and friends who share my memorable joy and pain for
the last four years. Thank you all.
4
List of Publications
The following publications have been written during the course of this work:
• Z. Li, E. Mazars, Z. Zhang and I. Jaimoukha, “State-space solution to the H− /H∞ fault-
detection problem,” International Journal of Robust and Nonlinear Control, Jan. 2011.
• Z. Zhang and I.M. Jaimoukha, “An optimal solution to an H− /H∞ fault detection problem,”
In Proceedings of the 50th IEEE Conference on Decision and Control, Orlando, America,
Dec. 2011.
• Z. Zhang and I.M. Jaimoukha, “Optimal state space solution to the fault detection prob-
lem at single frequency,” In Proceedings of the 18th World Congress of the International
Federation of Automatic Control, Milan, Italy, Aug. 2011.
• Z. Zhang and I.M. Jaimoukha, “A measurement analysis approach to online fault detection
and isolation for linear discrete-time uncertain systems,” In Proceedings of the 49th IEEE
Conference on Decision and Control, Atlanta, America, Dec. 2010.
• Z. Zhang and I.M. Jaimoukha, “Optimal state space solution to the fault detection problem
at DC,” In Proceedings of Conference on Control and Fault Tolerant Systems, Nice, France,
Oct. 2011.
• Z. Zhang and I.M. Jaimoukha, “Fault detection and isolation for linear discrete-time systems
using input/output measurement analysis,” In Proceedings of the 48th IEEE Conference on
Decision and Control, Shanghai, China, Dec. 2009.
CONTENTS 6
Submitted:
• Z. Zhang and I.M. Jaimoukha, “An approach to on-line robust fault detection and isolation
for linear discrete-time uncertain systems,” Automatica, 2011, under review.
• Z. Zhang and I.M. Jaimoukha, “An optimal solution to an H− /H∞ fault detection problem,”
System and Control Letters, 2010, under review.
• Z. Zhang and I. M. Jaimoukha, “A finite frequency domain approach to the fault detection
problem for linear time-invariant systems,” IEEE Transactions on Automatic Control, 2011,
under review.
• Z. Zhang, Y. Yuan, “Decentralized consensus value computation over network with uncer-
tainties,” IEEE Transactions on Automatic Control, 2011, under review.
7
Chapter 1
Introduction
Control systems, which are nowadays exploited to improve the overall performance of industrial
processes, involve both sophisticated digital system design techniques and complex hardware
(sensors, actuators, components and processing units). Such complexity results in an increased
probability of failure, consequently, control systems require automatic supervision of the closed-
loop operation for detecting and isolating malfunctions as early as possible. Hence, the problems
of FDI and diagnosis in dynamic processes have attracted great attention in both theoretical
research and industrial applications.
1.1 Background
1.1.1 Importance of fault detection
As many control systems and engineering processes become more and more complex, widespread
and integrated, the effects of system failures can be simply devastating to the infrastructure of
modern society for extensive commercial, industrial and safety reasons. Therefore, fault detec-
tion and diagnosis is necessary in many control engineering applications such as: safety-critical
systems (nuclear reactors, aircraft, missile guidance systems), cost-critical systems (large space
structures, space vehicles, autonomous underwater vehicles).
As an example of safety-critical systems, in nuclear power plants, the occurrence of a fault may
lead to extreme human and environmental problems. A proper control action could be a safe shut-
down for the plant in the presence of emergency. In cost-critical systems, the occurrence of a fault
may cause a tremendous economical loss. For example, autonomous underwater and unmanned
space vehicles must be able to cope with unexpected events or malfunctions in a predictable man-
1.1 Background 8
ner. In the worst case, the vehicle should be able to return a safe condition without necessarily
completing the mission. In industrial applications or manufacture processes, minor faults could
cause a shutdown of production systems or even a damage to equipment which could result in a
large loss of profit.
Among many examples of catastrophic faulty situations, two cases are cited here:
1. On 26th of April 1986, there occurred near the Ukrainian town of Chornobyl a tremendous
explosion at a huge nuclear power plant, followed by a gradual meltdown of the reactor No.
4. The main cause for this tragedy was the faulty outdated technology and the lack of a fault
handling mechanism [96].
2. The Ariane 5 rocket exploded on June 4th, 1996 thirty seven seconds after lift-off. The
reason was a software exception in the Inertial Reference Unit (IRU) that provides attitude
and trajectory information for the control system. The exception caused the normal attitude
information to be replaced by some diagnostic information that the control system was not
designed to understand [97].
The catastrophes of these faulty situations could have been avoided or at least alleviated if the con-
trol system is designed with certain abilities of fault detection and diagnosis. Hence, since the early
1970s, the areas of fault detection, diagnosis and tolerance has become increasingly important and
been actively studied due to reliability, security and automatic self-supervision requirements.
1.1.2 Definitions
It should be remarked that there is no consistent terminology in the field of FDI, which causes
difficulties for engineers and researchers to understand the contributions and to compare different
approaches. To avoid these misunderstandings, the IFAC Technical Committee SAFEPROCESS
(Fault Detection, Supervision and Safety for Technical Processes) has initiated a common
terminology related to the field of fault diagnosis. Further discussions regarding terminology can
be found in [78]. The terminology used in this thesis is listed below and is consistent with that
defined by the committee SAFEPROCESS.
1.1 Background 9
Note that the difference between fault and failure is that the term fault is to represent a malfunction
in the process of system operation while failure denotes the situation that system functions suffer
a complete breakdown. A fault may be tolerable via a proper system design. However, it must be
detected and diagnosed as soon as possible even though it has been tolerated at its early stage in
order to prevent any serious or dangerous consequences in the future.
The terminology of tasks of a monitoring system for the purpose of fault detection and diagnosis
is defined as follows:
• Fault detection: to determine whether there is an occurrence of a fault and the time of its
occurrence.
• Fault isolation: to determine the kind, location (which process, sensor or actuator) and time
of detection of a fault, following fault detection.
• Fault identification: to determine the size and time-variant behaviour of a fault. Follows
fault detection and isolation.
• Supervision: monitoring a physical system and taking appropriate actions to maintain the
operation in the case of fault.
These tasks can be chosen to satisfy the special safety and reliability requirements of modern
control systems, however fault detection is absolutely necessary for any practical system.
1.2 A brief history of fault detection and diagnosis 10
Research on fault diagnosis [11] has been carried out both in theory and application since the
early 1970s. The well-known failure detection filter approach for linear systems was proposed in
[3] and the application of identification methods to the fault detection of jet engines was considered
in [68]. Correlation methods were applied to a practical problem of leak detection in [82].
Since the 1980s, model-based methods for fault detection and diagnosis have attracted extensive
attention due to the increasing requirements of reliability and security. Model-based fault detection
and diagnosis with specific application to chemical processes was proposed in [43] and sensor
failure detection based on the inherent analytical redundancy of multiple observers was given in
[11].
The use of parameter estimation techniques for fault detection of technical systems was given in
[28] and the development of process fault detection methods based on modelling and parameter es-
timation was summarized in [45]. Parity equation-based methods were proposed by [10] and then
further developed in [63]. Some contributions solve FD problem by using probabilistic approaches
[17, 79]. It is noticed that one important approach, namely, the observer–based state estimation
approach, has been actively researched over the last two decades and many contributions have
been made to research the design of fault detection observers [19, 32, 35, 66, 65, 64, 77]. Also,
some important work proposed different approaches to achieve FD for nonlinear systems. See
[15, 42, 104] for details.
Recently, FD in a finite frequency setting has gained increasing attention from the theoretical re-
search community. The sensitivity optimization problem over a finite frequency range for fault
detection is studied to further improve the performance of fault detection or relax certain assump-
tions in the problem formulation [90, 92, 58, 30].
The development of H∞ filtering techniques provides potential solutions to FD problems [86, 99].
Recently developed linear matrix inequality (LMI) [6, 26] approaches offer promising and numer-
ically attractive techniques for formulating robust FDI decoupling problems. Hou and Patton gave
a basic realization of fault detection observer design based on the bounded real lemma [44]. Zhong
1.3 Outline of the thesis 11
et al. proposed a new performance index by introducing a reference residual model, which can be
formulated using LMI techniques [111].
The focus of this thesis is on the FDI problem for linear dynamic systems subject to faults, dis-
turbances and model uncertainties. Three approaches are proposed to solve the problem. The first
uses measurement analysis to detect and isolate a fault by evaluating its upper and lower bounds
given all available information, while the other two achieve FD by designing static observers to
generate fault indicators, namely, the residual signals; the first over the whole frequency range
while the second over a finite range. The work in this thesis is only one step towards practiced FD
systems. It is also needed to consider thresholds for residual signals, etc.
The thesis is organized as follows. The rest of this chapter will define the notation and introduce
related background material on matrix calculations.
Chapter 2 introduces the concept and techniques of analytical redundancy and also reviews the
main residual generation techniques in general model-based FD schemes such as parity-space and
observer-based approaches.
Chapter 3 proposes a new on-line approach to solve the robust FDI problem for linear discrete-
time systems subject to faults, disturbances and norm-bounded uncertainties by using input/output
measurement analysis. FDI is achieved by computing upper and lower bounds on the fault signal
such that a fault is regarded as having occurred when its upper bound is smaller than zero or
lower bound is larger than zero. LMI optimization techniques are employed to obtain the bounds.
Furthermore, a subsequent-state-estimation technique, together with an estimation horizon update
step are proposed to allow the on-line FDI process to be repeated in a moving horizon procedure.
Also, it is shown that the proposed approach reduces to the solution of linear programs when there
are no uncertainties in the model description.
In Chapter 4, an optimal state-space solution is given to the H− /H∞ FD problem for linear time-
invariant systems subject to faults and disturbances by designing a static observer. A novel per-
formance index is developed such that both the requirements of fault detection and disturbance
1.4 Notation and preliminaries 12
rejection are captured, and this index, unlike corresponding ones in the literature, is particularly
suitable for handling model uncertainties. Then, a class of solutions to the FD problem is given in
the form of simple LMIs with two degrees of freedom. The optimal solution and further specifica-
tions such as fault isolation can be achieved by appropriately choosing these degrees of freedom.
Furthermore, it is shown that the FD performance may be improved and certain assumptions can
be relaxed by applying a weighting filter.
In order to remove a widely made restrictive rank assumption on fault dynamics and to further
improve the FD performance, Chapter 5 first considers the FD problem at a given frequency. A
simple parameterization of all optimal solutions is derived with two degrees of freedom which
can be used to achieve further specifications. In particular, at DC (when ω = 0), the resulting
degrees of freedom allow the FD observer to be used simultaneously for control system design.
Moreover, an initial research is carried out for the FD problem over a finite frequency range. A
solution is derived in the form of a BMI optimization using the generalized KYP lemma followed
by a linearization procedure. Also, conditions under which this solution is optimal are derived.
Chapter 6 summarizes the results in this thesis, highlights the contribution and considers possible
improvements and extensions to the proposed approaches.
This section gives the notation as well as some preliminary results that will be used later in the
thesis. The notation used is fairly standard and is summarized as follows.
The set of real (complex) n × m matrices is denoted by Rn×m (C n×m ). For A ∈ C n×m we use the
notation AT and A∗ to denote the transpose and complex conjugate transpose, respectively. For
A ∈ C n×n , λi (A) denotes the i-th eigenvalue of A. For A = A∗ ∈ C n×n , A 0 (A ≻ 0) denotes
that A is positive semidefinite (positive definite), that is, λi (A) ≥ 0 (λi (A) > 0) for all i. Similar
definitions apply to A 0 and A ≺ 0. For A = A∗ ∈ C n×n , λ̄(A) denotes the largest and λ(A) the
p
smallest eigenvalue of A, respectively. For A ∈ C n×m , σ̄(A) = λ̄ (AA∗ ) denotes the largest, and
p
σ(A) = λ(AA∗ ) the smallest singular value of A, respectively.
The n × n identity matrix is denoted as In and the n × m null matrix is denoted as 0n,m with the
1.4 Notation and preliminaries 13
The open left half of the complex plane is denoted by C− and the closed right half of the complex
plane is denoted by C¯+ . The notation A = diag (a1 , . . . , an ) denotes that A is a diagonal matrix
with diagonal entries a1 , . . . , an .
Continuous and discrete time state–space descriptions of a linear time–invariant dynamic system
are given as follows, respectively.
where x(t), x(k) ∈ Rn , u(t), u(k) ∈ Rnu and y(t), y(k) ∈ Rny are the state, input and output
vectors, respectively, and where the real matrices A, B, C and D have appropriate dimensions.
Taking Laplace transform, assuming zero initial conditions, the input/output description is given
by y(s) = G(s)u(s) where G(s) = D + C(sI − A)−1 B is the system transfer matrix function.
s
Also, use the notation that G(s) = (A, B, C, D) and
s A B
G(s) = .
C D
s s
If G1 (s) = (A, B1 , C, D1 ) and G2 (s) = (A, B2 , C, D2 ), it is written as
s A B1 B2
G1 (s) G2 (s) = ·
C D1 D2
The matrix A ∈ Rn×n is called stable if λi (A) ∈ C− , ∀ i. The pair (A, C) is called detectable if
there exists a real matrix L such that (A + LC) is stable. The pair (A, C) is detectable if and only
if the matrix pencil
A − µI
C
1.4 Notation and preliminaries 14
m×p
The set R(s)m×p denotes the set of all m × p proper, real–rational matrix functions of s. L∞
denotes the space of m × p matrix functions with entries bounded on the extended imaginary
m×p m×p
axis. The subspace H∞ ⊂ L∞ denotes matrix functions analytic in the closed right–half
m×p
of the complex plane. A prefix R denotes a real–rational function, so that RL∞ denotes the
space of m × p real rational matrix functions with entries bounded on the extended imaginary
m×p m×p
axis jRe . The subspace RH∞ ⊂ RL∞ denotes real rational matrix functions analytic in the
closed right–half of the complex plane, i.e., all m × p stable real–rational matrix functions of s.
For G(s) ∈ R(s)m×p dependence on s is occasionally suppressed when it can be inferred by the
m×p m×p
context. For the spaces R(s)m×p , H∞ , L∞ , etc., the dimensions are occasionally suppressed
if they can be inferred by the context.
m×p
For G(s) ∈ RH∞ , define
kGk∞ = sup σ̄ (G(jω))
ω∈R
and
kGk− = inf σ (G(jω)) .
ω∈R
and
kGk− ≥ γ ⇐⇒ G∗ (jω)G(jω) − γ 2 I 0, ∀ ω ∈ R. (1.2)
Using these definitions and the singular value inequality σ̄(AB) ≥ σ(A)σ̄(B) [16], it is straight-
forward to show that
kG1 G2 k∞ ≥ kG1 k− kG2 k∞ (1.3)
Remark 1.4.1. The H− -norm k•k− for FDI problems was initially introduced by Hou and Patton
[44]. Note that kGk− is defined to characterise (1.2) and is not a real norm. However, the author
follows convention in the literature and calls it a norm. 2
s m×p
For G(s) = (A, B, C, D) ∈ RL∞ , define
s
G∗ (s) = GT (−s) = (−AT , −C T , B T , DT )
1.4 Notation and preliminaries 15
The following result gives an LMI formulation for H∞ -norm based synthesis and design. It is a
statement of the Bounded Real Lemma as derived in Theorem 2.41 in [75] for the strict inequality
case.
Lemma 1.4.2. [75] Let G(s) = D + C(sI − A)−1 B be real–rational and let γ > 0. Then A is
stable and G∼ (s)G(s) − γ 2 I < 0 on the jω–axis if and only if there exists P = P T such that
P ≻0
AT P + P A + C T C P B + C T D (1.4)
≺ 0.
(P B + C T D)T DT D − γ 2 I
s
Lemma 1.4.3. [75, 76]. Suppose that G(s) = (A, B, C, D) where A ∈ Rn×n and let γ > 0 be
given . Then A is stable and kGk∞ < γ if and only if there exists P = P T ∈ Rn×n such that
P ≻0
AT P + P A P B C T
(1.5)
BT P −γI DT ≺ 0.
C D −γI
2
1.4 Notation and preliminaries 16
The following lemma can be found in [86]. It is widely used in control theory and in connection
with trust region methods in optimization [84].
If ∆ is unstructured (that is, if ∆ = {∆ ∈ Rn∆ ×n∆ : k∆k ≤ 1}), then (1.6) becomes
R + λE T E F + λE T H
≺0 (1.7)
F T + λH T E λ(H T H − I)
for some scalar λ > 0. In this case, condition (1.7) is both necessary and sufficient.
Lemma 1.4.4. Given matrices A, B and a Hermitian matrix Θ, the following finite frequency
domain inequality
∗
(jωI − A)−1 B (jωI − A)−1 B
Θ ≺0 (1.8)
I I
holds if and only if there exist hermitian matrix Π such that the following LMI
∗
A B A B
Π +Θ≺0 (1.9)
I 0 I 0
is satisfied, where Π as shown below are specified for the cases of low, middle and high frequency
ranges.
−Q P
Low frequency range |ω| ≤ ωl : Π = ,
P ωl2
1.4 Notation and preliminaries 17
−Q P + jωc Q
Middle frequency range ω1 ≤ ω ≤ ω2 : Π = , ωc = (ω1 + ω2 )/2,
P − jωc Q −ω1 ω2 Q
Q P
High freqency range |ω| ≥ ωh : Π = ,
P −ωh2 Q
where Q ≻ 0 and P are Hermitian matrices and ωl , ω1 , ω2 and ωh are all non-negative
scalars.
18
Chapter 2
In the last two decades, the FD problem for dynamic systems has attracted great attention and
various FD approaches have been proposed and applied in practice. Approaches which are based
on mathematical models of the processes or systems under investigation are usually referred to as
Model-based FD.
Section 2.1 firstly introduces the definition and key points of analytical redundancy. Then Section
2.2 illustrates the modelling of a system with disturbances and faults which are to be detected.
Section 2.3 provides a basic introduction to the different methods of residual generation and Sec-
tion 2.4 illustrates the issues of residual evaluation. Finally, Section 2.5 considers the robustness
issue in model-based FD schemes.
In order to increase the reliability and safety and to avoid the additional cost of redundant hard-
ware, the so-called analytical or functional redundancy approach are proposed. It employs the
values of the variables from an analytical or mathematical model of the system to cross-check with
the real measured values generated from the monitored system, that is, the consistency checking
[64, 8, 33]. The resulting difference generated from the checking is called a residual signal. Ide-
ally, the residual should be zero when the system is in normal operation and should be different
from zero when a fault has occurred. This property of the residual is used to determine if there is
an occurrence of a fault [8, 62].
The major advantage of analytical redundancy schemes is that no additional hardware compo-
nents are required to implement an FD algorithm. Analytical redundancy can be implemented
via software on a process control computer. In many cases, the measurements which are used to
control the process are also sufficient for the FD algorithm so that no additional sensors have to be
installed [45, 8, 50].
Generally, FD approaches frequently used in practice monitor the level of the residual signal and
the occurrence of a fault is detected when the residual reaches a given level, that is, the threshold.
This method is easy to implement but has obvious drawbacks. The most serious one is that the
presence of disturbances, input variations and change of operating points of the monitored systems
can cause false alarms.
Analytical redundancy employs a mathematical model of the monitored system, hence it is often
referred as the model-based approach to FD.
where uR and yR are the effective inputs and outputs, that is, the post-actuator inputs and pre-
sensor outputs. Concerning the locations of different faults, the system under investigation can be
separated into the following different parts that can be affected by faults
• Actuators,
• Sensors.
Note that the control action is usually performed by a digital computer or controller while the
actuator is considered as a part or a component of the whole system, therefore, in model-based
FD schemes, faults are usually assumed to occur in actuators while the controller can be regarded
as fault-free and neglected. By assuming linearity, the system dynamics can be described by the
following time-invariant state-space form
where A, Bu , C and Du are system matrices with appropriate dimensions and x(t) ∈ Rn and
y(t) ∈ Rny are the system state vector and output vector, respectively, while u(t) is the input
2.2 Modelling of faulty systems 21
signal driven by the actuators but not directly readable. fa , fc , fu and fy denote the actuator,
component, input sensor and output sensor faults, respectively. The matrices W1 , W2 and W3 are
known as the entry matrices of faults which describe how the faults enter the system [64].
T
By taking f (t) = faT fuT fcT fyT ∈ Rnf , the state-space model of a faulty system in
(2.1) can be rewritten as
where Bf and Df are known as fault distribution matrices with appropriate dimensions. Since
practical systems are always affected by disturbances, for generality, the state-space model of a
linear time-invariant system with faults and disturbances can be described as
where Bd and Dd are also known as the disturbance distribution matrices [64].
Since the state-space description provides a general and mathematically rigorous tool for system
modelling and residual generation, it has been widely used in the design of FD schemes. Also, the
input-output model of a system is frequently applied since it directly shows the influence of faults
and disturbances on the system outputs.
where
n ×nu
Gu (s) = Du + C(sI − A)−1 Bu ∈ RL∞y ,
n ×nd
Gd (s) = Dd + C(sI − A)−1 Bd ∈ RL∞y ,
and
n×nf
Gf (s) = Df + C(sI − A)−1 Bf ∈ RL∞
2.3 Residual generation techniques 22
are the process, disturbance and fault transfer matrices, respectively [64].
As shown above, general schemes of model-based FD use the model given in (2.2) and (2.3) and
are based on the availability of the input and output signals u(t) and y(t), respectively.
The parity space approach applies to FD designs for linear discrete-time systems [101]. The
basic idea is to carry out a proper check of the parity or consistency between the input and output
measurements collected from the monitored system to obtain the redundancy. Then, a parity vector
or matrix is introduced to cancel the effects of the system state and to generate the residual signals
which are sensitive to faults and insensitive to disturbances and system errors.
In the early stage of research on parity-space-based FD, this approach was applied to static or
parallel redundancy schemes, in which input and output measurements could be directly obtained
from hardware (hardware redundancy) or analytical relations (analytical redundancy) [70, 8]. In
the case of hardware redundancy, two methods can be used to obtain the redundant signals. One
requires several sets of sensors with identical or similar functions to measure the same variable
while the other approach employs different sets of sensor to measure different variables but with
their outputs being related to each other. In the rest of this section, the attention is focused on
analytical forms of redundancy.
In analytical schemes, parity space approaches are used to implement the consistency check based
2.3 Residual generation techniques 23
xk+1 = Axk + Bd dk + Bf fk + Bu uk ,
yk = Cxk + Dd dk + Df fk + Du uk , (2.4)
where xk ∈ Rn , uk ∈ Rnu and yk ∈ Rny are the state, input and output vectors, respectively, and
dk ∈ Rnd and fk ∈ Rnf are the disturbance and fault vectors, respectively. Here, Bf ∈ Rn×nf
and Df ∈ Rny ×nf are the component and instrument fault distribution matrices, respectively,
while Bd ∈ Rn×nd and Dd ∈ Rny ×nd are the corresponding disturbance distribution matrices
[32, 35].
Let s be an integer denoting the length of a moving horizon k − s + 1, k . Input measurement
vector u, disturbance vector d and fault vector f for the estimation horizon can then be expressed as
uk−s+1 yk−s+1 dk−s+1 fk−s+1
uk−s+2 yk−s+2 dk−s+2 fk−s+2
u(k) = .. , y(k) = .. , d(k) = .. , f (k) = .. .
. . . .
uk yk dk fk
Using an iterative computation, the output measurements over the horizon from k − s + 1 to k can
be expressed as
Cy0 Dyu
z }| { z }| {
C Du 0 ··· 0
CA
CBu Du ··· 0
y(k) = .. x +
k−s+1 .. .. .. .. u
. . . . .
CAs−1 CAs−2 Bu CAs−3 Bu · · · Du
Dyd Dyf
z }| { z }| {
Dd 0 ··· 0 Df 0 ··· 0
CBd Dd ··· 0
CBf Df ··· 0
+ .. .. .. .. d+ .. .. .. .. f (2.5)
. . . . . . . .
CAs−2 Bd CAs−3 Bd · · · Dd CAs−2 Bf CAs−3 Bf ··· Df
In order to achieve FD effects, a parity (row) vector vs is introduced so that the dynamics
of the residual can be described by
where the design parameter vs is used to modulate the residual dynamics. Usually, it is required
that vs Cy0 = 0 to eliminate the influence of the initial state. In some contributions, a parity
matrix Vs is introduced and designed to achieve FD instead of using a parity vector to improve the
sensitivity of residuals to faults.
As shown above, the parity space approach transforms the FD problem to an optimization problem
such that vs is designed to increase the sensitivity of the residual to faults and the robustness to
disturbances. Also, it can be seen that if and only if the condition
rank Cy0 Dyd Dyf > rank Cy0 Dyd (2.7)
is satisfied, then a full decoupling from both the initial state and disturbances can always be
achieved by choosing a vs , which means the following condition is satisfied
vs Cy0 Dyd = 0, vs Dyf 6= 0 (2.8)
In other words, vs should be chosen to lie in the intersection between the left null space of Cy0 and
Dyd . However, a full-decoupling from the initial state and disturbances is usually not possible in
the sense of achieving optimal FD performance. Therefore, the FD problem in parity-space-based
schemes is often formulated as the following optimization problem
T vT
vs Dyf Dyf s
max g(vs ), g(vs ) = T vT
(2.9)
vs Cy0 =0 vs Dyd Dyd s
Generally, a high dimensional vs will improve the FD performance index Js (vs ) at the expense of
increasing the burden of on-line computational effort. Many contributions have been devoted to
the design of the parity vector vs in the literature. See [100, 103, 21, 110] for details.
In the last two decades, observer-based FD for complex dynamic systems exploiting analytic re-
dundancy have received increasing attention both in literature and applications [112, 49, 54, 58].
This approach involves the design of a dedicated observer to effectively cancel the nominal process
dynamics and generate residual signals to provide fault signatures.
Basically, there are two kinds of observer-based FD approaches for achieving this objective,
namely, perfect and approximate disturbance decoupling. In the former, the aim is to decouple
the residual signal from disturbances exactly, while the latter requires the transfer matrix from dis-
turbances to the residual signal to be small in either the H2 or H∞ norm sense. Both approaches
have been widely studied and implemented using several methods [9, 65, 33].
Although FD schemes using perfect disturbance decoupling can achieve reasonable FD perfor-
mance, its detection ability is often restricted and, in most cases of achieving perfect decoupling,
solvability conditions are generically difficult to be satisfied. Therefore, much research effort and
many contributions have been devoted to approximate the disturbance decoupling method [34].
In [49], a matrix factorization method is developed to obtain an optimal fault detection filter.
Sadrnia utilized a Riccati equation iteration method to construct an extended H∞ filter [72]. See
[27, 113, 54, 48] for more details.
Generally, there are two ways of constructing observers to generate residual signals, that is, static
and dynamic observers. The aim of both observers is not to estimate the state of the system, but
rather to respond promptly to the occurrence of a fault via residual signals. In this thesis, we
only use FD schemes based on static observers. In the following, the mathematical model of static
observer-based FD is presented and the dynamics of residuals are formulated in a state-space form.
The idea of static observer-based FD is to use the duplication of the system to cancel the input
system dynamics and use the degrees of freedom in the observer to optimize the dynamics of
residuals for FD specifications. For the FD dynamics described in (2.2), the static FD observer is
2.3 Residual generation techniques 26
described as:
˙
x̂(t) = Ax̂(t) + Bu u(t) − L (y(t) − C x̂(t)) ,
assuming Du = 0, where x̂(t) ∈ Rn is the observer state and r(t) ∈ Rnf is the residual signal.
Here L ∈ Rn×ny and H ∈ Rnf ×ny are the observer and residual gain matrices, respectively, and
are to be determined. The static FD observer which is employed in this thesis, is illustrated in
Figure 2.2.
d f d ? f
? ? ?
Bd Bf Dd Df
B B
B B
u B B
- Bu -BN f ẋ- x- -BN f
R
C
6
y
A
?- r-
L f
− H
6
−
- Bu ?x̂˙ -
-f
R x̂ -
C
6
A
By taking Laplace transforms to 2.11, the closed-loop form of the residual dynamics is hence
given by
r(s) = Trf (s)f (s) + Trd (s)d(s), (2.12)
2.4 Residual evaluation 27
where
s A+LC Bd +LDd Bf + LDf
Trd (s) Trf (s) = ∈ R(s)nf ×(nd +nf ) , (2.13)
HC HDd HDf
are the transfer matrices from faults and disturbances to residuals, respectively. Then a simple
calculation shows that
r(s) = F (s)[Gd (s)d(s) + Gf (s)f (s)],
where
s A + LC L
F (s) = ∈ R(s)nf ×ny , (2.14)
HC H
so that
Trd (s) Trf (s) = F (s) Gd (s) Gf (s) ·
with Gd and Gf defined in Section 2.2. Figure 2.3 shows the structure of a general static observer–
based FD residual generator in the frequency domain.
d(s)
Gd (s)
u(s) f (s)
G(s) Gf (s)
y(s)
G(s)
-
F (s)
r(s)
From the discussion above, the FD problem using a static observer is then transformed to solve a
sensitivity optimization problem, which seeks to increase the sensitivity of the residual to faults
and simultaneously to reduce the sensitivity to disturbances. In an ideal but not usual situation
2.4 Residual evaluation 28
faults can be completely decoupled from disturbances, which means that residuals are only sensi-
tive to faults and completely insensitive to disturbances.
However, the complete elimination of the effects of the disturbances on residuals may not be
possible due to the lack of degrees of freedom in the observer design. Hence, an appropriate
criterion is needed to examine the residuals in order to determine whether any faults have occurred
and where the faults are present.
Usually, the decision process for detecting a fault consists of a threshold test on a residual eval-
uation function. Define J(r) to be the residual evaluation function, ideally, J(r) should be zero
for fault free cases and non-zero for faulty cases. However, due to the presence of disturbances
and other uncertainties in the system, a perfect residual is almost impossible to obtain in prac-
tice. Therefore, for the aim of fault detection, the residual evaluation function is compared with a
threshold Jth such that
In many cases [95, 94], the threshold is obtained by recording the value of residuals in the situation
that no faults have occurred. For example, an on-line time-domain evaluation function [35] which
is formed by r(t) is given as follows
s Z
1 τ T
J(r) = r (t)r(t)dt,
τ 0
where τ is a user-defined time constant.
Many contributions [35, 59, 102] define thresholds using on-line or off-line functions which,
generally, are the energy of the unknown signals (or disturbances) since it is reasonable to
assume that bounds on the unknown inputs such as disturbances are available. The selection
of thresholds often involves a trade-off between false alarm rates and missed faults so that the
thresholding highly depends on the safety requirements on the plant and there are no general selec-
tion procedures which can be adopted universally. For example, in [93] the threshold is taken to be
s Z
1 τ
Jth = (F Gd d(t))T F Gd d(t)dt.
τ 0
If J(r) > Jth at some time interval, the occurrence of a fault is indicated.
An important requirement of residual generation is to achieve fault isolation, that is, to indicate
the position of the fault corresponding to some sensor or component [81, 80, 14]. One approach to
fault isolation [32, 98] is the so called dedicated observer scheme, which uses a bank of residual
signals where each of the residuals is designed to be sensitive to a specific fault while insensitive
to the other potential faults. The decision logic of residual evaluation is
In this thesis, time-domain based residual evaluation and thresholding for either fault detection or
isolation is employed in the examples given in subsequent chapters.
Although model-based approaches to FD have been accepted as an effective technique for de-
tecting and even isolating faults, a critical problem of reliability due to unavoidable modeling
2.5 Robustness in model-based FD 30
In practice, model uncertainties can be introduced by process noises, parameter variations and non-
linear dynamics. On the other hand, the mathematical models used in model-based FD schemes
never perfectly describe the supervised system. The mismatch between the actual system and its
mathematical model can also be regarded as model uncertainties. These uncertainties can cause
serious technical difficulties and lead to false alarms or missed faults which severely deteriorate
the performance and reliability of the FD designs.
The rest of this section describes the mathematical representation of model uncertainties which
are usually considered in the literature. Basically, there are two kinds of well-defined model
uncertainties, that is, norm-bounded and polytopic uncertainties. The former is firstly illustrated.
Let
A B Bd Bf
M=
C D Dd Df
denote the set of system matrices of an uncertain system. In the case of norm-bounded uncertain-
ties, the Linear Fractional Transformation (LFT) representation of uncertainty [115] is described
as follows. Here,
Mo
z
}| {
Ao B o Bdo Bfo FA
M∈{ o o o o + ∆H EA EB Ed Ef : ∆ ∈ ∆} =: M∆,
C D Dd Df FC
and where FA , FC , EA , EB , Ed , Ef and H are known and constant matrices with appropriate
dimensions and where it is assumed that ∆ is well-posed, that is, det(I − H∆) 6= 0 for all
∆ ∈ ∆. The LFT representation of norm-bounded uncertainties has great generality and is widely
used in control theory [115]. The robustness issue of FD involved in this thesis mainly considers
norm-bounded uncertainties.
32
Chapter 3
In this chapter, a new on-line approach is proposed to solve robust FDI problems for linear discrete-
time systems subject to faults, bounded additive disturbances and norm-bounded uncertainties
[105, 106]. This scheme uses a dynamic system model as well as input/output measurements over
a finite estimation horizon to compute upper and lower bounds on the fault signals at each sampling
instant. Then a fault is regarded as having ocurred when its upper bound is smaller than zero or
lower bound is larger than zero. LMI optimization techniques are employed to obtain the bounds.
Furthermore, a subsequent-state-estimation technique together with an estimation horizon update
procedure is proposed, which allows the on-line FDI process to be repeated in a moving horizon
procedure.
One feature of the proposed approach is that the upper and lower bounds on the initial state, if
available, can be used and the initial state is not necessarily decoupled. This is in contrast to the
parity space approach by which the effect of the initial state has to be removed. Furthermore, this
approach to the robust FDI problem is achieved by directly computing upper and lower bounds on
the fault signals rather than designing a residual generator as in observer-based and parity-space-
based schemes.
Moreover, it is shown that the bounds on the faults can be obtained by solving optimization prob-
lems subject to the constraints of the bounds on the initial state, disturbances and norm-bounded
uncertainties by using the system model and input/output measurements over a finite estimation
horizon. This approach can therefore be regarded as a dual to model predictive control (MPC)
3.1 FDI problem setting 33
schemes which use the system model to solve an on-line optimization problem to determine an
optimal control strategy subject to hard constraints at each sampling instant, usually over a finite
horizon [69, 60]. Also, the proposed approach reduces to the solution of a set of linear programs
when there are no uncertainties in the model description.
This chapter is organized as follows. After defining the notation, Section 3.1 defines the problem
settings. Then Section 3.2 gives the problem formulation and derives upper and lower bounds on
the fault signals in the form of solutions to LMI problems. In Section 3.4, a subsequent-state-
estimation as well as an estimation horizon update procedures are described, which complete the
description of the FDI algorithm. Examples are given to demonstrate the effectiveness of the
proposed scheme in Section 3.5. Finally, Section 3.6 summarizes this work.
Consider a linear time-invariant (LTI) discrete-time system subject to disturbances, process, actu-
ator and sensor faults and uncertainties of the form
xk
xk+1 A + ∆A Bd + ∆Bd Bf + ∆Bf Bu + ∆Bu dk
= (3.1)
yk Cy +∆Cy Dyd +∆Dyd Dyf +∆Dyf Dyu +∆Dyu fk
uk
It is assumed the norm-bounded uncertainty matrices in (3.1) are of form given by [61]:
∆A ∆Bd ∆Bf ∆Bu Bp
= ∆ Cq Dqd Dqf Dqu ,
∆Cy ∆Dyd ∆Dyf ∆Dyu Dyp
where Bp , Dyp , Cq , Dqd , Dqf and Dqu are known and ∆ is an n∆ × n∆ unknown matrix with
∆ ∈ ∆ where
Pl+r
Note that ∆ has l repeated diagonal blocks and r full block diagonal blocks so that i=1 n∆i =
n∆ .
Remark 3.1.1. Most contributions to the problem of robust FDI consider the uncertainty matrix
without a specific structure, that it, ∆ ∈ Rn∆ ×n∆ [111, 92, 37, 73]. In this work, the structure of
the uncertainty matrix has a more general form with both diagonal and full blocks, which is more
suitable and flexible for dealing with specific forms of system uncertainties in practice.
By introducing the new variables pk and qk , the state space equations (3.1) can be rewritten as:
x k
xk+1 A Bd Bf Bu Bp dk
yk = Cy Dyd Dyf Dyu Dyp fk ,
qk Cq Dqd Dqf Dqu 0n∆ uk
pk
pk = ∆qk . (3.3)
Assume that upper and lower bounds x̄0 , x0 and d¯k , dk on the initial state and disturbances,
respectively, are available and have of the form
x0 ≤ x0 ≤ x̄0 , dk ≤ dk ≤ d¯k , ∀k ∈ N .
Also assume that measurements of the input and output signals uk and yk are available for all
k ∈ N . For the estimation horizon, let
x0 d0 d¯0 ξ0
x1 d d¯1 ξ1
1
x = . , d = . , d¯ = . ∈ RN ξ
, ξ = .. (3.4)
.. .. .. .
xN dN −1 d¯N −1 ξN −1
where N∆ = N × n∆ . Using an iterative computation, the system algebraic formulation over the
estimation horizon is described as
x0
x A Bd Bf Bu Bp d
y = Cy Dyd Dyf Dyu Dyp f
. (3.6)
q Cq Dqd Dqf Dqu Dqp u
p
3.1 FDI problem setting 35
where
I Cy 0 0 ··· 0
A
Cy A
Bζ 0 ··· 0
A=
.. .. .. .. .. ..
. , Cy0 = , Bζ =
. . . . .
N −1 N −2
A Cy AN−2 A Bζ AN −3 Bζ ··· 0
A N Cy AN−1 AN −1 Bζ AN −2 Bζ ··· Bζ
Cq 0 0 ··· 0 0
Cq A
C q Bζ
0 ··· 0 0
Cq0 = .
.. .
.. .
.. .. .. .. ,
, Dqζ = . . .
Cq AN−2 Cq AN−3 Bζ Cq AN−4 Bζ · · · 0 0
Cq AN−1 Cq AN−2 Bζ Cq AN−3 Bu · · · C q Bζ 0
Dyζ 0 ··· 0 0
C y Bζ Dyζ ··· 0 0
Dyζ = .
.. .
.. . .. .
.. .. ,
.
N−3 N−4
Cy A Bζ Cy A Bζ · · · Dyζ 0
Cy AN−2 Bζ Cy AN−3 Bζ · · · Cy Bζ Dyζ
Applying some matrix manipulations to (3.5) and (3.6), the output over the estimation horizon is
given as
where w ≤ w ≤ w̄ and
∆Dyw ∆Dyf ∆Dyu = Dyp ∆a (I − Dqp ∆a )−1 Dqw Dqf Dqu , ∆a ∈ ∆a . (3.9)
For notational convenience, assume that the uncertainty data are re-sequenced so that the uncer-
tainty set has the form
Note that this re-sequencing can be simply affected by an appropriate permutation of the variables.
Note also that the uncertainty description involves repeated diagonal blocks where the repeated
units are full blocks.
As mentioned in the last section, FDI in this chapter is achieved by obtaining upper and lower
bounds on the fault signals over the estimate horizon. This section gives the formulation of the
problem of obtaining these upper and lower bounds. Then, an approach for computing these
bounds is proposed by solving an LMI optimization. The cases for upper and lower bounds are
dealt with separately.
Firstly, the FDI problem for linear discrete-time uncertain systems is described as follows.
Problem 3.2.1. Let all data and variables be as given above. Then for i = 1, 2, . . . , Nf find
Note that the interval [φi , φ̄i ] is our best estimate for eTi f given all the information about the
system.
Remark 3.2.1. It is possible to combine our method and parity space methods by decoupling x0
and then carrying out the on-line optimization proposed above, but nothing, as far as accuracy,
can be gained since our solution is optimal (subject to relaxation error), however, it might reduce
the on-line computation burden by performing offline computations. 2
The following theorem shows that evaluating an upper bound on each eTi f can be obtained by
solving an LMI problem.
3.2 Fault signal bounds 37
Theorem 3.2.1. Let all data be as defined previously. Let Dm denote the set of all real m × m
diagonal matrices and define
Then for i = 1, . . . , Nf ,
and where the dependence of the Lijk s on the variables is suppressed for convenience.
Proof. Take p = ∆a (I − Dqp ∆a )−1 Dqw w Dqf f Dqu u , then (3.8) can be represented as
the following equalities,
It follows that the problem formulation for obtaining the upper bound on fi in (3.11) can be
rewritten as
Define Q := {Q : Q∆a = ∆a Q, ∀∆a ∈ ∆a } which is the set of matrices which commute with
all matrices in ∆a . Note that since q = Dyw w + Dyf f + Dyu u + Dyp p then Q̄i p = ∆a Q̄i q for
any Q̄i ∈ Q. According to the specific structure of ∆a given in (3.10), it can be verified that
Note that the constraints in (3.15) imply that pT Q̄Ti Q̄i p ≤ q T Q̄Ti Q̄i q, ∀Q̄i ∈ Q. Also, note that Ḡi
is skew-symmetric with the structure as defined above, then a computation shows that pT Ḡi q = 0.
−(Dqw w + Dqf f + Dqu u + Dqp p)T Q̄Ti Q̄i (Dqw w + Dqf f + Dqu u + Dqp p)
−τ̄i (y − Dyw w − Dyf f − Dyu u − Dyp p)T (y − Dyw w − Dyf f − Dyu u − Dyp p)
−pT Ḡi (Dqw w + Dqf f + Dqu u + Dqp p) + (Dqw w + Dqf f + Dqu u + Dqp p)T Ḡi p
w
f
− wT f T pT 1 Li (f¯i , D̄i , τ̄i , Q̄i , Ḡi )
p (3.16)
1
is valid for all f¯i ∈ R , D̄i ∈ RNw ×Nw , τ̄i ∈ R, Q̄i ∈ RN∆ ×N∆ and Ḡi ∈ RN∆ ×N∆ . From
(3.16), it can be seen that f¯i is an upper bound on eTi f if there exist 0 D̄i ∈ DNw , τ̄i ∈ R,
Q̄i ∈ Q and Ḡi ∈ G such that Li (f¯i , D̄i , τ̄i , Q̄i , Ḡi ) 0. An inspection of the sets Q and S
verifies that Q̄i ∈ Q if and only if S̄i := Q̄Ti Q̄i ∈ S and so an upper bound on eTi f is given by
solving the LMI optimization in (3.13). 2
3.2 Fault signal bounds 39
Remark 3.2.2. In the case of unstructured uncertainty on the data, the uncertainty set in (3.2)
is simply ∆ = Rn∆ ×n∆ . However, defining the variables in (3.4) over the horizon gives the
uncertainty set over the horizon as
S = {S ⊗ In∆ : 0 S = S T ∈ RN ×N }, G = 0.
It follows that even in the case of unstructured uncertainty description in the original problem
data, the uncertainty description over the horizon is structured. Thus the uncertainty description
in Theorem 3.2.1 for both structured and unstructured uncertainties is more general than the one
usually used in the literature [111, 92, 37, 73]. 2
Similarly, the next theorem derives lower bounds on the faults. The proof is omitted since it
follows similar lines to that of Theorem 3.2.1.
Theorem 3.2.2. Let all data and variables as defined in Theorem 3.2.1. Then for i = 1, . . . , Nf ,
Remark 3.2.3. From Theorems 3.2.1 and 3.2.2, it can be seen that the upper and lower bounds
on eTi f are given in the form of the LMI optimizations (3.13) and (3.17), respectively. Moreover,
since LMI problems are relatively easy to solve, it is possible to implement the proposed method
on-line. 2
Remark 3.2.4. Similar to the way of dealing with constraints in MPC, in our approach, the bounds
on the disturbances are explicitly included in the optimization. This is in contrast to observer-
based methods which deal with the energy transfer levels from disturbances/faults to residuals. 2
Remark 3.2.5. It can be seen that our approach uses the same information as the parity space
approach to achieve FD except that the approach in this work can exploit any available bounds
on the initial state as a constraint rather than decouple the state completely. 2
3.3 Fault signal energy bounds 40
Remark 3.2.6. Since, by this proposed approach, lower and upper bounds on eTi f are evaluated,
then provided that the assumptions are satisfied, that is, the bounds on the initial state and dis-
turbances are satisfied and the actual model falls within the description of the uncertainty, then
the algorithm in this work is immune to false alarms. However, it is pointed out that the proposed
approach may fail to identify a fault in the following sections. 2
The previous section gives a solution to the discrete-time FDI problem by providing upper and
lower bounds on each component of the fault signal at each time instant. In the case that the
interest is only in fault detection, then it is only needed to decide whether the energy of fault
signal f T f > 0. This section gives the formulation of the problem of obtaining a lower bound on
f T f over the estimation horizon, then an approach is proposed to evaluate the bound by using an
LMI optimization. Then a fault is regarded as having occurred if the lower bound on f T f > 0.
Firstly, the problem is described as follows.
Problem 3.3.1. Let all data and variables be as given above. Find
η := min fT f (3.18)
w ≤ w ≤ w̄
y = (Dyw +∆Dyw)w+(Dyf +∆Dyf)f +(Dyu +∆Dyu)u
∆Dyf , ∆Dyu , ∆Dyw given by (3.9) and (3.10)
y, u given
The theorem below is proposed to obtain the lower bound on f T f . Since the proof is similar to
that of Theorem 3.2.1 it is omitted.
Theorem 3.3.1. Let all data and variables as defined in Theorem 3.2.2. Then
f T f ≥ η ≥ η := min γ
L(γ, D, τ, S, G) 0
0 D ∈ D Nw
−S ∈ S
G∈G
τ ∈R
where
L11 ⋆ ⋆ ⋆
L21 M22 ⋆ ⋆
L(γ, D, T, S, G) =
L31 L32 L33 ⋆ ,
with
3.4 Subsequent state estimation 41
T D
M22 = −τ Dyf T
yf − Dqf SDqf − INf ,
As shown in Section 3.1, the initial state x0 is assumed to be in a given interval. After obtaining
the bounds on each fi within the estimation horizon from 0 to N − 1, it is required to predict the
interval of the next state x1 to carry on the proposed method for the next horizon which is from
1 to N . In this section, a modified version of the method proposed in the last section is given to
obtain upper and lower bounds on the subsequent state x1 .
Next, the problem of obtaining upper and lower bounds on x1 is formulated, given all information
during the estimation horizon.
Problem 3.4.1. Let all data be as defined previously. Then for i = 1, 2, . . . , n find
Next, the upper and lower bounds on x1 are obtained separately. Firstly, the upper bound on each
element of x1 can be evaluated by solving the LMI problem in the following theorem.
Theorem 3.4.1. Let all data and variables as defined in Theorem 3.2.1. Then for i = 1, . . . , n,
eTi x1 ≤ χ̄i ≤ χ̄
¯i := min γ
Ki (γ, D̄i , τ̄i , S̄i , Ḡi ) 0
0 D̄i ∈ DNw
S̄i ∈ S
Ḡi ∈ G
τ̄i ∈ R
3.4 Subsequent state estimation 42
where
L11 ⋆ ⋆ ⋆
L21 L22 ⋆ ⋆
Ki (γ, D̄i , τ̄i , S̄i , Ḡi ) =
L31 L32 L33
,
⋆
i
K41 i
K42 i
K43 i
K44
(3.22)
with
i = − 1 (w̄ + w)T D̄ + τ̄ (y − D u)T D T T 1 T
u
K41 2 i i yu yw − u Dqu S̄i Dqw − 2 ei A Bd Id ,
i = τ̄ (y − D u)T D T T 1 T u
K42 i yu yf − u Dqu S̄i Dqf − 2 ei Bf If ,
i = τ̄ (y − D u)T D − uT D T S̄ D + uT D T Ḡ − 1 eT B u I ,
K43 i yu yp qu i qp qu i 2 i p p
i = w̄ T D̄ w − τ̄ (y − D u)T (y − D u) − uT D T S̄ D u − eT B u I u + γ.
K44 i i yu yu qu i qu i u u
Inψ 0 . . . 0
u 0 0 ... 0
u B AB u . . . AN −2 B u AN −1 B u
where Bψ = ψ ψ ψ ψ , Iψ = . . . .
.. .. . . ..
0 0 ... 0
and where ψ stands for u, d, f and p. Note that the entries denoted by Lij are as defined
in (3.14).
Proof. Similar to (3.15), the problem formulation for obtain the upper bound on eTi x1 can be
rewritten as
eTi x1 − γ = eTi Ax0 + eTi Bdu Id d + eTi Bfu If f + eTi Buu Iu u + eTiBpu Ip p − γ
−(Dqw w + Dqf f + Dqu u + Dqp p)T Q̄Ti Q̄i (Dqw w + Dqf f + Dqu u + Dqp p)
−τ̄i (y − Dyw w − Dyf f − Dyu u − Dyp p)T (y − Dyw w − Dyf f − Dyu u − Dyp p)
−pT Ḡi (Dqw w + Dqf f + Dqu u + Dqp p) + (Dqw w + Dqf f + Dqu u + Dqp p)T Ḡi p
w
f
− wT f T pT 1 Ki (γ, D̄i , τ̄i , Q̄i , Ḡi )
p (3.24)
1
3.4 Subsequent state estimation 43
is valid for all γ ∈ R , D̄i ∈ RNw ×Nw , τ̄i ∈ R, Q̄i ∈ RN∆ ×N∆ and Ḡi ∈ RN∆ ×N∆ . From (3.24),
it can be seen that γ is an upper bound on eTi x1 if there exist 0 D̄i ∈ DNw , Q̄i ∈ Q and Ḡi ∈ G
such that Ki (γ, D̄i , τ̄i , Q̄i , Ḡi ) 0. It can be verified that Q̄i ∈ Q if and only if S̄i := Q̄Ti Q̄i ∈ S
and therefore an upper bound on eTi x1 is given by solving the LMI optimization in (3.23). 2
Similarly, it is proposed the following theorem which solves the lower bound on each element of
x1 .
Theorem 3.4.2. Let all data and variables as defined in Theorem 3.2.2. Then for i = 1, . . . , n,
eTi x1 ≥ χi ≥ χ := max γ
i Ki (γ, Di , τ i , S i , Gi ) 0
0 D i ∈ D Nw
−S i ∈ S
Gi ∈ G
τi ∈ R
where Ki (γ,Di , τ i , S i , Gi ) is as defined in (3.22).
The proof is omitted since it can be obtained just by following the lines in the proof of Theo-
rem 3.4.1.
Remark 3.4.1. Note that there is no observer or residual generator design involved in the pro-
posed approach, and FDI is achieved by directly computing the upper and lower bounds on the
fault signal. This implies that FDI could be implemented on-line based on the input and output
data, system model and uncertainty description. 2
Remark 3.4.2. Observer-based FD methods generate residuals which seek to indicate faults for
all plants within the uncertainty set [111, 61], while the approach in this work is proposed for
the actual system with its particular uncertainty since it is implemented by using on-line input
and output data of the system. It follows that our approach can in fact be used to identify the
uncertainty of a system model. This is described in more detail in Section 6.2.5. 2
Remark 3.4.3. Since the optimization problem to obtain the bounds is solved on-line, it could
be computationally expensive, hence our approach can only be used to monitor sufficiently slow
systems. To some extent, this can be ameliorated by varying the horizon length but at the expense
of precision. 2
3.4 Subsequent state estimation 44
As illustrated above, the upper and lower bounds on the fault signal and subsequent state can
simply be obtained by solving LMI problems. Next, a procedure is given to illustrate the steps by
which the proposed scheme is applied to real systems.
Algorithm 3.4.1.
1. Set i = 0.
5. Set j = 1.
7. Obtain uj+N −1 , and yj+N −1 , compute upper and lower bounds on fj , fj+1 , . . . , fj+N −1 ,
and xj+1 .
Note that steps 1 and 2 only involve data collections: the computations start after the estimation
horizon N is reached. If fault estimate before the horizon is reached are required, the algorithm
can be easily modified so that the faults can be estimated using an increasing horizon length.
By applying the procedure, the upper and lower bounds on faults can be computed over the moving
horizon. A fault is regarded as having occurred at some sampling instant when its upper and lower
bounds obtained are both either greater or smaller than zero.
Remark 3.4.4. In some sense, the threshold for FDI in this work is implicitly built into the algo-
rithm since the upper and lower bounds on the fault signal are computed. The need for threshold
in the proposed approach only arises if the actual system and disturbances fall outside our model
description. 2
Remark 3.4.5. On the other hand, if the upper bound on eTi f is greater than zero and the lower
bound is less than zero, then there is insufficient information to determine whether a fault has
occurred and faults may be missed, although the availability of the lower and upper bounds on
the faults may still be useful. 2
3.5 Numerical examples 45
Remark 3.4.6. Although the procedure presented above is for linear time-invariant systems, they
can be easily modified for linear time-varying systems. 2
Remark 3.4.7. Similarly, the scheme proposed in this work can be used to solve FDI problems
for linear discrete-time systems without model uncertainties. Optimal upper and lower bounds on
the fault signal and the subsequent state can be easily obtained by solving the following Linear
Programming problems:
η := min fT f
x0 ≤ x0 ≤ x̄0
d ≤ d ≤ d¯
y = Cy0 x0 +Dyd d+Dyf f +Dyu u
y, u given
Remark 3.4.8. Many observer-based FDI schemes [49, 59, 53] are based on the assumption that
Df or Dd has full column rank, which could be restrictive in some circumstances. In this work,
this rank assumption is not necessary for applying the proposed FDI approach. However, loss of
column rank on Df could make the upper and lower bounds on the fault signal at the last sampling
instant of each horizon not reliable. 2
In this section, two examples are given to illustrate the effectiveness of the proposed scheme.
3.5.1 Example 1
Firstly, consider a discrete-time model (with no uncertainties) of a vertical takeoff and landing air-
craft in the vertical plane used in [90], where the states are the horizontal velocity (knot), vertical
3.5 Numerical examples 46
velocity (knot), pitch rate (degree/second) and pitch angle (degree), respectively. The actuator in-
puts are collective pitch control and longitudinal cyclic pitch control, respectively. The parameters
of the linearized model are given as
0.9828 0.0083 −0.0454 −0.2461
0.0117 0.5813 −0.3898 −1.6662
A = 0.0458 0.1274 0.8230
,
0.4803
0.0117 0.0358 0.4433 1.1361
0.2666 0.0365 0.2666 0.0365
1.7629 −3.2664 1.7629 −3.2664
−2.3152 1.7209 , Bf =−2.3152 1.7209 ,
Bu =
0 1 1 1 0.1144 0.0063
0 0.2
0 0.1
0.3 0 ,
Dd = Du = 04×2 , Df = 04×2 .
0 0
T
The same faults and disturbances are used as given in [90]: f (k) = 1.2 0.8 , for k ≥ 20
T T
and f (k)= 0 0 elsewhere. Also, d(k) = 0.5 cos(k)e−0.05k 0.5 sin(k) for all k and take
the horizon length N = 10. Generally, FDI schemes give better results as the number of outputs
increases. However, to demonstrate the effectiveness of the proposed method, the system model is
modified by removing the first output. Applying the LP optimization given in (3.25), the following
figures are obtained by using the last three outputs showing the upper and lower bounds on the
fault signal from sampling instants 12 to 21 and 15 to 24, respectively.
1.5
Lower bound
1.2 Fault1
Upper bound
1
Magnitude
0.5
−0.5
12 13 14 15 16 17 18 19 20 21
Time
1.5
Lower bound
Fault2
Upper bound
1
0.8
Magnitude
0.5
−0.5
12 13 14 15 16 17 18 19 20 21
Time
1.5
Lower bound
1.2 Fault1
Upper bound
1
Magnitude
0.5
−0.5
16 17 18 19 20 21 22 23 24 25
Time
1.5
Lower bound
Fault2
Upper bound
1
0.8
Magnitude
0.5
−0.5
16 17 18 19 20 21 22 23 24 25
Time
From the graphs shown in Figures 3.1, 3.3, 3.2 and 3.4 and the values of the bounds, it can be
seen at the 21st sampling instant that both the upper and lower bounds were greater than zero at
the 20th sampling instant, which indicates fault occurrence in the system. The band between the
upper and lower bounds on the fault signal is narrow and the fault is well-located. Note that since,
in this example Df = 0, the bounds on the faults at the last sampling instant of each estimation
horizon are not reliable. Also, using Theorem 3.4.1 and 3.4.2, it can be shown that state x16 is
3.5 Numerical examples 48
T
x16 = −0.9649 0.7665 −0.1595 −0.3022
T
x16 = 0.2508 0.8626 −0.1283 −0.2579
T
x16 = 1.4179 0.8627 −0.0081 −0.2064
Furthermore, the proposed method can give even better results when it is applied using all four
outputs as in [90]. Figures 3.5−3.8 illustrate the FDI results with the same faults and disturbances.
It shows that both the upper and lower bounds are tight so that the faults are accurately detected
at the 21st sampling instant.
1.5
Lower bound
1.2 Fault1
Upper bound
1
Magnitude
0.5
−0.5
12 13 14 15 16 17 18 19 20 21
Time
1.5
Lower bound
Fault2
Upper bound
1
0.8
Magnitude
0.5
−0.5
12 13 14 15 16 17 18 19 20 21
Time
1.5
Lower bound
1.2 Fault1
Upper bound
1
Magnitude
0.5
−0.5
16 17 18 19 20 21 22 23 24 25
Time
1.5
Lower bound
Fault2
Upper bound
1
0.8
Magnitude
0.5
−0.5
16 17 18 19 20 21 22 23 24 25
Time
3.5.2 Example 2
To demonstrate the effectiveness of the proposed robust FDI scheme in a real system with
uncertainties, a jet engine example is considered in this section. By using zero order hold
equivalent method [2] with a sample period T = 0.1s, the discrete-time state-space model of the
GE-21 jet engine [48, 67, 89] is given as follows.
0.7119 0.1257 0.0525 −0.1127 0.1072
A = , Bu = ,
−0.0250 0.8252 0.0365 0.1037 0.0109
−0.0252 0.0027 0.0525 −0.1127
Bd = , Bf = ,
0.0013 −0.0166 0.0365 0.1037
1 0 0 0 0
Cy = 0 1 , Dyu = 0 0 0 ,
0.731 0.786 0.267 −0.025 −0.146
−0.0008 0.0006 1 0
Dyd = 0.0008 0.0007 , Dyf = 0 1 .
−0.0005 −0.0008 0 0
Since the example in [67, 89] does not involve disturbances, the matrices Bd and Dyd have been
3.5 Numerical examples 50
assumed as above. Two fault signals are considered in this example. f1 (k) and f2 (k) are simulated
as positive and negative step signals with magnitude 1 which are connected from k = 7 and 9, re-
T
spectively. The disturbances are assumed to be given as d(k) = 0.5 cos(k)e−0.05k 0.5 sin(k)
and the inputs are taken as unit step signals starting from the initial sampling instant. Also, as-
sume that the upper and lower bounds on the initial state x0 and disturbances are 0.5 and −0.5,
respectively, and take the estimation horizon length N = 5.
Assuming no uncertainties, by applying the algorithm given in (3.25), upper and lower bounds on
the fault signals are obtained as follows.
1.5
Lower bound
Fault1
Upper bound
1
Magnitude
0.5
−0.5
3 4 5 6 7
Time
0.5
0
Magnitude
−0.5
−1
Lower bound
Fault2
Upper bound
−1.5
5 6 7 8 9
Time
1.5
Lower bound
Fault1
Upper bound
1
Magnitude
0.5
−0.5
5 6 7 8 9
Time
0.5
Lower bound
Fault2
Upper bound
0
Magnitude
−0.5
−1
−1.5
7 8 9 10 11
Time
According to the decision logic of the proposed FDI scheme, it can be concluded that
there are two faults connected from the 7th and 9th sampling instant, respectively. Moreover, as
can be seen from Figure 3.9−3.12, the band between the upper and lower bounds on each fault
signal is narrow and its corresponding fault is well-located in it.
In order to illustrate the robustness of the FDI approached given in Theorems 3.2.1 and 3.2.2,
assume that uncertainty parameters are introduced into the same system model. Here consider an
unstructured norm-bounded uncertainty with the parameters as follows.
0.8063 −0.8364 −0.0330 −0.0211
∆ = , Bp = ,
0.1002 −0.6527 −0.0520 0.0570
−0.020 0.026 0.0124 −0.0233 0.0106
Cq = , Dqu = ,
−0.029 −0.032 −0.063 −0.0123 −0.011
0.038 0.011 −0.015 −0.016
Dqf = , Dqd =
0.0084 −0.05 0.011 0.023
−0.1012 0.0550
Dp = 0.0231 −0.1100 .
−0.0501 0.0282
3.5 Numerical examples 52
By solving the LMIs in (3.14) and (3.17), the figures and values of the bounds are obtained as
follows. It can be concluded that Fault 1 and 2 are well detected and isolated at the 7th and 9th
sampling instants, respectively. Note that since there exist uncertainties in the system model, the
bounds shown in Figure 3.13−3.16 are less tight than the ones shown above when no uncertainties
are considered.
1.5
Lower bound
Fault1
Upper bound
1
Magnitude
0.5
−0.5
3 4 5 6 7
Time
0.5
0
Magnitude
−0.5
−1
Lower bound
Fault2
Upper bound
−1.5
5 6 7 8 9
Time
1.5
Lower bound
Fault1
Upper bound
1
Magnitude
0.5
−0.5
5 6 7 8 9
Time
0.5
Lower bound
Fault2
Upper bound
0
Magnitude
−0.5
−1
−1.5
7 8 9 10 11
Time
Also, using Theorem 3.4.1 and 3.4.2, the lower and upper bounds on the subsequent state x21 are
solved and given as follows. It shows that x21 is well-estimated and within its lower and upper
bounds.
T
x21 = 0.5592 0.9404
T
x21 = 0.6050 1.0182
T
x̄21 = 0.6812 1.0622
It is noted that, although the band between the upper and lower bounds is narrow, there is still the
possibility that the fault is missed if its magnitude is very small. The following simulation shows
that when there is only an abrupt pulse fault at 7s with value 0.1, the lower and upper bound on
Fault 1 is negative and positive, respectively at 7s. Hence, no conclusion can be drawn on the
result of FDI according to Theorem 3.2.1 and 3.2.2.
In this case, the method in Theorem 3.3.1 is used such that the objective is only to achieve fault
detection. By solving (3.19), the lower bounds on f T f over the estimation horizon from 3 to 7
is 0.0044 which is larger than zero. Therefore, a conclusion can be drawn that a fault occurred
during the estimation horizon from 3s to 7s. Thus Theorem 3.3.1 provides a better reliability for
fault detection or, in other words, fault detection is relatively easier to achieve than fault isolation.
3.6 Concluding remarks 54
0.15
Lower bound
0.1 Fault1
Upper bound
0.05
Magnitude
−0.05
−0.1
−0.15
−0.2
3 4 5 6 7
Time
This chapter considers the FDI problem for linear discrete-time uncertain systems subject to faults,
bounded additive disturbances and norm-bounded uncertainties. A novel approach is proposed to
solve the FDI problems by using a system model, together with input/output measurements over a
finite estimation horizon. First, the proposed method is used to compute upper and lower bounds
on the fault signals on-line. Then a fault is regarded as detected and isolated if, at some sampling
instant, its upper bound is lower than zero or the lower bound is larger than zero. LMI optimization
techniques are used to obtain the bounds. Similarly, FD can be achieved by solving the lower
bounds on the fault energy f T f .
In this work, the bounds on faults are obtained by solving optimization problems subject to the
constraints of the bounds on the initial state, disturbances and norm-bounded uncertainties by
using system model and input/output measurements over a finite estimation horizon. Therefore, it
can be regarded as a dual to model predictive control (MPC) schemes which use the system model
to solve an on-line optimization problem to determine an optimal control strategy subject to hard
constraints at each sampling instant, usually over a finite horizon.
One advantage of the proposed approach over the parity space approaches is that in this work
the upper and lower bounds on the initial state, if available, can be used and the initial state is
not necessarily decoupled, which can give more information in computing the bounds on faults.
Furthermore, by applying the proposed approach, FDI is achieved by computing upper and lower
bounds on the actual fault signals rather than producing a residual signal. So that the occurrence
of a fault is detected if its upper bound is lower than zero or the lower bound is larger than zero.
Hence, this approach is free of false alarms.
56
Chapter 4
In [93], an H− /H∞ FD problem is investigated and a suboptimal solution is given, which is due to
the fact that a bilinear matrix inequality formulation is employed so that some variables are needed
to be pre-defined to solve the optimization problem. Wang et al [91] consider the H∞ /H∞ FD
problem and give a solution which is suboptimal due to the fact that a common Lyapunov variable
is used to solve the matrix inequalities, which results in a conservative FD observer design. In
[87], both H∞ /H∞ and H− /H∞ FD problems are considered and a suboptimal solution is given
since the FD problem is formulated as a quasi-linear matrix inequality problem that necessitates
an iterative method to approximate the optimal solution. In [49, 57, 54] an optimal solution to the
H− /H∞ is presented. However, the solution depends on the availability of an exact model of the
system and therefore may not have good robustness properties against model uncertainties.
This work gives an optimal solution to the robust H− /H∞ FD problem for linear time-invariant
systems subject to faults and disturbances. First, a new performance index is developed to capture
the requirements of fault detection and disturbance rejection which is particularly suitable for
4.1 Fault detection problem formulation 57
dealing with model uncertainties. The disturbance rejection performance is measured, using the
H∞ norm, by the size of the disturbance to residual dynamics. Using a static observer, a class of
solutions to a novel FD problem is given in the form of a simple LMI with two degrees of freedom.
Then, the degrees of freedom in this class of solutions are used to derive the optimal filter to the
H− /H∞ FD problem. Moreover, it is shown that, under certain conditions, fault isolation can also
be achieved. Finally, it is shown that, with the use of weighting filters, the results can be improved
and certain, commonly held, rank assumptions can be relaxed.
The structure of this chapter is organized as follows. A novel H− /H∞ FD problem is formulated
in Section 4.1. Section 4.2 gives a class of solutions to the FD problem. In Section 4.3, the optimal
filter design for the H− /H∞ FD problem is given and further specifications are investigated.
Section 4.4.1 gives a technique to improve the results through the use of weighting filters, while
Section 4.4.2 considers the removal of certain common rank assumptions on the fault dynamics.
Section 4.5 extends the scheme to uncertain systems. Finally, numerical examples are presented
in Section 4.6 to validate the proposed approaches, and Section 4.7 summarizes the results in this
chapter.
Consider a linear time-invariant dynamic system subject to disturbances, modeling errors and
process, sensor and actuator faults modeled as
where x(t) ∈ Rn , u(t) ∈ Rnu and y(t) ∈ Rny are the process state, input and output vectors,
respectively, and where f (t) ∈ Rnf and d(t) ∈ Rnd are the fault and disturbance vectors, respec-
tively. Here, Bf ∈ Rn×nf and Df ∈ Rny ×nf are the component and instrument fault distribution
matrices, respectively, while Bd ∈ Rn×nd and Dd ∈ Rny ×nd are the corresponding disturbance
distribution matrices [35]. As mentioned in Chapter 2, a residual signal in an FD system repre-
sents the inconsistency between the system variables and the mathematical model. Here, consider
a residual generator using a static observer of the form
˙
x̂(t) = Ax̂(t)+Bu u(t)−L (y(t)−C x̂(t)−Du u(t)) ,
where x̂(t) ∈ Rn is the observer state and r(t) ∈ Rnr is the residual signal. Here L ∈ Rn×ny and
H ∈ Rnr ×ny are the observer and residual gain matrices, respectively, and are to be determined.
Define the state estimation error signal as e(t) = x(t) − x̂(t). It follows that the residual dynamics
are given by
By taking Laplace transforms, r(s) = Trf (s)f (s) + Trd (s)d(s), where
s A+LC Bf +LDf Bd +LDd
Trf (s) Trd (s) =
HC HDf HDd
= F (s) Gf (s) Gd (s) (4.3)
are the transfer matrices from faults and disturbances to the residual, respectively, and where
s A + LC L s A Bf Bd
F (s) = , Gf (s) Gd (s) = (4.4)
HC H C Df Dd
are the FD filter and the fault and disturbance dynamics, respectively.
Remark 4.1.1. Greater flexibility can be introduced to the design of the FD schemes based on
residual generation by considering a dynamic observer. However, this causes a couple of difficul-
ties. See comments in Section 6.2.2. 2
With the preliminaries above, the H− /H∞ FD problem under investigation is formulated as fol-
lows.
A2. Gf (s) has full column rank over the extended imaginary axis (see Remarks 4.1.4 and Sec-
tion 4.4.)
Find
kTrd k∞
γ0 := min { : A + LC is stable}.
kTrf k−
4.1 Fault detection problem formulation 59
Remark 4.1.2. An interpretation of Problem 4.1.1 is that σ̄(Trd (jω)) is required to be smaller
than or equal to γ0 σ(Trf (jω)) for all ω ∈ R. This captures the requirement of ensuring insensi-
tivity to the disturbance and sensitivity to faults. 2
Since
−1 −1
Trf Trd ≤ Trf kTrd k∞ = kTrd k∞ /kTrf k− ,
∞ ∞
then,
−1
γ0 ≥ min { Trf Trd : A + LC is stable} =: γ1 . (4.5)
∞
Problem 4.1.2. Let all variables and assumptions be as given in Problem 4.1.1. For a given γ > 0
find L and H, if they exist, such that A + LC is stable and
−1
Trf Trd < γ. (4.6)
∞
Remark 4.1.3. Assumption A1 is needed to guarantee the existence of at least one L such that
A + LC is stable. 2
−1 ∗ (s)T −∗ (s) ≺ γ 2 I
Remark 4.1.4. The cost function in (4.6) can be expressed as Trf (s)Trd (s)Trd rf
or, equivalently, as
∗
Trd (s)Trd (s) ≺ γ 2 Trf (s)Trf
∗
(s), (4.7)
If Gf (s) loses rank over the extended imaginary axis, then so does Trf (s) = F (s)Gf (s) and (4.7)
cannot be satisfied for any γ > 0. In particular, the assumption implies that Df has full column
−1
rank which is necessary for Trf (s) to have a proper state-space realization. This assumption can
∗ (s)−γ 2 T (s)T ∗ (s)
be (somewhat) relaxed if the cost function in (4.7) is changed to Trd (s)Trd rf rf
0 and if Gd (s) loses rank over the imaginary axis whenever Gf (s) does. Dealing with this situa-
tion will necessitate an intricate analysis of spectral factorizations with imaginary axis zeros [13],
which is outside the scope of this work. In Section 4.4, this assumption is relaxed by modifying the
cost function in (4.6) and introducing weighting functions. 2
Remark 4.1.5. An alternative approach in the case that Gf (s) loses rank over the external imag-
inary axis is to design the FD scheme over a specific finite frequency range inside which Gf (s)
has full rank [90, 92, 58], although the solution involves non-convex optimization. 2
4.2 A class of optimal FD filters for Problem 4.1.2 60
Remark 4.1.6. Suppose that ny < nf , then assumptions A1 and A2, together with Theorem 13.32
n ×ny n ×nf
in [114] imply that Gf (s) = Ĝf (s)Gi (s) where Ĝf (s) ∈ RL∞y and Gi (s) ∈ RL∞y such
that Gi (s)G∗i (s) = Iny . Furthermore, Ĝf (s) can be chosen to have the same A and C matrices
as Gf (s). Using (4.4) and (4.7) it gives
∗
Trd (s)Trd (s) ≺ γ 2 Trf (s)Trf
∗
(s) ⇔ F (s)Gd (s)G∗d (s)F ∗ (s) ≺ γ 2 F (s)Gf (s)G∗f (s)F ∗ (s)
Thus Gf (s) may be replaced by Ĝf (s) and so there is no loss of generality in assuming that
nf ≤ ny . 2
1. There is no need for nr > nf since the interest of this work is in increasing the sensitivity of
the residual to faults by increasing the singular values of Trf (s) relative to those of Trd (s),
and there are at most nf singular values of Trf (s) since ny ≥ nf .
2. Since the set of all optimal filters when nr < nf is a subset of those when nr = nf , maximal
sensitivity to faults is obtained when nr = nf .
3. This will prove useful when fault isolation is considered in Section 4.3 below. 2
In this section, a set of (not necessarily stable) filters of the form given in (4.4) is characterized
such that (4.6) is satisfied. Then, the approach is to use the bounded real lemma to derive necessary
and sufficient conditions for (4.6) in the form of a (nonlinear) matrix inequality. Then, a change
of variables is introduced to derive an equivalent LMI.
Since Df ∈ Rny ×nf has full column rank, there exist Df† ∈ Rnf ×ny and Df⊥ ∈ R(ny −nf )×ny
such that " # " #!
Df† Inf Df†
Df = , rank = ny .
Df⊥ 0 Df⊥
Let
H = U1 Df† + U2 Df⊥ , L = (Z1 − Bf )Df† + Z2 Df⊥ (4.8)
4.2 A class of optimal FD filters for Problem 4.1.2 61
where U1 ∈ Rnf ×nf , U2 ∈ Rnf ×(ny −nf ) , Z1 ∈ Rn×nf and Z2 ∈ Rn×(ny −nf ) are free parame-
ters, with U1 nonsingular to ensure that HDf is invertible. Then
Then
−1 s Ac Bc
Trf (s)Trd (s) = (4.9)
Cc D c
where
where
A − Bf Df† C Bd − Bf Df† Dd
A1 Bd,1
Df† C Df† Dd
C1 Dd,1 = (4.11)
C2 Dd,2 ⊥
Df C ⊥
Df Dd
Theorem 4.2.1. Let all variables and assumptions be as given in Problem 4.1.2. For some γ > 0,
there exist L and H such that A + LC is stable and (4.6) is satisfied if and only if there exist
P = P T , Q and U0 such that the following LMI
AT1 P +C2TQT+PA1+QC2 ⋆ ⋆
T P +D T QT
Bd,1 −γI ⋆ ≺ 0,
d,2 (4.12)
C1 +U0 C2 Dd,1 +U0 Dd,2 −γI
Proof. From a generalized bounded real lemma in [52], there exist L and H such that (4.6) is
satisfied if and only if there exists P = P T (not necessarily positive definite) such that
ATc P +P Ac P Bc CcT
BcT P −γI DcT ≺ 0. (4.13)
Cc Dc −γI
Substituting (4.10) in (4.13) and then making the change of variables Q = P Z0 shows that there
exist L and H such that (4.6) is satisfied if and only if there exist P = P T , Q and U0 such that
(4.12) is satisfied. Furthermore, if (4.12) has feasible solutions P, Q = P Z0 and U0 and defining
L = L 0 + Z1 H 0 , H = U 1 H0 , (4.15)
then a class of (not necessarily stable) FD filters F (s) of the form (4.4) is parameterized by free
variables U1 and Z1 , subject to U1 nonsingular. It can be seen that A + LC = (A + L0 C) +
Z1 (H0 C). Then it remains to prove the stability of the filter F (s). The proof below shows that
it is always possible to stabilize A + LC by appropriately choosing Z1 . Consider the following
three cases:
has full column rank for all s such that s + s′ > 0. Using the fact that A21 = −L0,2 C11 , it
follows that
A11 − sI A12 I 0 0 A11 − sI A12
0 A22 + L0,2 C12 − sI = 0 I L0,2 A21 A22 − sI
C11 C12 0 0 I C11 C12
has full column rank for all s such that s + s′ > 0. Thus the pair (A11 , C11 ) is detectable. It
follows that there exists L̃0,1 such that A11 + L̃0,1 C11 is stable. Define L̃0 = [ L̃T0,1 LT0,2 ]T .
Then
A11 + L̃0,1 C11 A12 + L̃0,1 C12
A+ L̃0 C
= 0 A22 +L0,2 C12
H0 C
0 H0 C12
s
and so (A + L̃0 C, H0 C) is detectable. Define L̃ = L̃0 + Z1 H0 and F̃ (s) = (A +
L̃C, L̃, HC, H). Then a calculation shows that F̃ (s) = F (s). It follows that
(F̃ Gf )−1 (F̃ Gd ) < γ. Redefining L̃0 as L0 and F̃ (s) as F (s), then since (A +
∞
L0 C, H0 C) is detectable, choose Z1 such that A+L0 C +Z1 H0 C is stable. 2
Remark 4.2.1. If P is singular, then, because of the strict inequality in (4.12), we can perturb P
such that the perturbed P is nonsingular and inequality (4.12) is still satisfied. This shows that we
can always recover Z from P and Q. 2
In the previous section, a set of FD filters is derived and residual dynamics are parameterized by
free Z1 and U1 , with U1 nonsingular that solve Problem 4.1.2. This section shows that, through
a suitable choice of Z1 and U1 , the optimal H− /H∞ FD filter for Problem 4.1.1 is obtained and
further specifications such as fault isolation can also be achieved in some circumstances.
Theorem 4.3.1. Let all variables and assumptions be as given in Problem 4.1.1 and Theorem 4.2.1
and define F (s) using (4.14), (4.15) and (4.4). Then
1. If A + L0 C is stable, let Z1 = 0 and U1 = Inf . Then Trf (s) = Inf so that F (s) is a
solution of Problem 4.1.1 and furthermore achieves fault isolation.
2. If A + L0 C is unstable, then there always exist a Z1 ∈ Rn×nf and a unitary U1 ∈ Rnf ×nf
∗ (s)T (s) = I . It follows that F (s) solves Problem 4.1.1.
such that F (s) is stable and Trf rf nf
4.3 Design of an optimal H− /H∞ FD filter 64
Proof. Firstly, note that since Bf + L0 Df = 0 and H0 Df = Inf from (4.14), then
s A+L0 C +Z1 H0 C Z1
Trf (s) = (4.16)
U 1 H0 C U1
from (4.3) and (4.15). To ensure fault isolability for all faults, it is required that the transfer
matrix Trf (s) is diagonal. For example, if Trf (s) = Inf (ignoring disturbances), r = f and each
individual fault only affects its corresponding residual signal.
2. If A + L0 C is not stable we choose suitable Z1 and U1 such that F (s) is stable and
∗ (s)T (s) = I . A simple calculation using (4.10) and (4.14) shows that
Trf rf nf
A + L0 C = Ac . (4.17)
It follows from the (1,1) block of (4.13) and the assumed nonsingularity of P that A + L0 C
has no eigenvalues on the imaginary axis. It follows that a similarity transformation can be
effected on the data, if necessary, such that
As 0
A + L0 C
= 0 Aa
H0 C
Cs Ca
Note that Pa ≻ 0 follows from the facts that −Aa is stable and (Aa , Ca ) is observable,
which in turn follows from the fact that (A + L0 C, H0 C) is detectable. Then (4.18) implies
4.4 Extensions of the fault detection problem 65
that
s Aa + Za Ca Za
Trf (s) =
U 1 Ca U1
Aa − Pa−1 CaT Ca −Pa−1 CaT
= ·
U 1 Ca U1
Hence, Trf (s) and so that Trd (s) are stable, which follows from (Aa +Za Ca )T Pa +Pa (Aa +
Za Ca ) + CaT Ca = 0 and Pa ≻ 0. Since U1T U1 = I and given (4.20) is satisfied , by taking
∗ (s)T (s) = I.
matrix calculations, it shows that Trf rf
∗ (s)T (s) = I
Therefore, the above result shows that a suitable choice of Z1 and U1 ensures Trf rf
and then γ0 = γ1 in (4.5). Therefore, this proves that our scheme gives the optimal solution of
Problem 4.1.1. 2
Remark 4.3.1. Note that if A + L0 C is not stable then due to the specific choice of Z1 in (4.19)
to stabilize F (s), Trf (s) is no longer diagonal, which implies that it is not possible to achieve
fault isolation. However, it is always possible to choose the free variable U1 suitably such that
T
F (s) achieves fault isolation at DC (ω = 0). Set U1 = Inf − Ca (Aa + Za Ca )−1 Za , then
Trf (0) = U1 − U1 Ca (Aa + Za Ca )−1 Za = Inf , which ensures optimal FDI at DC. 2
Remark 4.3.2. Note that if A + L0 C stable, the optimal FD filter for the H− /H∞ problem also
achieves fault isolation as shown in Part 1 of Theorem 4.3.1. Now suppose in the LMI in (4.13),
the constraint P = P T ≻ 0 is added in instead of just P = P T . This will, in general, result in
a larger optimal γ but will ensure that A + L0 C is stable from (4.17). Hence, it follows from the
preceding argument that the fault isolation capability of the filter can be improved at the expense
of its fault detection ability. 2
In the previous sections, an optimal FD scheme with some fault isolation properties is presented to
solve Problem 4.1.1. In this section, two extensions are briefly outlined to the proposed scheme.
4.4 Extensions of the fault detection problem 66
First, a frequency weighting filter is introduced to improve the fault detection effectiveness and
also to remove assumption A2 in Problem 4.1.1. Then, an approach is given to extend the scheme
in this work to uncertain systems.
In the fault detection scheme developed in the previous section, the performance of the FD filter
depends on the value of γ. If γ < 1, then the smallest singular value of Trf (s) is larger than
the largest singular value of Trd (s) over all frequencies and the performance of the FD scheme
is expected to be satisfactory. On the other hand, if γ >> 1, then the fault detection may not be
effective.
for all frequencies ω ∈ R. In practice, for the purpose of fault detection, it is only needed
that (4.22) is achieved over a limited frequency range, or even at a single frequency. Let
Wω0 be a stable and proper minimum-phase single-input single-output bandpass filter with a
peak gain of one at frequency ω0 ∈ R. Suppose the cost function in (4.6) is replaced by
(F Gf Wω−1
0
)−1 (F Gd Wω0 ) ∞ < γω0 . This can be recast as Problem 4.1.2 by replacing
[Gf (s) Gd (s)] by Gf Wω−1 0
(s) Gd Wω0 (s) , which, because of our assumptions on Wω0 , in-
herits the assumptions of Problem 4.1.2.
A major assumption that is widely made is that Gf (s) has no zeros over the extended imaginary
axis, and in particular that Df has full column rank. This excludes important situations such as
no sensor faults (Df = 0). The problem is caused by the choice of the cost function in (4.7):
4.5 Fault detection for uncertain systems 67
in the case that Gf (jω1 ) loses rank at ω1 ∈ R, then so does Trf (jω1 ), so that (4.7) cannot in
general be satisfied at jω1 . In the following, a discussion is given for relaxing Assumption A2 and
ameliorating the problem of loss of rank of Gf (s).
Suppose that Gf (s) has zeros on the extended imaginary axis. Then Gf (s) can be written as
Gf (s) = Ĝf (s)Gz (s) where Ĝf (s) has no zeros over the extended imaginary axis and Gz (s) has
all its zeros on the extended imaginary axis [54]. Suppose now that the cost function in (4.6) is
−1
replaced by T̂rf Trd < γ where T̂rf (s) = F (s)Ĝf (s) and the we enforce that T̂rf (s) is inner
∞
as shown in Section 4.3. Then σi (Trf (jω)) = σi (T̂rf (jω)Gz (jω)) = σi (Gz (jω)), since T̂rf (s)
is inner, and σ̄(Trd (jω)) < γ. Note that the singular values of Trf (jω) are set equal to those of
Gz (jω) instead of being set equal to 1 in the original scheme, therefore γ is the fault detection
level relative to σi (Gz (jω)) rather than 1 as given in (4.21).
This section shows that the FD scheme developed in the chapter can also be applied to uncertain
systems. Two different type of uncertainties, i.e. unstructured and structured uncertainties, are
dealt with separately. For each case, a suboptimal solution is derived in the form of bilinear matrix
inequality (BMI).
The matrices ∆A, ∆Bu , ∆C, ∆Du represent modeling errors and are given by:
∆A ∆Bu Fa
= ∆H E a E b ,
∆C ∆Du Fc
(4.24)
4.5 Fault detection for uncertain systems 68
This LFT representation of uncertainty, which is assumed to be well-posed over ∆ (that is, det(I −
H∆) 6= 0 for all ∆ ∈ ∆), has great generality and is widely used in control theory [115]. The
same static observer scheme of Section 4.1 is used to get:
d(s)
r(s) = Trf (s)f (s) + Trdu (s) ,
u(s)
where
A+LC ∆A+L∆C Bf +LDf Bd +LDd ∆Bu +L∆Du
s
[Trf (s) || Trdu (s)] = 0 A+∆A Bf Bd Bu +∆Bu .
HC H∆C HDf HDd H∆Du
and Trdu (s) is the transfer function from the disturbances and input to the residual.
Remark 4.5.1. Note that the residual signal in the case of uncertain systems depends on the input
signal u and so the observer does not cancel the system dynamics. This makes the FD problem
much harder than the case when there are no uncertainties and the solution will involve BMI
optimizations. 2
By following a similar derivation to (4.9) and separating the terms involving modeling uncertain-
−1
ties from the other terms, the state space representation of Trf (s)Trdu (s) is given as follows.
−1 s
Trf (s)Trdu (s) = (Aoc + A∆ o ∆ o ∆ o ∆
c , Bc + Bc , C c + C c , D c + D c ) (4.25)
where
A1 + Z0 C2 0 Bd,1 + Z0 Dd,2 0
Ac Bc
o o
= −Bf (C1 + U0 C2 ) A Bd − Bf (Dd,1 + U0 Dd,2 ) Bu .
Cco Dco
C1 + U 0 C2 0 Dd,1 + U0 Dd,2 0
and where
∆ ∆
Fa + Z0 Df⊥ Fc −Bf Df† Fc
Ac Bc F
= r1 ∆H Er1 Er2 = Fa −Bf (Df† Fc +U0 Df⊥ Fc ) ∆H 0 Ea 0 Eb .
Cc∆ Dc∆ Fr 2
Df† Fc +U0 Df⊥ Fc
and where the changes of variables in (4.8), (4.9) and (4.11) were applied. It follows from the
generalized bounded real lemma in [52] that there exists a (not necessarily stable) filter of the
form given in (4.4) such that
−1
Trf Trdu <γ
∞
4.5 Fault detection for uncertain systems 69
where
(Aoc )T Pc + Pc Aoc Pc Bco (Cco )T
Tco = (Bco )T Pc −γI (Dco )T
Cco Dco −γI
T
(A∆c ) Pc + Pc Ac
∆
Pc Bc∆ (Cc∆ )T
Tc∆ = ∆ T
(Bc ) Pc 0 (Dc∆ )T ·
∆
Cc Dc∆ 0
When the system uncertainty is unstructured, by applying Lemma 1.4.1 from [86], it follows that
(4.26) ⇔ Tco + F̃ ∆H Ẽ + Ẽ T ∆H T F̃ T ≺ 0
o
Tc + λẼ T Ẽ F̃ + λẼ T H
⇔ ≺0
F̃ T + λH T Ẽ λ(H T H − I)
o
Tc F̃ λẼ T
⇔ T − T (−λ−1 I) λẼ λH ≺ 0 (4.28)
F̃ −λI λH
for some λ > 0. Using a Schur complement argument in (4.28), it is obtained that
(Aoc )T Pc + ∗ Pc Bco (Cco )T T
Pc Fr1 λEr1
∗ −γI (Dco )T 0 T
λEr2
(4.26) ⇔
∗ ∗ −γI Fr2 0 ≺ 0,
∗ ∗ ∗ −λI λH T
∗ ∗ ∗ ∗ −λI
Then the following lemma is obtained to solve the FD problem for systems with unstructured
uncertainties.
4.5 Fault detection for uncertain systems 70
Lemma 4.5.1. Let all variables be as in (4.23) and (4.25) and given that ∆ = {∆ ∈ Rn∆ ×n∆ :
−1 P11 P12
k∆k ≤ 1}. For given γ > 0, Trf Trdu < γ if and only if there exist P = T =PT
∞ P12 P22
and U0 such that M ≺ 0 is satisfied, where
M11 M12 M13 M14 M15
⋆
−γI M23 0 M25
M= ⋆ ⋆ −γI M34 0
⋆ ⋆ ⋆ −λI λH T
⋆ ⋆ ⋆ ⋆ −λI
and
" #
P11 (A1 + Z0 C2 ) − P12 Bf (C1 + U0 C2 ) + ⋆ ⋆
M11 = T ,
(A1 + Z0 C2 )T P12 − (C1 + U0 C2 )T BfT P22 P22 A + ⋆
P11 (Bd,1 +Z0 Dd,2 )+P12 (Bd −Bf (Dd,1 +U0 Dd,2 )) P12 Bu
M12 = T (B ,
P12 d,1 +Z0 Dd,2 )+P22 (Bd −Bf (Dd,1 +U0 Dd,2 )) P22 Bu
(C1 +U0 C2 )T (Dd,1 +U0 Dd,2 )T 0
M13 = , M23 = , M15 = ,
0 0 λEaT
P11 (Fa +Z0 Fc,2 −Bf Fc,1 ) + P12 (Fa − Bf (Fc,1 +U0 Fc,2 ))
M14 = T (F +Z F −B F ) + P (F − B (F +U F )) ,
P12 a 0 c,2 f c,1 22 a f c,1 0 c,2
0
M34 = Fc,1 + U0 Fc,2 , M25 = .
λEbT
It follows that a solution can be obtained by solving a BMI using the global optimaization algo-
rithms given in [40, 99], as well as local numerical searching algorithms that converge much faster
[38, 5]. A related discussion of the solution techniques for BMIs can be found in [71].
However it is possible to obtain a suboptimal solution that requires the solution of a quasi-LMI.
−1
If set P12 = 0, then by pre- and post-multiplying M by T = diag(I, P22 , I), and setting Q1 =
−1
P11 Z0 and Q2 = P22 it gives that
−1 P11 0
Trf Trdu < γ ⇐ ∃λ > 0, ∃P = = P T such that Msub ≺ 0,
∞ 0 P22
where
Msub
11 Msub
12 Msub
13 Msub
14 Msub
15
⋆ −γI M23 M24 M25
sub
M =
⋆ ⋆ −γI 0 0 ,
⋆ ⋆ ⋆ −λI λH T
⋆ ⋆ ⋆ ⋆ −λI
4.5 Fault detection for uncertain systems 71
and where
P11 A1 +Q1 C2 +⋆−(C1 +U0 C2 )T BfT
Msub
11 = ,
⋆ AQ2 + ⋆
P11 Bd,1 +Q1 Dd,2 0 (C1 + U0 C2 )T
Msub
12 = sub
, M13 = ,
Bd −Bf (Dd,1 +U0 Dd,2 ) Bu 0
P11 (Fa −Bf Fc,1 )+Q1 Fc,2 0
Msub
14 = , sub
M15 = .
Fa − Bf (Fc,1 +U0 Fc,2 ) λQ2 EaT
−1
The solution can be retrieved by noting that Z0 = P11 Q1 . Although Msub
15 contains a nonlinear
term, it is possible to get the optimal solution since it is linear for a fixed λ. So that λ ∈ [0, +∞)
is tuned until the minimum value of γ is found. It follows that a class of filters that solves Prob-
lem 4.1.1 is obtained for uncertain models. Z1 and U1 can then be chosen to get a stable filter as
has been done in the case of no model uncertainty.
Remark 4.5.2. Note that the results given in [56, 55] are optimal, however, not suitable for models
with uncertainties since the requirement that the transfer matrix Trf (s) = I is essential, which
cannot be achieved for uncertain system model. 2
When ∆ is structured as in (4.24) we proceed as follows. By using (1.6) from Lemma 1.4.1, we
have
(4.26) ⇔ Tco + F̃ ∆H Ẽ + Ẽ T ∆H T F̃ T ≺ 0
where
Tco F̃ + Ẽ T G Ẽ T S
TSG = + S −1 [ S Ẽ SH ]. (4.29)
F̃ T + GT Ẽ H T G + GT H − S HT S
∗ ∗ ∗ ∗ −S
where Ẽ and F̃ are as defined in (4.27). The following lemma derives a solution for the FD
4.6 Numerical examples 72
Lemma 4.5.2. Let all variables be as in (4.23) and (4.25) and given
that ∆ ∈ ∆. For some γ > 0,
−1 P11 P12
Trf Trdu < γ if and only if there exist P = T = P T and U0 such that L ≺ 0 is
∞ P12 P22
satisfied, where
M11 M12 M13 L14 L15
∗ −γI M23 L24 L25
L= ∗ ∗ −γI L34 0 ≺ 0. (4.30)
∗ ∗ ∗ T T
H G+G H −S H S T
∗ ∗ ∗ ∗ −S
and
P11 (Fa +Z0 Fc,2 −Bf Fc,1 ) + P12 (Fa − Bf (Fc,1 +U0 Fc,2 ))
L14 = T (F +Z F −B F ) + P (F − B (F +U F )) + E T G ,
P12 a 0 c,2 f c,1 22 a f c,1 0 c,2 a
0 0 0
L15 = T , L24 = T , L25 = T , L34 = Fc,1 + U0 Fc,2 .
Ea S Eb G Eb S
2
Similar to the case of unstructured uncertainties, the optimal solution is obtained by solving a BMI
by using the algorithms guaranteeing global convergence given in [40] and [99].
In this section, two examples are considered to illustrate the effectiveness of the proposed FD
scheme. The first one, from the literature, compares our FD filter with other design techniques.
The second example is randomly generated, and illustrates the extensions given in Section 4.3 and
Section 4.4.
Consider example 2 given in [91] and also considered in [87]. Both approaches employed an
H∞ /H∞ FD formulation. The state-space model of the LTI system is described as follows.
−10 0 5 0 0.8 0.04 4
0 −5 0 2.5 −2.4 0.08 4
A Bd Bf 0 0 −2.5 0 1.6 0.08 8
= .
C Dd Df
0 5 0 −3.75 0.8 0.08 −8
1 0 0 1 0.2 0.04 2
1 0 1 1 0.4 0.06 −1
4.6 Numerical examples 73
By following the proposed algorithm, we get γ0 = 0.2392 and a stable A + L0 C. Set Z1 = 0 and
U1 = 2.6I so that
−2.0034 −1.0735
0.7778 5.4963
−4.8810 −1.6664 ,
L= H = 1.0385 −0.5231 .
2.1436 −2.8805
This setting is optimal for the H− /H∞ FDI following Remark 4.3.2. Next, consider two faulty
scenarios for the example
Scenario 1:
As given in [91]: the disturbance is taken as d(t) = [1.2 sin(2t)e−0.1t 1.5 cos(2t)e−0.1t ]T and the
fault f1 is defined as f1 (t) = 0.5, 5s ≤ t ≤ 10s and f1 (t) = 0 elsewhere. Figure 4.1 shows the
residuals and highlights the fact that our residual based on H− /H∞ is more robust to disturbances.
Also, note that the residuals in [91] and [87] are of dimension 2 since in their scheme the dimension
of the residual is the same as the number of outputs.
1.5
our residual
residual in [87]
residual in [91]
1
0.5
Magnitude
−0.5
−1
0 2 4 6 8 10 12 14 16 18 20
Time(secs)
Scenario 2:
Next, consider the same system but this time subject to two white noise disturbances of mean zero
and covariance 7 connected at t = 0. A fault is simulated by a unit positive jump connected at
t = 22s. Figure 4.2 shows the residuals. This example justifies the use of H− /H∞ schemes over
H∞ /H∞ .
5
our residual
4 residual in [87]
residual in [91]
1
Magnitude
−1
−2
−3
−4
−5
0 5 10 15 20 25 30 35
Time(secs)
−30.4438 4.8715 −0.3643 −23.6674 −27.0508
9.8449 −7.1563 6.7607 5.6055 1.1667
A= −0.4460 −2.0473 −5.2344 0.0314
0.5579 ,
23.0620 3.3600 5.1258 19.5067 27.8028
−2.9175 −8.1727 −6.1518 −4.5557 −10.6214
−0.9800 −2.1080 −0.6891 −0.4471 0.3988
C = −0.3647 −0.8006 0.4282 0.2585 1.4254 ,
0.9081 0.5181 −1.1179 0.2686 0.9972
−0.8999 1.7625
0.1478 0.1409
−0.8032 −0.1423
−1.3598 1.1176 , Df = −0.9889 1.4751 ,
Bf =
−1.3020 0.7315 −1.0297 −0.4817
0.8723 −0.4119
0.0880 1.2801
−0.9907 0.2294 0.0905 −0.3783
1.3235 0.2285 , Dd = 0.8423 −1.8179 .
Bd =
1.1952 0.5443 0.4173 1.1086
−0.9682 0.9220
Our algorithm gives γ = 0.8170 and an unstable A + L0 C. It follows that fault isolation is
not possible; however, from section 4.3, the optimal solution and fault isolation at DC are both
achieved by setting
0.9353 0.2914
−0.3267 −0.1976
0.9970 −0.0778
−0.1134 −0.0265 ,
Z1 = U1 = .
−2.3823 −0.9739 0.0778 0.9970
1.1268 0.6113
Suppose that the system is subject to a disturbance, which is a white noise with mean zero and
standard deviation 1. Two faults are simulated by a unit negative and positive jump respectively,
connected from the 12th and 35th second. Figure 4.3 shows the residuals. It is clear that fault
isolation is reached at DC and the impact of disturbances on the residuals has been attenuated.
1.5
r1
r2
0.5
Magnitude
−0.5
−1
−1.5
0 5 10 15 20 25 30 35 40 45 50
Time(secs)
centered around a frequency ω0 = 10 rad/s is used. The optimal value of γ is now γ = 0.5898.
The modified residual responses are also provided in Figure 4.4. The result validates the
employment of a weighting filter. Note that with the design incorporating the weights, it is easier
to determine a residual threshold for fault detection.
4.7 Concluding remarks 77
1.5
r1
r2
0.5
Magnitude
−0.5
−1
−1.5
0 5 10 15 20 25 30 35 40 45 50
Time(secs)
In this chapter, the H− /H∞ FD problem for linear time-invariant systems is considered subject
to faults and disturbances. A new performance index is proposed to capture both performance
requirements, namely, fault detection and disturbance rejection. The disturbance rejection perfor-
mance is measured by the size of the disturbance to residual dynamics where the H∞ norm is used
as the measure. Then, by using a static observer, a class of solutions to a FD problem is given in
the form of a simple LMI with two degrees of freedom.
It is shown that the degrees of freedom in this class of solutions can be used to derive the optimal
filter to the H− /H∞ FD problem. Under certain conditions, further specifications such as fault
isolation can also be achieved. Moreover, a technique is given to improve the results and remove
certain assumptions with the use of weighting filters and it is shown that the proposed scheme can
be extended to uncertain systems, although the resulting solutions are suboptimal.
78
Chapter 5
As mentioned in Chapter 4, many contributions [49, 54, 92, 58, 48] implemented FD approaches to
carry out the sensitivity optimization over the whole frequency range . However, a major drawback
of these schemes is that a restrictive rank assumption on the fault dynamics is needed to ensure a
nonzero level of sensitivity to faults. A second drawback is that, for many systems, the resulting
sensitivity to disturbances is still large compared with the sensitivity to faults.
As a way of ameliorating these problems, more recent FD schemes carry out the sensitivity op-
timization over a finite frequency range within which the rank assumption on the fault dynamics
is satisfied [90, 92, 58]. However, the major difficulty of these schemes is that the solution al-
gorithms are difficult to implement because the optimization problems are nonlinear which may
result in suboptimal solutions.
In principle, for the purpose of fault detection, it is only needed that the sensitivity to disturbances
is small compared with the sensitivity to faults over a small frequency range or even at a single fre-
quency. This chapter considers the observer-based FD problem for linear time invariant systems at
a single frequency and over a finite frequency range. For the FD problem at a single frequency, an
optimal solution is derived such that the FD observer generates a residual signal which minimizes
the sensitivity of the residual to disturbances while maintaining a minimum level of sensitivity to
faults at the given frequency. LMI optimization techniques are used to obtain the optimal solu-
tion. Moreover, we derive a simple parameterization of all optimal solutions which can be used to
achieve other specifications. In particular, it is shown that, when ω = 0, the resulting degrees of
5.1 Fault detection problem at a single frequency 79
freedom allow the FD observer to be used simultaneously for control system design. Conversely,
this work demonstrates that observers used for control system design can be also employed to
implement fault detection schemes at DC with minor modifications.
For the FD problem over a frequency range, an initial research is carried out such that a novel
problem formulation is developed, by which the FD problem is transformed to a sensitivity op-
timization problem over a finite frequency range. Then, a solution is derived in the form of an
BMI optimization using the generalized KYP lemma followed by a linearization procedure. Also,
conditions under which this solution is optimal are also derived. However, the resulting observer
parameters may be complex so that they cannot be directly used in practice.
This chapter is organized as follows. Section 5.1 considers the fault detection problem at a single
frequency. The optimal LMI solution to the single frequency FD problem is given and further
results are derived for the FD problem at DC. Examples are given to demonstrate the effectiveness
of the proposed scheme. In Section 5.2, an initial study is given for the FD problem over a
frequency range. Finally, Section 5.3 summarizes this work.
It is well known that FD at DC is a very important specification which is usually required for
practical systems, therefore this section first gives a general investigation for the FD problem at
a single frequency and then specializes the result to FD at DC. It also shows that, at DC, fur-
ther specifications such as control system design can be achieved while maintaining optimal FD
properties.
Consider a linear time-invariant (LTI) dynamic system subject to disturbances and process, sensor
and actuator faults modeled as
where x(t) ∈ Rn , u(t) ∈ Rnu and y(t) ∈ Rny are the process state, input and output vectors,
respectively, and where f (t) ∈ Rnf and d(t) ∈ Rnd are the fault and disturbance vectors, re-
5.1 Fault detection problem at a single frequency 80
spectively. Here, Bu ∈ Rn×nu and Du ∈ Rny ×nu are input distribution matrices, respectively.
Bf ∈ Rn×nf and Df ∈ Rny ×nf are the component and instrument fault distribution matrices,
respectively, while Bd ∈ Rn×nd and Dd ∈ Rny ×nd are the corresponding disturbance distribution
matrices [35]. Taking Laplace transforms, the system input/output behavior can be described by
where
nu nf nd
s A Bu Bf Bd
ny [ Gu (s) Gf (s) Gd (s) ] = (5.2)
C Du Df Dd
are the system process, disturbance and fault transfer matrices, respectively.
In this chapter, the residual signal generator is given using a static observer of the form
˙
x̂(t) = Ax̂(t) + Bu u(t) − L (y(t) − C x̂(t) − Du u(t)) ,
where x̂(t) ∈ Rn is the observer state and r(t) ∈ Rnf is the residual signal. Here L ∈ Rn×ny and
H ∈ Rnf ×ny are the observer and residual gain matrices, respectively, and are to be determined.
Define the state estimation error signal as e(t) = x(t) − x̂(t). It follows that the residual dynamics
are given by
Taking Laplace transforms, r(s) = Trf (s)f (s) + Trd (s)d(s), where
s A + LC Bf + LDf Bd + LDd
Trf (s) Trd (s) =
HC HDf HDd
= F (s) Gf (s) Gd (s) (5.4)
are the transfer matrices from disturbances and faults to the residual, respectively, and where
s A + LC L
F (s) = (5.5)
HC H
Problem 5.1.1. Let the system and filter dynamics be as defined in (5.1) and (5.5) and let ω > 0
be given. Assume that:
Find a stable filter F (s), of the form given in (5.5) such that the optimum
is achieved.
Remark 5.1.1. Assumption A1 is needed to guarantee there exists at least one L such that A+LC
is stable. 2
Remark 5.1.2. In the case that ny < nf , a bank of observers may be used such that for each
observer ny ≥ nf is satisfied. In particular, in the case that nf > 1, the banks of observers can
also be used to isolate faults, namely, each observer is dedicated to detect and isolate one of the
faults while the other faults are regarded as disturbances for that observer. See [1, 34, 12] for
more details. 2
Remark 5.1.3. The KYP lemma [47] shows that given system matrices (A, B, C, D), then
kG(jω)k < γ for a single frequency ω if and only if there exist Hermitian matrices P , Q ∈ C n×n
such that Q ≻ 0
T
−A QA−jωQA+jωATQ−ω 2 Q+P A+ATP ⋆ ⋆
−B TQA+B TP +jωB TQ −B TQB −γI ⋆ ≺ 0 (5.7)
C D −γI
By rearranging the (1,1) entry, the LMI in (5.7) can be equivalently expressed as
−(−jωI−A)TQ(jωI−A)−(−jωI−A)TP −P (jω−A) ⋆ ⋆
B TQ(jωI−A)+B TP −B TQB −γI ⋆ ≺ 0 (5.8)
C D −γI
Take a congruence transformation by post and pre multiplying T and T ∗ to (5.8), where
(jωI −A)−1 (jωI −A)−1B 0
T = 0 I 0 ,
0 0 I
5.1 Fault detection problem at a single frequency 82
it is obtained that
−Q−(−jωI−A)−TP −P (jωI−A)−1 ⋆ ⋆
−B T(−jωI−A)−TP −γI ⋆ ≺ 0. (5.9)
C(jωI −A) −1 −1
D+C(jωI −A) B −γI
Since both P and Q are free parameters, by following a Schur complement [6] it is easy to see
that
−γI GT (−jω) −γI DT +B T(−jωI −A)−TC T
= −1 ≺0 (5.10)
G(jω) −γI D+C(jωI −A) B −γI
is necessary and sufficient to guarantee (5.7) is satisfied. Hence, in the following sections only
(5.10) is used to achieve the infimum of the norm of a transfer matrix rather than by using the
GKYP lemma since it gives relatively easy formulation and avoids unnecessary computation of
the variables P and Q. 2
This section derives equivalent problem formulations to Problem 5.1.1 which are used to remove
or relax constraints and to give a simple LMI solution to the FD filter design at a single fre-
quency. First, an equivalent formulation to Problem 5.1.1 is given for the purpose of removing
n ×ny
the constraint on the stability of F (s). It shows that if there exists an F (s) ∈ RL∞f that
achieves the minimum of the cost function, it is always possible to find a left coprime factoriza-
n ×nf n ×ny
tion F (s) = (M (s))−1 F̃ (s) such that M (s) ∈ RH∞f is inner and F̃ (s) ∈ RH∞f .
(a) Suppose that the pair (A + LC, HC) is detectable. Then it follows from Theorem 13.34
in [115] that there exists a left coprime factorization F (s) = (M (s))−1 F̃ (s) such that
n ×nf n ×n
M (s) ∈ RH∞f is inner and F̃ (s) ∈ RH∞f y where
s A + LC + L0 HC L0 L + L0 H
M (s) F̃ (s) =
HC Inf H
A + L̃C L0 L̃
=
HC Inf H
where
L̃ = L + L0 H, L0 = −XC T H T (5.12)
and where X = X T ∈ Rn×n is the stabilizing solution of the algebraic Riccati equation
Then kF̃ (jω)Gd (jω)k = kM (jω)F (jω)Gd (jω)k = γ and σ(F̃ (jω)Gf (jω)) =
σ(M (jω)F (jω)Gf (jω)) ≥ 1 since M (s) is inner. Thus γo = γ.
(b) Suppose that the pair (A + LC, HC) is not detectable. Then, introducing a similarity trans-
formation on the data if necessary,
s A + LC L
F (s) =
HC H
A11 + L1 C1 A12 + L1 C2 L1
= 0 A22 + L2 C2 L2 (5.14)
0 HC2 H
where A11 + L1 C1 is antistable (that is, all its poles are in the open right half-plane), the
pair (A22 + L2 C2 , HC2 ) is detectable, HC1 = 0 and A21 + L2 C1 = 0 and where the
partitioning of A and C is induced by the partitioning of the realization of F (s) in (5.14).
Next we prove that the pair (A11 , C1 ) is detectable. Since we assume that the pair (A, C) is
detectable, it follows that
A11 − sI A12
A21 A22 − sI
C1 C2
has full column rank for all s such that s + s∗ > 0. Using the fact that A21 = −L2 C1 , it
follows that
A11 − sI A12 I 0 0 A11 − sI A12
0 A22 + L2 C2 − sI = 0 I L2 A21 A22 − sI
C1 C2 0 0 I C1 C2
5.1 Fault detection problem at a single frequency 84
has full column rank for all s such that s + s∗ > 0. Thus the pair (A11 , C1 ) is
detectable. It follows that there exists L̃1 such that A11 + L̃1 C1 is stable. Define
s
F̃ (s) = (A + L̃C, L̃, HC, H) where L̃ = [ L̃T1 LT2 ]T . Then
A11 + L̃1 C1 A12 + L̃1 C2 L̃1
s
F̃ (s) = 0 A22 + L2 C2 L2
0 HC2 H
The next result shows that we may restrict F (s) such that F (jω) is the set of all left–inverses of
Gf (jω).
s
Proof. Note first that γo ≤ γ since σ(Inf ) ≥ 1. Let F̃ (s) = (A + L̃C, L̃, H̃C, H̃) be a minimizer
for (5.15) and define Q(s) = F̃ (s)Gf (s) so that σ(Q(jω)) ≥ 1 and kF̃ (jω)Gd (jω)k = γo .
n ×nf n ×nf
Then Q−1 (s) ∈ RL∞f and kQ−1 (jω)k ≤ 1. Let F (s) = (Q(s))−1 F̃ (s) ∈ RL∞f .
s
Then a state–space manipulation verifies that F (s) = (A + LC, L, HC, H) where L = L̃ −
(Bf + L̃Df )(H̃Df )−1 H̃ and H = (H̃Df )−1 H̃. Furthermore, F (jω)Gf (jω) = Inf . Now,
kF (jω)Gd (jω)k = kQ−1 (jω)F̃ (jω)Gd (jω)k ≤ kF̃ (jω)Gd (jω)k = γ0 since kQ−1 (jω)k ≤ 1.
This proves that γ ≤ γ0 and since γ ≥ γo it follows that γo = γ. 2
Remark 5.1.4. As can be concluded from Lemma 5.1.2 that the optimal cost function is obtained
by achieving the minimum of kF (jω)Gd (jω)k where F (s) is such that F (jω) is the set of all left
inverses of Gf (jω). Hence, the constraint on F (jω) can be expressed as an equality constraint
such that F (jω)Gf (jω) = Inf . 2
n ×ny
Remark 5.1.5. Note that the equality constraint F (jω)G(jω) = I where F (s) ∈ RL∞f is the
condition to ensure fault isolation at ω, which constructs an identity transfer matrix from faults to
residuals. It means that if the obtained F (s) is stable, optimal fault detection filter also achieves
5.1 Fault detection problem at a single frequency 85
fault isolation. In the case that F (s) is not stable, by using the techniques in Lemma 5.1.1, it is
always possible to find an F̃ (s) that achieves optimal fault detection, however F̃ (jω)Gf (jω) 6=
Inf but is unitary. 2
As stated in Lemma 5.1.2, the FD problem is reduced to the solution to the optimization prob-
lem in (5.15). In this section, a parameterization of all L and H satisfying the equality constraint
F (jω)Gf (jω) = Inf is derived. Then, Problem 5.1.1 is solved by using LMI optimization tech-
niques in which stability constraints are not imposed.
Firstly, let all variables and assumptions be as given in Problem 5.1.1 and define
(jωI − A − LC)−1
−1
= −(A − jωI)−1 I + LC(A − jωI)−1
= −(A − jωI)−1 I − L(I + C(A − jωI)−1 L)−1 C(A − jωI)−1
= −(A − jωI)−1 I − (I + LC(A − jωI)−1 )−1 LC(A − jωI)−1
Next, let F (s) and Gf (s) be defined as in (5.4) and (5.5). By using (5.16) and (5.17),
F (jω)Gf (jω) can be expressed as
Since it is assumed that Gf (jω) has full column rank, there exist G†f ∈ C nf ×ny and G⊥
f ∈
C (ny −nf )×ny which can be obtained from a singular value decomposition of Gf (jω), such that
" # " #!
G†f Inf G†f
Gf (jω) = , rank = ny . (5.19)
G⊥f
0 G⊥
f
5.1 Fault detection problem at a single frequency 86
It follows that all H(I + CY ) satisfying the constraint F (jω)Gf (jω) = Inf can be parameterized
by
H(I + CY ) = G†f + SG⊥
f (5.20)
where S ∈ C nf ×(ny −nf ) is a free parameter and S = S1 + jS2 where S1 , S2 ∈ Rnf ×(ny −nf ) .
The next theorem gives an LMI solution to the optimization problem proposed in Lemma 5.1.2 by
rearranging the variables as given in (5.16) and (5.17).
where
−γI ḠTf1 + G̃Tf1 S1T − G̃Tf2 S2T
W1 = ,
Ḡf 1 + S1 G̃f 1 − S2 G̃f 2 −γI
0 −ḠTf2 − G̃Tf2 S1T − G̃Tf1 S2T
W2 = ,
Ḡf 2 + S1 G̃f 2 + S2 G̃f 1 0
Ḡf 1 and Ḡf 2 are the real and imaginary part of G†f (Dd + C(jωI − A)−1 Bd ), respectively while
G̃f 1 and G̃f 2 are the real and imaginary part of G⊥ −1
f (Dd + C(jωI − A) Bd ), respectively.
where F (jω)Gd (jω) = HDd + HC(jωI − A − LC)−1 (Bd + LDd ). By using (5.16) and (5.17),
the nonlinear matrix inequality in (5.22) can be written as
−γI GTd (−jω)(I + CY )∗ H T
≺0 (5.23)
H(I + CY )Gd (jω) −γI
To ensure F (jω)Gf (jω) = Inf , the parameterization given in (5.20) is used in the matrix inequal-
ity to give h i∗
−γI G∗d (jω) G†f + SG⊥
f
h i ≺0 (5.24)
G†f + SG⊥
f Gd (jω) −γI
It can be seen that (5.24) is linear but complex. To numerically solve this LMI problem, it is
5.1 Fault detection problem at a single frequency 87
necessary to convert the complex LMI to a real LMI with larger dimensions by noting that if
W = W1 + jW2 where W1 and W2 are real, then
W1 W2
W ≺0⇔ ≺ 0. (5.25)
−W2 W1
2
Since S can be obtained by solving the LMI optimazation problem in (5.21), then by following
(5.20) the relation between L and H can be expressed as
Then
It follows that L should be chosen such that the imaginary part of H is zero. Hence, L should
satisfy
(Z2 CA + ωZ1 C)(A2 + ω 2 I)−1 L = EL = −Z2 (5.29)
By taking a singular value decomposition of E, it follows that all L satisfying (5.29) is of the form
L = −E † Z2 + E ⊥ K (5.30)
where K ∈ R(n−nf )×ny is free. E † ∈ Rn×nf and E ⊥ ∈ Rn×(n−nf ) are obtained from a singular
value decomposition of E such that
E E † E ⊥ = I nf 0 , rank E † E ⊥ = n. (5.31)
5.1 Fault detection problem at a single frequency 88
To summarize the results obtained above, a procedure is given to illustrate the steps to solve
Problem (5.1.1).
n ×nd n ×nf
Algorithm 5.1.1. Let Gd (s) ∈ RL∞y and Gf (s) ∈ RL∞y be as given in (5.2) and let
ω > 0 be given. Assume that ny ≥ nf , (A, C) is detectable and that Gf (jω) has full rank.
5. Parameterize L as shown in (5.30) and choose K, say K = 0, find L from (5.30) and
compute H from (5.32).
(b) Else, if the pair (A + LC, HC) is detectable, let X = X T ∈ Rn×n be the stabilizing
solution of the algebraic Riccati equation (5.13) and redefine L := L − XC T H T H.
(c) Else, redefine L using the procedure in part (2) of the proof of Lemma 5.1.1 so that the
pair (A + LC, HC) is detectable and go to Step 5(b).
s n ×ny
6. The optimal FD filter is given by F (s) = (A + LC, L, HC, H) ∈ RH∞f .
Remark 5.1.6. Since L can be parameterized as shown in (5.30), it follows that there remains
the degrees of freedom in K to ensure other design specifications. For example, to achieve fault
isolation requires the stability of A + LC (see Remark 5.1.5). In other words, it is required to find
a K such that A + LC = A − E † Z2 C + E ⊥ KC is stable. However, this is equivalent to the
problem of stabilization using static output feedback, which is known to be a difficult problem. See
[83] for a recent technique for solving this problem. 2
5.1 Fault detection problem at a single frequency 89
5.1.4 FD problem at DC
This section considers the special case when ω = 0, that is, we consider the FD problem at
DC. In order to exploit the available degrees of freedom, kF Gd k∞ is minimized rather than
kF (0)Gd (0)k.
Problem 5.1.2. Let the variables and assumptions as defined in Problem (5.1.1). Find a stable
filter F (s) such that the optimum
γo = inf kF Gd k∞ , (5.33)
F (0)Gf (0)=Inf
nf ×ny
F (s)∈H∞
Next, a parameterization of L and H is given such that the equality constraint in Problem 5.1.2 is
always satisfied. Let Y = (A + LC)−1 L, then (A + LC)−1 = A−1 − Y CA−1 . By setting w = 0
and following similar lines of the derivations in Section 5.1.3, it follows that
" # " #!
G†f I nf G†f
Gf (0) = , rank = ny , (5.35)
G⊥
f
0 G⊥
f
H̃ can be parameterized such that all H̃ satisfying (5.34) are of the form
H̃ = H(I − CY ) = G†f + S0 G⊥
f (5.36)
Note that S0 ∈ Rnf ×(ny −nf ) is a free parameter. Following from the application of a congruence
transformation to the bounded real lemma [6], the theorem below gives an LMI solution to the
optimization problem in (5.33) without imposing stability constraint.
where
P A−1 + A−TP − RCA−1 − A−T C T RT ⋆ ⋆
W0 = −BdT A−T P − Gd T (0)RT −γI ⋆ (5.38)
(G†f + S0 G⊥
f )CA
−1 † ⊥
(Gf + S0 Gf )Gd (0) −γI
Proof. Using the Bounded Real Lemma [6], kF Gd k∞ < γ if and only if there exists a symmetric
P such that
P (A + LC) + (A + LC)T P P (B + LDd ) (HC)T
(B +LDd )T P −γI (HDd )T ≺ 0 (5.39)
HC HDd −γI
Take a congruence transformation by post and pre multiplying T and its transpose to (5.39), where
(A + LC)−1 −(A + LC)−1 (Bd + LDd ) 0
T = 0 I 0
0 0 I
This gives
(A + LC)−T P + P (A + LC)−1 ⋆ ⋆
−(Bd + LDd )T (A + LC)−T P −γI ⋆ ≺ 0 (5.40)
HC(A + LC) −1 −1
H[Dd − C(A + LC) (Bd + LDd )] −γI
Redefine (A + LC)−1 L = Y and (A + LC)−1 = A−1 − Y CA−1 and apply the parameterization
given in (5.36), then the nonlinear matrix inequality in (5.40) is equivalent to the LMI in (5.37).2
Remark 5.1.7. As noted, to achieve optimal fault detection, the stability condition P ≻ 0 is
not imposed as in [52]. This implies that the design variable L in (5.37) may not be stabilizing.
However, by following similar procedures as given in Lemma 5.1.1, the filter can be stabilized
without losing any optimality. 2
Remark 5.1.8. For a special case of Theorem 5.1.1, namely, at DC, the variable S can be solved
from the LMI optimization in (5.21) by setting ω = 0. However, by following (5.20), only the
product H(I − C(A + LC)−1 L) should be fixed for achieving the optimal solution. Hence, the
filter parameter L can be freely chosen firstly and H can be then obtained from L. It follows
that the observer parameter L can be used to implement other control specifications such as
LQG control or Kalman filter [115] first, then with little extra effort, an H from (5.26) can be
implemented and optimal fault detection at DC can also be achieved. 2
5.1 Fault detection problem at a single frequency 91
To demonstrate the effectiveness of the proposed fault detection filter scheme in a real system, a
jet engine example is considered in this section. The state-space model of the GE-21 jet engine
[48, 67] is given as
−3.370 1.636 0.586 −1.419 1.252
A = , Bu = ,
−0.325 −1.896 0.410 1.118 0.139
1 0 0 0 0
C = 0 1 , Du = 0 0 0 .
0.731 0.786 0.267 −0.025 −0.146
The system has n = 2 and ny = nu = 3. Suppose that this system is subject to two potential
faults and two disturbances (nf = 2 and nd = 2). The distribution matrix of faults are given by
1 0
0.586 −1.419
Bf = , Df = 0 1 ,
0.410 1.118
0 0
Since the example in [67] does not involve disturbances, Bd and Dd are randomly generated as
follows:
0.0865
0.1177
−0.2973 0.0467
Bd = , Dd = 0.3331 −0.4366 .
0.0094 −0.1814
−0.0251 0.0273
−0.5860 1.4190 13.1302 1 0 −0.0040
L= , H= .
−0.4100 −1.1180 −11.3298 0 1 10.3635
Note that since A + LC is stable, then F (jω)Gf (jω) = Inf and the given filter achieves fault
isolation at w = 1.
Suppose that disturbance 1 is a band-limited white noise with power 0.2 (zeroth-order-hold
with sampling time 0.1s) and disturbance 2 is a constant biases of amplitude 1 occurred from 2
5.1 Fault detection problem at a single frequency 92
second. The two faults are assumed to be f1 = 0.5sin(t) and f2 = sin(t) connected from the
10th and 15th seconds, respectively. Figure 1 gives the residual responds before the filter, while
Figure 2 shows the filtered residual responds. The example demonstrates that the designed filter
gives satisfactory fault detection and isolation performance and also shows its robustness against
disturbances.
0.8
0.6
0.4
0.2
Magnitude
−0.2
−0.4
−0.6
−0.8
0 5 10 15 20 25 30 35 40
Time(secs)
1.5
r1
r2
1
0.5
Magnitude
−0.5
−1
−1.5
0 5 10 15 20 25 30 35 40 45 50
Time(secs)
This section gives an initial research on the FD problem over a frequency range. By applying the
result in Chapter 4, a novel problem formulation is proposed, which transforms this problem to an
optimization problem which is to, over a frequency range, minimize the sensitivity to disturbances
with respect to the sensitivity to faults. Then by using a generalized GKYP lemma, a solution is
derived in the form of an LMI optimization followed by a linearization procedure.
Consider a linear time-invariant dynamic system subject to disturbances and faults as given in
(4.1) and suppose the residual generator using a static observer be of the form in (4.2). For G(s) ∈
m×p
RH∞ , define
kGkω = sup σ̄ (G(jωl )) . (5.41)
|ωl |<ω
Problem 5.2.1. Let all variables and assumptions be as given in Problem 4.1.2. For a given
5.2 Fault detection problem over a frequency range 94
−1
Trf Trd < γo . (5.42)
ω
where
−1 s A1 + Z0 C2 Bd,1 + Z0 Dd,2
Trf (s)Trd (s) = , (5.43)
C1 + U 0 C2 Dd,1 + U0 Dd,2
Remark 5.2.1. The major difference of this work from Problem 4.1.2 is that the cost function
(5.42) which combines both fault detection and disturbance attenuation requirements is minimized
only over a finite frequency range such that the ’ratio’ between these two objectives is optimized
over a limited frequency domain rather than over the whole frequency range. 2
The previous section proposes a single cost function which captures both fault detection and dis-
turbance attenuation, then the FD problem is transformed to a minimization problem using the
norm as a measure as defined in (5.41). In this section, an LMI solution to the problem is given by
using the following generalized KYP lemma [47].
(jωI −A1 −Z0 C2 )−1 = (jωI −A1 )−1 +Yω C2 (jωI −A1 )−1
(jωI −A1 −Z0 C2 )−1 (Bd,1 +Z0 Dd,2 ) = (jωI −A1 )−1Bd,1+Yω Dd,2 +C2 (jωI −A1 )−1Bd,1
(5.44)
(C1 + U0 C2 )(jωI −A1 −Z0 C2 )−1 = C1 (jωI − A1 )−1 + H̃C2 (jωI − A1 )−1 ,
To simply the notations, in the following we take (jωI − A1 )−1 = Ar1 + jAi1 where Ar1 , Ai1 ∈
Rn×n . The next result shows that the FD problem over a frequency range can be solved by using
the following form of the generalized KYP lemma [47].
5.2 Fault detection problem over a frequency range 95
Lemma 5.2.1. Given matrices (A, B, C, D), for ∀ |ωl | < ω, the following finite frequency domain
inequality
D + C(jωl I − A)−1 B < γo (5.46)
holds if and only if there exist hermitian matrices Q ≻ 0 and P such that the following LMI
−AT QA + P A + AT P + ω 2 Q ⋆ ⋆
−B T QA + B T P −B T QB − γo I ⋆ ≺0 (5.47)
C D −γo I
2
The next theorem gives a solution to Problem 5.2.1 using Lemma 5.2.1 followed by a series of
linearizations.
where
r
M11 ⋆ ⋆ i
M11 −M21i T −M31i T
M r = M21
r ⋆ , M i = M21 i T .
−γo I i 0 −M32
r r
M31 M32 −γo I i i
M31 M32 0
and where
r X2 + X2T
M11 = − (X1 Ar1 − X2 Ai1 ) − (X1 Ar1 − X2 Ai1 )T − (Z1 C2 Ar1 − Z2 C2 Ai1 )
2ω
−(Z1 C2 Ar1 − Z2 C2 Ai1 )T ,
r
M21 = −(X1 Ar1 Bd,1 − X2 Ai1 Bd,1 )T − [Z1 (Dd,2 + C2 Ar1 Bd,1 ) − Z2 C2 Ai1 Bd,1 ]T
r
M31 = H̃1 C2 Ar1 − H̃2 C2 Ai1 + C2 Ar1 ,
r
M32 = H̃1 (Dd,2 + C2 Ar1 Bd,1 ) − H̃2 C2 Ai1 Bd,1 + (Dd,2 + C2 Ar1 Bd,1 ),
i X1T − X1
M11 = − (X2 Ar1 + X1 Ai1 ) + (X2 Ar1 + X1 Ai1 )T − (Z1 C2 Ai1 + Z2 C2 Ar1 ),
2ω
+(Z1 C2 Ai1 + Z2 C2 Ar1 )T ,
i
M21 = (X2 Ar1 Bd,1 + X1 Ai1 Bd,1 )T + [Z2 (Dd,2 + C2 Ar1 Bd,1 ) + Z1 C2 Ai1 Bd,1 ]T ,
i
M31 = H̃1 C2 Ai1 + H̃2 C2 Ar1 + C2 Ai1 ,
i
M32 = H̃2 (Dd,2 + C2 Ar1 Bd,1 ) + H̃1 C2 Ai1 Bd,1 + C2 Ai1 Bd,1 .
5.2 Fault detection problem over a frequency range 96
Proof. Firstly, by rearranging the entries, the LMI in (5.47), applied to the realisation in (5.43),
can be equivalently expressed as
−(−jωI−A)TQ(jωI−A)−(−jωI−A)T(P −jωQ)−(P +jωQ)(jωI−A) ⋆ ⋆
B TQ(jωI−A)+B T(P − jωQ) −B TQB −γI ⋆ ≺ 0
C D −γI
Apply a congruence transformation by post- and pre-multiplying (5.39) by T and its Hermitian
where
(jωI − A)−1 (jωI − A)−1 B 0
T = 0 I 0
0 0 I
Then,
−Q−(−jωI−AT )−1 (P −jωQ)−(P +jωQ)(jωI−A)−1 ⋆ ⋆
B T(jωI −AT )−1 (P + jωQ) −γI ⋆ ≺ 0
C(jωI − A) −1 −1
D + C(jωI − A) B −γI
By substituting the state space model formulated in (5.43) into (5.49), a nonlinear matrix inequality
is obtained as follows.
M11 ⋆ ⋆
M = M21 −γI ⋆ ≺ 0, (5.49)
M31 M32 −γI
where
X −X ∗
M11 = −X(jωI−A1 )−1−Z0 C2 (jωI −A1 )−1−(−jωI−AT1 )−1X ∗−(−jωI −AT1 )−1C2TZ0∗ ,
2ωj
T
T
M21 = Bd,1 (−jωI − AT1 )−1 X ∗ + Dd,2 T
+ Bd,1 (−jωI − AT1 )−1 C2T Z0∗ ,
As can be seen from (5.49), the LMI is complex. To numerically solve this LMI problem, it is
necessary to transform the complex LMI to a real LMI but with larger dimensions as stated in
(5.25). Take M = M r + jM i where M r and M i are real, then
" #
T
Mr Mi
M ≺0⇔ ≺ 0.
Mi Mr
2
By solving the LMI in (5.48), it is obtained that X = X1 +jX2 , Z = Z1 +jZ2 and H̃ = H̃1 +j H̃2
so that Y = X −1 Z. By using (5.44), it follows that
So that by following the arrangements defined in (4.14) and (4.15), the observer parameters L and
H satisfying (5.48) are calculated as follows.
It can be seen that L and H from (5.51) may be complex since Z0 and U0 calculated from (5.50)
may not be real so that the observer is not practically implementable. Take
Although there are remain degrees of freedom in Z1 ∈ Rn×nf and U1 ∈ Rnf ×nf in (5.51) which
can be chosen to achieve
Z0i + Z1 U0i = 0, U1 U0i = 0 (5.53)
to make the design parameters L and H real, however, U1 = HDf means that U1 should always
have full rank, which implies there may not always exist a full rank U1 such that (5.53) is satisfied.
Therefore, the optimal pair (L, H) from (5.51) could be complex and the resulting filter is not
applicable. Further research is needed to find a real and optimal pair (L, H) or to find a systematic
algorithm to get a suboptimal solution to guarantee (L, H) can be always real.
In this chapter, the observer-based FD problem for linear time invariant systems at a single fre-
quency and over a finite frequency range are investigated, respectively. For the FD problem at a
single frequency, an optimal solution is derived such that at a given frequency, the residual signal
minimizes the sensitivity of the residual to disturbances while maintaining a minimum level of
5.3 Concluding remarks 98
sensitivity to faults. LMI optimization techniques are used to obtain the optimal solution. More-
over, a simple parameterization of all optimal solutions is derived such that other specifications
can also be achieved by choosing the degrees of freedom remaining in the parameterization. In
particular, it is shown that, when ω = 0, the resulting degrees of freedom allow the FD observer
to be used simultaneously for control system design. Conversely, our work demonstrates that ob-
servers used for control system design can be also employed to implement fault detection schemes
at DC with minor modifications.
Moreover, an initial investigation is carried out on the FD problem over a finite frequency range
whose solution turns out to involve a difficult nonlinear optimization. By applying the General-
ized KYP lemma and using the results given in Chapter 4, a novel problem formulation is pro-
posed such that the FD problem is transformed to a sensitivity optimization problem over a finite
frequency range, where the ratio of the sensitivity of disturbances over the sensitivity of faults is
minimized. By following a linearization procedure, a state space solution is obtained by solving an
LMI optimization and conditions under which this solution is optimal are also derived. However,
this problem is still not fully resolved since the resulting observer parameters could be complex
and may not be usable in practice. Further work is needed to find an algorithm to obtain real and
optimal solutions.
99
Chapter 6
Conclusion
6.1 Contributions
This thesis presents research work on the problem of model-based FDI for linear dynamic systems
subject to faults, disturbances and model uncertainties. It contains three different approaches
to solve the FDI problem: an on-line moving horizon estimation procedure using measurement
analysis, a robust static observer design for a mixed H− /H∞ FD problem and a FD procedure for
a finite frequency range etc.
• First, in Chapter 3 a novel on-line approach to the robust FDI problem is proposed for
linear discrete-time systems subject to faults, disturbances and norm-bounded uncertainties.
The upper and lower bounds on the fault signal are computed at each sampling instant.
Then a fault is regarded as having occurred when its upper bound is smaller than zero or
lower bound is larger than zero. LMI optimization techniques are employed to obtain the
bounds. Furthermore, a subsequent-state-estimation technique together with an estimation
horizon update procedure is proposed, which allows the on-line fault detection and isolation
process to be repeated in a moving horizon. Also, it is shown that the proposed approach
reduces to the solution of linear programs when no uncertainties are considered in the model
description. The details of this approach have been published in [105, 106].
• Chapter 4 gives an optimal state-space solution to the H− /H∞ FD problem for linear time-
invariant systems subject to faults, disturbances and model uncertainties. First, a novel
performance index is developed to capture both the requirements of fault detection and
disturbance rejection which is particularly suitable for handling model uncertainties. Then,
6.2 Future Research 100
by using a static observer, a class of solutions to the FD problem is given in the form of
a simple LMI with two degrees of freedom. These degrees of freedom are then used to
derive the optimal filter to the H− /H∞ FD problem and achieve further specifications such
as fault isolation. Moreover, it is shown that by introducing a weighting filter, the results
are further improved and certain widely held assumptions can be removed. Finally, the
proposed approach is extended to implement FD for uncertain systems. The details of this
approach have been published in [108].
Moreover, a further study and result related to the on-line FDI approach in Chapter 3 are given
at the end of this chapter. It provides a new idea of reconfiguring system models and estimating
model uncertainties by using only input and output measurements. This result has already been
applied to the area of networked system control as stated in [116].
The approach in Chapter 3 achieves FDI by computing upper and lower bounds on faults over
a finite estimation horizon for linear discrete-time systems with model uncertainties. However,
one main drawback is that solving LMIs over a finite estimation horizon of a certain length could
be unnecessary and computationally expensive, which may limit its application to fast systems.
Hence, it would be useful to propose an algorithm to choose a flexible horizon length to save
6.2 Future Research 101
The approach proposed in Chapter 4 is based on static observer design. In fact, in the literature
there is another kind of observer , that is, the dynamic observer which has already been used to
achieve FDI [61]. Generally, a dynamic observer can be described as the following form:
x̃(t)
˙x̃(t) Ak Bku Bky
= u(t) , (6.1)
r(t) Ck Dku Dky
y(t)
where x̂(t) ∈ Rnk is the observer state and r(t) ∈ Rnf is the residual signal. Figure 6.1 shows a
dynamic observer in the residual generation scheme. By defining the following augmented state
T
z(t) = x(t)T x̃(t)T ,
where Trf (s), Trd (s) and Tru (s) denote the dynamics from faults, disturbances and inputs to the
residual, respectively. It follows that
Since dynamic observer-based FDI schemes have more degrees of freedom in the observer param-
eters, it may be possible to obtain better performance for fault detection and disturbance attenu-
ation. Therefore, it is of interest to use a dynamic observer and derive an optimal solution to the
FD problem. Also, it would be useful to carry out a comparison of FD performance between the
static and dynamic observer-based approaches.
Chapter 3 gives a novel idea for FDI by evaluating the fault signal bounds while most contributions
in the literature usually achieve FD by generating the residuals. Also, there are also approaches
from the model control literature [85] concerning state constraints which are also considered in
Chapter 3. Hence, it is needed and of interest to compare the on-line FDI results proposed in
Chapter 3 with the results by using other FD approaches from the literature. Further improvement
on the proposed FDI approach may be achieved from the comparison.
FD problems over a finite frequency range have received increasing attention from the literature
[90, 92, 58] since they can relax the restrictive rank assumptions on fault dynamics and improve
the FD performance. However, many contributions give sub-optimal solutions and are difficult
to implement due to the involved nonlinear optimization. In Chapter 5, an initial research is
presented for this problem over a low frequency range |ω| < ωl where a solution is given in the
form of LMIs and conditions under which the solution is optimal are also derived. However, the
resulting parameters may be complex and not usable in practice. Hence, it is of interest to find
a systematic algorithm to obtain real and optimal solutions to the problem. Moreover, it would
6.2 Future Research 103
also be useful to investigate the FD problem over the middle (ω1 < |ω| < ω2 ) and high frequency
ranges (|ω| > ωh ).
In practice, system model uncertainties can be inevitably introduced by process noises, parameter
variations and non-linear dynamics. A precise uncertainty estimation and description is impor-
tant for designing control systems and achieving other specifications such as FD. Recently, much
research effort has been devoted to the uncertainty estimation problem for linear systems. See
[51, 74, 41, 4] for details.
In this section, a novel approach is proposed to solve the uncertainty estimation problem for linear
discrete-time systems subject to bounded additive disturbances and norm-bounded uncertainties.
Similar to the FDI scheme in given Chapter 3, this approach achieves uncertainty estimation by
computing upper and lower bounds on the entries of the uncertainty matrix using a dynamic system
model as well as input/output measurements over a finite horizon.
Consider an LTI discrete-time system subject to disturbances and uncertainties of the form
xk
xk+1 A + ∆A Bd + ∆Bd Bu + ∆Bu
= dk (6.4)
yk Cy +∆Cy Dyd +∆Dyd Dyu +∆Dyu
uk
where Bp , Dyp , Cq , Dqd , and Dqu are known and ∆ is an n∆ × n∆ unknown matrix with ∆ ∈ ∆
where
In this description δ 0i and δ̄i0 represent initial bounds on the uncertainty. In the following the
problem of refining the lower and upper bounds for this simple uncertainty set is presented while
the problem for more complex uncertainty descriptions is left for future research.
6.2 Future Research 104
Note that ∆ has l repeated diagonal blocks. Assume that upper and lower bounds x̄0 , x0 and d¯k ,
dk on the initial state and disturbances, respectively, are available and define
x0 x0 x̄
w= ,w = , w̄ = ¯0 , Dyw = Cy0 Dyd , Nw = n + Nd .
d d d
Similarly to (3.10), the uncertainty data are re-sequenced so that the uncertainty set has the form:
Using an iterative computation, the system algebraic formulation over the estimation horizon N is
described as:
where w ≤ w ≤ w̄ and
∆Dyw ∆Dyu = Dyp ∆e (I − Dqp ∆e )−1 Dqw Dqu , ∆e ∈ ∆ e . (6.7)
Then, the uncertainty estimation problem for linear discrete-time systems is described as follows.
Problem 6.2.1. Let all data and variables be as given above. Then for i = 1, 2, . . . , l find
Note that the interval [ϕi , ϕ̄i ] is our best estimate for δi given all the information about the system.
The following theorem shows that evaluating an upper bound on each eTi δ can be obtained by
solving an LMI problem.
6.2 Future Research 105
Theorem 6.2.1. Let all data be as defined previously. Let Dm denote the set of all real m × m
diagonal matrices and define
where
Li11 ⋆ ⋆ ⋆
Li21 Li22 ⋆ ⋆
Li (δ̄i , K̄i , Ēi , D̄i , τ̄i , Q̄i , Ḡi ) =
Li31 Li32 Li33 ⋆ ,
(6.11)
Li41 Li42 Li43 Li44
and
1 T
Li11 = D̄i − τ̄i Dyw
T D T i i
yw − Dqw S̄i Dqw , L21 = 2 K̄i Dqw , L22 = Ēi ,
1 T
T D
Li31 = −τ̄i Dyp T i
yw − Dqp S̄i Dqw − Ḡi Dqw , L32 = 2 Dqp K̄i ,
It follows that the problem formulation for obtaining the upper bound on δi in (6.8) can be rewritten
6.2 Future Research 106
as
Define Q := {Q : Q∆e = ∆e Q, ∀∆e ∈ ∆e } which is the set of matrices which commute with all
matrices in ∆e . Note that since q = Dyw w + Dyu u + Dyp p then Q̄i p = ∆e Q̄i q for any Q̄i ∈ Q.
According to the specific structure of ∆e given in (3.10), it can be verified that
Note that the constraints in (6.12) imply that pT Q̄Ti Q̄i p ≤ q T Q̄Ti Q̄i q, ∀Q̄i ∈ Q. Also, note that Ḡi
is skew-symmetric with the structure as defined above, then a computation shows that pT Ḡi q = 0.
T
eTi δ − δ̄i = − (w̄ − w)T D̄i (w − w) − δ̄ 0 − δ Ēi δ − δ 0
+pT Q̄Ti Q̄i p − (Dqw w + Dqu u + Dqp p)T Q̄Ti Q̄i (Dqw w + Dqu u + Dqp p)
−pT Ḡi (Dqw w + Dqu u + Dqp p) + (Dqw w + Dqu u + Dqp p)T Ḡi p
is valid for all δ̄i ∈ R , K̄i ∈ RN∆ ×l , D̄i ∈ RNw ×Nw , Ēi ∈ Rl×l , τ̄i ∈ R, Q̄i ∈ RN∆ ×N∆
and Ḡi ∈ RN∆ ×N∆ . From (6.13), it can be seen that δ̄i is an upper bound on eTi δ if there
exist 0 D̄i ∈ DNw , 0 Ēi ∈ Dl , K̄i ∈ K, τ̄i ∈ R, Q̄i ∈ Q and Ḡi ∈ G such that
Li (δ̄i , K̄i , R̄i , D̄i , τ̄i , Q̄i , Ḡi ) 0. An inspection of the sets Q and S verifies that Q̄i ∈ Q if and
only if S̄i := Q̄Ti Q̄i ∈ S and so an upper bound on eTi δ is given by solving the LMI optimization
in (6.10). 2
Similarly, the next theorem derives lower bounds on the faults. The proof is omitted since it
follows similar lines to that of Theorem 6.2.1.
6.2 Future Research 107
Theorem 6.2.2. Let all data and variables as defined in Theorem 6.2.1. Then for i = 1, . . . , nl ,
Example
T
δ= 0.1286 −0.5332
T
δ= 0.1363 −0.5327
T
δ̄ = 0.1440 −0.5204
It shows that the difference between the upper and lower bounds on δ are small, which means δ is
well-estimated.
108
Bibliography
[1] E. Alcorta-Garcia and P.M. Frank. Analysis of a class of dedicated observer schemes to
sensor fault isolation. In UKACC International Conference on Control ’96, pages 60–65,
1996.
[2] K.J. Astrom and B. Wittenmark. Computer-Controlled Systems: Theory and Design, Sec-
ond ed. Prentice-Hall, Inc., 1990.
[3] R.V. Beard. Failure accommodation in linear systems through self–reorganization. PhD
thesis, Massachusetts Institute of Technology, 1971.
[4] G. Belforte, B. Bona, and V. Cerone. Parameter estimation algorithms for a set-membership
description of uncertainty. Automatica, 26:887–898, 1990.
[5] E. Beran and K. Grigoriadis. A combined alternating projections and semidefinite program-
ming algorithm for low-order control design. In In Proc. Prep. 13th IFAC World Congr.,
pages 85–90, San Francisco, CA, 1996.
[6] S.P. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan. Linear Matrix Inequalities in
Systems and Control Theory. SIAM Publications, Philadelphia, PA, 1994.
[7] F.M. Callier and C.A. Desoer. Linear System Theory. Springer–Verlag, New York, 1991.
[8] J. Chen and R. J. Patton. Robust model–based fault diagnosis for dynamic systems. Boston:
Kluwer Academic Publishers, 1999.
[9] J. Chen, R. J. Patton, and H.Y. Zhang. Design of unknown input observers and robust fault
detection filters. Int. J. Control, 63(1):85–105, 1996.
[10] E. Y. Chow and A. S. Willsky. Analytical redundancy and the design of robust failure
detection systems. IEEE Trans. Automatic Control, 29(7):603–614, 1984.
[11] R.N. Clark. Instrument fault detection. IEEE trans. aero. and electronic Systems, 14(3),
1978.
[12] R.N. Clark, D.C. Fosth, and V.M. Walton. Detecting instrument malfunctions in control
systems. IEEE Trans. on Aerosp. and Elec. Sys., AES–11(4):465–473, 1975.
[13] D.J. Clements, B.D.O. Anderson, A.J. Laub, and J.B. Matson. Spectral factorization with
imaginary–axis zeros. Lin. Alg. Appl., 250:225–252, 1997.
[14] Tramone D. Curry and Emmanuel G Collins. Robust fault detection and isolation using
robust l1 estimation. Journal of Guidance, Control and Dynamics, 28(6):1131–1139, 2005.
BIBLIOGRAPHY 109
[15] C. De Persis and A. Isidori. A geometric approach to nonlinear fault detection and isolation.
IEEE Trans. Automatic Control, 46(6):853–865, Jun. 2001.
[16] S. X. Ding, T. Jeinsch, P.M. Frank, and E. L. Ding. A unified approach to the optimization
of fault detection systems. Int. J. Adapt. Control Signal Process, 14:725–745, 2000.
[18] X. Ding and P.M. Frank. An approach to robust residual generation and evaluation. In Proc.
IEEE Conf. Dec. & Control, pages 656–661, Brighton, England, 1990.
[19] X. Ding and P.M. Frank. Fault detection via factorization approach. Systems & Control
Letts., 14:431–436, 1990.
[20] X. Ding, P.M. Frank, and L. Guo. An approach to residual generator and evaluator design
and synthesis. In Proc. of IFAC World Congress, Sydney, Australia, 1993.
[21] X. Ding, L. Guo, and P. Shi. A characterization of parity space and its application to robust
fault detection. IEEE Trans. Automatic Control, 44(2):337–343, 1999.
[22] J.C. Doyle. Structured uncertainty in control system design. IEE Proceedings, 129-
D(6):242–250, 1982.
[23] J.C. Doyle. Lecture notes in advances in multivariable control. ONR/Honeywell Workshop,
1984.
[24] J.C. Doyle, K. Glover, P. Khargonekar, and B. Francis. State space solutions to standard
H2 and H∞ control problems. IEEE Trans. Automatic Control, 24(8):731–747, 1989.
[25] A. Emami-Naeini, M.M. Akhter, and S.M. Rock. Effect of model uncertainty on failure
detection - the threshold selector. IEEE Trans. Automatic Control, 33(2):1106–1115, 1988.
[26] Y.E. Faitakis and J.C. Kantor. Residual generation and fault detection for discrete–time
systems using l∞ technique. Int. J. Control, pages 155–174, 1996.
[27] H. Fang, H. Ye, and M. Zhong. Fault diagnosis of networked control systems. Annual
Reviews in Control, 31:55–68, 2007.
[28] D. Filbert and K. Metzger. Quality test of systems by parameter estimation. In 9th IMEKO
congress, Berlin, 1982.
[29] M. Fliess, C. Join, and H. Sira-Ramirez. Robust residual generation for linear fault diagno-
sis: an algebraic setting with examples. Int. J. Control, 77(14):1223–1242, 2004.
[30] K. F. Fong and A. P. Loh. A frequency domain approach for fault detection. In Proc. IEEE
Conf. Dec. & Control, pages 1023–1028, Orlando, Florida, USA, 2001.
[32] P. M. Frank. Fault diagnosis in dynamic systems using analytical and knowledge–based
redundancy: A survey and some new results. Automatica, 26(3):459–474, 1990.
BIBLIOGRAPHY 110
[33] P. M. Frank. Analytical and qualitative model-based fault diagnosis - a survey and some
new results. European Journal of Control, 2(1):6–28, 1996.
[34] P. M. Frank and X. Ding. Frequency domain approach to optimally robust residual genera-
tion and evaluation for model–based fault diagnosis. Automatica, 30(5):789–804, 1994.
[35] P. M. Frank and X. Ding. Survey of robust residual generation and evaluation methods in
observer–based fault detection systems. J. Proc. Cont., 7(6):403–424, 1997.
[36] E. Frisk and L. Nielsen. Robust residual generation for diagnosis including a reference
model for residual behavior. In Proceedings of the 14th IFAC World Congress, pages 55–
60, Beijing, China, 1999.
[37] E. Frisk and L. Nielsen. Robust residual generation for diagnosis including a reference
model for residual behavior. Automatica, 42:437–445, 2006.
[38] J.C. Geromel, P.L.D. Peres, and S.R. de Souza. Output feedback stabilization of uncertain
systems through a min/max problem. In Preprints of Papers, 12th World Congress, I.F.A.C.,
July, 1993, Volume 6, pages 35–38. The Inst. of Engineers, Australia, 1993.
[39] L. El Ghaoui and J.P. Folcher. Multi-objective robust control of LTI systems subject to
unstructured perturbations. Systems & Control Letts., 28(1):23–30, 1996.
[43] D.M. Himmelblau. Fault diagnosis in chemical and petrochemical processes. Elsevier,
1978.
[44] M. Hou and R.J. Patton. An LMI approach to H− /H∞ fault detection observers. In UKACC
International Conference on Control. ’96, pages 305–310, England, 1996.
[45] R. Isermann. Supervision, fault detection and fault diagnosis methods: an introlduction.
Control engineering practice, 5(5):639–652, 1997.
[46] R. Isermann and P. Balle. Trends in the application of model-based fault detection and
diagnosis of technical processes. Control engineering practice, 5(5):709–719, 1997.
[47] T. Iwasaki and S. Hara. Generalized kyp lemma: unified frequency domain inequalities
with design applications. IEEE Trans. Automatic Control, 50(1):41–59, 2005.
[48] I.M. Jaimoukha, Z. Li, and E. Mazars. Filter design to the optimal weighted H− /H∞ fault
detection problem. In UKACC International Conference on Control. ’06, Glasgow, UK,
Aug 2006.
BIBLIOGRAPHY 111
[49] I.M. Jaimoukha, Z. Li, and V. Papakos. A matrix factorization solution to the H− /H∞ fault
detection problem. Automatica, 42(11):1907–1913, 2006.
[50] M. Kinnaert. Design of redundancy relations for failure detection and isolation by con-
strained optimization. Int. J. Control, 63(3):609–622, 1996.
[51] R. Kosut. Adaptive robust control via transfer function uncertainty estimation. In Proc.
Amer. Control Conf., pages 349–354, 1988.
[52] A. Laszlo. A generalization of the bounded real lemma. Lin. Alg. Appl., 372:79–103, 2003.
[53] X. Li and K. Zhou. A time domain approach to robust fault detection of linear time-varying
systems. Automatica, 45:94–102, 2009.
[54] Z. Li and I. M. Jaimoukha. Observer-based fault detection and isolation filter design for
linear time-invariant systems. Int. J. Control, 82:171–182, 2009.
[55] Z. Li, I. M. Jaimoukha, and E. Mazars. State space solution to the H− /H∞ fault detection
problem. In Proc. IEEE Conf. Dec. & Control, San Diego, CA, USA, Dec 2006.
[56] Z. Li, E. Mazars, and I. M. Jaimoukha. State space solution to the H− /H∞ fault detection
problem. International Journal of Robust and Nonlinear Control, 2010.
[57] Z. Li, E. Mazars, Z. Zhang, and I. Jaimoukha. Statespace solution to the H− /H∞ fault-
detection problem. Int. J. Robust & Nonlinear Control, 2011. accepted.
[58] J. Liu, J. L. Wang, and G. H. Yang. An LMI approach to minimum sensitivity analysis with
application to fault detection. Automatica, 41:1995–2004, 2005.
[59] N. Liu and K. Zhou. Optimal robust fault detection for linear discrete time systems. Journal
of control science and engineering, 2008(829459), 2008.
[60] D.Q. Mayne, J.B. Rawlings, C.V. Rao, and P.O.M. Scokaert. Constrained model predictive
control: stability and optimality. Automatica, 36:789–814, 2000.
[61] E. Mazars, I.M. Jaimoukha, and Z. Li. Computation of a reference model for robust fault
detection and isolation residual generation. Journal of control science and engineering,
2008:1–12, 2008.
[62] R.J. Patton. Fault detection and diagnosis in aerospace systems using analytical redundancy.
Computing and Control Engineering, 2(3):127–136, 1991.
[63] R.J. Patton and J. Chen. A review of parity space approaches to fault diagnosis for aerospace
systems. AIAA journal of guidance, control and dynamics, 17(2):278–285, 1994.
[64] R.J. Patton, P.M. Frank, and R.N. Clark (Eds). Fault diagnosis in dynamic systems, theory
and applications. Prentice Hall, 1989.
[65] R.J. Patton, P.M. Frank, and R.N. Clark (eds). Issues of Fault Diagnosis for Dynamic
Systems. Springer, 2000.
[66] R.J. Patton and S. M. Kangethe. Robust fault diagnosis using the model–based approach.
In IEE Colloquium on Condition Monitoring and Fault Diagnosis – Part 1, pages 2/1–2/13,
1988.
BIBLIOGRAPHY 112
[67] M. L. Rank and H. Niemann. Norm based design of fault detectors. Int. J. Control,
72(9):773–783, 1999.
[69] J.B. Rawlings and D.Q. Mayne. Model predictive control: Theory and Design. Nob Hill
Publishing, Madison, 2009.
[70] A. Ray and R. Luck. An introduction to sensor signal validation in redundanct measurement
systems. IEEE Control Systems Magazine, 11(2):44–49, 1991.
[71] A. Saberi, P. Sannuti, and B. M. Chen. h2 optimal control system and control engineering.
London, UK, 1995.
[72] M.A. Sadrnia, J. Chen, and R.J. Patton. Robust fault diagnosis observer design using H∞
optimisation and µ synthesis. IEE Colloquium on Modelling and Signal Processing for
Fault Diagnosis, pages 9/1–9/16, 1996.
[73] D. Sauter and F. Hamelin. Frequency-domain optimization for robust fault detection and
isolation in dynamic systems. IEEE Trans. Automatic Control, 44(4):878–882, 1999.
[74] A. V. Savkin and I. R. Petersen. Robust state estimation and model validation for discrete-
time uncertain systems with a deterministic description of noise and uncertainty. Automat-
ica, 34:333–356, 1989.
[75] C. Scherer. The Riccati inequality and state–space H∞ optimal control. PhD thesis, Univ.
Würzburg, Germany, 1990.
[76] C. Scherer, P. Gahinet, and M. Chilali. Multi-objective output-feedback control via LMI
optimization. IEEE Trans. Automatic Control, 42(7):896–911, 1997.
[77] D. Van Schrick. Investigation of reliability for instrument fault detection state-estimator
scheme. European Journal of diagnosis and Safety in Automation, 1(1):63–78, 1991.
[78] D. Van Schrick. Remarks on terminology in the field of supervision, fault detection and
diagnosis. In Proceedings of IFAC Symposium on Dault Detection, Supervision and Safety
for Technical Processes, pages 959–964, SAFEPROCESS’97, Pergamon, UK, 1997.
[79] M. L. Schwall and J. C. Gerdes. A probabilistic approach to residual processing for vehicle
fault detection. In Proc. Amer. Control Conf., pages 2552–2557, 2002.
[80] L. Shen, S. Chang, and P. Hsu. Robust fault detection and isoaltion with unstructured
uncertainty using eigenstructure assignment. Journal of guidance control and dynamics,
21(1):50–57, 1998.
[81] L. Shen and P. Hsu. Robust design of fault isolation observers. Automatica, 34(11):1421–
1429, 1998.
[82] H. Siebert and R. Isermann. Fault diagnosis via on-line correlation analysis. Technical
report, 1976.
BIBLIOGRAPHY 113
[83] J. Stefanovski and G. Dimirovski. A new approach to static output feedback stabilization of
linear dynamic systems. International Journal of Systems Science, 37(9):643–662, 2006.
[84] R. J. Stern and H. Woldkowicz. Indefinite trust region subproblems and nonsymmetric
eigenvalue perturbations. SIAM J. Optimization, 5:286–313, 1995.
[85] D. Sui, T. Johansen, and L. Feng. Linear moving horizon estimation with pre-estimating
observer. IEEE Trans. Automatic Control, 55(10):2363–2368, 2010.
[86] K. P. Sun and A. Packard. Robust H2 and H∞ filters for uncertain lft systems. IEEE Trans.
Automatic Control, 50(5):715–720, 2005.
[87] F. Tao and Q. Zhao. Fault detection observer design with unknown inputs. In Proceeding
of the conference on control application, pages 1275–1280, Toronto, Canada, 2005.
[88] M. Vidyasagar. Control Systems Synthesis: A Factorization Approach. MIT Press, Cam-
bridge, MA, 1985.
[89] N. Viswanadham, J.H. Taylor, and E.C. Luce. A frequency domain approach to failure
detection and isolation with application. Control–theory and Advanced Technology, 3:45–
72, 1987.
[90] H. Wang and G. H. Yang. A finite frequency domain approach to fault detection for linear
discrete-time systems. Int. J. Control, 81:1162–1171, 2008.
[91] H. B. Wang, L. Lam, S. X. Ding, and M. Y. Zhong. Iterative linear matrix inequality
algorithms for fault detection with unknown inputs. Journal of Systems and Control Engi-
neering, 219(2):161–172, August 2005.
[92] H.B. Wang, J.L. Wang, and J. Lam. Robust fault detection observer design: iterative LMI
approach. Journal of Dynamic Systems, Measurements and Control, 129(1):77–82, 2007.
[93] J. L. Wang, G. H. Yang, and J. Liu. An LMI approach to H− index and mixed H− /H∞
fault detection observer design. Automatica, 43:1656–1665, 2007.
[94] Y. Q. Wang, S. X. Ding, Y. Hao, and G. Z. Wang. A new fault detection scheme for
networked control systems subject to uncertain time-varying delays. IEEE Trans. Signal
Processing, 56(10):5258–5268, 2008.
[95] Y. Q. Wang, H. Ye, S. X. Ding, Y. Cheng, P. Zhang, and G. Z. Wang. Fault detection of
networked control systems with limited communication. Int. J. Control, 82(7):1344–1356,
2009.
[96] Wikipedia. http://en.wikipedia.org/wiki/Chernobyl disaster.
[97] Wikipedia. http://en.wikipedia.org/wiki/Ariane 5 Flight501.
[98] J. Wunnenberg. Observer-based fault detection in dynamic systems. PhD thesis, University
of Duisburg, 1990.
[99] Y. Yamada, S. Hara, and H. Fujioka. Global optimization for constantly scaled h∞ control
problem. In Proc. American Contr. Conf, pages 427–430, 1995.
[100] H. Ye, G. Wang, and S. X. Ding. A new parity space approach for fault detection based on
stationary wavelet transform. IEEE Trans. Automatic Control, 49(2):281–287, 2004.
BIBLIOGRAPHY 114
[101] P. Zhang and S. X. Ding. On fault detection in linear discrete-time, periodic, and sampled-
data systems. Journal of control science and engineering, 2008(849546), 2008.
[102] P. Zhang, S. X. Ding, G. Z. Wang, and D. H. Zhou. Fault detection of linear discrete-time
periodic systems. IEEE Trans. Automatic Control, 50(2):239–244, 2005.
[103] P. Zhang, H. Ye, S. X. Ding, G. Z. Wang, and D. H. Zhou. On the relationship between
parity space and h2 approaches to fault detection. Systems & Control Letts., 55:94–100,
2006.
[104] X. Zhang, M. Polycarpou, and T. Parisini. A robust detection and isolation scheme
for abrupt and incipient faults in nonlinear systems. IEEE Trans. Automatic Control,
47(4):576–593, 2002.
[105] Z. Zhang and I.M. Jaimoukha. Fault detection and isolation for linear discrete-time systems
using input/output measurement anaysis. In Proc. IEEE Conf. Dec. & Control, pages 4908–
4913, Shanghai, China, 2009.
[106] Z. Zhang and I.M. Jaimoukha. A measurement analysis approach to online fault detec-
tion and isolation for linear discrete-time uncertain systems. In Proc. IEEE Conf. Dec. &
Control, Atlanta, USA, 2010.
[107] Z. Zhang and I.M. Jaimoukha. Optimal state space solution to the fault detection problem
at dc. In Conference on Control and Fault Tolerant Systems, Nice, France, 2010.
[108] Z. Zhang and I.M. Jaimoukha. An optimal solution to an H− /H∞ fault detection problem.
In Proc. IEEE Conf. Dec. & Control, Orlando, USA, 2011.
[109] Z. Zhang and I.M. Jaimoukha. Optimal state space solution to the fault detection problem
at single frequency. In Proc. of IFAC World Congress, Milan, Italy, 2011.
[110] M. Zhong, Q. Ding, and P. Shi. Parity space-based fault detection for markovian jump
systems. International Journal of Systems Science, 40(4):421–428, 2009.
[111] M. Zhong, S.X. Ding, J. Lam, and H. Wang. An LMI approach to design robust fault
detection filter for uncertain LTI systems. Automatica, 39(3):543–550, 2003.
[112] M. Zhong, H. Ye, S. X. Ding, and G. Wang. Observer-based fast rate fault detection for a
class of multirate sampled-data systems. IEEE Trans. Automatic Control, 52(3):520–525,
2007.
[113] D.H. Zhou and P.M. Frank. Fault diagnostics and fault tolerant control. IEEE Transactions
on Aerospace and Electronic Systems, 34(2):420–427, 1998.
[114] K. Zhou and J. Chen. Performance bounds for coprime factor controller reductions. Systems
& Control Letts., 26(2):119–127, 1995.
[115] K. Zhou, J.C. Doyle, and K. Glover. Robust and Optimal Control. Prentice-Hall Inc., NJ,
1996.
[116] Z.Zhang and Y. Yuan. Decentralized consensus value computation over network with un-
certainties. IEEE Trans. Automatic Control, 2011. submitted.