!"#$%&'())
)*+*,-).**
// -000 1 - 0 2" !2& // 34%4*3424 -
#"45*##62
,(-000
7 /889 8$# $$#:81-0 "#$% "6$4%#$
;/889 8$# $$#:81-0 "#$% "6$4%#$<
1 )
/88 88/8""64:8
7(/!&
7(=/ /(>)
(!&8/( /
))//(/)/(
))/(*)*/?//
/ / / ) /
(/87/ 1
(( @((
))/)/( 1)/(
/88 8/
1(A)?/)
// 18)
/ ) / /= ! /
(+ &
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS
Abstract—With the ever increase of complexity and expense of
industrial systems, there is less tolerance for performance
degradation, productivity decrease and safety hazards, which
greatly stimulates to detect and identify any kinds of potential
abnormalities and faults as early as possible, and implement real-
time fault-tolerant operation for minimizing performance
degradation and avoiding dangerous situations. During the last
four decades, fruitful results were reported about fault diagnosis
and fault-tolerant control methods and their applications in a
variety of engineering systems. The three-part survey paper aims
to give a comprehensive review for real-time fault diagnosis and
fault tolerant control with particular attention on the results
reported in the last decade. In the first-part review, fault
diagnosis approaches and their applications are reviewed
comprehensively from model-based and signal-based
perspectives, respectively.
Index Terms—Analytical redundancy, model-based fault
diagnosis, signal-based fault diagnosis, real-time monitoring,
fault tolerance
I. INTRODUCTION
S is known, many engineering systems, such as aero
engines, vehicle dynamics, chemical processes,
manufacturing systems, power network, electric machines,
wind energy conversion systems, and industrial electronic
equipment and so forth, are safety-critical systems. There is an
ever increasing demand on reliability and safety of industrial
systems subjected to potential process abnormalities and
component faults. As a result, it is paramount to detect and
Manuscript received November 13, 2014; revised January 24, March 8,
2015; accepted March 19, 2015
Copyright © 2015 IEEE. Personal use of this material is permitted.
However, permission to use this material for any other purposes must be
obtained from the IEEE by sending a request to pubs-permission@ieee.org.
Z. Gao is with the Faculty of Engineering and Environment, University of
Northumbria at Newcastle, Newcastle upon Tyne, NE1 8ST, United Kingdom
(Tel: +441912437832; Fax: +441912274397; e-mail:
zhiwei.gao@northumbria.ac.uk).
C. Cecati is with the Department Information Engineering, Computer
Science and Mathematics, University of L’Aquila, 67100 L’Aquila, Italy (e-
mail: c.cecati@ieee.org ).
S. X. Ding is with the Institute of Automatic Control and Complex
Systems, University of Duisburg-Essen, 47057 Duisburg, Germany (e-mail:
steven.ding@uni-due.de).
identify any kinds of potential abnormalities and faults as
early as possible, and implement fault-tolerant operation for
minimizing performance degradation and avoiding dangerous
situations.
A fault is defined as an unpermitted deviation of at least one
characteristic property or parameter of the system from the
acceptable/usual/standard condition [1]. Examples of such
malfunctions are the blocking of an actuator, the loss of a
sensor (e.g., a sensor gets stuck at a particular value or has a
variation in the sensor scalar factor), or the disconnection of a
system component. Therefore, the faults are often classified as
actuator faults, sensor faults and plant faults (or called
component faults or parameter faults), which either interrupt
the control action from the controller on the plant, or produce
substantial measurement errors, or directly change the
dynamic input/output properties of the system, leading to
system performance degradation, and even the damage and
collapse of the whole system. In order to improve the
reliability of a system concerned, fault diagnosis is usually
employed to monitor, locate and identify the faults by using
the concept of redundancy, either hardware redundancy or
software redundancy (or called analytical redundancy). The
basic idea of the hardware redundancy is to use identical
components with the same input signal so that the duplicated
output signals can be compared leading to diagnostic decision
by a variety of methods such as limit checking, and majority
voting etc. The hardware redundancy is reliable, but expensive
and increasing weights and occupying more space. It is
necessary for key components to equip with the redundant
duplicate, but would not be applicable if the hardware
redundancy is applied to the whole system due to the cost or
the difficulty for physical installing when the space and/or
weight are strictly constrained. With the mature of modern
control theory, the analytical redundancy technique has
become the main stream of the fault diagnosis research since
the 1980s, whose schematic diagram can be depicted by Fig.1.
For a controlled system subjected to actuator fault
process/component fault
and sensor fault
, the input and
output are used to construct a fault diagnosis algorithm,
which is employed to check the consistency of the feature
information of the real-time process carried by the input and
output data against the pre-knowledge on a healthy system,
and a diagnostic decision is then made by using diagnostic
A Survey of Fault Diagnosis and Fault-Tolerant
Techniques Part I: Fault Diagnosis with Model-
Based and Signal-Based Approaches
Zhiwei Gao, Senior Member, IEEE, Carlo Cecati, Fellow IEEE, and Steven X. Ding
A
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS
logics. Compared with hardware redundancy methods,
analytical redundancy diagnostic methods are more cost
effective, but more challenging due to environmental noises,
inevitable modelling error, and the complexity of the system
dynamics and control structure.
Fig. 1. Analytical redundancy based fault diagnosis
Fault diagnosis includes three tasks, that is, fault detection,
fault isolation and fault identification. Fault detection is the
most basic task of the fault diagnosis, which is used to check
whether there is malfunction or fault in the system and
determine the time when the fault occurs. Furthermore, fault
isolation is to determine the location of the faulty component,
and fault identification is to determine the type, shape and size
of the fault. Clearly, the locations of the faulty components
and their severe degrees of the malfunctions described by the
types, shapes and sizes of the faults are vital for the system to
take fault-tolerant responses timely and appropriately to
remove the adverse effects from the faulty parts to the system
normal operation.
Fig. 2. Schematic diagram of fault tolerant control
The schematic of fault-tolerant control is depicted by Fig. 2,
which is shown that fault tolerant control is integrated with
fault diagnosis in essence. Real-time fault diagnosis can detect
whether the system is faulty, and tell where the fault occurs
and how severe the malfunction is. Based on the valuable
information, the supervision system can thus take appropriate
fault-tolerant actions such as off-setting the faulty signals by
actuator/sensor signal compensation, tuning or reconfiguring
the controller, and even replacing faulty components by
redundant duplicates, so that the adverse effects from faults
are accommodated or removed.
During the last four decades, fruitful results have been
reported on fault diagnosis methods, fault-tolerant control
techniques and their applications in various industrial
processes and systems. A number of survey papers were
written, for instance [2-38], which are depicted by Table 1,
categorized in terms of years and methods/applications.
TABLE 1
SURVEY PAPERS FOR FAULT DIAGNOSIS AND FAULT
TOLERANCE
Specifically, in 1976, Willsky presented the key concepts of
analytical redundancy for model-based fault detection and
diagnosis in the early survey paper [2]. More comprehensive
model-based fault diagnosis methods such as parity space
approaches, observer-based methods and parameter estimation
techniques are reviewed by [3-9]. A three-part survey paper
[10-12] on fault diagnosis was presented in 2003, respectively
from the viewpoint of quantitative mode-based methods,
qualitative model-based methods and process history based
methods. In [13], a structured and comprehensive overview of
the research on anomaly detection is provided, which is
referred to the problem of finding patterns in data that do not
conform to expected behavior, and has an extensive use in a
wide variety of applications such as intrusion detection for
cyber-security, military surveillance for enemy activities, as
well as fault detection in safety critical systems. In [14-16],
comprehensive fault diagnostic methods were reviewed
respectively from the data-driven perspective. In [17], a short
review on fault detection in sensor networks was provided.
With respect to fault diagnosis methods for various
processes/systems applications, a couple of survey papers
were addressed for mining equipment [18], electric motors
[19-21], building systems (such as heating, ventilating, air-
conditioning and refrigeration) [22, 23], machinery system
[24, 25], and swarm systems (consisting of multiple intelligent
interconnected nodes and possessing swarm capability) [26],
respectively.
For fault-tolerant control, an early review paper was
presented by [27] in 1991, which introduced the basic
concepts of fault-tolerant control and analyzed the
applicability of artificial intelligence (e.g., neural network and
expert systems) to fault-tolerant control systems. In 1997, an
overview of fault-tolerant control was given from the system
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS
development view [28]. In the same year, a comprehensive
review was contributed by [29], which presented the key
issues of the fault-tolerant control systems and outlined the
state of the art in this field. Reconfigurable fault-tolerant
control systems are reviewed extensively respectively by [30-
32]. Some results on fault-tolerant control for nonlinear
systems were reviewed by [33]. Along with fault diagnosis,
brief reviews on data-driven fault-tolerant control and model-
based fault-tolerant reconfiguration were presented by [34, 35]
respectively. From the viewpoint of industrial applications,
fault tolerance techniques were reviewed for electric drive
systems [36] and power electronics systems [37, 38],
respectively.
The three-part survey paper aims to give a comprehensive
overview for real-time fault diagnosis and fault tolerant
control with particular attention on the results reported in the
last decade. Generally, fault diagnosis methods can be
categorized into model-based methods, signal-based methods,
knowledge-based methods, hybrid methods (the combination
methods of at least two methods) and active fault diagnosis
methods. In the first-part survey paper, fault diagnosis
techniques will be reviewed from the model-based and signal-
based perspectives, and the knowledge-based, hybrid, and
active fault diagnosis techniques will be reviewed in the
second-part survey paper. The first-part and second-part
survey papers aim to review the existing fault diagnosis
methods and applications within a framework by using the up-
to-date references.
The rest of this paper is organized as follows. Following the
introduction session, model-based fault diagnosis techniques
are reviewed in Section II. Signal-based fault diagnosis is
reviewed in Section III. The paper is ended by Section IV with
conclusions.
II. MODEL-BASED FAULT DIAGNOSIS METHODS
Model-based fault diagnosis was originated by Beard [39]
in 1971 in order to replace hardware redundancy by analytical
redundancy, and comprehensive results were documented in
some well-written books (i.e., see [40, 41]). In model-based
methods, the models of the industrial processes or practical
systems are required to be available, which can be obtained by
using either physical principles or systems identification
techniques. Based on the model, fault diagnosis algorithms are
developed to monitor the consistency between the measured
outputs of the practical systems and the model predicted
outputs. In this section, model-based fault diagnosis methods
are reviewed following the four categories: deterministic fault
diagnosis methods, stochastic fault diagnosis methods, fault
diagnosis for discrete-events and hybrid systems, and fault
diagnosis for networked and distributed systems, which are
classified in terms of types of the models used.
A. Deterministic fault diagnosis methods
Observer plays a key role in model-based fault diagnosis for
monitored systems/processes characterized by deterministic
models. The schematic diagram of the observer-based fault
diagnosis is depicted by Fig. 3, which includes fault detection,
fault isolation and fault identification (or called fault
reconstruction or fault estimation).
Fig. 3. Scheme of model-based fault diagnosis
For simplicity, the model of the process in Figure 3 is
assumed to be linearized state-space model, which is described
by the following:
(1)
where
and
are the
system state, control input, measured output, unexpected
actuator fault, component/parameter fault, sensor fault,
process disturbance and measurement noises, respectively.
and
are known parameter matrices,
and and are unknown modelling parameter errors.
An observer-based fault detection filter is given in the
following form:
(2)
where (k) and (k) are the estimates of the state and output,
respectively; is the residual signal and is the observer
gain to be designed. Let
the frequency-
domain residual signal can be described by
(3)
where
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS
.
It is indicated from (3) that the residual signal is subjected
to both fault signals and disturbance signals (including
modelling errors, process disturbances and measurement
noises). In order that the residual signal is sensitive to faults,
but robustness against disturbances, the observer gain can be
designed by solving an optimization problem over a specific
frequency range:
(4)
In order to solve (4), the parametric eigenstructure
assignment approach for fault diagnosis was initialized by [42]
and further revisited in [43], in which the observer gain is
formulated as the function of the eigenvalues and
eigenvectors, therefore seeking an optimal is transformed to
the problem of finding optimal eigenvalues and eigenvectors.
Recently, eigenstructure assignment based fault diagnosis
approaches have been applied to vehicles [44], gas turbine
engines [45], spacecraft [46] and wind turbine systems [47].
Alternatively, the multi-object optimization problem described
by (4) can be reformulated by linear matrix inequality (LMI),
which has been a popular method for fault diagnosis research
and applications owing to its wide applicability to a variety of
dynamic systems. Recent development of the LMI-based fault
diagnosis can be found for various systems such as Lipschitz
nonlinear systems [48], TS fuzzy nonlinear systems [49, 50],
time-delay systems [51], switching systems [52], and
application to structure damage detection [53], and shaft crack
detection [54] etc.
A bank of observer-based residuals is generally required in
order to realize fault isolation. A nature idea is to make a
single residual is sensitive to the fault concerned but
robustness against other faults, disturbances and modelling
errors, which is called structure residual fault isolation [4].
Alternative fault isolation logic is to make each residual signal
sensitive to all but one fault, and robustness against modelling
errors and disturbances, which is called generalized residual
fault isolation [5]. Recent results on robust fault isolation are
developed for nonlinear systems [55,56], and various
applications such as for aircraft engine [57], robotic
manipulators [58], and lithium-ion batteries [59]. The
unknown input observer, proposed by [60], is another fault
isolation tool by decoupling input disturbance, modelling
errors and other faults in the corresponding residuals.
Recently, the unknown input observer based fault isolation
techniques are extended to nonlinear systems [61, 62] and
applied to aircraft systems [63], inductor motors [64] and
waste water treatment plant [65].
Fault identification (or called fault reconstruction/fault
estimation) is to determine the type, size and shape of the fault
concerned, which is vital information for fault tolerant
operation. Advanced observer techniques such as proportional
and integral (PI) observers [66, 67], proportional multiple-
integral (PMI) observers [68-70], adaptive observer [71-73],
sliding mode observers [74, 75], and descriptor observers [76,
77] are usually utilized for fault estimation/reconstruction.
The essence of the advanced observers is to construct an
augmented system by introducing the concerned fault as an
additional state and the extended state vector is thereafter
estimated, leading to the estimates of the concerned fault
signal together with original system states. Therefore, the
advanced observers are also called simultaneous state and
fault observers. The above advanced observer techniques are
in an advantage position either for reconstructing slow-varying
additive faults (PI, and PMI observers), slow-varying
parameter faults (adaptive observers), actuator faults with
sinusoidal waveforms (sliding mode observers), and high-
frequency sensor faults (descriptor system approaches).
Actually the above observer techniques may be integrated or
combined in order to solve engineering-oriented problems. For
instance in [78], integral observer, sliding observers and
adaptive observers are combined to reconstruct sensor faults
for satellite control systems. In [79], PI observer and
descriptor observer techniques are integrated to estimate
parameter faults for an aero engine system.
Another well-known model-based fault diagnosis is parity
relation approach, which was developed in the early of 1980s
[80, 81]. The parity relation approach is to generate residuals
(parity vector) which is employed to check the consistency
between the model and process outputs. The parity relation
approach can be applied to either time-domain state-space
model or frequency-domain input-output model, which is well
revisited by the books [40,41,82]. Recently, the parity relation
method is extended for fault diagnosis for more complex
models such as TS fuzzy nonlinear systems [83] and fuzzy
tree models [84], and applied to various industrial systems
such as aircraft control surface actuators [85] and
electromechanical brake systems [86].
Stable factorization approach is frequency-domain fault
diagnosis method, which was initiated in 1987 by [87] and
further extended by [88] in 1990. The basic idea is to generate
a residual, based on the stable coprime factorization of the
transfer function matrix of the monitored system, which is
made sensitive to the fault, but robustness against disturbances
by selecting an optimal weighting factor. Recent
developments of stable factorization approach can be found in
[89] for nonlinear systems, and [90, 91] for applications in
auto-balancing two-wheeled cart and thermal process,
respectively.
It is worthy to point out that the parity relation method and
stable factional approach both have some kind of connections
with observers. For instance, the parity relation approach is
equivalent to the use of a dead-beat observer, and the coprime
factorization realization includes the design of observer gain
(together with state-feedback gain).
B. Stochastic Fault diagnosis methods
In parallel with the development of the fault diagnosis for
deterministic systems, stochastic approaches were also
developed for fault diagnosis in the early 1970s. A general