Annual Reviews in Control 42 (2016) 63–81

Contents lists available at ScienceDirect

Annual Reviews in Control

journal homepage: www.elsevier.com/locate/arcontrol

Review

Bridging data-driven and model-based approaches for process fault

diagnosis and health monitoring: A review of researches and future
challenges
Khaoula Tidriri∗, Nizar Chatti, Sylvain Verron, Teodor Tiplica
LARIS, Avenue Notre Dame du Lac, 49000 Angers, France

a r t i c l e i n f o a b s t r a c t

Article history: Fault Diagnosis and Health Monitoring (FD-HM) for modern control systems have been an active area of
Received 4 August 2016 research over the last few years. Model-based FD-HM computational approaches have been extensively
Accepted 13 September 2016
developed to detect and locate faults by considering logical or mathematical description of the mon-
Available online 16 September 2016
itored process. However, because of parametric, measurement and model uncertainties, applicable ap-
Keywords: proaches that endeavor to locate faults with great accuracy are likely to give false alarms. Recently, many
Fault detection research works have been conducted in order to tackle this issue by making a tradeoff between accuracy
Fault diagnosis and robustness during the fault detection phase. Due to the recent advances in sensor technology, com-
Data-driven methods putational capabilities and dedicated software/hardware interfaces, data-driven FD-HM approaches have
Model-based methods demonstrated that highly accurate fault detection is possible when the system monitoring data for nom-
Hybrid methods inal and degraded conditions are available. Therefore, it seems that more than one approach is usually
required for developing a complete robust fault detection and diagnosis tool. In this paper, the features
of different model-based and data-driven approaches are investigated separately as well as the existing
works that attempted to integrate both of them. In this latter context, there have been only few works
published in the literature and hence reviewing and discussing them is strongly motivated by providing
a good reference for those interested in developing hybrid approaches for FD-HM.
© 2016 International Federation of Automatic Control. Published by Elsevier Ltd. All rights reserved.

1. Introduction
The increasing complexity of industrial systems and their re-
lated performance requirements have induced the need to develop
new approaches for supervising them. Supervision includes moni-
toring tasks which aim to determine the system’s operating state
at each time. This issue can be divided into two distinct but com-
plementary steps, namely:
• Detection which aims to identify the presence of an eventual
fault in the system.
• Diagnosis which aims to determine the root causes of the de-
tected fault. This task encompasses the fault isolation and iden-
tification steps which enable to characterize the type of fault,
its size and its profile.
A system is considered in normal operating mode when it pro-
vides a set of desired functions i.e. the system is in fault free case.
∗
Corresponding author.
E-mail addresses: khaoula.tidriri@univ-angers.fr, tidririkhaoula@gmail.com
(K. Tidriri), nizar.chatti@univ-angers.fr (N. Chatti), sylvain.verron@univ-angers.fr
(S. Verron), teodor.tiplica@univ-angers.fr (T. Tiplica).
The occurrence of a fault in the system can generate a failure. The
term fault is defined by Isermann (2006) as an unpermitted de-
viation of at least one characteristic property or parameter of the
system from the acceptable condition while failure is a permanent
interruption of a system’s ability to perform a required function
under specified operating conditions. In general, faults that occur
in a physical system can be classified into three categories i.e. ac-
tuator faults (parameters changes), sensor faults (bias for example)
and plant faults (leak in a tank for example).
These faults can lead to critical failures if they remain unde-
tected. Therefore, there is a need to generate a set of fault indi-
cators that should be significantly sensitive to these faults. In re-
cent years, there has been an increasing interest in fault detec-
tion and diagnosis approaches in order to cope with such issue.
Among these approaches, one can distinguish between data-driven
approaches, model-based approaches and expert knowledge ones
(Venkatasubramanian et al., 2003a; Venkatasubramanian et al.,
2003b; Venkatasubramanian et al., 2003c).
Data-driven approaches (Ding et al., 2011; Joe Qin, 2003) con-
sider the detection and the diagnosis as classification tasks. This
classification can be either supervised or unsupervised. Among the

http://dx.doi.org/10.1016/j.arcontrol.2016.09.008
1367-5788/© 2016 International Federation of Automatic Control. Published by Elsevier Ltd. All rights reserved.
64 K. Tidriri et al. / Annual Reviews in Control 42 (2016) 63–81
most popular data-driven methods, one can cite Neural Networks
(Zhang, 20 0 0), Bayesian Networks (Nielsen, 20 07; Pearl, 2014),
Control Charts (Joekes & Barbosa, 2013; Montgomery, 2007),
Principal Component Analysis (Yu, 2012), Partial Least Squares
(Kruger, Wang, Chen, & Qin, 2001). These methods play an impor-
tant role in modern monitoring systems especially for large-scale
industry applications since they do not require a lot of computa-
tions and hence they are compatible with real-time constraints of
dynamic complex systems. However, a preprocessing step is nec-
essary to extract information before applying the data-driven ap-
proaches.
Model-based approaches (Ding, 2008; Isermann, 2005) are a
good alternative when an access to a significant amount of data
is not possible but an accurate analytical model is available. This
model is generally build on the physics of the process to be mon-
itored. For each time step, residuals, which depict the differences
between the measured process variables and their estimates, are
generated. Then, a decision rule is chosen to evaluate those resid-
uals in order to detect a fault.
The generation of residuals can be accomplished by various
methods: diagnostic observers (Luenberger, 1966; Yang, Ding, & Li,
2015), parity relations (Gertler, 1997; Zhong, Song, & Ding, 2015),
Bond Graph (Ould-Bouamama, El Harabi, Abdelkrim, & Gayed,
2012; Paynter, 1961). A major advantage of these methods con-
sists on their ability to provide a description of the dynamic be-
havior and a physical understanding of the system. However, in
practice, it is very difficult to develop an accurate mathematical
model that takes into account modeling errors and uncertainties,
because some sources of uncertainty are not quantifiable. In order
to address such issue, statistics and interval approaches have been
developed. The former ones have represented the uncertainty with
Gaussian stochastic variables while the latter supposed the uncer-
tainties to be unknown but to stay within known and acceptable
bounds.
Thereby, each approach has its own advantages and drawbacks.
The development of hybrid approaches could improve the FDI per-
formances and overcome the limitations of individual methods
used separately. Besides, the state of the art encourages defin-
ing a common framework that enables the fusion of different ap-
proaches (Venkatasubramanian et al., 2003a; Ding, Zhang, Naik,
Ding, & Huang, 2009).
Accordingly, in Sections 2 and 3, an overview of the methods
from data-driven and model-based approaches respectively, is pre-
sented. Section 4 will highlight the benefits and limits of each ap-
proach through a comparative analysis and emphasize the interest
of hybrid approaches. In Section 5, a review on the existing re-
searches that aimed to bridge different approaches is carried out.
Finally, the last section will discuss some perspectives about FD-
HM with hybrid approaches and conclude the paper.
2. Data-driven approaches
Actual processes are increasingly automated, allowing the ac-
cess to a sizeable amount of data. Therefore, it is natural to moni-
tor the process using methods based on these data. For data-driven
approaches, fault diagnosis can be considered as a two stages pro-
cedure that encompasses (1) fault detection and (2) classification.
The first stage aims at detecting whether the system behavior
matches with the expected one while the second stage concerns
the determination of the class (type) of fault. These two stages can
be performed independently or combined to each other.
Within the data-driven approaches category, one can distin-
guish between supervised and unsupervised classification. In su-
pervised classification, it is necessary to define the classes and la-
bel the training data i.e. provide the category label for each of
them before the training procedure. This latter consists on fea-
ture extraction step which maps the high dimensional vectors to
feature space in order to find certain projecting vectors with low-
dimension. A common way to perform fault diagnosis is to employ
statistical models which aim at classifying the data acquired from
the monitored process into a normal operating condition class and
a faulty (out of control) class or distinguishing between different
fault classes. However, it is difficult to anticipate a priori all the
possible ways in which faults can occur. In order to tackle this is-
sue, different unsupervised approaches have been developed.
In the following, a review of supervised and unsupervised ap-
proaches, as well as discussions on their main advantages and
drawbacks are carried out.
2.1. Supervised classification
Supervised classification uses historical data to construct a
learning model, which is used for the fault detection and diag-
nosis of the new data. Among the most used methods, one can
cite Bayesian Networks (Pearl, 2014) and Artificial Neural Networks
(Zhang, 20 0 0).
2.1.1. Bayesian Networks (BN)
A BN is a directed acyclic probabilistic graphical model in-
troduced by Pearl (2014). BN has been successfully applied in
various application domains including clinical decision support
(Sesen, Nicholson, Banares-Alcantara, Kadir, & Brady, 2013), diag-
nostic diseases (Antal, Fannes, Timmerman, Moreau, & De Moor,
2003), genotype data analysis (Yan & Cercone, 2010), cancer metas-
tasis modeling and prediction (Wang, Makond, & Wang, 2014),
fault detection and diagnosis (Atoui, Verron, & Kobi, 2015b; Ver-
ron, Li, & Tiplica, 2010a; Verron, Tiplica, & Kobi, 2008; 2010b;
Zhao, Wen, & Wang, 2015; Zhao, Xiao, & Wang, 2013). In particu-
lar, BN provides a powerful tool for knowledge representation and
reasoning in presence of uncertainties (Correa, Bielza, & Pamies-
Teixeira, 2009; Gaymard & Tiplica, 2014; Lu, Bai, & Zhang, 2009;
Weidl, Madsen, & Israelson, 2005). Within this scope, BN has been
used to represent Gaussian Mixture Model (GMM) which is a non-
regular statistical model (Tao, Li, Zhu, & Li, 2012).This strategy en-
abled to deal with non-Gaussianity problem since the regular sta-
tistical models cannot be applied in such case. Another problem
concerns modeling of temporal relationships between variables. It
has been addressed by developing the Dynamic Bayesian Network
(DBN) which combines static network with temporal information
(Murphy, 2002).
The DBN has been applied to improve the quality of Internet
service (Li, Cheng, Qiu, & Wu, 2009), to detect transient faults
(Jha, Li, & Seshia, 2009), to identify the fault propagation path-
ways, and diagnose the root cause variables (Yu & Rashid, 2013).
In Zhang and Dong (2014), the authors proposed a multi-time-slice
DBN with a mixture of the Gaussian output to handle two princi-
pal issues: the missing data samples and the non-Gaussian process
data.
Beyond its ability to reason with uncertain information, BN can
use historical data and expert knowledge to complete the lack
of data (Zhao et al., 2013). Moreover, multivariate control charts
and other techniques as principal component analysis were mod-
eled by a BN classifier which enabled to detect and isolate faults
within the same framework (Atoui, Verron, & Kobi, 2015c; Verron
et al., 2010a). This strategy had proved its robustness and good
performance. However, the network structure is designed depend-
ing on the prior process knowledge and requires a large amount
of training data. Furthermore, the prior probability determination
and conditional probability table (CPT) computation are still chal-
lenging issues.
Thus, the effectiveness of BN depends on the various assump-
tions or conditions required for developing an accurate model. In
 K. Tidriri et al. / Annual Reviews in Control 42 (2016) 63–81
general, BN and statistical classification methods are built on the
basis of the Bayesian decision theory. Artificial Neural Networks
are a good alternative since they can adjust themselves to the data
without any explicit specification of the distribution in the input
data (Zhang, 20 0 0).
2.1.2. Artificial Neural Networks (ANN)
In recent years, ANNs have been used successfully in pattern
recognition in order to handle noisy or incomplete data, and for
prediction and classification problems (Zhang, 20 0 0). This method
has been widely used for different applications such as bankruptcy
prediction and credit scoring (Tsai & Wu, 2008), medical diagno-
sis (Mazurowski et al., 2008), transportation (Garrido, De Oa, &
De Oa, 2014), image classification (Barat & Ducottet, 2016), senti-
ment analysis (Vinodhini & Chandrasekaran, 2016). The classifica-
tion problem in ANN is reduced to the estimation of the connec-
tion weights, which are learned by using the mismatch between
the desired and actual outputs. Among the most common net-
works used for classification, one can cite the single-layer and the
multilayer perceptron network (Sorsa, Koivo, & Koivisto, 1991). The
use of the single-layer perceptron is restricted since it can produce
only linearly separable decision regions while the multilayer per-
ceptron network, which consists of three or more layers (an input
layer, hidden layers, an output layer), can achieve more general de-
cision regions (Sorsa et al., 1991). It is worth noting that the ANN’s
performances are characterized by the ANN’s structure, the learn-
ing algorithm, and the transfer function. Furthermore, an inappro-
priate structure may lead to a weak ANN classifier.
2.2. Unsupervised classification
Unlike supervised methods, unsupervised ones do not make any
assumption on the knowledge about the individuals belonging to
different classes as for control charts, principal component analysis
or partial least squares.
2.2.1. Control charts
The first monitoring method based solely on data was proposed
in 1931 by Shewhart and called control chart (Shewhart, 1931).
This graphical method, based on statistical hypothesis testing, is
used to monitor a process quality characteristic in order to iden-
tify its special causes of variability (Montgomery, 2007).
One has to distinguish between two phases during the design
of control charts: the retrospective analysis in Phase I and the
process monitoring in Phase II (Montgomery, 2007). The objec-
tive of the retrospective analysis is to determine the in-control
state of the process and to estimate the parameters when they
are unspecified. The most suitable performance index in this step
is the False Alarm Probability (FAP) which is the error that oc-
curs when a free-fault measurement is considered as a fault
(Bernal-de Lázaro, Llanes-Santiago, Prieto-Moreno, Knupp, & Silva-
Neto, 2016) [See Jones-Farmer, Woodall, Steiner, and Champ (2014),
Chakraborti, Human, and Graham (2008), and Coelho, Chakraborti,
and Graham (2015) for Phase I control charting review]. In Phase II,
it is assumed that a reference sample is available from the analysis
performed in Phase I and hence the control chart is used to mon-
itor the process in order to detect changes. The performances dur-
ing this step are generally quantified by the Average Time to Sig-
nal (ATS) and/or the Average Run Length (ARL) (Riaz, Ahmad et al.,
2016). This latter represents the number of samples until the de-
tection of an out-of-control signal.
According to Montgomery (Montgomery, 2007), Shewhart con-
trol charts are useful for Phase I, since they are effective in de-
tecting large changes. To improve the sensitivity of the control
chart to small process shifts, several methods have been pro-
posed such as The Western Electric Run Rules (Electric, 1956) ,
65
the Cumulative Sum control chart (CUSUM) [See the pioneering
work of Page (1954), Lucas (1982), and Liu, Xie, and Goh (2006)],
the Exponentially Weighted Moving Average chart (EWMA) [See
the introduction by Roberts (Lucas & Saccucci, 1990; Roberts,
20 0 0) for a performance study], the Weighted CUSUM (WCUSUM)
(Yashchin, 1989) etc.
From the 1920s to the present, there have been many develop-
ments in control chart methodologies. An important one consists
of taking variable sample size (VSS) from the process using a vari-
able sampling interval (VSI). VSS and VSI control charts can detect
process changes faster than conventional fixed sampling interval
control charts [for VSI chart (Amin & Miller, 1993; Chou, Chen, &
Liu, 2006), for VSI-CUSUM (Reynolds, Amin, & Arnold, 1990)].
Based on the same principle as the univariate control charts,
multivariate control charts (Lowry & Montgomery, 1995) were de-
veloped in order to handle a large number of variables, to consider
the correlation between these variables and to control the False
Alarm Rate (FAR). Among the multivariate control charts, one can
cite the Hotelling’s T2 control chart (Lowry & Montgomery, 1995;
Mason, Tracy, & Young, 1995), the MEWMA (Multivariate EWMA)
(Bernal-de Lázaro et al., 2016; Lowry, Woodall, Champ, & Rigdon,
1992; Prabhu & Runger, 1997; Stoumbos & Sullivan, 2002; Sullivan
& Jones, 2002), the MCUSUM (Multivariate CUSUM) (Crosier, 1988).
Thus, the proposed methods based on multivariate control charts
have become increasingly sensitive to small changes, more robust
to the normality assumption and increasingly fast in detection. But
in return, their development and design have become more com-
plex. However, detecting a small shift in a process very quickly
with tolerable false alarms is still a challenging problem in vari-
ous areas.
Even though the multivariate control charts are much bet-
ter than the univariate control charts for monitoring modern in-
dustrial processes, they are suffering from two inconveniences:
they lose their effectiveness when the number of variables be-
comes very large and the diagnosis remains a complex task (i.e.
identifying what happened wrong in the process when a fault
has been detected is not obvious) (Li, Jin, & Shi, 2008; Ma-
son et al., 1995). In order to cope with dimensionality reduc-
tion, statistical approaches such as Principal Component Anal-
ysis (PCA) (Montgomery, 2007) and Partial Least Squares (PLS)
(Venkatasubramanian et al. (2003a)) are generally used.
2.2.2. PCA and PLS
PCA (Montgomery, 2007) is a multivariate statistical analysis
method that reduces data dimensionality so that essential informa-
tion is retained and analyzed more easily than in the original data
set. It also captures the maximum variance of the data and the cor-
relation between the process parameters. PCA has been widespread
used in process monitoring especially in real-time control of con-
tinuous processes (Kourti & MacGregor, 1996; Lu, Wang, & Gao,
2003; Zhou, Guo, Celler, & Su, 2014).
However, PCA is a linear technique while most real processes
are nonlinear. Therefore, many extended methods have been de-
veloped to address this issue such as Nonlinear PCA based on auto-
associative neural networks (Kramer, 1992), Kernel PCA which can
efficiently compute PCs in a high-dimensional feature space using
nonlinear kernel functions (Lee, Yoo, Choi, Vanrolleghem, & Lee,
2004; Phillpotts, 2007)...etc.
In order to monitor time-varying systems, adaptive or recursive
approaches have been exploited such as the recursive algorithm
for KPCA (Zhang, Li, & Teng, 2012), the moving window kernel
PCA (Liu, Kruger, Littler, Xie, & Wang, 2009), the adaptive multi-
block and multi-scale PCA which updates the model structure to
deal with changing process (Lee, Park, & Vanrolleghem, 2005), the
weighted adaptive recursive PCA for the slow and normal process
changes (Portnoy, Melendez, Pinzon, & Sanjuan, 2016)...etc.
66 K. Tidriri et al. / Annual Reviews in Control 42 (2016) 63–81
Many researches have been conducted to address fault isolation
issues using PCA, and they aim to determine the type, location and
detection time of a fault. Contribution charts and multi-block ap-
proaches have been proposed to improve fault interpretation using
only process data or using some process knowledge for block par-
titioning when it is available (MacGregor, Jaeckle, Kiparissides, &
Koutoudi, 1994). Sparse PCA and shrinking PCA (Xie, Lin, & Zeng,
2013) can also enhance fault isolation by resolving an optimization
problem.
Similarily to PCA, PLS is another powerful statistical tool widely
used for fault detection and diagnosis purposes. This method con-
sists on projecting the information contained in high dimension
data spaces onto low dimension spaces, defined by a number of la-
tent variables. These new latent variables summarize all the impor-
tant information contained in the original data sets. Equivalently to
PCA, Multi-block and Multiway PLS (Kourti, Nomikos, & MacGregor,
1995) were developed to facilitate the diagnostic procedure.
It is clear that data-driven approaches are suitable for complex
and large-scale systems and can reduce time and cost since they
do not require the development of models (Alzghoul, Backe, Lf-
strand, Bystrm, & Liljedahl, 2014). However, they mainly depend
on the quantity and the quality of data (Chiang, Braatz, & Rus-
sell, 2001). This brings out a crucial point of distinction between
approaches based on historical process data and model-based ap-
proaches.
3. Model-based approaches
Model-based approaches compare available measurements with
a prior information, represented by a mathematical model. The in-
consistencies between the real and predicted behavior are consid-
ered as fault indicators. They are called residuals and they can be
generated by using observer based techniques (Narasimhan, Vach-
hani, & Rengaswamy, 2008), parity space (Odendaal & Jones, 2014),
parameter estimation (Mulumba, Afshari, Yan, Shen, & Norford,
2015), graph theoretic approaches (Bouamama, Biswas, Loureiro, &
Merzouki, 2014) and so on. Those residuals are then evaluated in
order to detect a fault and to reveal abnormal situations on the
monitored process.
3.1. Observers
Since the system is observable (i.e. it is possible to determine
the behavior of the entire system from the system’s outputs) and
the process parameters are known, the output of a process can
be estimated by an observer and residuals can be generated by
reconstructing the unknown intern states of the system. Various
fault detection and diagnosis observers have been proposed such
as the Luenberger in a deterministic setting (Luenberger, 1966;
Tarantino, Szigeti, & Colina-Morles, 20 0 0), the Kalman Filter (KF)
in a stochastic setting (Rusinov, Vorobiev, & Kurkina, 2013), the
H∞ based observer which enables the involvement of frequency
specifications as additional criteria for better fault discrimination
(Pierri, Paviglianiti, Caccavale, & Mattei, 2008).
According to Patton and Chen (1997), an accurate and complete
mathematical model of a process is never available in practice.
Therefore, model uncertainties have to be considered since they
may lead to a bias in the residuals, which can be interpreted as a
faulty situation. Moreover, the residuals may be influenced by dis-
turbances which represent unknown and uncontrolled inputs act-
ing on the system. Thus, for the purpose of fault detection and iso-
lation, Patton and Chen (1997) proposed to compute the residuals
by decoupling the effect of a fault from the effects of different in-
puts (the disturbance decoupling principle) and from other faults.
On the one hand, different approaches have been proposed to
achieve disturbance decoupling such us the Unknown Input Ob-
server (UIO) or the Unknown Input Filter (UIF) (Hwang, Kim, Kim,
& Seah, 2010; Liu, Gao, & Han, 2015; Sotomayor & Odloak, 2005).
On the other hand, to deal with the robust fault isolation, a set
of robust structured or directional residuals have been designed
(Patton & Chen, 1997). These enhanced residuals can be obtained
from consistency relations, or from a full PCA model (Gertler &
Cao, 2005).
Thus, the observer based techniques have become very robust
to model uncertainties and disturbances but in return, their sensi-
tivity to faults with slow time constants has decreased. To over-
come this limitation, Bastin and Gevers (1988) proposed to use
adaptive observers. This solution is applicable not only to linear
but also to a class of nonlinear systems (Caccavale, Pierri, Iamarino,
& Tufano, 2009). Basically, the linear observers theory for residual
generation was extended to certain classes of nonlinear systems
(Zarei & Shokri, 2014) by using fuzzy techniques (Li, Ding, Yang,
& Zhang, 2016), H− and H∞ formulation (Chadli, Abdo, & Ding,
2013) etc. Nevertheless, many physical systems cannot be modelled
in this way.
Furthermore, filtering approaches have been investigated to
address the problem of robust fault detection and diagnosis in
the presence of uncertainties. He, Liu, and Hua (2015) proposed
a strategy based on an adaptive Extended KF (EKF) which en-
ables to reduce the measurement noises. However, this approach
presents a limitation since it deals only with sensors fault, includ-
ing the bias fault in current and voltage sensors. Recently, another
strategy was proposed by Xiong, Jauberthie, Trave-Massuyes, and
Le Gall (2013) which combines the interval analysis and the sta-
tistical behavior of the noise based on the improved Interval KF
and hence it leads to an extension of conventional Kalman filtering.
The output prediction of the KF is considered as the reference of
the healthy system and the standard deviation is used to build the
fault detection thresholds. This approach, used for detecting addi-
tive sensor faults, considers bounded parameter uncertainties and
centered Gaussian noise. A comparison between a statistical deci-
sion and an interval-based approach can be found in Gelso, Biswas,
Castillo, and Armengol (2008).
3.2. Parity space
Among the approaches based on residual generation, parity
space has been extensively investigated (Chow & Willsky, 1984).
Parity space is based on the transformation of the state-space
model of the plant to obtain the parity relations by observing the
system on a finite horizon (Gertler, 1991). The purpose of the gen-
eration of parity relations is to provide equations which only de-
pend on known or measured variables namely the inputs and the
outputs. To obtain the parity relations, redundancies between dif-
ferent variables of the system are used.
Parity space was initially developed for static systems (Potter
& Sunman, 1977), then extended to dynamic systems (Chow &
Willsky, 1984). As it has been shown by Gertler, any model-based
residual generator can be expressed in the form of parity relations.
The author demonstrated the complete equivalence of diagnostic
observers and parity relations in Gertler (1991). Therefore, as for
observers, the issue of fault isolation and robustness to noise, dis-
turbances and model errors has been considered in many studies
(Chan, Hua, & Hong-Yue, 2006; Ding, Guo, & Jeinsch, 1999; Gertler,
1997). In the last reference, Ding et al. have demonstrated that in-
creasing the order of the parity equations may improve the sys-
tem robustness. However, parity space approach generally does not
consider significant model uncertainties, unmodeled disturbances
or multiplicative faults. It is also limited in most cases to linear
time-invariant systems.
It is worth noting that there is a known tradeoff between
the performance and the computational cost in the design of a
 K. Tidriri et al. / Annual Reviews in Control 42 (2016) 63–81
parity relation-based residual generation (Hwang et al., 2010). In
order to obtain a good performance index with a low computa-
tional cost, Ye, Zhang, Ding, and Wang (20 0 0) proposed a wavelet-
based parity vector of low order but this approach presents a
problem of misdetection. The authors introduced then an infi-
nite impulse response filter to improve the performance index
without increasing the computational effort. However the intro-
duction of the filter leads to slow response when faults occur
(Ye, Wang, Ding, & Su, 2002). To overcome these issues, Ye, Wang,
and Ding (2004) used a stationary wavelet transform (SWT) which
presents a low misdetection rate, a satisfactory response speed to
faults and a good performance index.
Since observer and parity space based schemes require the de-
velopment of the complete model and given that a model cannot
be easily obtained due to nonlinearity, complexity, and high di-
mensionality of a process, an alternative data-driven approach has
been proposed recently, using only the primary form of the resid-
ual generator (Ding, Zhang, Huang, & Ding, 2005).
3.3. Parameter estimation
Since the basic model structure is known, the process supervi-
sion can be fulfilled by monitoring the evolution of its structural
parameters (Isermann, 1984; Venkatasubramanian et al., 2003b).
Indeed, the non-measurable quantities such as process parameters
and state variables can be estimated based on the measurable in-
put and output variables. This estimation can be performed by an
algorithm based on measured data such as the KF (Kalman, 1960),
the EKF (Pan et al., 2016), the particle filter (Walker, Rayman, &
White, 2015), the Bayesian method and other derived methods
(Ljung, 1987).
Parameter estimation for fault detection and diagnosis relies on
a comparison between nominal parameters as computed in fault-
free operation and parameters estimated during online operation
over a given time horizon. Therefore, a fault is detected once the
values estimated online are inconsistent with the nominal ones
(Isermann, 1993). This approach has been extensively studied in
system identification (Ljung, 1987), then applied to fault detection
and diagnosis as reviewed in Isermann (2006). Parameter estima-
tion can be very simple when the model parameters are directly
related to the physical coefficients (such as resistance, loads etc.)
(Dai & Gao, 2013). However, this technique requires a process in-
put excitation and it is especially suitable for the detection of mul-
tiplicative faults.
3.4. Bond Graph (BG)
Among the graphical approaches, the BG has proved its ade-
quacy to represent energy exchanges in mixed systems and has
been used to generate residuals in a systematic and generic way.
BG is a unified graphical description that presents a domain-
independent and energy-based methodology for modeling the dy-
namic behavior of physical systems from different domains (elec-
trical, mechanical, hydraulic, thermodynamic ...etc.).
The BG methodology encompasses various kinds of informa-
tion due to its causal and structural features that enable to de-
duce directly a set of fault indicators. Recently, the BG model has
been extended by incorporating the parameters uncertainties in
order to generate robust and adaptive thresholds for the resid-
uals evaluation stage. The procedure for obtaining residuals is
based on covering causal paths (Samantaray & Bouamama, 2008)
and is implemented in dedicated software (Chatti, Gehin, Ould-
Bouamama, & Merzouki, 2013; Xiaotian & Anlin, 2014). Notice that
Djeziri, Merzouki, Bouamama, and Dauphin-Tanguy (2007) pro-
posed BG modeling approach in the LFT (Linear Fractional Transfor-
mation) configuration that enables to take into account parameter
67
uncertainties. However, this approach cannot address the measure-
ments uncertainties. To overcome the limitations of the BG and
to generate qualitative and quantitative diagnosis on the basis of
a single representation, Chatti, Ould-Bouamama, Gehin, and Mer-
zouki (2014) proposed a new formalism called Signed Bond Graph
(SBG).
A Signed Bond Graph G(X,A) is a directed labeled and signed
graph built from the BG. This formalism exploits its qualitative
and quantitative structural properties enabling the generation of
multiple behavior predictions (i.e. possible conflicts). In order to
enhance the existing fault isolation algorithms based on residuals
generation, it incorporates qualitative features to the BG that take
into account temporal orders of measurement deviations within a
specific sliding time window, to address multiple faults. A refined
description of the SBG can be found in Chatti et al. (2014).
Although research on model-based approaches for fault detec-
tion and diagnostic has been actively carried out with noteworthy
results (Isermann, 2006), it is still a challenging task, especially in
researches related to robustness in presence of different sources of
uncertainties (Keliris, Polycarpou, & Parisini, 2015), nonlinear sys-
tems and hybrid system issues. During the last years, different fil-
tering and parameters estimation approaches have been developed
and they were based on statistical hypothesis tests (Li, Chan, &
Zhang, 2005). However, such statistical tests require to define the
statistical distributions of the residuals in faulty situation. This is
why various works focused on the so-called set-membership ap-
proach such as the robust fault detection strategy proposed by
Tornil-Sin, Ocampo-Martinez, Puig, and Escobet (2012) and formu-
lated as a set-membership state estimation problem, for both time-
variant and time-invariant system models.
Furthermore, special attention has been given to fault detection
approaches for Linear Parameters Varying (LPV) systems since they
represent an attractive alternative to consider nonlinear dynamics
(López-Estrada, Ponsart, Theilliol, Zhang, & Astorga-Zaragoza, 2015;
Varga & Ossmann, 2014). Many efforts have also been made to deal
with fault detection and diagnosis in decentralized and distributed
systems (Boem, Ferrari, Parisini, & Polycarpou, 2013; Keliris et al.,
2015). Finally, it is worth mentioning that various works focused
recently on the design of consistency tests for dynamical systems
with additive and multiplicative parameter uncertainties by deal-
ing with intervals analysis (Puig, Oca, & Blesa, 2013).
4. Comparison between data-driven and model-based
approaches
Despite numerous studies published in the literature, no single
approach is adapted to all the requirements for a diagnostic sys-
tem.
On the one hand, data-driven approaches rely on a huge
amount of data and focus on the analysis of the systems output
signals. These approaches can handle high dimensional and corre-
lated process variables (Venkatasubramanian et al., 2003a). Thus,
they are suitable for complex and large-scale systems and can re-
duce time and cost since they do not require the development of
models (Alzghoul et al., 2014). Although data-driven approaches
are simpler than model-based ones, their application require a pre-
processing step to extract the useful information from data with a
high computational cost. Therefore, the performances rely on train-
ing data and may degrade once the system is working in an un-
known condition (i.e. outside of the training data) or is affected by
unknown faults. Fig. 1 shows a comparison study of various data-
driven methods in terms of desirable criteria for a diagnostic sys-
tem.
On the other hand, model-based approaches need small amount
of online data but require an explicit mathematical model, avail-
able from first principles or identified through system estimation
68 K. Tidriri et al. / Annual Reviews in Control 42 (2016) 63–81
Fig. 1. A comparison of data-driven fault detection and diagnosis methods.
techniques. As the system complexity increases, the estimation be-
comes more difficult. In addition, a model validation step is neces-
sary since most models rely on a set of assumptions and approxi-
mations (Sargent, 2013). Indeed, these approaches have focused on
linear system description and some very specific class of nonlin-
earities. According to Chiang et al. (2001), the model-based ap-
proaches are not suitable for large-scale and complex systems and
their performances depend on how good and precise the model
is. Because it integrates the physical understanding of the system
and incorporates the interactions with the environment into the
fault detection and diagnosis process, these approaches can per-
form better than the data-driven approaches when an accurate
model is available (Chiang et al., 2001).
The main features of existing model-based approaches are sum-
marized in Fig. 2.
However, most of the fault diagnosis applications in process in-
dustries are based on data-driven approaches since they are able to
detect faulty situations faster, are easier to implement and require
less a priori knowledge.
Furthermore, integrating various strategies of fault detection
and diagnostic within the same framework seems to be interesting.
For example, parameter estimation method has been integrated to
observers and parity space (Dai & Gao, 2013). Indeed, the state es-
timation based schemes are efficient for fault detection because
they are inherently fast (Patton & Chen, 1997) when compared to
parameter estimation method.
Statistical techniques like PCA and PLS have been combined
with supervised methods like ANN (Ghate & Dudul, 2011) or with
the time-frequency analysis (Lu et al., 2003) to extract statistical
features, multiple classifier systems have been mixed to improve
the performance (Ghosh, Ng, & Srinivasan, 2011)...etc.
In recent years, there has been a growing interest in inte-
grating different methods to each other. The main motivation for
developing hybrid frameworks is that no single method is able
to satisfy all the requirements of an accurate diagnostic system.
The purpose of the brief comparison presented above is to high-
light and reveal the need for a hybrid approach. Furthermore, ac-
cording to several studies (Ding et al., 2011; Ding et al., 2009;
Venkatasubramanian et al., 2003a), there is a need for integrat-
ing data-driven and model-based approaches to benefit from the
advantages while avoiding the drawbacks of both approaches. This
integration has many interests:
• It overcomes the weakness of a diagnostic method with the
strength of another method to achieve a better performance.
• It enables to use a variety of information sources such as ser-
vice history, operation and maintenance records, historical and
on-line data, mathematical models, causal relationships...etc.
when they are available. Indeed, the more knowledge and in-
formation are merged, the more efficient and reliable the diag-
nosis would be.
• It enables to benefit from all the progress and achievements
made by each community.
5. Hybrid approaches
Recently, there has been a growing interest in data-driven
and model-based diagnosis to understand and integrate both ap-
proaches in order to provide better diagnostic systems and im-
prove health monitoring. In the following, a review of hybrid ap-
proaches is carried out.
 K. Tidriri et al. / Annual Reviews in Control 42 (2016) 63–81 69

Fig. 2. A comparison of model-based fault detection and diagnosis methods.

5.1. Hybrid approaches based on SDG Table 1

Comparison between SDG, QTA and PDF.

One of the earliest examples of a hybrid approach is the hybrid SDG QTA PDF
framework for large scale process fault diagnosis called DKit, pro-
Root cause analysis ++ − −−
posed in 1997 (Mylaraswamy & Venkatasubramanian, 1997). The Large scale, complex, nonlinear process − + ++
motivation for designing such an alternative approach relies on the Robustness to process noise, modeling uncertainties + + ++
fact that a perfect method does not exist. The essential compo- Quickness of response −− ++ ++
nents of DKit are illustrated in Fig. 3. It consists of an input-output Ease of deployment − + ++
Adaptability to new steady states + − −−
interface which receives process measurements, a collection of di-
agnostic methods, a scheduler (responsible for resolving the con-
flicts in results) and an operator interface that presents diagnostic
results to the operator. Assuming that there are n diagnostic methods and that each
The design of Dkit depends on three parameters: method (j) will compute the probability Pij of the hypothesis Hi
given the sensor evidence e, the conflict resolution can be done by
• The collection of diagnostic methods using the weighted voting scheme to estimate the probability of
• The scheme for triggering diagnostic methods the overall hypothesis Hi :
• The algorithm for conflict resolution.  1 n−1
P a j Pi j or
In this paper, the framework combined causal model-based di- P (Hi ) = n−1 i,SDG j=1
 −1
Pi,SDG if nj=1 Pi j = 0
agnosis (Signed Digraph) with Probability Density Function-based
statistical classifier (PDF) and qualitative trend analysis (QTA). With 0 ≤ aj ≤ 1.
Thus, a comparison between SDG, PDF and QTA is presented to ex- It implies that SDG is used as a validation for the hypothesis
plain the need for a new hybrid diagnostic system. The Table 1 proposed by QTA and PDF. Indeed, according to the authors, pro-
summarizes this comparison study. cess history based methods are often short-sighted, hence a fault
70 K. Tidriri et al. / Annual Reviews in Control 42 (2016) 63–81
Fig. 3. The essential components of DKit.
Table 2
Comparison between SDG and QTA.
SDG QTA
Completeness ++ +
Diagnostic resolution + ++
Computational complexity + −
hypothesis that could not be explained by SDG will be rejected.
Therefore, Mylaraswamy and Venkatasubramanian choose to com-
bine the quickness of PDF and QTA to quickly narrow down the
scope of search with the knowledge based SDG to identify and an-
alyze the root causes. They tested the performance of DKit on the
industrial fluid catalytic cracking unit (FCCU) process, by simulat-
ing a set of different fault scenarios (such as simple sensor biases
or complex process faults). The evaluation took into account the
accuracy of diagnosis, the detection ability, the root cause expla-
nation and the timely response. It showed that DKit was able to
identify the root cause and to provide an explanation for the prop-
agation of the fault in all cases.
Unlike the previous approach, Maurya et al. proposed in
Maurya, Rengaswamy, and Venkatasubramanian (2007) to use SDG
first as a filter to reduce the set of possible faults and then QTA
to identify the actual fault. Table 2 shows the comparison between
these methods.
First, SDG is used to reduce the set of candidate faults by com-
paring the measured response pattern with the response predicted
by the SDG. Then, QTA is applied to improve the resolution by as-
signing ranks to the faults from Level 1. The description is given by
Fig. 4.
This approach addressed the problem of inaccurate diagnosis
due to noise and model uncertainties, by including safety-critical
faults within the set of fault candidates, using the so called de-
noising techniques and overriding the SDG-based candidate fault
set when the confidence index is low. Another important issue ad-
dressed in this study is the diagnosis of novel faults, which is dif-
ficult with QTA. In this case, the SDG-based candidate fault set
is presented to the operator for a manual diagnosis. This hybrid
framework only focused on incipient faults in the industrial bench-
mark of Tennessee Eastman process (TEP) and proved that the
diagnostic resolution is improved while the computational com-
plexity is reduced. However, a mathematical model is needed to
develop the SDG which limits the industrial applicability of this
method but note that one can use another type of qualitative
causal model for level 1 instead of a SDG model.
In addition to the DKit, an Operator Decision Support Sys-
tem, called Op-Aide, was also applied in the FCCU in 1999
(Vedam, Dash, & Venkatasubramanian, 1999). Op-Aide has been
developed to assist the operator in quantitative diagnosis and as-
sessment of faulty situations. The presence of faults is indicated
using PCA while the diagnosis is performed by SDG, as shown in
Fig. 5. Another diagnosis method, the B-splines based adaptive sys-
tem for trend analysis, is applied to identify the fault and confirm
the diagnosis. It uses the piecewise approximation of the data gen-
erated by a data compression algorithm (Vedam et al., 1999), in
order to extract qualitative features of the sensor trend.
Thus, the Op-Aide was able to correctly identify the nature of
root causes but the issue of robustness to noise and adaptability
were not discussed in this study.
The three previous approaches have used SDG in their work,
but in different ways. The first and the third approaches (D-Kit and
Op-Aid), as depicted by Figs. 3 and 5, applied the different meth-
ods in parallel unlike the second approach, which is used in a se-
rial fashion as shown in Fig. 4. This strategy requires the presence
of a separate scheduler module to resolve the conflicts between
the results of individual methods. Moreover, the real advantage of
 K. Tidriri et al. / Annual Reviews in Control 42 (2016) 63–81 71

Fig. 4. The proposed framework combining SDG and QTA for fault diagnosis.

Fig. 5. The proposed framework combining PCA and SDG for fault detection and diagnosis.

the combination of SDG and QTA in the second approach is that
both methods are sharing the diagnosis task. Indeed, SDG is used
as a filter to reduce the possible root-causes and QTA is used there-
after to improve the diagnostic resolution while in the PCA-SDG
approach, PCA is only used for detection and SDG for diagnosis.
5.2. Identifying a primary form of residual generator from data
Some researches focused on combining the ease of deploy-
ment of statistical methods with the powerful isolation capa-
bilities of analytical redundancy. Indeed, Gertler, Li, Huang, and
McAvoy (1999) showed that there is equivalence between PCA and
parity relations and proved this close duality strictly. The idea be-
hind this work is to transfer to PCA, which is a powerful tool in
detecting faults in complex systems, some concepts of analytical
redundancy like the fault isolation capability.
Thus, the paper presented how structured residuals with good
isolation properties can be obtained by PCA. The overall method-
ology is depicted in Figs. 6 and 7. The performance of this study
is evaluated on the TEP for static linear systems. Extension to dy-
namic and nonlinear systems is not discussed in details in this pa-
per.
Ding et al. (2009) were also interested in identifying a pri-
mary form of residual generator from data, using parity space
or observers. The main objective consisted in proving that ob-
server and parity space residuals can be obtained directly from
test data, without using the process model. This approach is based
on the work of Wang and Qin (2002) which combines the ideas
of PCA/DPCA and observer based schemes. The main contributions
of the study are the algorithms for the identification of the noise
model and for the construction of parity space and observer based
residuals. Ding et al. assumed that the processes under consider-
ation can be modelled as linear time invariant (LTI) systems and
proved the applicability of their idea from the TEP.
The two previous approaches (Ding et al., 2009; Gertler et al.,
1999) benefit from isolation abilities of model-based approaches
72 K. Tidriri et al. / Annual Reviews in Control 42 (2016) 63–81
Fig. 6. Enhanced PCA.
Fig. 7. Methodology to obtain an enhanced PCA.
while overcoming the difficult task of analytical model develop-
ment. Indeed, no special knowledge of control and system theory
is needed to generate structured residuals, which is a significant
practical advantage.
5.3. Improving the classification performance
It has been proven that combining different approaches in a
specific way could improve the classification performances. The au-
thors in Siswantoro, Prabuwono, Abdullah, and Idrus (2016) pro-
posed a linear model, whose parameters are estimated by a KF,
that brings the predicted output of the ANN closer to the desired
output. Therefore, the error related to ANN outputs is reduced and
the classification accuracy is improved. This approach was em-
pirically validated using data sets from the UCI Machine Learn-
ing Repository. However, the performance depends on the training
data since the proposed method consists of training phase followed
by a testing phase. This technique is depicted in Fig. 8:
During the testing phase, the NN-LMKF model predicts the out-
put and classifies the objects. However, it is assumed that the sys-
tem is stationary and that the model has a normally distributed
error term and depends on the inputs (object features) and pre-
dicted outputs of the ANN as shown in Fig. 8. Thus, developing
Fig. 8. Combination of ANN and KF.
such model requires a linear relationship between the different in-
volved variables. This method can be applied to other classifiers
and KF can be replaced with EKF for example to deal with nonlin-
ear systems.
An example of a nonlinear system that has been the subject of
an increasing interest in fault detection and diagnosis is the chem-
ical reactor. In Benkouider, Buvat, Cosmao, and Saboni (2009), the
authors focused on the fault detection in semi-batch reactors by
using statistical methods and the EKF to estimate the overall heat
transfer coefficient U of the reactor. In 2012, they completed their
work by proposing a hybrid fault detection and diagnosis strategy
based on the EKF and ANN (Benkouider, Kessas, Yahiaoui, Buvat, &
Guella, 2012). The EKF uses the reactor model as a predictor and
enables the online fault detection by estimating the heat transfer
coefficient of the reactor. The detection step is formulated as a hy-
pothesis test problem while the diagnosis is performed using the
ANN classifier, which depends on the input-output measurements
of the reactor and the estimated parameter U. The ANN identifies
then the type and location of the fault detected (See Fig. 9).
The performance was demonstrated through two types of
exothermic reactions and showed good results. However, an accu-
rate model of the reactor is required for the EKF algorithm.
On the one hand, the first approach integrates the KF and the
ANN in order to improve the performance of diagnosis. Indeed, a
training phase sets the KF in order to improve the classification.
Thereafter, the combined NN-LMKF is used for fault detection and
diagnosis. On the other hand, the EKF is only used to estimate the
desired overall heat transfer coefficient in the second approach.
Detection is performed using a classical hypothesis test that de-
termines whether or not a change has occurred. Then, the ANN
perform the diagnosis taking into account the estimated parame-
ter.
5.4. Unified scheme for hybrid fault detection and diagnosis
Developing unified and generic frameworks that enable the in-
tegration of different methods or various sources of knowledge has
met considerable interest. Zhao et al. (2013) proposed a generic in-
telligent fault detection and diagnosis strategy for chiller experts.
The objective was to integrate more useful information and expert
 K. Tidriri et al. / Annual Reviews in Control 42 (2016) 63–81
Fig. 9. Combination of EKF, statistical test and ANN for online estimation, fault de-
tection and diagnosis.
knowledge such as service history and maintenance records, in the
presence of uncertain, incomplete and conflicting information, in
order to improve the reliability of the results. The framework con-
sists in a three-layer diagnostic BN: additional information (repair-
ing service, abnormal operation records, routine maintenances, the
healthy states of equipment...etc. which help to diagnose the faults
more accurately) in Layer 1, faults in Layer 2 and fault symptoms
(sensor measurements, performance indices) in Layer 3. The objec-
tive is to identify a fault according to one or multiple symptoms
observed and other qualitative information. The details are shown
in Fig. 10.
In this article, the independence assumption was introduced to
reduce the number of required parameters and to simplify the in-
ference procedure, making the computation load affordable. The
main advantage of this approach is that it takes into account the
uncertainties in various information by using probabilities to quan-
tify the occurrence possibilities of faults instead of Boolean. The
framework was evaluated using experimental data and showed
that the difficulty rely on the determination of the BN parameters.
The BN was also used by Atoui, Verron, and Kobi (2015a) as a
framework for hybrid fault detection and diagnosis. Their objective
was to associate a data-driven method, a T2 statistic, to the resid-
ual generated by a suitable model-based method, under a common
tool: the conditional Gaussian BN. In order to handle this issue, the
two methods are first modeled under a BN and then combined to
make better decisions, as described in Fig. 11.
On the one hand, the data-driven method can provide a de-
cision at each time step, by taking the state with the maximum
posterior probability, given a new observation. On the other hand,
the model-based method evaluates the outputs of the residuals
and gives at each instant the occurrence probability of each fault.
The decision is given by the added node Srm which indicates the
probability of occurrence for each state. This framework was sim-
ulated on a water heater process under different hypothesis (ac-
curate model, degraded model, more or less data) and showed a
better reliability in decision than the two methods used indepen-
dently. However, it only handles single faults, sensor and compo-
nents faults. To further improve the decision making, other meth-
ods and information can be modeled and integrated to the BN.
73
Both approaches were designed to improve the reliability of de-
tection and diagnosis results using a BN. This latter enables to in-
tegrate different types of information. The first approach aims to
use all the system’s available knowledge. The second approach re-
quires calculated residuals and measured data and considers sev-
eral scenarios that can be encountered in practice. Using a hybrid
approach was justified especially when a big amount of data and
a specific model were not available. In such case, hybrid approach
showed good results comparing with the two methods used sepa-
rately.
Two other works discussed the difficulty of having a unified
scheme for fault detection and diagnosis.
First, Schubert, Kruger, Arellano-Garcia, de S Feital, and
Wozny (2011) proposed a unified scheme for fault detection and
isolation that integrates model-based and data-driven methods.
This work used subspace model identification and state-observers
to identify the dynamic model from data and unknown input esti-
mation to generate residuals. A statistic (T2 or SPE) is then applied
to detect abnormal behavior. The fault diagnosis task was carried
out by a set of enhanced residuals. This framework is described by
Fig. 12. It is designed to detect simple sensor and actuator faults
as well as complex process faults. It was applied to three industrial
processes: heating furnace, distillation column, reactive distillation
unit. It proved its ability to detect simultaneous faults, to avoid the
need for an accurate model and to use convenient statistics that
simplify the monitoring task. During the study, it is assumed that
the system is LTI of finite dimension and that the model is iden-
tified. This implies that the model can only be employed for cer-
tain regions. It is also supposed that the noise is described by a
zero mean multivariate Gaussian distribution under normal oper-
ating conditions.
Second, Zhang et al. proposed in Zhang and Hoo (2011) to han-
dle the fault detection and diagnosis problem by a hierarchical
strategy, which consists of narrowing successively the search space
of potential faults, as shown in Fig. 13.
This strategy is defined by the following four main steps:
• First, the system is decomposed based on its BG and domain
boundaries.
• Second, PCA is applied to reduce the data dimension and to
eliminate the multivariate outliers.
• Third, a discrete wavelet transform (DWT) is applied to the
scores obtained by PCA to provide a multiple time-scale de-
composition. The DWT consists in a signal processing with a
high-pass and a low-pass filters. The coefficients obtained by
the filters describe the details and a smooth approximation of
the signal. The slowest dynamic time constant present in the
system defines the length of the DWT window.
• Finally, the Mahalanobis Distance (MD) is determined and the
confidence level is calculated from the χ 2 distribution.
If the occurrence of a fault is detected (i.e. if there is a viola-
tion of the MD at any scale), the BG and the BN are used to lo-
cate the domain of the root cause. In fact, a comparison is made
between the time scales at which the violation of the nominal
confidence level occurred, and the natural time constants associ-
ated with each physical domain determined by the BG. After that,
a DBN uses the conditional dependence between fault signatures
and faulty domains to find the possible fault origin. The DBN, as
shown in Fig. 14, consists of two nodes for each slice with St rep-
resenting the fault signatures (all time scales) and Dt representing
the candidate faulty domains (all the physical domains obtained by
the BG).
Regarding the conditional dependence between Dt et St , the au-
thors assigned higher conditional probability to a fault signature
with time scales closer to that of the given domain. Furthermore,
if there is a bi-directional power exchange in the BG between two
74 K. Tidriri et al. / Annual Reviews in Control 42 (2016) 63–81

Fig. 10. Framework for fault detection and diagnosis using a three-layer BN.

Fig. 11. Framework using measurements and residuals for fault detection and diagnosis.

domains, the conditional probability for a fault to propagate from
a domain to the other has a high probability.
This idea was simulated on a continuous-stirred tank reactor
and on a biochemical wastewater treatment plant to prove its effi-
ciency. Indeed, slow drifts, intermittent and multiple faults were
detected and associated to their domains. Hence, the reliability
of isolation was improved by using different methods. However,
since PCA is used first to reduce the data dimension, the remaining
methods are only applied to the resulting scores. Thus, a fault that
occurs in the residual space may not be detected, which represents
a major limit.
5.5. Integrating the reliability data
In addition to providing an interesting framework for fault de-
tection and diagnosis, BN were also used to improve the decision
making in ARR. In Weber, Theilliol, Aubrun, and Evsukoff (2007),
the authors integrated reliability data in a DBN with nodes rep-
resenting the residuals and others associated to the failure of
the components. However, this method only considers components
with exponential probability distribution functions to simplify the
computational tasks.
Within this scope, Zaidi, Tagina, and Bouamama (2010) pro-
posed to combine the BG method for residual generation with
 K. Tidriri et al. / Annual Reviews in Control 42 (2016) 63–81 75

Fig. 12. Framework for model identification, residual generation, fault detection and diagnosis.

Fig. 13. Hierarchical strategy for fault detection and diagnosis.

BN for data integration. BG is a powerful graphical tool for the
generation of ARR which are explicitly associated to components
faults. This information about the reliability of each component is
used to improve the decision-making step, making it possible to
handle unknown and identical signatures of failures. The residu-
als and the components reliability are integrated into a hybrid BN
(with continuous and discrete nodes) which calculates the poste-
rior probabilities of the components faults. The discrete part of the
BN consists of the residuals and the failures of the components.
This structure depends on the failures signatures: if a residual ri
is sensitive to the failure of the component Cj , then an arc joins
Fj to ri . The continuous part represents the reliability associated to
every component and gives the prior information on the failure of
the component.
76 K. Tidriri et al. / Annual Reviews in Control 42 (2016) 63–81
Fig. 14. Dynamic Bayesian Network representing fault signatures and faulty do-
mains.
Fig. 15. The scheme for fault detection and diagnosis using decision fusion strategy.
5.6. Improving the decision-making step
Combining multiple methods for fault detection and diagnosis
has met a considerable interest. However, this combination may
generate conflicts in results.
Ghosh et al. (2011) addressed this important issue by imple-
menting both model-based (EKF) and data-driven (PCA, ANN and
Self-organizing maps (SOM)) methods for online process monitor-
ing. The input to each method is online process data while the
output from each method is an assigned class (normal or a faulty
class). These individual outputs are then combined to solve the
conflicts and to obtain a consolidated result. Several decision fu-
sion strategies can be used as utility-based and evidence-based
strategies. In this paper, the authors chose evidence-based fusion
strategies such as weighted voting, Bayesian and Dempster-Shafer
based fusion to deal with this issue. Fig. 15 shows their proposed
scheme.
The authors made simulations on a lab-scale distillation col-
umn, with two different scenarios. The first scenario used well de-
signed and well trained methods, which lead to a quick detection
and identification of all faults. In this case, the improvement real-
ized through decision fusion is marginal. However, in practice, it
is difficult for a single method to quickly detect and identify all
faults. Thus, in the second scenario, the methods are redesigned in
order to detect and identify only certain faults while being insen-
sitive to others, resulting in significant disagreement among their
results. In this situation, evidence-based fusion strategies such as
weighted voting, Bayesian and Dempster-Shafer, achieved a com-
plete fault coverage (i.e. all the faults were detected and identified
quickly), a high prediction accuracy, an improvement in monitor-
ing performance and a reduction in fault detection and diagnosis
delays.
6. Discussion and conclusion
In this paper, an overview of different methods from data-
driven and model-based approaches for fault detection and diag-
nosis is presented and aims to provide the main concepts pro-
posed by both diagnosis communities. Thus, the features of differ-
ent data-driven and model-based approaches are highlighted and
the benefits and limits of each approach are pointed out through a
comparative analysis. Furthermore, a review of existing works that
attempted to bridge methods from both sides, in order to build
more robust and effective diagnostic systems, is carried out. This
last section will discuss the main conclusions that can be drawn,
emphasize the future challenges and conclude the paper.
Fault detection and diagnosis represent a complex task that re-
quires the system model and/or data and a variety of knowledge
or information sources. A suitable integration of multiple hetero-
geneous methods can benefit from all the knowledge available on
the system and improve the diagnosis performance. In this paper,
we highlighted the advantages that result from deploying various
methods simultaneously. Indeed, hybrid approaches perform gen-
erally better than the involved individual methods considered sep-
arately for detection and diagnosis tasks. Table 3 summarizes the
research works that have been considered in this review.
Most of these works used the data-driven approaches for fault
detection step while the model-based approaches were used for
fault diagnosis (Gertler et al., 1999; Schubert et al., 2011; Vedam
et al., 1999).
However, only few works have attempted to provide a frame-
work to integrate data-driven and model-based concepts through
the BN framework as it offers a probabilistic framework able to
merge information and uncertain knowledge from various sources.
Despite the obvious advantages that hybrid approaches can of-
fer for fault detection and diagnosis, their potential still weakly ex-
ploited.
Actually, the main obstacle which is faced when developing hy-
brid approaches is the lack of a generic framework. Some studies
have suggested solutions to address this issue but it still remains
an important topic for further research.
Besides, once the results from various methods are available, an
important step is to apply an effective strategy to combine these
results. However, the development of decision fusion strategies to
combine the results remains largely unexplored.
Moreover, a set of criteria for hybrid approaches is not yet de-
fined. Though, the desired characteristics are necessary to analyze,
evaluate and compare hybrid approaches.
Here are the points to be addressed during the design of a hy-
brid approach.
6.1. Main objective
The main purpose of combining different approaches is to im-
prove the performance of the fault detection and diagnosis system.
In such issue, results are expected in order to answer to real in-
dustrial problems such as: Does the combination of such methods
really improve the performance and the resolution of the overall
system? However, an effective and well-suited combination of ap-
proaches for monitoring is difficult to obtain because it requires
 K. Tidriri et al. / Annual Reviews in Control 42 (2016) 63–81 77

Table 3
Hybrid approaches considered in this review.

Methods Application case Improvements Reference

SDG-QTA-PDF FCCU Detection ability Mylaraswamy and Venkatasubramanian (1997)

Root cause explanation
SDG-QTA TEP Diagnostic resolution Maurya et al. (2007)
Computational complexity
SDG-PCA FCCU Detection ability Vedam et al. (1999)
Root cause explanation
PCA-parity relations TEP Isolation capabilities Gertler et al. (1999)
PCA-observer TEP Modeling requirements Ding et al. (2009)
ANN-Kalman UCI Machine Learning Repository Classification accuracy Siswantoro et al. (2016)
ANN-EKF Batch chemical reactors Online estimation of parameters Benkouider et al. (2009)
Semi-batch chemical reactors Classification accuracy Benkouider et al. (2012)
BN Different type Water-cooled centrifugal chiller Reliability of the results using Zhao et al. (2013)
various sources of knowledge of
information
BN Residuals and measurements TEP Reliability of the decision when an Atoui et al. (2015a)
accurate model or data are not
available
SMI- observer- UIE Heating furnace, Modeling requirements Schubert et al. (2011)
Distillation column Simplification of the monitoring
Reactive distillation unit task
PCA-DWT-MD-BG-BN Continuous-stirred tank reactor Detection ability of different faults Zhang and Hoo (2011)
Biochemical wastewater treatment Reliability of isolation
plant
BN-BG Controlled two-tank system Reliability of the decision by Weber et al. (2007)
integrating reliability of Zaidi et al. (2010)
components
EKF-PCA-ANN-SOM Lab-scale distillation column Decision fusion strategy Ghosh et al. (2011)
TEP Prediction accuracy
Detection and diagnosis delay

bridging different theories and concepts. This issue requires to in-
vestigate the possibilities for modifying some features of the exist-
ing approaches in order to make them compatible with others.
6.2. Main assumptions
Integrating various methods seems to be a straightforward exer-
cise but it is never the case in practice, since each one of them re-
lies on its own assumptions. Indeed, a method that performs well
in one condition may not provide a satisfactory performance in an-
other condition. Thus, taking into account several assumptions may
lead to reduce the scope of the hybrid approaches to be consid-
ered.
6.3. Desirable characteristics (criteria)
The utility of hybrid approaches is to arrive at a more com-
prehensive solution in order to increase the overall confidence in
the diagnosis and the performance. But, how to measure such im-
provement quantitatively?
Obviously, it is useful to have a set of criteria to compare vari-
ous hybrid methods between each other. Diagnostic systems based
on hybrid approaches should meet the desirable characteristics of
classical diagnostic systems, such as quick detection and diagnosis
with a tolerable false alarm rate, robustness to noise and uncer-
tainties, novelty and multiple fault identifiability, isolability, stor-
age and computational requirements...etc. A detailed discussion in
these criteria can be found in Venkatasubramanian et al. (2003a).
In addition to these criteria, the design of an effective hybrid fault
detection and diagnosis method can include other desirable fea-
tures such as:
an effective way to combine the results in order to keep the
correct decisions while removing incorrect ones.
To our knowledge, a work that compares and discusses the per-
formance of hybrid approaches was not published yet in the liter-
ature. Hence, the development of a database to measure and com-
pare the performance of hybrid approaches is worth investigating.
This represents another research opportunity for the FD-HM com-
munity.
6.4. Generic framework for implementation
It seems important to point out the fact that once the hybrid
approach is designed, one needs to implement it properly. Indeed,
a hybrid approach represents a combination of several methods
that are implemented separately and in different ways. Thus, en-
countering interface issues when transposing it into computer code
is inevitable. Therefore, it is useful to propose a generic framework
in order to simplify the implementation of hybrid approaches.
This work has pointed out several research opportunities that
must be addressed for the development of effective hybrid ap-
proaches. It is worth mentioning that detection and diagnosis
could interact with other modules of fault management such as
prognosis module which is devoted to estimate the remaining
useful life of the critical component/subsystem/system. Prognosis
helps to avoid the occurrence of abnormal events and to reduce
production loss. Combined with fault detection and diagnosis, it
could provide global and comprehensive solutions to fault manage-
ment. Thus, achieving unified frameworks for detection, diagnosis
and prognosis, based on hybrid approaches will represent an active
field of research for the next years.

• Completeness, which means a total coverage of the faults.

• Few requirements and a priori information. The objective is to Funding
provide a better solution with fewer requirements.
• Effective decision fusion strategies, to avoid conflict in results Khaoula TIDRIRI is supported by a PhD grant from the Ministry
and to reach an agreement quickly. This explains the need for of Higher Education and Research.
78 K. Tidriri et al. / Annual Reviews in Control 42 (2016) 63–81
References
Alzghoul, A., Backe, B., Lfstrand, M., Bystrm, A., & Liljedahl, B. (2014). Comparing a
knowledge-based and a data-driven method in querying data streams for sys-
tem fault detection: A hydraulic drive system application. Computers in Industry,
65(8), 1126–1135.
Amin, R. W., & Miller, R. W. (1993). A robustness study of X charts with variable
sampling intervals. Journal of Quality Technology, 25, 36.
Antal, P., Fannes, G., Timmerman, D., Moreau, Y., & De Moor, B. (2003). Bayesian
applications of belief networks and multilayer perceptrons for ovarian tumor
classification with rejection. Artificial Intelligence in Medicine, 29(1), 39–60.
Atoui, M. A., Verron, S., & Kobi, A. (2015a). A Bayesian network dealing with mea-
surements and residuals for system monitoring. Transactions of the Institute of
Measurement and Control. 0142331215581446.
Atoui, M. A., Verron, S., & Kobi, A. (2015b). Fault detection and diagnosis in
a Bayesian Network classifier incorporating probabilistic boundary 1. IFAC-
PapersOnLine, 48(21), 670–675.
Atoui, M. A., Verron, S., & Kobi, A. (2015c). Fault detection with Conditional Gaussian
Network. Engineering Applications of Artificial Intelligence, 45, 473–481.
Barat, C., & Ducottet, C. (2016). String representations and distances in deep con-
volutional neural networks for image classification. Pattern Recognition, 54,
104–115.
Bastin, G., & Gevers, M. R. (1988). Stable adaptive observers for nonlinear time-vary-
ing systems. Automatic Control, IEEE Transactions on, 33(7), 650–658.
Benkouider, A. M., Buvat, J. C., Cosmao, J. M., & Saboni, A. (2009). Fault detection in
semi-batch reactor using the EKF and statistical method. Journal of Loss Preven-
tion in the Process Industries, 22(2), 153–161.
Benkouider, A. M., Kessas, R., Yahiaoui, A., Buvat, J. C., & Guella, S. (2012). A hybrid
approach to faults detection and diagnosis in batch and semi-batch reactors by
using EKF and neural network classifier. Journal of Loss Prevention in the Process
Industries, 25(4), 694–702.
Boem, F., Ferrari, R. M., Parisini, T., & Polycarpou, M. M. (2013). Distributed fault
diagnosis for continuous-time nonlinear systems: The inputoutput case. Annual
Reviews in Control, 37(1), 163–169.
Bouamama, B. O., Biswas, G., Loureiro, R., & Merzouki, R. (2014). Graphical meth-
ods for diagnosis of dynamic systems: Review. Annual Reviews in Control, 38(2),
199–219.
Caccavale, F., Pierri, F., Iamarino, M., & Tufano, V. (2009). An integrated approach to
fault diagnosis for a class of chemical batch processes. Journal of process control,
19(5), 827–841.
Chadli, M., Abdo, A., & Ding, S. X. (2013). H-/Hâê¤ fault detection filter design for
discrete-time TakagiSugeno fuzzy system. Automatica, 49(7), 1996–2005.
Chakraborti, S., Human, S. W., & Graham, M. A. (2008). Phase I statistical process
control charts: an overview and some results. Quality Engineering, 21(1), 52–62.
Chan, C. W., Hua, S., & Hong-Yue, Z. (2006). Application of fully decoupled parity
equation in fault detection and identification of DC motors. Industrial Electronics,
IEEE Transactions on, 53(4), 1277–1284.
Chatti, N., Gehin, A., Ould-Bouamama, B., & Merzouki, R. (2013). Functional and be-
havior models for the supervision of an intelligent and autonomous system. Au-
tomation Science and Engineering, IEEE Transactions on, 10(2), 431–445.
Chatti, N., Ould-Bouamama, B., Gehin, A.-L., & Merzouki, R. (2014). Signed bond
graph for multiple faults diagnosis. Engineering Applications of Artificial Intelli-
gence, 36, 134–147.
Chiang, L. H., Braatz, R. D., & Russell, E. L. (2001). Fault detection and diagnosis in
industrial systems. Springer Science & Business Media.
Chou, C.-Y., Chen, C.-H., & Liu, H.-R. (2006). Economic design of EWMA charts with
variable sampling intervals. Quality and Quantity, 40(6), 879–896.
Chow, E. Y., & Willsky, A. S. (1984). Analytical redundancy and the design of ro-
bust failure detection systems. Automatic Control, IEEE Transactions on, 29(7),
603–614.
Coelho, M. L. I., Chakraborti, S., & Graham, M. A. (2015). A comparison of Phase I
control charts. South African Journal of Industrial Engineering, 26(2), 178–190.
Correa, M., Bielza, C., & Pamies-Teixeira, J. (2009). Comparison of Bayesian networks
and artificial neural networks for quality detection in a machining process. Ex-
pert systems with applications, 36(3), 7270–7279.
Crosier, R. B. (1988). Multivariate generalizations of cumulative sum quality-control
schemes. Technometrics, 30(3), 291–303.
Dai, X., & Gao, Z. (2013). From model, signal to knowledge: A data-driven perspec-
tive of fault detection and diagnosis. Industrial Informatics, IEEE Transactions on,
9(4), 2226–2238.
Ding, S. (2008). Model-based fault diagnosis techniques: Design schemes, algorithms,
and tools. Springer Science & Business Media.
Ding, S. X., Zhang, P., Huang, B., & Ding, E. L. (2005). Subspace method aided data–
driven design of observer based fault detection systems. In Proceedings of the
16th IFAC world congress: vol. 16 (pp. 167–172).
Ding, S. X., Zhang, P., Jeinsch, T., Ding, E. L., Engel, P., & Gui, W. (2011). A survey
of the application of basic data-driven and model-based methods in process
monitoring and fault diagnosis. In Proceedings of the 18th IFAC world congress
(pp. 12380–12388).
Ding, S. X., Zhang, P., Naik, A., Ding, E. L., & Huang, B. (2009). Subspace method
aided data-driven design of fault detection and isolation systems. Journal of pro-
cess control, 19(9), 1496–1510.
Ding, X., Guo, L., & Jeinsch, T. (1999). A characterization of parity space and its ap-
plication to robust fault detection. IEEE Transactions on Automatic Control, 44(2),
337–343.
Djeziri, M. A., Merzouki, R., Bouamama, B. O., & Dauphin-Tanguy, G. (2007). Robust
fault diagnosis by using bond graph approach. Mechatronics, IEEE/ASME Transac-
tions on, 12(6), 599–611.
Electric, W. (1956). Statistical quality control handbook. Easton: The Mack Printing
Company.
Garrido, C., De Oa, R., & De Oa, J. (2014). Neural networks for analyzing service qual-
ity in public transportation. Expert Systems with Applications, 41(15), 6830–6838.
Gaymard, S., & Tiplica, T. (2014). Conditional respect towards the pedestrian: Dif-
ference between men and women and risk modeling by the Bayesian approach.
Quality & Quantity, 48(1), 91–110.
Gelso, E. R., Biswas, G., Castillo, S. M., & Armengol, J. (2008). A comparison
of two methods for fault detection: A statistical decision, and an inter-
val-based approach. In 19th international workshop on principles of diagnosis dx
(pp. 261–268). Citeseer.
Gertler, J. (1991). Analytical redundancy methods in fault detection and isolation. In
Preprints of IFAC/IMACS symposium on fault detection, supervision and safety for
technical processes SAFEPROCESS91 (pp. 9–21).
Gertler, J. (1997). Fault detection and isolation using parity relations. Control engi-
neering practice, 5(5), 653–661.
Gertler, J., & Cao, J. (2005). Design of optimal structured residuals from partial prin-
cipal component models for fault diagnosis in linear systems. Journal of Process
Control, 15(5), 585–603.
Gertler, J., Li, W., Huang, Y., & McAvoy, T. (1999). Isolation enhanced principal com-
ponent analysis. AIChE Journal, 45(2), 323–334.
Ghate, V. N., & Dudul, S. V. (2011). Cascade neural-network-based fault classifier for
three-phase induction motor. Industrial Electronics, IEEE Transactions on, 58(5),
1555–1563.
Ghosh, K., Ng, Y. S., & Srinivasan, R. (2011). Evaluation of decision fusion strategies
for effective collaboration among heterogeneous fault diagnostic methods. Com-
puters & chemical engineering, 35(2), 342–355.
He, H., Liu, Z., & Hua, Y. (2015). Adaptive extended kalman filter based fault detec-
tion and isolation for a lithium-ion battery pack. Energy Procedia, 75, 1950–1955.
Hwang, I., Kim, S., Kim, Y., & Seah, C. E. (2010). A survey of fault detection, isolation,
and reconfiguration methods. Control Systems Technology, IEEE Transactions on,
18(3), 636–653.
Isermann, R. (1984). Process fault detection based on modeling and estimation
methodsa survey. Automatica, 20(4), 387–404.
Isermann, R. (1993). Fault diagnosis of machines via parameter estimation and
knowledge processingtutorial paper. Automatica, 29(4), 815–835.
Isermann, R. (2005). Model-based fault-detection and diagnosisstatus and applica-
tions. Annual Reviews in control, 29(1), 71–85.
Isermann, R. (2006). Fault-diagnosis systems: An introduction from fault detection to
fault tolerance. Springer Science & Business Media.
Jha, S., Li, W., & Seshia, S. A. (2009). Localizing transient faults using dynamic
bayesian networks. In High level design validation and test workshop, 2009. HLDVT
2009. IEEE international (pp. 82–87). IEEE.
Joe Qin, S. (2003). Statistical process monitoring: Basics and beyond. Journal of
chemometrics, 17(8–9), 480–502.
Joekes, S., & Barbosa, E. P. (2013). An improved attribute control chart for moni-
toring non-conforming proportion in high quality processes. Control Engineering
Practice, 21(4), 407–412.
Jones-Farmer, L. A., Woodall, W. H., Steiner, S. H., & Champ, C. W. (2014). An
overview of Phase I analysis for process improvement and monitoring. Journal
of Quality Technology, 46(3), 265.
Kalman, R. E. (1960). A new approach to linear filtering and prediction problems.
Journal of basic Engineering, 82(1), 35–45.
Keliris, C., Polycarpou, M. M., & Parisini, T. (2015). A robust nonlinear observer-based
approach for distributed fault detection of inputoutput interconnected systems.
Automatica, 53, 408–415.
Kourti, T., & MacGregor, J. F. (1996). Multivariate SPC methods for process and prod-
uct monitoring. Journal of quality technology, 28(4), 409–428.
Kourti, T., Nomikos, P., & MacGregor, J. F. (1995). Analysis, monitoring and fault diag-
nosis of batch processes using multiblock and multiway PLS. Journal of process
control, 5(4), 277–284.
Kramer, M. A. (1992). Autoassociative neural networks. Computers & chemical engi-
neering, 16(4), 313–328.
Kruger, U., Wang, X., Chen, Q., & Qin, S. J. (2001). An alternative PLS algorithm for
the monitoring of industrial process. In American control conference, 2001. Pro-
ceedings of the 2001: vol. 6 (pp. 4455–4459). IEEE.
Bernal-de Lázaro, J., Llanes-Santiago, O., Prieto-Moreno, A., Knupp, D., & Sil-
va-Neto, A. (2016). Enhanced dynamic approach to improve the detection of
small-magnitude faults. Chemical Engineering Science, 146, 166–179.
Lee, D. S., Park, J. M., & Vanrolleghem, P. A. (2005). Adaptive multiscale principal
component analysis for on-line monitoring of a sequencing batch reactor. Jour-
nal of Biotechnology, 116(2), 195–210.
Lee, J.-M., Yoo, C., Choi, S. W., Vanrolleghem, P. A., & Lee, I.-B. (2004). Nonlinear
process monitoring using kernel principal component analysis. Chemical Engi-
neering Science, 59(1), 223–234.
Li, J., Chan, C., & Zhang, H.-Y. (2005). Asymptotic local approach in fault detec-
tion based on predictive filters. Journal of guidance, control, and dynamics, 28(6),
1112–1122.
Li, J., Jin, J., & Shi, J. (2008). Causation-based t2 decomposition for multivariate pro-
cess monitoring and diagnosis. Journal of Quality Technology, 40(1), 46.
Li, L., Ding, S. X., Yang, Y., & Zhang, Y. (2016). Robust fuzzy observer-based fault de-
tection for nonlinear systems with disturbances. Neurocomputing, 174, 767–772.
 K. Tidriri et al. / Annual Reviews in Control 42 (2016) 63–81
Li, Z., Cheng, L., Qiu, X.-s., & Wu, L. (2009). Fault diagnosis for high-level applications
based on dynamic Bayesian network. In Management enabling the future Internet
for changing business and new computing services (pp. 61–70). Springer.
Liu, J. Y., Xie, M., & Goh, T. N. (2006). CUSUM chart with transformed exponential
data. Communications in StatisticsTheory and Methods, 35(10), 1829–1843.
Liu, X., Gao, X., & Han, J. (2015). Robust unknown input observer based fault detec-
tion for high-order multi-agent systems with disturbances. ISA transactions.
Liu, X., Kruger, U., Littler, T., Xie, L., & Wang, S. (2009). Moving window kernel
PCA for adaptive monitoring of nonlinear processes. Chemometrics and Intelli-
gent Laboratory Systems, 96(2), 132–143.
Ljung, L. (1987). System identification: Theory for the user. Englewood Cliffs.
Lowry, C. A., & Montgomery, D. C. (1995). A review of multivariate control charts.
IIE transactions, 27(6), 800–810.
Lowry, C. A., Woodall, W. H., Champ, C. W., & Rigdon, S. E. (1992). A multivariate ex-
ponentially weighted moving average control chart. Technometrics, 34(1), 46–53.
Lu, J., Bai, C., & Zhang, G. (2009). Cost-benefit factor analysis in e-services using
Bayesian networks. Expert Systems with Applications, 36(3), 4617–4625.
Lu, N., Wang, F., & Gao, F. (2003). Combination method of principal component and
wavelet analysis for multivariate process monitoring and fault diagnosis. Indus-
trial & engineering chemistry research, 42(18), 4198–4207.
Lucas, J. M. (1982). Combined Shewhart-CUSUM quality control schemes. Journal of
Quality Technology, 14(2), 51–59.
Lucas, J. M., & Saccucci, M. S. (1990). Exponentially weighted moving average control
schemes: Properties and enhancements. Technometrics, 32(1), 1–12.
Luenberger, D. G. (1966). Observers for multivariable systems. Automatic Control,
IEEE Transactions on, 11(2), 190–197.
López-Estrada, F. R., Ponsart, J.-C., Theilliol, D., Zhang, Y., & Astorga-
Zaragoza, C.-M. (2015). LPV model-based tracking control and robust sen-
sor fault diagnosis for a quadrotor UAV. Journal of Intelligent & Robotic Systems,
1–15.
MacGregor, J. F., Jaeckle, C., Kiparissides, C., & Koutoudi, M. (1994). Process monitor-
ing and diagnosis by multiblock PLS methods. AIChE Journal, 40(5), 826–838.
Mason, R. L., Tracy, N. D., & Young, J. C. (1995). Decomposition of T2 for multivariate
control chart interpretation. Journal of quality technology, 27(2), 99–1108.
Maurya, M. R., Rengaswamy, R., & Venkatasubramanian, V. (2007). A signed directed
graph and qualitative trend analysis-based framework for incipient fault diagno-
sis. Chemical Engineering Research and Design, 85(10), 1407–1422.
Mazurowski, M. A., Habas, P. A., Zurada, J. M., Lo, J. Y., Baker, J. A., &
Tourassi, G. D. (2008). Training neural network classifiers for medical decision
making: The effects of imbalanced datasets on classification performance. Neu-
ral networks, 21(2), 427–436.
Montgomery, D. C. (2007). Introduction to statistical quality control. John Wiley &
Sons.
Mulumba, T., Afshari, A., Yan, K., Shen, W., & Norford, L. K. (2015). Robust
model-based fault diagnosis for air handling units. Energy and Buildings, 86,
698–707.
Murphy, K. P. (2002). Dynamic bayesian networks: Representation, inference and learn-
ing. University of California, Berkeley Ph.D. thesis..
Mylaraswamy, D., & Venkatasubramanian, V. (1997). A hybrid framework for large
scale process fault diagnosis. Computers & Chemical Engineering, 21, S935–S940.
Narasimhan, S., Vachhani, P., & Rengaswamy, R. (2008). New nonlinear residual
feedback observer for fault diagnosis in nonlinear systems. Automatica, 44(9),
2222–2229.
Nielsen, T. D. (2007). Bayesian Networks and decision graphs. Springer.
Odendaal, H. M., & Jones, T. (2014). Actuator fault detection and isolation: An opti-
mised parity space approach. Control Engineering Practice, 26, 222–232.
Ould-Bouamama, B., El Harabi, R., Abdelkrim, M. N., & Gayed, M. B. (2012). Bond
graphs for the diagnosis of chemical processes. Computers & chemical engineer-
ing, 36, 301–324.
Page, E. S. (1954). Continuous inspection schemes. Biometrika, 41(1/2), 100–115.
Pan, S., Xiao, D., Xing, S., Law, S. S., Du, P., & Li, Y. (2016). A general extended Kalman
filter for simultaneous estimation of system and unknown inputs. Engineering
Structures, 109, 85–98.
Patton, R. J., & Chen, J. (1997). Observer-based fault detection and isolation: robust-
ness and applications. Control Engineering Practice, 5(5), 671–682.
Paynter, H. M. (1961). Analysis and design of engineering systems. MIT press.
Pearl, J. (2014). Probabilistic reasoning in intelligent systems: Networks of plausible in-
ference. Morgan Kaufmann.
Phillpotts, D. (2007). Nonlinear fault detection and diagnosis using kernel PCA applied
to a pilot distillation column. university of Pretoria Master’s thesis.
Pierri, F., Paviglianiti, G., Caccavale, F., & Mattei, M. (2008). Observer-based sensor
fault detection and isolation for chemical batch reactors. Engineering Applica-
tions of Artificial Intelligence, 21(8), 1204–1216.
Portnoy, I., Melendez, K., Pinzon, H., & Sanjuan, M. (2016). An improved weighted
recursive PCA algorithm for adaptive fault detection. Control Engineering Practice,
50, 69–83.
Potter, I. E., & Sunman, M. C. (1977). Threshold-less redundancy management with
arrays of skewed instruments.
Prabhu, S. S., & Runger, G. C. (1997). Designing a multivariate EWMA control chart.
Journal of Quality Technology, 29(1), 8.
Puig, V., Oca, S. M., & Blesa, J. (2013). Adaptive threshold generation in robust
fault detection using interval models: Time-domain and frequency-domain ap-
proaches. International Journal of Adaptive Control and Signal Processing, 27(10),
873–901.
Reynolds, M. R., Amin, R. W., & Arnold, J. C. (1990). CUSUM charts with variable
sampling intervals. Technometrics, 32(4), 371–384.
79
Riaz, M., Ahmad, S., Khaliq, Q., Riaz, M., & Ahmad, S. (2016). On designing a new
tukey-ewma control chart for process monitoring. The International Journal of
Advanced Manufacturing Technology, 82(1–4), 1–23.
Roberts, S. W. (20 0 0). Control chart tests based on geometric moving averages. Tech-
nometrics, 42(1), 97–101.
Rusinov, L. A., Vorobiev, N. V., & Kurkina, V. V. (2013). Fault diagnosis in chemical
processes and equipment with feedbacks. Chemometrics and Intelligent Labora-
tory Systems, 126, 123–128.
Samantaray, A. K., & Bouamama, B. O. (2008). Model-based process supervision: A
bond graph approach. Springer Science & Business Media.
Sargent, R. G. (2013). Verification and validation of simulation models. Journal of
simulation, 7(1), 12–24.
Schubert, U., Kruger, U., Arellano-Garcia, H., de S Feital, T., & Wozny, G. (2011). Uni-
fied model-based fault diagnosis for three industrial application studies. Control
Engineering Practice, 19(5), 479–490.
Sesen, M. B., Nicholson, A. E., Banares-Alcantara, R., Kadir, T., & Brady, M. (2013).
Bayesian networks for clinical decision support in lung cancer care. PloS one,
8(12), e82349.
Shewhart, W. A. (1931). Economic control of quality of manufactured product. ASQ
Quality Press.
Siswantoro, J., Prabuwono, A. S., Abdullah, A., & Idrus, B. (2016). A linear model
based on Kalman filter for improving neural network classification performance.
Expert Systems with Applications, 49, 112–122.
Sorsa, T., Koivo, H. N., & Koivisto, H. (1991). Neural networks in process fault diag-
nosis. Systems, Man and Cybernetics, IEEE Transactions on, 21(4), 815–825.
Sotomayor, O. A., & Odloak, D. (2005). Observer-based fault diagnosis in chemical
plants. Chemical Engineering Journal, 112(1), 93–108.
Stoumbos, Z. G., & Sullivan, J. H. (2002). Robustness to non-normality of the multi-
variate EWMA control chart. Journal of Quality Technology, 34(3), 260.
Sullivan, J. H., & Jones, L. A. (2002). A self-starting control chart for multivariate
individual observations. Technometrics, 44(1), 24–33.
Tao, J., Li, Q., Zhu, C., & Li, J. (2012). A hierarchical naive Bayesian network clas-
sifier embedded GMM for textural image. International Journal of Applied Earth
Observation and Geoinformation, 14(1), 139–148.
Tarantino, R., Szigeti, F., & Colina-Morles, E. (20 0 0). Generalized Luenberger ob-
server-based fault-detection filter design: An industrial application. Control En-
gineering Practice, 8(6), 665–671.
Tornil-Sin, S., Ocampo-Martinez, C., Puig, V., & Escobet, T. (2012). Robust fault de-
tection of non-linear systems using set-membership state estimation based on
constraint satisfaction. Engineering Applications of Artificial Intelligence, 25(1),
1–10.
Tsai, C.-F., & Wu, J.-W. (2008). Using neural network ensembles for bankruptcy pre-
diction and credit scoring. Expert Systems with Applications, 34(4), 2639–2649.
Varga, A., & Ossmann, D. (2014). LPV model-based robust diagnosis of flight actuator
faults. Control Engineering Practice, 31, 135–147.
Vedam, H., Dash, S., & Venkatasubramanian, V. (1999). An intelligent operator deci-
sion support system for abnormal situation management. Computers & Chemical
Engineering, 23, S577–S580.
Venkatasubramanian, V., Rengaswamy, R., Yin, K., & Kavuri, S. N. (2003a). A review
of process fault detection and diagnosis: Part I: Quantitative model-based meth-
ods. Computers & Chemical Engineering, 27(3), 293–311.
Venkatasubramanian, V., Rengaswamy, R., & Kavuri, S. N. (2003b). A review of pro-
cess fault detection and diagnosis: Part II: Qualitative models and search strate-
gies. Computers & Chemical Engineering, 27(3), 313–326.
Venkatasubramanian, V., Rengaswamy, R., Kavuri, S. N., & Yin, K. (2003c). A review
of process fault detection and diagnosis: Part III: Process history based methods.
Computers & Chemical Engineering, 27(3), 327–346.
Verron, S., Li, J., & Tiplica, T. (2010a). Fault detection and isolation of faults in a
multivariate process with Bayesian network. Journal of Process Control, 20(8),
902–911.
Verron, S., Tiplica, T., & Kobi, A. (2008). Fault detection and identification with a
new feature selection based on mutual information. Journal of Process Control,
18(5), 479–490.
Verron, S., Tiplica, T., & Kobi, A. (2010b). Fault diagnosis of industrial systems by
conditional Gaussian network including a distance rejection criterion. Engineer-
ing applications of artificial intelligence, 23(7), 1229–1235.
Vinodhini, G., & Chandrasekaran, R. M. (2016). A comparative performance evalua-
tion of neural network based approach for sentiment classification of online re-
views. Journal of King Saud University-Computer and Information Sciences, 28(1),
2–12.
Walker, E., Rayman, S., & White, R. E. (2015). Comparison of a particle filter and
other state estimation methods for prognostics of lithium-ion batteries. Journal
of Power Sources, 287, 1–12.
Wang, J., & Qin, S. J. (2002). A new subspace identification approach based on prin-
cipal component analysis. Journal of Process Control, 12(8), 841–855.
Wang, K.-J., Makond, B., & Wang, K.-M. (2014). Modeling and predicting the occur-
rence of brain metastasis from lung cancer by Bayesian network: A case study
of Taiwan. Computers in biology and medicine, 47, 147–160.
Weber, P., Theilliol, D., Aubrun, C., & Evsukoff, A. (2007). Increasing effectiveness of
model-based fault diagnosis: A dynamic bayesian network design for decision
making. In In proceedings of the conference on fault detection, supervision and
safety of technical processes (pp. 90–95).
Weidl, G., Madsen, A. L., & Israelson, S. (2005). Applications of object-oriented
Bayesian networks for condition monitoring, root cause analysis and decision
support on operation of complex continuous processes. Computers & chemical
engineering, 29(9), 1996–2009.
80 K. Tidriri et al. / Annual Reviews in Control 42 (2016) 63–81
Xiaotian, L., & Anlin, W. (2014). Definitions of causality in bond graph model for
efficient simulation mechanism. Mechanism and Machine Theory, 80, 112–124.
Xie, L., Lin, X., & Zeng, J. (2013). Shrinking principal component analysis for en-
hanced process monitoring and fault isolation. Industrial & Engineering Chem-
istry Research, 52(49), 17475–17486.
Xiong, J., Jauberthie, C., Trave-Massuyes, L., & Le Gall, F. (2013). Fault detection using
interval kalman filtering enhanced by constraint propagation. In Decision and
control (cdc), 2013 IEEE 52nd annual conference on (pp. 490–495). IEEE.
Yan, L. J., & Cercone, N. (2010). Bayesian network modeling for evolutionary genetic
structures. Computers & Mathematics with Applications, 59(8), 2541–2551.
Yang, Y., Ding, S. X., & Li, L. (2015). On observer-based fault detection for nonlinear
systems. Systems & Control Letters, 82, 18–25. doi:10.1016/j.sysconle.2015.05.004.
Yashchin, E. (1989). Weighted cumulative sum technique. Technometrics, 31(3),
321–338.
Ye, H., Wang, G., & Ding, S. X. (2004). A new parity space approach for fault detec-
tion based on stationary wavelet transform. Automatic Control, IEEE Transactions
on, 49(2), 281–287.
Ye, H., Wang, G. Z., Ding, S. X., & Su, H. Y. (2002). An IIR filter based parity space
approach for fault detection. In Proceedings of the 15th Triennial world congress
of the international federation of automatic control.
Ye, H., Zhang, P., Ding, S. X., & Wang, G. Z. (20 0 0). A time-frequency domain
fault detection approach based on parity relation and wavelet transform. In
Decision and control, 2000. Proceedings of the 39th IEEE conference on: vol. 4
(pp. 4156–4161). IEEE.
Yu, J. (2012). Local and global principal component analysis for process monitoring.
Journal of Process Control, 22(7), 1358–1373.
Yu, J., & Rashid, M. M. (2013). A novel dynamic Bayesian network-based networked
process monitoring approach for fault detection, propagation identification, and
root cause diagnosis. AIChE Journal, 59(7), 2348–2365.
Zaidi, A., Tagina, M., & Bouamama, B. O. (2010). Reliability data for improvement
of decision-making in Analytical Redundancy Relations Bond Graph based di-
agnosis. In Advanced intelligent mechatronics (AIM), 2010 IEEE/ASME international
conference on (pp. 790–795). IEEE.
Zarei, J., & Shokri, E. (2014). Robust sensor fault detection based on nonlinear un-
known input observer. Measurement, 48, 355–367.
Zhang, G. P. (20 0 0). Neural networks for classification: A survey. Systems, Man,
and Cybernetics, Part C: Applications and Reviews, IEEE Transactions on, 30(4),
451–462.
Zhang, X., & Hoo, K. A. (2011). Effective fault detection and isolation using bond
graph-based domain decomposition. Computers & chemical engineering, 35(1),
132–148.
Zhang, Y., Li, S., & Teng, Y. (2012). Dynamic processes monitoring using recursive
kernel principal component analysis. Chemical Engineering Science, 72, 78–86.
Zhang, Z., & Dong, F. (2014). Fault detection and diagnosis for missing data systems
with a three time-slice dynamic Bayesian network approach. Chemometrics and
Intelligent Laboratory Systems, 138, 30–40.
Zhao, Y., Wen, J., & Wang, S. (2015). Diagnostic Bayesian networks for diagnosing
air handling units faultsPart II: Faults in coils and sensors. Applied Thermal En-
gineering, 90, 145–157.
Zhao, Y., Xiao, F., & Wang, S. (2013). An intelligent chiller fault detection and di-
agnosis methodology using Bayesian belief network. Energy and Buildings, 57,
278–288.
Zhong, M., Song, Y., & Ding, S. X. (2015). Parity space-based fault detection for linear
discrete time-varying systems with unknown input. Automatica, 59, 120–126.
Zhou, J., Guo, A., Celler, B., & Su, S. (2014). Fault detection and identification span-
ning multiple processes by integrating PCA with neural network. Applied Soft
Computing, 14, 4–11.
 K. Tidriri et al. / Annual Reviews in Control 42 (2016) 63–81 81

K. Tidriri was born in 1992. She received her Engineering degree in Electronic and Automatic Control from Institut Polytechnique de Grenoble
(Grenoble INP) in the year of 2015. Since october 2015, she is performing a PhD in the monitoring of dynamic systems in the Laboratoire Angevin
de Recherche en Ingenierie des Systemes (LARIS, EA 7315). Her research activities are mainly devoted to data-driven, model-based and hybrid
approaches for fault detection and diagnosis.

N. Chatti was born in 1985. He obtained a M.Sc. from Ecole Centrale de Lille and a PhD in Automatic Control From Ecole Polytechnique de Lille in
2010, 2013 respectively. Since september 2014, he is associate professor in the public engineering school ISTIA (University of Angers) and researcher
in LARIS laboratory EA 7315. His research interests include Safety System Design, Robust Fault Diagnosis and interval analysis.

S. Verron received the M.S. degree in Industrial Engineering from Institut National Polytechnique de Toulouse (INPT) in 2004. He received his
doctoral degree in Industrial Engineering from University of Angers in 2007. He is currently associate professor at LARIS Laboratory of ISTIA of the
University of Angers (France). His research interests are in the monitoring of complex processes (fault detection and root cause diagnosis).

T. Tiplica, Ph.D., is an associate professor at the Institute of Engineering Sciences and Technologies of LUNAM University, France. He is a member
of the Laboratoire Angevin de Recherche en Ingnierie de Systemes (LARIS, EA 7315). His research interests include statistical process control, fault
detection and isolation, and Bayesian networks.