Computers & Industrial Engineering 139 (2020) 106188

Contents lists available at ScienceDirect

Computers & Industrial Engineering

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

A novel automata and neural network based fault diagnosis system for PLC T
controlled manufacturing systems
Arup Ghosha, Gi-Nam Wangb, Jooyeoun Leeb,
⁎

a
School of Information Technology and Engineering, Vellore Institute of Technology, Vellore, India
b
Department of Industrial Engineering, Ajou University, Suwon, South Korea

ARTICLE INFO ABSTRACT

Keywords: The Fault and Anomaly Detection and Isolation (FADI) in Programmable Logic Controller (PLC) controlled
Autoassociative neural network (AANN) systems is an important and challenging problem. In this paper, we present an automated tool, called the
Deterministic finite-state automaton (DFA) Manufacturing Process Failure Diagnosis Tool (MPFDT) that can detect and isolate the faults and anomalies in
Electrical energy usage monitoring the PLC controlled manufacturing systems effectively. MPFDT utilizes two independent knowledge-based pro-
Fault detection and isolation
cess behaviour models of the manufacturing system to satisfy the FADI purpose. The fundamental idea is to
Programmable logic controller (PLC)
Signal processing
detect the inconsistencies between the modelled and the observed manufacturing process behaviour. The first
model is a Deterministic Finite-state Automaton (DFA) based model of the PLC control process that is used to
determine whether the observed state transition behaviour of the PLC control process is consistent with the
modelled state transition behaviour or not. The second model is basically a set of Artificial Neural Network
(ANN) based one-class classifiers that are used to identify whether any significant difference exist between the
observed and the reference electrical power consumption profile of the manufacturing system or not. The ex-
perimental results show that the FADI accuracy rate of the proposed tool is very high (more than 98%).

1. Introduction In this paper, we present an automated tool, called the

Manufacturing Process Failure Diagnosis Tool (MPFDT) that can detect
The Programmable Logic Controllers (PLCs) are a special type of and isolate the faults and anomalies in a PLC controlled manufacturing
microprocessor-based digital computers that are used for the automa- system effectively. MPFDT uses two independent knowledge-based
tion of various electro-mechanical processes in industries. Even after process behaviour models of the manufacturing system to solve the
more than four decades since their introduction, most of the existing Fault and Anomaly Detection and Isolation (FADI) problem. The first
manufacturing plants are still controlled by PLCs (Ghosh, Qin, Lee, model i.e., the Deterministic Finite-state Automaton (DFA) based pro-
Wang, 2016; Ghosh, Qin, Lee, Wang, 2017). However, the PLC device cess behaviour model is used to define the fault-free state transition
has a very limited fault and anomaly detection capability (Ghosh et al., behaviour of the PLC control process. Most of the times, if a failure
2016; Ghosh, Qin, Lee, et al., 2017; Hu, Starr, & Leung, 1999). For this occurs in any device or part of the manufacturing system, then it
reason, it is absolutely necessary to have an additional software module completely alters the normal state transition behaviour of the PLC
or tool that can detect and isolate different kinds of faults and control process. So, MPFDT can easily detect and isolate it by using the
anomalies in the PLC controlled manufacturing system automatically DFA based control process model. The log data records of the PLC
and accurately. This can drastically shorten the standard debugging or control signals are utilized to fulfil this particular objective.
diagnosis time and hence, can greatly increase the availability and Unfortunately, not all types of faults and anomalies in the manu-
productivity of the manufacturing system. Moreover, in general, the facturing system (especially that do not affect much the state transition
detection of a large number of anomalies within a short period of time behaviour of the PLC control process) can be detected by using the
and/or the detection of an anomaly that exhibits a significant deviation control process model. So, another independent and complementary
from the normal process behaviour indicate a potential future fault by knowledge-based process behaviour model is introduced. The second
which the preventive maintenance tasks can be appropriately scheduled process behaviour model is composed of a set of Artificial Neural
(Ghosh et al., 2016; Ghosh, Qin, Lee, et al., 2017). This can save the Network (ANN) based one-class classifiers. Each ANN classifier is in-
organization a hefty amount of money. dividually trained using the feature vectors extracted from a specific

⁎
Corresponding author.
E-mail addresses: [email protected] (A. Ghosh), [email protected] (G.-N. Wang), [email protected] (J. Lee).

https://doi.org/10.1016/j.cie.2019.106188
Received 12 February 2019; Received in revised form 2 September 2019; Accepted 15 November 2019
Available online 25 November 2019
0360-8352/ © 2019 Elsevier Ltd. All rights reserved.
A. Ghosh, et al. Computers & Industrial Engineering 139 (2020) 106188

region or segment of the reference electrical energy or power con- Lunze & Schroder, 2004; Roth, Lesage, & Litz, 2009, 2011; Roth,
sumption signals of the manufacturing system in such a way that the Schneider, Lesage, & Litz, 2012; Sampath, Lafortune, & Teneketzis,
classifier set can precisely determine any significant changes between 1998; Sampath, Sengupta, Lafortune, Sinnamohideen, & Teneketzis,
the observed and the reference power consumption profile of the 1996; Schneider, Litz, & Danancher, 2011). The centralized FDI ap-
manufacturing system. We have experimentally found that the combi- proaches basically differ in the choice of modelling formalism. Briefly,
nation of the above two techniques always yields a very high FADI their modelling techniques can be grouped into three general cate-
accuracy (above 98%). gories: (i) Automata (and their timed and probabilistic extensions)
The organization of the remainder of the paper is as follows. Section based modelling (see Ghosh, Qin, Lee, et al., 2017; Ghosh et al., 2016;
2 presents a detailed review of the existing literature in this field. A Bao et al., 2014; Ghosh, Qin, Wang, et al., 2017; Hashtrudi Zad et al.,
brief overview of how MPFDT tool works is given in Section 3. We 2005; Lunze & Schroder, 2004; Roth et al., 2009, 2011, 2012; Sampath
thoroughly discuss the FADI procedure of MPFDT based on the first et al., 1998, 1996; Schneider et al., 2011); (ii) Petri nets based mod-
process behaviour model (i.e., the DFA based control process model) elling (see Dotoli et al., 2009; Cabasino, Giua, & Seatzu, 2010;
and the second process behaviour model (i.e., the set of ANN based Cabasino, Giua, & Seatzu, 2013); and (iii) Mathematical and temporal
signal deviation detectors) in Section 4 and Section 5, respectively. logic based modelling (see Hu et al., 1999; Hu et al., 2003; Jiang &
Section 6 contains our conclusive remarks of the work followed by a list Kumar, 2004). Among the aforementioned articles, the studies in re-
of relevant and state-of-the-art references. Further technical details of ferences (Bao et al., 2014; Ghosh et al., 2016; Ghosh, Qin, Lee, et al.,
some proposed methodologies and the experimental results (omitted 2017; Ghosh, Qin, Wang, et al., 2017; Hu et al., 1999, 2003; Roth et al.,
due to space considerations) are presented in Appendix A and Appendix 2009, 2011, 2012; Schneider et al., 2011) are specifically proposed for
B, respectively (the appendix sections can be found in the supplemen- the PLC controlled manufacturing systems. It is easy to see that the vast
tary file ‘MPFDT-Supplementary.docx’). majority of the centralized FDI approaches follow the automata based
modelling formalism (see for instance Ghosh et al., 2016; Ghosh, Qin,
2. Background study and literature review Lee, et al., 2017; Ghosh, Qin, Wang, et al., 2017; Bao et al., 2014; Roth
et al., 2009, 2012, 2011; Schneider et al., 2011). The main advantages
The PLC controlled manufacturing systems are a special type of of the automata based centralized FDI approaches are their diagnosis
Discrete Event System (DES) (Cassandras & Lafortune, 2008). Over the precision, speed and conceptual simplicity (Ghosh, Qin, Lee, et al.,
past three decades, several research works have been conducted on the 2017; Roth et al., 2011, 2012). Therefore, below we restrict our at-
Fault Detection and Isolation (FDI) in DESs. All those approaches can tention to only the automata based approaches.
broadly be divided into two major categories: the decentralized and the
centralized FDI approaches. 2.2.1. The drawbacks of the automata based centralized FDI approaches
Although these approaches use different types of automata models,
2.1. Decentralized fault diagnosis approaches the underlying fundamental modelling concept and fault diagnosis
mechnism are almost the same. The accuracy of the centralized FDI
These FDI approaches are proposed for large modular dynamic DESs approaches does not differ much based on the type and specification of
(for example, the manufacturing plants) that are composed of several the automata model, but it actually varies depending on how much
coupled subsystems. Important examples of the decentralized FDI ap- control process information is incorporated into the model (this is be-
proaches include (Cabasino, Giua, Paoli, & Seatzu, 2013; Kumar & cause the faults and anomalies are detected from the differences in the
Takai, 2009; Qiu & Kumar, 2006; Sayed-Mouchaweh & Lughofer, control process information acquired before and after their occur-
2015). rences). So, instead of designing a completely new automata model,
from the point of view of the authors, it is more apposite to include
2.1.1. The drawbacks of the decentralized FDI approaches additional control process information into the current (possibly) best
The decentralized FDI approaches are designed to deal with regular existing model. Among the aforementioned fault diagnosis approaches
DESs. Unlike other DESs, if a PLC controlled manufacturing system and software tools, FBMTP (introduced by Ghosh, Qin, Lee, et al.
becomes large enough, then it is partitioned into several (both physi- (2017) – website: http://www.udmtek.com/ksub05_03/articles/view/
cally and operationally) disjoint subsystems (Ghosh, Qin, Lee, et al., tableid/ksub05_03/id/40) is the most accurate and precise one (for
2017). This is because the PLC device has a very limited input/output solving the FADI problem in PLC controlled manufacturing systems).
(I/O) capability. Each subsystem controlled by a separate PLC does not This is primarily because of the following reasons:
have any interaction or communication with the other subsystems (i.e.,
all the subsystems are independent and standalone subsystems) (Ghosh, i. FBMTP is specifically devised and implemented for the PLC con-
Qin, Lee, et al., 2017). So, a PLC controlled manufacturing system does trolled manufacturing systems.
not ‘truly’ function as a decentralized DES. For this reason, the decen- ii. It introduces a novel DFA based control process model, called the
tralized FDI approaches are not really suitable for solving the FDI DFA with Signal-status Vector and Time-out Function (DSVTF)
problem in the PLC controlled manufacturing systems. model that incorporates all the best features or characteristics of the
earlier models into a single unified model (for example, the in-
2.2. Centralized fault diagnosis approaches formation associated with the state transition time, control signal
event sequence, multi-state transition execution pattern, etc. is in-
In centralized DESs, there exists a single central computer system cluded in the model – also see Ghosh et al., 2016; Ghosh, Qin,
that is responsible for controlling or monitoring the complete system Wang, et al., 2017; Bao et al., 2014; Roth et al., 2009, 2012, 2011;
operation (in other words, has access to all the control process in- Schneider et al., 2011).
formation). So, these approaches propose a centralized fault diagnosis iii. Unlike other existing approaches, in FBMTP, the distinctive opera-
framework (or a global diagnoser) that takes the final or global diag- tional characteristics of different types of extreme control process
nosis decision based on a single control process model of the whole situations (such as the soft fault propagation, infinite looping at a
system. Important examples of such works include (Bao, Wu, & Yan, state, etc.) are taken into account at the time of fault diagnosis; and
2014; Cabasino, Giua, & Seatzu, 2010, 2013; Dotoli, Fanti, Mangini, & sufficient actions are taken to deal with the log data inconsistency
Ukovich, 2009; Ghosh et al., 2016; Ghosh, Qin, Lee, et al., 2017; Ghosh, issues associated with large manufacturing systems (such as the
Qin, Wang, Lee, & Jang, 2017; Hashtrudi Zad, Kwong, & Wonham, event sequence disorder problem, critical signal miss event issues,
2005; Hu et al., 1999; Hu, Starr, & Leung, 2003; Jiang & Kumar, 2004; etc.).

2
A. Ghosh, et al. Computers & Industrial Engineering 139 (2020) 106188

All the above features make FBMTP best suitable to solve the FADI analog output signals and are transmitted to the physical actuators. In
problem in the PLC controlled manufacturing systems (for more details, the PLC I/O memory table, these analog inputs and outputs are gen-
see Ghosh, Qin, Lee, et al., 2017). However, FBMTP has several in- erally stored as 12-bit integer numbers (ranging between 0 and 4000)
herent shortcomings (the reader should note that the limitations dis- (Jack, 2010; Laughton & Warne, 2003).
cussed here are equally valid for all other aforementioned FDI ap- As can be seen in Fig. 1, the data logging module of MPFDT con-
proaches). Firstly, in FBMTP, it is simply assumed that the PLC I/O tinually records the I/O signal data from the PLC memory table during
signals i.e., the sensor and actuator signals can take only two possible the communication phase and inserts them into the log database for
values: 1 and 0 (implies that every PLC I/O signal is essentially a binary further analysis (the details of the PLC I/O signal data logging proce-
digital signal). However, this is not always the case. Most of the modern dure can be found in Section III of Ghosh, Qin, Lee, et al. (2017) – for
PLC devices have an analog I/O card that does accept the continuous or the convenience of the readers, we have briefly discussed it in Appendix
analog I/O signals (Jack, 2010; Laughton & Warne, 2003). Un- A.I). Each log data record is basically an integer vector (or array) that
fortunately, FBMTP cannot handle such analog I/O signals. Secondly, contains the (integer) status values of all the PLC I/O signals, where the
FBMTP detects the faults and anomalies based solely on a nominal ith element of the vector represents the status value of the ith PLC I/O
control process model and hence, it cannot identify all types of faults signal (please note that the boolean ON/OFF status of a binary digital I/
and anomalies in the manufacturing system (particularly ones that do O signal is also represented as an integer number). MPFDT inserts into
not alter much the state transition behaviour of the PLC control process) the database a new log data record only when at least one I/O signal
(Ghosh, Qin, Lee, et al., 2017). changes its status value (because, in the MDSVTF model, only the in-
In this paper, we present a novel fault and anomaly diagnosis tool, formation associated with the state transition behaviour of the PLC
called MPFDT that can easily overcome the above stated shortcomings. control process is included) (Ghosh, Qin, Lee, et al., 2017; Roth et al.,
MPFDT resolves (part of) the first issue (see the paragraph above) by 2011, 2012). MPFDT at first constructs a nominal DFA based control
incorporating the relevant control process information associated with process model (i.e., the MDSVTF model) from the reference log data
the analog I/O signals into the DSVTF model (the control process model records taken from the fault-free manufacturing system (see Fig. 1).
of FBMTP (Ghosh, Qin, Lee, et al., 2017), while keeping the basic Once the control process model is built, it can be continuously used to
structure of the DSVTF model unaltered (the rationale is discussed detect and isolate the faults and anomalies in the manufacturing
earlier in this section). The second issue is addressed by using a set of system. For simplicity, if any system state transition is found in the
ANN based signal deviation detectors (they are basically one-class observed log data records that is not similar to any state transition in
outlier or novelty detectors). As discussed previously, these signal de- the MDSVTF model, then we can conclude that a fault or anomaly has
viation detectors are employed to determine whether there is any sig- occurred in the manufacturing system (the details of this procedure are
nificant deviation in the power consumption behaviour of the manu- described in Section 4) (Bao et al., 2014; Ghosh et al., 2016; Ghosh,
facturing system. In general, the occurrence of a fault or anomaly Qin, Lee, et al., 2017; Ghosh, Qin, Wang, et al., 2017; Roth et al., 2009,
causes significant changes in the power consumption behaviour (even 2011, 2012; Schneider et al., 2011). This checking is performed for
though it does not necessarily affect the control process behaviour) of each and every log data record in the PLC I/O signal log database (or for
the manufacturing system and hence, can easily be identified using the each observed system state transition). At this point, we should mention
signal deviation detectors. It is easy to perceive that these measures can that the term ‘analog signal’ is simply being used to indicate the original
dramatically increase the accuracy and completeness of the proposed continuous or analog form of the signal. Please note that even though
FADI approach. At this point, we should mention that the ANN based we refer to a PLC I/O signal as an analog signal, it is always processed
signal deviation detectors are also being utilized to find out whether after signal quantization (in other words, as a non-binary digital signal).
there is any substantial change in the behaviour of the analog I/O Fig. 2 provides a schematic overview of the formulation and the
signals (not discussed earlier to avoid confusing the reader). This is training procedure of the second process behaviour model (i.e., the set
because it is very difficult to include all the relevant structural in- of signal deviation detectors). Each signal deviation detector is basically
formation or statistical properties of the analog I/O signals into a single an Autoassociative Neural Network (AANN) based one-class classifier
process behaviour model. In addition, doing so can result in an ex- (or outlier detector) which is designed and trained in such a way that it
ponential increment of the computation time and memory-space re- can detect any significant changes in the corresponding time series
quirements. The reader should also know that all these features are signal (i.e., a particular segment of the power consumption or analog I/
retroactively added to FBMTP software and hence, MPFDT will always O signal). The data logging module of MPFDT continuously collects the
provide greater FADI accuracy and precision than FBMTP. (active or real) electric power usage data of the (fault-free) manu-
facturing plant from the power meter and inserts it into the power usage
3. MPFDT: system overview and working principle log database (see Fig. 2). The log data records stored in the power usage
and the PLC I/O signal log database are then used for training the signal
An overview of the (modified) DSVTF control process model deviation detectors. As can be seen in Fig. 2, there are two types of
building procedure of MPFDT is shown in Fig. 1 (we will refer to this AANN based signal deviation detectors. The first kind is proposed for
model as the Modified DSVTF (MDSVTF) model in order to avoid the power consumption signal, whereas the second one is proposed for
confusion with the original DSVTF model proposed in Ghosh, Qin, Lee, the analog I/O signals. As can be seen in Fig. 2, the power consumption
et al. (2017). In a PLC controlled manufacturing system, most of the signal corresponding to a particular manufacturing system cycle is
control I/O signals are the binary digital signals (they take alternatively partitioned into several non-overlapping regions based on the distinct
two status values: ON and OFF or 1 and 0). However, in order to handle set of actuator operations (a manufacturing system cycle or a produc-
some critical control process operations, most of the manufacturing tion cycle starts with the receipt of the raw materials and ends with the
systems do need a few analog sensors and/or actuators that are con- release of the finished product). Each signal segment of interest basi-
trolled via analog output signals. Almost all the modern PLC devices cally denotes the electric power usage during a particular set of actuator
have a built-in analog I/O card thus they can handle the analog I/O operations (please note that the power consumption rates of the sensors
signals. However, the PLC program cannot process the continuously are typically very low as compared to that of the actuators). A certain
varying analog I/O signals directly in their analog form. The status number of features are then extracted from each power consumption
value of the analog sensor signal is converted into a digital equivalent signal segment and are inserted (packed into a vector) to the respective
or an integer number thus it can be stored in the PLC input memory feature database (see Fig. 2). This feature extraction process is repeated
table (Jack, 2010; Laughton & Warne, 2003). Similarly, the integer for the power consumption signal corresponding to each production
numbers stored in the PLC output memory table are converted into the cycle. In the same way, the feature databases are prepared for the

3
A. Ghosh, et al. Computers & Industrial Engineering 139 (2020) 106188

Fig. 1. An overview of the Modified DSVTF (MDSVTF) control process model building procedure.

analog I/O signals. However, the feature values are extracted from the does not affect the overall FADI accuracy of the proposed tool. The
signal segment corresponding to each energization cycle of the analog faults and anomalies are eventually going to be detected (supposedly),
I/O signal (please note that an energization cycle starts when the I/O however, only after a short period of time. Anyhow, if the fault or
signal is turned ON or activated and ends when it is turned OFF or anomaly is ‘severe’, then it can definitely be detected online (instantly)
deactivated – also see Fig. 2). by using the MDSVTF model (Ghosh et al., 2016; Ghosh, Qin, Lee, et al.,
The feature vectors of a given feature database are then normalized 2017; Roth et al., 2011). In the next section, we discuss the MDSVTF
and are fed as inputs to the corresponding AANN (see Fig. 2). The model design and implementation process in detail (also give a brief
AANNs are actually a special class of feedforward neural networks that overview of the FADI method based on it).
are trained to produce an approximation of the identity mapping be-
tween network inputs and outputs (Kramer, 1991, 1992). If there is any
significant change in the input time series signal, then the AANN pro- 4. MDSVTF model formulation and the FADI procedure
duces an output vector that is not quite similar to the input vector by
which the faults and anomalies can be identified (we will discuss this This section is split into two subsections. In Section 4.1, we formally
process further in Section 5). This fault and anomaly searching task is define the MDSVTF model and illustrate it with an example manu-
carried out periodically (by default, after every five production cycles – facturing system scenario. In Section 4.2, some features of the primary
since the computational cost of this method is little high). However, this MDSVTF model are redefined thus the state-space or memory-space
explosion problem can be avoided (at the end of this subsection, we

Fig. 2. A schematic diagram of the architecture and training procedure of the second process behaviour model.

4
A. Ghosh, et al. Computers & Industrial Engineering 139 (2020) 106188

briefly discuss the FADI procedure of MPFDT based on the proposed So, the transition can be formally expressed as: x → e (T1+δ1, T2+δ2) x',
MDSVTF model). where δ1 and δ2 are the permitted time delay periods. During the fault
and anomaly searching phase, the reference or corresponding TTO time
instance (i.e., T1+δ1 or T2+δ2) is selected either based on the tran-
4.1. The theoretical MDSVTF automaton model and its practical
sition execution pattern characteristics or depending on the proximity
applications
with the observed transition time instance (further explained later in
this section).
In MPFDT, the control process behaviour of a PLC controlled man-
ufacturing system is defined by using a single MDSVTF automaton. The Definition 3. A state belongs to a MDSVTF automaton GM is
MDSVTF model is similar in notion to the timed automata or timed represented by the integer vector: [IO1, IO2,…, ION], where N is the
finite-state machine (Cassandras & Lafortune, 2008; Ghosh, Qin, Lee, total number of PLC I/O signals and IOi denotes the integer status value
et al., 2017). The primary difference between the DSVTF and the of the ith PLC I/O signal (∀i = 1, 2,…, N).
MDSVTF model is that the MDSVTF model can handle the analog I/O
Definition 4. A signal-status vector that induces a transition between
signals. In addition, the concept of time-out function is further revised
the two states of a MDSVTF automaton, is comprised of the signal status
so that a transition can have multiple transition times (Ghosh, Qin, Lee,
change events that need to be executed during the state transition
et al., 2017). The MDSVTF automaton is defined as follows:
process. It can be formally expressed as follows. Let’s assume that:
Definition 1. A DFA, denoted by G, is a six-tuple automaton of the
form: G = (X, E, f, Γ, x0, Xm) i. the integer vectors associated with state x and state x', denoted by
where: IVx and IVx', are of length N;
ii. the signal status value located at the ith index position of the integer
x is the finite set of states and E is the finite set of events vector IV is denoted by IV[i]; and
f: X × E → X is the transition function iii. the actual name of the ith signal is symbolized by Sigi (∀i = 1, 2,…,
[f(x,e) = x' signifies that there is a transition with event label e from N)
state x to x']
Γ: X → PS(E) is the active event function or the feasible event then, the signal-status vector e that causes a transition from
function, where PS(E) defines the power set of event set E state x to state x' is represented by:
x0 ∈ X is the initial state and Xm ⊆ X is a non-empty set of final
states or marked states e = [ (IVx ([1]), IVx' ([1])), (IVx ([2]), IVx ' ([2])), ..., (IVx [N ], IVx ' [N ])]

Definition 2. A MDSVTF automaton, denoted by GM, is a seven-tuple where is the signal status change detection function and is formally
automaton of the form: GM = (X, E, TO, f, Γ, x0, Xm) defined as follows:
where:
A new function i.e., time-out function TO is included and the notion Sigi _IVx ' [i] if IVx [i] IVx ' [i]
(IVx [i], IVx ' [i]) = ( i = 1 ,.., N)
of the event is replaced by that of the ‘signal-status vector’ (other n ull** if IVx [i] = IVx ' [i]
symbols have the same meanings as in Definition 1). Please note that a
signal-status vector actually characterizes the signal status change op- [*the symbol Sigi_IVx'[i] denotes the concatenation of the following two
erations that need to be executed while transiting from a MDSVTF strings: the signal name Sigi and the updated status value IVx'[i] (the un-
model state to the next one (a formal account is given in Definition 4). derscore symbol ‘_’ is just used as a separator character between the element
strings)]
The time-out function TO is formally specified as:
TO: X × E PS( 0 ) where ℕ0 is the set of all natural numbers and [**the element is not included in the signal-status vector]
PS(ℕ0) denotes the power set of number set ℕ0.
So, if f(x,e) = x' and TO(x,e) = t, denoted as x → e (t) x', then we can We further discuss the MDSVTF model building procedure with the
state that the MDSVTF automaton GM, being at state x, accepts the help of an example manufacturing system scenario (shown in Fig. 3). As
signal-status vector e within the associated time-bound period t and can be seen in Fig. 3, the manufacturing system has in total two sensors
advances to the next state x'. (one analog and one digital sensor) and three actuators. The analog
[the above specified time period is hereinafter referred to as the sensor AWS is used to measure the weight of the package and the digital
Transition Time-Out (TTO) time (a state transition needs to be completed sensor DS is used to detect the presence of the package at a pre-
within the corresponding TTO time)] determined set point. The whole system works as follows (also see
The TTO time associated with a particular transition is generally Fig. 3): (i) a package is placed on the conveyor belt (the weight is
estimated as a little longer (user-defined) than the (maximum) observed measured by the analog sensor AWS – we assume that there are two
transition time. A transition can have multiple transition times (or TTO types of packages: one with weight 10 kg (KG) and other with weight
times – please check the definition of the time-out function TO). Let’s 15 KG); (ii) then the conveyor belt starts to move towards the loader
assume that the transition x → e x' has two transition times t1 and t2. crane system (this motion is driven by the conveyor belt actuator CBA);

Fig. 3. An example manufacturing system scenario (for simplicity, we assume that the crane system can be driven by a single actuator).

5
A. Ghosh, et al. Computers & Industrial Engineering 139 (2020) 106188

Fig. 4. The signal status versus time chart of the PLC I/O signals associated with the manufacturing system of Fig. 3.

(iii) when the conveyor belt motion starts, the analog sensor AWS gets determined based only on the aforementioned model which transition
detached from the belt (it returns back to the original position after a the control process is going to execute at a given point of time. In order
while); (iv) the conveyor belt motion stops after the package reaches to overcome this limitation, a regular expression formula is attached to
the end of the conveyor belt (detected by the digital sensor DS); (v) if each MDSVTF model state from which multiple outgoing transition arcs
the package weight is 10 KG, then the loader crane picks it up and puts emanate (see Fig. 5). The regular expression formula is used to com-
it into the container 1 (the loader crane actuator LCA carries out this pactly encode the transition execution pattern (or order) information.
task); (vi) otherwise, if the package weight is 15 KG, then the gantry The details of how to derive the regular expression formula and how it
crane picks it up and places it into the container 2 (the gantry crane can be utilized to detect and isolate the ‘invalid transition path’ faults
actuator GCA performs this operation); and (vii) this process cycle starts can be found in Ghosh, Qin, Lee, et al. (2017) (see Section IV-C – also
again when another package is set on the conveyor belt. The signal briefly explained in Appendix A.II). Please note that the order or se-
status versus time chart of the above specified sensor and actuator quence in which the transitions from a state are executed may not al-
signals is given in Fig. 4 (we simply assume that in the PLC memory ways follow a repetitive pattern. In this type of situation, the transition
table, the weight values 0, 10 and 15 KG are represented by the integer execution behaviour is characterized by using the transition probability
numbers 0, 100 and 150, respectively). It should be noted that in Fig. 4, values (for details, see Section IV-C of Ghosh, Qin, Lee, et al., 2017).
it is assumed that the 10 KG and 15 KG packages are presented alter- By following exactly the same procedure as outlined above, a reg-
nately to the system. The sequence of integer vectors associated with ular expression formula defining the transition time pattern informa-
the signal status versus time chart (see lower portion of Fig. 4) basically tion is attached to each transition that has multiple TTO times (see
characterizes the state transition sequence of the manufacturing system Fig. 5), thus during the FADI phase, the reference TTO time instance
of Fig. 3. can accurately be determined (note that this information is not included
in the DSVTF model (Ghosh, Qin, Lee, et al., 2017). If no regular ex-
The MDSVTF automaton model of the above stated system is given
pression formula is found that can properly define the sequence pattern
in Fig. 5 (also see Definition 1 to Definition 4). As described in
of the transition time instances, then the transition time pattern in-
Definition 3, a particular state of a MDSVTF automaton basically de-
formation is characterized by the probability values. It is easy to realize
notes a unique integer vector (the vector contains the integer status
that in such a case, the reference TTO time instance is selected based on
values of all the PLC I/O signals). In practice, the state ID of a MDSVTF
the proximity with the observed transition time instance (for detailed
model state (see Figs. 4 and 5) is determined by hashing the corre-
discussions about how to calculate the transition times and how to se-
sponding integer vector (in order to reduce the memory space re-
lect the reference TTO time instance, the interested readers are referred
quirement). As can be seen in Fig. 5, each transition belongs to a
to Ghosh et al. (2016) – please see Section 2.4.2 and Section 2.6). It is
MDSVTF automaton model has two types of information attached to it:
obvious that by using the transition probability values, it cannot be
(i) the TTO times (given inside the parentheses) and (ii) the signal-
explicitly determined which transition the control process is going to
status vector (given as a list of space-separated signal status change
execute at a given point of time. However, it can certainly be identified
events or elements). It is easy to see from Fig. 5 and Definition 4, a
from the changes in the transition probability values whether the
signal-status vector that induces a transition between two states of a
overall transition execution behaviour can be classified as normal or not
MDSVTF automaton is actually consisted of the Signal Status Change
(the same is also applicable to the transition time probabilities) (Ghosh
(SSC) events that take place during the corresponding state transition
et al., 2016; Ghosh, Qin, Lee, et al., 2017).
process (the signal name and the integer status value are separated by a
It is easy to perceive that the MDSVTF automaton basically defines a
‘_’). As an example, the SSC event AWS_100 (implies that the status
set of system state transition rules. If any violation of such rules (in-
value of the signal AWS is updated to 100) is solely responsible for the
cluding regular expression rules) is observed, then we can conclude that
transition X1 → X2 (see Figs. 4 and 5). It is easy to see from this dis-
a fault has occurred in the manufacturing system (Bao et al., 2014;
cussion that we are able to incorporate a great amount of control pro-
Ghosh et al., 2016; Ghosh, Qin, Lee, et al., 2017; Ghosh, Qin, Wang,
cess information (such as: the SSC event sequence information, transi-
et al., 2017; Roth et al., 2009, 2011, 2012; Schneider et al., 2011).
tion time information, etc.) into the proposed automaton model (by
Similarly, if significant variations in the transition times and the tran-
which the faults and anomalies can be accurately detected). However, if
sition and transition time probabilities attached to a MDSVTF model
a state has multiple outgoing transitions, then it cannot be explicitly
state or transition are observed, then it can be concluded that an

6
A. Ghosh, et al. Computers & Industrial Engineering 139 (2020) 106188

Fig. 5. The MDSVTF control process model of the manufacturing system of Fig. 3.

Fig. 6. A PLC controlled stamping press system.

anomaly has occurred in the system. It should be noted that the tran- However, if all the I/O signals are the analog signals, then the MDSVTF
sition and the transition time probabilities are verified periodically after model can have, in theory, RN number of possible states (the symbols R
the control process reaches the steady-state phase (Ghosh et al., 2016; and N retain the same meanings as above). Although, in practice, the
Ghosh, Qin, Lee, et al., 2017). total number of analog I/O signals cannot be too high, the above factor
can still lead to a state-space explosion problem. In addition, it can
immensely slow down the execution speed of the FADI approach. For
4.2. Integrating the trend information of the continuously varying analog I/
these reasons, it is absolutely necessary to restrict even further the
O signals into the MDSVTF model
status value domain of the analog I/O signals. For simplicity, a further
signal quantization (or data binning) step is performed in order to
The example analog signal of Fig. 4 is too simple to be of much
substantially reduce the state-space complexity of the MDSVTF model.
practical value. In general, an analog I/O signal changes its status too
Recall that as because the analog I/O signals are continuous in both
frequently and even after the signal quantization, it can draw the status
time and amplitude, we need to inspect whether the overall trend or
values from a quite large integer domain [0, R], where the value of R is
shape of the signal can be classified as normal or not (no need to inspect
typically 4000. If all the PLC I/O signals are the binary digital signals,
each and every sampled value of the signal). The above task is partly
then theoretically the MDSVTF model can have 2N number of possible
performed by means of the MDSVT model and more extensively per-
states, where N is the total number of PLC I/O signals (typically ranges
formed by using the signal deviation detectors (described in detail in
from a few hundreds to a few thousands). However, empirically, it has
Section 5 – also see Fig. 2).
been observed that the state complexity of the DSVTF model or its re-
We further illustrate the above idea with the help of a manu-
vised version, the MDSVTF model, is usually O(kN), where k is a small
facturing system scenario presented in Fig. 6 (a simulation video of this
positive number (Ghosh, Qin, Lee, et al., 2017; Roth et al., 2011).

7
A. Ghosh, et al. Computers & Industrial Engineering 139 (2020) 106188

Fig. 7. The signal status versus time chart of the PLC I/O signals associated with the stamping press system of Fig. 6.

Fig. 8. The MDSVTF automaton model of the stamping press system of Fig. 6.

virtual system is shown in the supplementary video file ‘Stamping_- contiguous intervals, namely, TI11 and TI12 (these two intervals basi-
Press.mp4’). The signal status versus time chart of the control I/O sig- cally represent the readings of the sensor ADS during the downward
nals of this system can be found in Fig. 7 (the graph is constructed based motion of the punch tools and the stamping operation, respectively).
on the computer simulation data). The MDSVT automaton model of this Similarly, the time interval TI2 is (arbitrarily) divided into two (equal-
system is given in Fig. 8. As can be seen, the MDSVT model is built length) intervals, namely, TI21 and TI22. Hereinafter, this type of dis-
following absolutely the same procedure as sketched in Fig. 5 (see joint and contiguous time interval divisions (identified either auto-
Section 4.1 for details). However, as the analog sensor signal ADS matically by MPFDT or manually by operator) is referred to as the
changes its status value frequently during the time intervals TI1 and TI2 ‘marked time interval’ and the signal segment within the marked time
in each signal cycle (the status value range is (Kramer, 1991) for both interval is referred to as the ‘marked signal segment’ (the reader should
the cases), the data points of the signal ADS within these two interval note that the data points belonging to a marked signal segment fluc-
windows need to be further quantized and encoded in order to reduce tuate markedly or exhibit a very high degree of variability). We should
the state-space complexity of the model (see Figs. 7 and 8). As can be stress here the fact that the boundaries of the marked time intervals or
seen in Fig. 7, the time interval TI1 is split into two disjoint and the marked signal segments are determined not only based on the

8
A. Ghosh, et al. Computers & Industrial Engineering 139 (2020) 106188

occurrence time of the disruptive SSC events (after which the analog I/ set in such a way that at least a few initial data points of the succeeding
O signal sharply changes its slope), but also based on the status value segment lie outside the status value range of the preceding segment –
range of the interval (the time interval and the corresponding status this is necessary because otherwise the corresponding SSC event cannot
value range should not be too wide, otherwise it may decrease the be identified automatically). In order to assist this decision-making
accuracy of the MDSVTF model based FADI method). MPFDT, by de- process, MPFDT automatically identifies the relatively linear and non-
fault, automatically determines the boundaries of the marked signal linear regions of the analog I/O signal (resolved by using a moving
segments based on certain predefined statistical criteria (if the parti- window algorithm) and the resulting information is then displayed in
tioning solution provided by MPFDT is not adequate, then the operator an intuitive graphical user interface window (for the convenience of the
should manually set the status value range of the partitions). users). In practice, most of the analog I/O signals follow the similar
All the status values of the analog I/O signal within a marked signal waveform pattern (or shape) as that of the sensor signal ADS of Fig. 7
segment (or marked time interval) constitute one group and are con- and hence, it is not too difficult to determine the appropriate boundary
sidered (or managed) as a single SSC event. As one can see in Fig. 8, the points of the marked signal segments.
SSC elements of the signal-status vectors are computed following ex- For the sake of correctness, we must mention that the previously
actly the same method as discussed in Section 4.1. However, in case of described MDSVTF models (see Figs. 5 and 8) are built upon several
analog I/O signals, the status values within a marked signal segment are assumptions (to keep the discussion simple and focused on the main
denoted (as a single unit) by the signal label extended by the status points). The first and foremost assumption is that in a PLC controlled
value range label (i.e., the concatenation of the signal name, and the system, the data logger is able to record all the SSC events accurately,
start and end values of the status value range – the tokens are separated which is not true for large PLC controlled systems (as can be seen in
by a ‘_’). For example, the signal segment of the analog signal ADS Fig. 1, the log data records can be collected only for a small amount of
within the time interval TI11 is a marked signal segment (see Fig. 7) time in each PLC scan cycle i.e., during the communication phase). This
and therefore, as can be seen in Fig. 8, the status values within this can lead to several log data inconsistency problems, such as the SSC
signal segment are denoted by the event label ADS_26_434 (note that event sequence disorder problem, the missing SSC event problem, etc.
the status values of the signal ADS always remain within the range These data inconsistency problems are handled in exactly the same way
(Kramer, 1991) during the time interval TI11 – the SSC event label as described in Ghosh, Qin, Lee, et al. (2017) (see Section IV-B). An-
ADS_26_434 simply expresses this in symbolic form). It is easy to realize other simplifying assumption we made is that a state transition always
that if this signal quantization step is not performed, then the MDSVTF takes a fixed amount of time to complete. It is easy to perceive that the
model of Fig. 8 may have additional 998 states (also see Fig. 7). So, by transition time of a particular state transition ultimately depends on the
applying the above signal quantization step, we are able to drastically performance of the corresponding physical devices and hence, cannot
reduce the state-space or memory-space complexity of the MDSVTF always be the same. For this reason, in the TTO time calculation, the
model. Please note that most of the faults and anomalies in the analog maximum value of the observed transition time instances is taken as the
sensor ADS (especially that have significant impact on system opera- actual transition time (the rationale is easy to understand). For the same
tion) can still be detected by using the MDSVTF model of Fig. 8. As an reason (and also because of the fluctuations in the status values of the
example, suppose that a failure occurs in the analog sensor ADS (im- analog I/O signals), the status value range associated with the SSC
plies that the sensor readings are always 0) just before the stamping event of the form sig_p_q (the term sig_p_q retains the same meaning as
operation begins (see Figs. 6 and 7). In this case, we get the transition before) is set slightly wider than that of all the corresponding observed
X2 →ADS_0 SPA1_0 X3 instead of the transition X2 → ADS_435_524 X3 by marked signal segments. We must also emphasize the fact that the
which the fault can easily be detected (see Fig. 8). However, if there is a marked time intervals or the marked signal segments need to be defined
minor change in a marked signal segment of the analog signal ADS, in such a way that the corresponding MDSVTF model states belong to
then it may not be detected by using the MDSVTF model. Anyhow, if the category of ‘stable system states’ (Ghosh, Qin, Lee, et al., 2017).
the change is major, then obviously the signal-status vector and/or the Informally speaking, the term stable system state refers to a state where
transition time properties of the corresponding MDSVTF model transi- the control process spends (or stops for) such a long period of time that
tion get affected and hence, can easily be detected by using the the data logger is (eventually) able to record the complete state in-
MDSVTF model (Bao et al., 2014; Ghosh et al., 2016; Ghosh, Qin, Lee, formation (i.e., the status values of all the I/O signals) accurately (in
et al., 2017; Ghosh, Qin, Wang, et al., 2017; Roth et al., 2009, 2011, other words, cannot, in any case, remain unobserved or partially ob-
2012; Schneider et al., 2011). served) (Ghosh, Qin, Lee, et al., 2017). This measure is necessary be-
We must clarify here that the SSC event of the form sig_p_q (where, cause the FADI procedure is (essentially) carried out based on the
sig is the name of the analog I/O signal, and p and q are the start and transitions between the stable system states (for more details, see Sec-
end points of the status value range) denotes that the updated status tion IV-B1 of Ghosh, Qin, Lee, et al. (2017).
value of the signal sig must fall within the range [p, q] (see Figs. 7 and MPFDT uses exactly the same FADI technique as proposed in Ghosh,
8). A transition whose signal-status vector includes such SSC events has Qin, Lee, et al. (2017) (see Section V). Thus, we have decided not to
slightly different characteristics (or meaning). For example, the tran- discuss it further here (for the sake of completeness, the MDSVTF model
sition x → sig_p_q x' defines that the MDSVTF automaton, being at based FADI procedure of MPFDT is briefly described in Appendix A.III).
state x, accepts the SSC event sig_p_q (implies that the updated status In the next section, we will discuss the working procedure of the signal
value of the signal sig is within the valid range) and advances to the deviation detectors in detail.
next state x' (same as before). However, until the MDSVTF automaton
leaves the state x' (and moves to the next valid state), the continually 5. The signal segmentation and the neural network based signal
changing status value of the analog signal sig must lie within the range deviation detection procedure
[p, q]. If we increase the number of signal partitions or marked signal
segments, then it may increase the accuracy of the MDSVTF model From the discussion in Section 4, it is easy to perceive that a fault or
based FADI procedure up to a certain extent. However, at the same anomaly that has significant impact on system operation (referred to as
time, it can considerably increase the state-space complexity of the the hard fault or anomaly) can definitely be detected by using the
MDSVTF model (and also the associated costs). This trade-off needs to MDSVTF model (for more details, see Appendix A.III or Section V of
be addressed subjectively. So, it is recommended that the system op- Ghosh, Qin, Lee, et al. (2017). However, some soft faults and anomalies
erator manually set the boundaries (the status value range) of the that do not have much effect on the system operation can remain un-
marked signal segments through experimentation (please note that the detected (please note that not all types of soft faults and anomalies are
boundaries of any two contiguous marked signal segments need to be non-hazardous (Ghosh, Qin, Lee, et al., 2017). There are two possible

9
A. Ghosh, et al. Computers & Industrial Engineering 139 (2020) 106188

reasons why a fault (or anomaly) does not have much impact on control corresponding time series database for further analysis. We must clarify
process operation. Firstly, if the fault occurs in a sensor which is in- here that originally the feature values are directly extracted from the
dependent and standalone (otherwise, it would have completely altered signal segments of the analog I/O and the power consumption signals
the SSC event sequence) (Ghosh, Qin, Lee, et al., 2017). This is more of and are inserted into the corresponding feature database (as depicted in
a system architecture design problem and hence, must be resolved in Fig. 2). However, in order to keep the discussion simple and easy to
the system architecture design phase. The second and most common follow, we decouple this process into two both independent and con-
reason is that an actuator fault has occurred. An actuator failure (unlike secutive steps (see Fig. 10): (i) the analog I/O and the power con-
sensor failure) incident never gets reflected in the PLC memory table sumption signals are segmented and the resulting segments are inserted
because, the actuator signals are the output signals (therefore, the into the corresponding time series databases; and (ii) the feature values
control process model may remain unchanged). In some cases, if an are extracted from the time series signals in the time series database and
actuator stops functioning, the control process gets stuck in an infinite are inserted into the corresponding feature database (this process is
loop and hence, can be detected by using the MDSVTF model (see repeated iteratively for each and every time series database).
Appendix A.III for details). However, this is not always, or even often, The boundary points of the energization cycles are determined
the case in practice. This is the reason why the second process beha- following the same procedure as that described for the marked time
viour model (a set of AANN based signal deviation detectors) is pro- intervals. Recall from Section 4.2 that each signal cycle of the analog I/
posed. The proposed signal deviation detectors are capable of identi- O signal is split into marked and un-marked time intervals depending
fying (also see Fig. 2): (i) whether there is any significant variation in on the fluctuations or variance in the status value of the signal (the
the behaviour of any analog I/O signal (the rationale is explained in status value of the signal varies frequently during the marked time in-
Section 4); and (ii) whether there is any significant variation in the tervals and remains constant during the un-marked time intervals).
power consumption profile of the manufacturing system (please note Similarly, by following the same procedure, each signal cycle of the
that usually an actuator failure causes significant changes in the power analog sensor signal APS is split into four marked time intervals (see
consumption profile of the manufacturing system and hence, can easily Fig. 10). Among all the marked and un-marked time intervals, a par-
be detected by analysing the corresponding power consumption signal ticular marked or (sometimes) un-marked time interval is selected as
segment). The remainder of this section is organized in three subsec- the waiting time interval (the time interval between two consecutive
tions. The signal segmentation procedure for the analog PLC I/O signals energization cycles – see Fig. 10) based on (mainly) the following cri-
and the power consumption signal are described in Section 5.1 and terions: (i) the status value of the signal does not fluctuate much during
Section 5.2, respectively. Section 5.3 is fully devoted to an in-depth the waiting time interval; and (ii) the waiting time interval is much
discussion on the signal deviation detectors. longer than the other marked and un-marked time intervals. Once the
waiting time interval is determined, the entire analog I/O signal is split
5.1. The analog PLC I/O signal segmentation procedure into individual energization cycles based on the status value range (the
boundary points) of the waiting time interval (see Fig. 10). The signal
The analog I/O and the power consumption signals need to be segment or region within an energization cycle window (respectively,
segmented into regions having similar statistics before they are passed waiting time interval window) is commonly referred to as the active
to the signal deviation detectors (in order to separate the regions of (respectively, inactive) signal segment or region. We partitioned the
interest and also to keep the lengths of the time series signals reason- analog I/O signals based on energization cycle because: (i) if a fault
able). Recall from Section 3 (also see Fig. 2) that in MPFDT, the entire occurs in a sensor or the control process of an actuator, then the active
analog I/O signal is segmented into individual energization cycles and regions of the corresponding I/O signal tend to deviate more from the
are passed to the corresponding signal deviation detector for training normal than the inactive regions of the signal; and (ii) the inactive
(or classification). An energization cycle starts when the I/O signal is regions of the PLC I/O signal are (generally) too long and hence, need
activated or switched ON and ends when it is deactivated or switched to be removed from the signal in order to better capture the cycle-to-
OFF. Here, it is necessary to clarify that an analog I/O signal is said to cycle difference in signal waveform.
be activated (from the control engineering viewpoint) when it starts to Here, we should emphasize the fact that a sensor or actuator can
move away from the baseline level and is supposed to be deactivated have more than one type of energization cycles or active signal seg-
when it returns back to the baseline level. We discuss the analog I/O ments (although extremely unlikely). For example, suppose that in the
signal segmentation procedure with the help of an Argon Oxygen first phase (during time interval TI1) of the ith and jth production cycle
Decarburization (AOD) steelmaking system shown in Fig. 9 (a simula- of the stamping press system of Fig. 6 (∀i = 1, 3, 5, … and ∀j = 2, 4, 6,
tion video of this virtual system is shown in the supplementary video …), the slide moves down a distance d1 and d2, respectively (where,
file ‘AOD_Steelmaking.avi’). The pressure inside the furnace chamber is d1 ≇ d2). In this case, as one can easily realize by looking at Fig. 7, the
measured by an analog sensor, named APS (other devices are not dis- analog sensor signal ADS will have two groups of energization cycles
cussed here for space reasons). The signal status versus time chart of the (one for odd and another for even production cycles). In MPFDT, each
analog sensor signal APS can be found in Fig. 10 (the graph is con- group of energization cycles is treated individually as a separate data
structed based on the computer simulation data). As can be seen in set (implies that if the analog I/O signal has K groups of energization
Fig. 10, the signal segment corresponding to each energization cycle is cycles or active signal segments, then MPFDT produces K different time
extracted from the analog I/O signal and is inserted into the series databases, one for each group). Each time series database is then

Fig. 9. An Argon Oxygen Decarburization (AOD) stainless steelmaking system.

10
A. Ghosh, et al. Computers & Industrial Engineering 139 (2020) 106188

Fig. 10. The signal status versus time chart of the analog sensor signal APS and the signal segmentation procedure of MPFDT.

processed in the same way as described above (also see Fig. 10). In is continuously recorded (or sampled) by a power meter and is inserted
practice, an analog I/O signal has a single or just a few (which is rarely into the corresponding power usage log database. So, basically, we can
the case) groups of energization cycles and hence, the above measure have several (instead of only one) power usage log databases (one for
does not really incur much extra computational overhead. each device cluster). We opted for the power consumption signal ana-
lysis technique to solve the FADI purpose mainly because: (i) the power
usage data can be collected at a very low cost; and (ii) it has been shown
5.2. The power consumption signal segmentation procedure to be a very reliable method of evaluating the health condition of the
manufacturing system (see ElMaraghy, Youssef, Marzouk, &
At this point, we should mention that the power usage data accu- ElMaraghy, 2017; Endo, Nakajima, & Hata, 2012). In MPFDT, the
sation procedure sketched in Fig. 2 is a very simplistic portrayal of how power consumption signal (or the power usage log database) corre-
the power usage data is collected from the manufacturing system (only sponding to each device cluster is processed separately and in-
feasible for small systems). A schematic diagram illustrating how the dependently (a device cluster can simply be viewed as a small manu-
power usage data is actually recorded is given in Fig. 11. Without en- facturing system). In practice, a device cluster contains only a few
tering into the technical details, we can summarize the whole data (typically ten to fifteen) devices. We should stress here the fact that in
accusation procedure as follows (see Fig. 11): (i) in a PLC controlled reality, the power consumption profile of a device cluster is more or less
manufacturing system, the devices are inherently organized into (from equivalent to the power consumption profile of the actuators in it
the electrical connection point of view) different device clusters; and (because the sensors consume a negligible amount of power). Recall
(ii) the power usage (instantaneous wattage) data of each device cluster

Fig. 11. A more detailed overview of the power usage data collection process.

11
A. Ghosh, et al. Computers & Industrial Engineering 139 (2020) 106188

Fig. 12. The robotic drilling subsystem: an example small device cluster (it has three sensors and five actuators).

that the main objective of the (power consumption) signal deviation system when the actuators DRA2 and CBA2 are working. Please note
detectors is to identify the actuator faults and breakdown incidents (not that the boundary points of the task-specific power segments can easily
the sensor faults and anomalies). We should also emphasize that when be determined by performing phase alignment (in time domain) be-
we talk about the power consumption profile of an actuator, we actu- tween the power consumption signal and the PLC I/O signals. In the
ally refer to the power consumption profile of the electric motor that next step, the task-specific power segments are merged or discarded
drives the actuator. depending on what type of actuators are active during that specific time
We further discuss the power consumption signal segmentation period (see Fig. 13).
procedure with the help of an example device cluster shown in Fig. 12 From the control engineering viewpoint, a manufacturing system
(as just stated above, the proposed signal segmentation process is de- can have two types of actuators (Ghosh, Qin, Lee, et al., 2017): (i)
vice cluster based and hence, needs to be repeated for each power usage loosely-coupled actuator: an actuator is said to be a loosely-coupled
log database). The power consumption profile of the device cluster of actuator if the operational status (i.e., normal or faulty) of the actuator
Fig. 12 is given in Fig. 13 (the actual devices that are used to generate does not have any (or a little) effect on the subsequent control process
the power usage data for the device cluster of Fig. 12 can be found in operations; and (ii) strongly-coupled actuator: an actuator is said to be a
Appendix A.IV). As can be seen in Fig. 13, at first, the power con- strongly-coupled actuator if without successfully finishing the corre-
sumption signal of the device cluster (corresponding to each production sponding physical operation, the rest of the control process operations
cycle) is partitioned into multiple smaller segments based on distinct set cannot be executed normally (as an example, in Fig. 12, the actuators
of actuator operations (each such signal segment is hereinafter referred CBA1, CBA2 and DRA1 are the strongly-coupled actuators). As ex-
to as the task-specific power segment). For example, the task-specific plained earlier, the MDSVTF model based FADI procedure may fail to
power segment TPS6 represents the power consumption profile of the detect the loosely-coupled actuator faults (the strongly-coupled

Fig. 13. The power consumption profile of the device cluster of Fig. 12 and the signal segmentation procedure of MPFDT.

12
A. Ghosh, et al. Computers & Industrial Engineering 139 (2020) 106188

actuator faults can definitely be detected by using the MDSVTF model – classification procedure also follows the same steps). As can be seen
see Appendix A.III) (Bao et al., 2014; Ghosh, Qin, Lee, et al., 2017; from Fig. 14, in MPFDT, in total seventeen features are extracted from
Ghosh, Qin, Wang, et al., 2017; Roth et al., 2009, 2011, 2012; each time series signal, where the first ten features are the (classical
Schneider et al., 2011). So, only the task-specific power segments cor- time-domain) statistical features (Caesarendra & Tjahjowidodo, 2017;
responding to the loosely-coupled actuators are needed to be inspected Wang, Zheng, Zhao, & Wang, 2015) and the remaining ones are the
further. As can be seen in Fig. 13, for each loosely-coupled actuator in singular value features (Hernandez-Vargas, Cabal-Yepez, & Garcia-
the device cluster, the task-specific power segments during which the Perez, 2014; Liu, Wang, & Lu, 2014). Among the statistical features, the
actuator is active are merged, and the resulting segment is inserted into first one is the length of the time series. The length of the time series is
the corresponding time series database. The remaining task-specific an important parameter for classification of the analog I/O and the
power segments are just simply ignored or discarded. The above process power consumption signals because, usually, if a fault occurs in a sensor
is repeated for the power consumption signal corresponding to each or actuator, the control system takes much longer (sometimes, shorter)
production cycle (see Fig. 13). Please note that if the energization cycles time to finish the corresponding physical operation. The other nine
of a loosely-coupled actuator overlap with that of the other actuators in statistical features are obtained by partitioning the time series signal
time domain (as happened in case of the actuator DRA2 – see Fig. 13), into three equal-length segments, and then by extracting three features
then it becomes more difficult to identify small changes in the power from each of those segments (see Fig. 14). Various time-domain sta-
consumption profile of the actuator (which as a result may decrease the tistical features have been employed in the past for time series classi-
accuracy of the FADI procedure). In order to largely avoid such situa- fication. Among them, the most commonly used ones are listed in
tions, the devices of the manufacturing system need to be partitioned Table 1 (Caesarendra & Tjahjowidodo, 2017; Wang et al., 2015). In
into device clusters in such a way that most of the loosely-coupled MPFDT, all the fourteen statistical features listed in Table 1 are ex-
actuators in each device cluster operate sequentially with the other tracted from each signal segment, however, among them, only three
actuators. features are selected as the ANN inputs (see Fig. 14). This feature se-
From the above discussions, one can easily conclude that the analog lection step is necessary because, no statistical feature has been proven
I/O and the power consumption signals are segmented into meaningful to be advantageous for classification of all types of time series signals
regions of interest in such a way that: i) all the signal segments in a (Fan, Hu, & Cheng, 2013). The features are selected based on their
particular time series database are of homogeneous type; and ii) all the discriminatory power with regard to the detection of faulty signals. We
signal segments in each time series database are of moderate lengths use a simple but very effective feature evaluation metric, called the
(so, an excellent trade-off between computational cost and accuracy is sensitivity factor for this purpose (Fan et al., 2013). As just stated
maintained). In the next section, we describe the algorithmic details of above, in MPFDT, along with ten statistical features, seven singular
the proposed signal deviation detector. value features are extracted (by applying the Singular Value Decom-
position (SVD) algorithm) from each time series signal. The SVD algo-
5.3. The FADI mechanism of the AANN based time series signal deviation rithm has been shown to be a highly effective and reliable feature ex-
detector traction method for classification of different types of time series signals
(see for instance (Hernandez-Vargas et al., 2014; Liu et al., 2014). In
As discussed in Section 5.1 and Section 5.2 (also see Fig. 2), the MPFDT, eleven singular values are extracted from each time series
signal segmentation procedure of MPFDT produces in total (N + M) signal and among them, the largest seven singular values are chosen as
number of time series databases, where N and M number of databases the final feature subset (see Fig. 14). This is because, usually, the large
are respectively generated by the power consumption and the analog I/ singular values reflect the fault characteristics more effectively. The
O signal segmentation procedure (it is easy to perceive that a real-world detailed justification of the proposed feature extraction and selection
manufacturing system can have only a few loosely-coupled actuators method (and also the definitions of the SVD, statistical features listed in
and analog I/O signals, and hence, the total number such databases Table 1, and sensitivity factor) is given in Appendix A.V.
cannot be very high in practice). The objective of the signal deviation The output of the feature extraction process is basically a feature
detectors is to check whether the time series signals in the time series database or an P × 17 matrix, where P is the number of time series
databases are normal or faulty (if a large fraction of the time series signals in the time series database (see Fig. 14). At first, the classifier
signals in a particular time series database are classified as faulty, then needs to be trained using the feature vectors in the reference feature
we can conclude that a fault or anomaly has occurred in the corre- database. Once the classifier is fully trained, it can be applied to classify
sponding sensor or actuator). MPFDT employs in total (N + M) AANN the time series signals in the query time series database. A wide variety
based signal deviation detectors, one for each time series database, to of approaches has been proposed in the literature to solve the one-class
satisfy this purpose. Each signal deviation detector works in- time series classification problem. A comprehensive survey of these
dependently and all of them work exactly in the same manner. From the approaches can be found in (Chandola, Banerjee, & Kumar, 2009;
computer science point of view, this problem can be viewed as a classic Hodge & Austin, 2004). Among them, we found the AANN to be the
one-class classification problem. To put it simply, we have two data- most suitable for our application. The AANN has been used extensively
bases: the reference and the query time series databases. The reference in various pattern recognition and time series classification applications
time series database contains a large number of normal time series (see for instance Chandola et al., 2009; Hodge & Austin, 2004;
signals (or positive samples). The classifier needs to be trained using the Krothapalli & Koolagudi, 2013; Sangeetha & Jothilakshmi, 2016; Rao,
positive samples from the reference time series database. On the other Nandi, & Koolagudi, 2014; Wang & Cui, 2013; Miranda, Castro, & Lima,
hand, the query time series database contains a presumably small 2012; Nazarko & Ziemianski, 2016; Bratina, Muškinja, & Tovornik,
number of samples (extracted from the observed analog I/O or power 2010). The AANN is essentially one kind of self-supervised feedforward
consumption signals) and the one-class classifier or the novelty detector neural network suitably designed to perform identity mapping (the
has to decide whether those time series signals are normal or faulty. inputs are approximated at the output layer) (Kramer, 1992; Wang &
The neural network training (or classification) is basically a two- Cui, 2013). Once the network is fully trained (using the feature vectors
step process. In the first step, the feature values are extracted from each extracted from the samples of a particular class), the reconstruction
time series signal and are inserted into the feature database (see error for an input feature vector that is extracted from a sample of the
Fig. 10). The feature vectors in the feature database are then normal- same class is expected to be much lower than that of a sample of a
ized and are fed as inputs to the neural network (for training or clas- different class (Miranda et al., 2012; Sangeetha & Jothilakshmi, 2016;
sification). An overview of the proposed feature extraction, feature Wang & Cui, 2013). So, the distance between the input and the output
selection and ANN training procedure is shown in Fig. 14 (the vector of the AANN (often referred in literature as the novelty index

13
A. Ghosh, et al. Computers & Industrial Engineering 139 (2020) 106188

Fig. 14. An overview of the feature extraction, feature selection, and neural network training procedure of MPFDT.

Table 1
Classical time-domain statistical feature list.
(1) Mean (2) Standard Deviation (3) Root Mean Square (4) Shape Factor (5) Peak Value (6) Peak-to-Peak Value (7) Crest Factor
(8) Impulse Factor (9) Margin or Clearance Factor (10) Skewness (11) Skewness Factor (12) Kurtosis (13) Kurtosis Factor (14) Entropy

(Nazarko & Ziemianski, 2016) can be used as an indicator to char- need to be trained with both positive and negative samples; and (iii) the
acterize (or evaluate) the status of the input sample and by determining output of the AANN classifier is simply a dissimilarity or novelty score
an appropriate threshold of being anomalous, the test or query samples that not only enables us to detect a faulty sample but to assess the se-
can be classified into normal and faulty groups (a more detailed dis- verity of the fault as well (Wang & Cui, 2013). However, the AANN (just
cussion about the AANN can be found in Appendix A.VI). like any other ANN) has an inherent disadvantage i.e., it is extremely
We opted to use the AANN to solve the one-class time series clas- hard to determine the optimal (or nearly optimal) neural network
sification problem mainly because of the following reasons: (i) the structure, learning or training algorithm and neural network para-
AANN has remarkable ability to learn complex linear and nonlinear meters. We adopted a different strategy to solve this issue. We have
relationships between the variables (Krothapalli & Koolagudi, 2013; thoroughly reviewed a wide range of literatures on this subject to find
Rao et al., 2014); (ii) unlike conventional ANN, the AANN does not out what type of AANN configuration tends to produce the best results

14
A. Ghosh, et al. Computers & Industrial Engineering 139 (2020) 106188

for different types of time series classification applications, and then Most of the faults and anomalies in the manufacturing system (speci-
simply adopted it in our approach. We chose this strategy because, the fically those which significantly affect the state transition behaviour of
proposed method needs be robust enough to handle different types of the control process) can be detected by using this process behaviour
time series signals. In MPFDT, a simple five-layer (fully-connected) model. However, there are several other types of faults and anomalies
AANN of the structure 17-34-6-34-17 is used for classification (note that (such as the loosely-coupled actuator faults) that cannot be detected by
it is theoretically sufficient for an AANN to have three hidden layers using this or any kind of control process model. For this reason, another
(Kramer, 1991; Kramer, 1992). We opted to use this particular network process behaviour model is introduced. The second process behaviour
structure because, most of the studies that we reviewed (whose input model is basically a set of AANN based signal deviation detectors that
vector dimension is similar to us – see for instance (Krothapalli & are utilized to detect whether there is any significant change (the faults
Koolagudi, 2013; Sangeetha & Jothilakshmi, 2016; Rao et al., 2014) and anomalies) in the analog PLC I/O signals and the power con-
have found the best classification results by using an AANN of the sumption profiles of the loosely-coupled actuators. MPFDT is more ef-
following structure: N-2N-kN-2N-N, where N is the number of nodes in fective and accurate than the other existing approaches mainly because
the input layer and the value of k varies from 0.2 to 0.5 (note that we of the following reasons: (i) MPFDT has been developed with real-world
simply set the k value to the middle value of the range [0.2, 0.5] i.e., applications in mind; (ii) MPFDT can handle the analog PLC I/O signals;
k = 0.35). and (iii) unlike other approaches, it can detect all types of sensor and
Most of the studies reported above have either used the standard actuator faults (such as the analog sensor faults, loosely-coupled ac-
backpropagation algorithm (see Krothapalli & Koolagudi, 2013; tuator faults, etc.) accurately. Future work planned is to develop a
Sangeetha & Jothilakshmi, 2016; Rao et al., 2014) or the Levenberg- cloud-based machine health condition monitoring and fault diagnosis
Marquardt (LM) algorithm (see Wang & Cui, 2013; Miranda et al., 2012; platform for large-scale manufacturing systems (using the state-of-the-
Nazarko & Ziemianski, 2016) as the AANN learning algorithm. We art Industrial Internet of Things (IIoT) technologies such as digital twin)
opted to use the LM algorithm for AANN training (the mean squared that can:
error is used as the performance function). The LM algorithm is actually
an interpolation between the Gauss-Newton and the gradient descent • diagnose the faults and anomalies in the senors and actuators ef-
algorithms (Wang & Cui, 2013). Informally speaking, the LM algorithm fectively and in real-time
searches for the minimum of an (nonlinear) error function by following • continuously assess the health condition (or the performance de-
the descent directions determined from both the first and second partial gradation statistics) of the industrial machines
derivatives of the error function, whereas the gradient descent method • detect and isolate the faults and anomalies (including incipient
searches for the minimum based only on the first derivatives of the faults and anomalies) in the electric motors accurately (using motor
error function (Wang & Cui, 2013). Therefore, the LM algorithm can voltage and current signature analysis techniques)
result in achieving both faster convergence and higher accuracy • can handle multiple large manufacturing systems simultaneously
(Bratina et al., 2010; Wang & Cui, 2013). The hyperbolic tangent sig- and efficiently.
moid function (respectively, pure linear function) is used as the acti-
vation function in the hidden layers (respectively, output layer) of the Funding
network (most of the studies reviewed in this paper used the same and
the study (Bratina et al., 2010) found the above combination to be the This work was supported in part by the Ministry of Trade, Industry
‘winning combination’). Please note that the input vectors are nor- and Energy (MOTIE), South Korea under Grant 20000432 and
malized into the range [−1, +1] before presenting to the neural net- 20003661; and in part by the Korea Institute for Advancement of
work. Technology (KIAT), Seoul, South Korea under Grant N0001083.
The experimental results of the MDSVTF model based FADI ap-
proach and the AANN based signal deviation detector are reported in Acknowledgement
Appendix B.I and Appendix B.II, respectively. In summary, we have
experimentally found that the FADI accuracy rate of the MDSVTF model The authors would like to thank Dong Yang Piston Co. Ltd., Ansan,
based FADI approach and the AANN based signal deviation detector are South Korea; Ajou University, Suwon, South Korea; and UDMTEK Co.
on average over 83% and 96%, respectively (recall that the AANN Ltd., Suwon, South Korea for the use of their research facilities during
based signal deviation detectors are utilized to detect and isolate the this study.
faults and anomalies that cannot be detected by using the MDSVTF
model based FADI approach). The overall FADI accuracy rate of MPFDT Appendix I. Supplementary material
is on average 98.48% (which is a quite high accuracy rate). So, the
combination of the above two fault and anomaly diagnosis techniques is Supplementary data to this article can be found online at https://
able to enhance the FADI accuracy considerably. Anyhow, as explained doi.org/10.1016/j.cie.2019.106188.
in Section 2, MPFDT will always produce better FADI accuracy than
those contained in this paper. This is because, among the aforemen- References
tioned fault and anomaly diagnosis approaches (see Section 2 for de-
tails), FBMTP is the most accurate one and MPFDT is basically a sig- Bao, J., Wu, H., & Yan, Y. (2014). A fault diagnosis system-PLC design for system relia-
nificantly improved and extended version of FBMTP (as discussed bility improvement. The International Journal of Advanced Manufacturing Technology,
75(1), 523–534.
above, unlike FBMTP, MPFDT can detect and isolate the faults and Bratina, B., Muškinja, N., & Tovornik, B. (2010). Design of an auto-associative neural
anomalies associated with the analog I/O signals and the loosely-cou- network by using design of experiments approach. Neural Computing & Applications,
pled actuators accurately). 19(2), 207–218.
Cabasino, M. P., Giua, A., Paoli, A., & Seatzu, C. (2013). Decentralized diagnosis of dis-
crete-event systems using labeled petri nets. IEEE Transactions on Systems, Man, and
6. Conclusions and future work Cybernetics: Systems, 43(6), 1477–1485.
Cabasino, M. P., Giua, A., & Seatzu, C. (2010). Fault detection for discrete event systems
using Petri nets with unobservable transitions. Automatica, 46(9), 1531–1539.
In this paper, we have presented a fault diagnosis tool, called Cabasino, M. P., Giua, A., & Seatzu, C. (2013). Diagnosis using labeled petri nets with
MPFDT that can detect and isolate the faults and anomalies in the PLC silent or undistinguishable fault events. IEEE Transactions on Systems, Man, and
controlled manufacturing systems effectively. MPFDT uses two process Cybernetics: Systems, 43(2), 345–355.
Caesarendra, W., & Tjahjowidodo, T. (2017). A review of feature extraction methods in
behaviour models to satisfy the above purpose. The first process be-
vibration-based condition monitoring and its application for degradation trend
haviour model is a MDSVTF automaton based control process model.

15
A. Ghosh, et al. Computers & Industrial Engineering 139 (2020) 106188

estimation of low-speed slew bearing. Machines, 5(4), 1–28. of Systems Science, 43(10), 1826–1841.
Cassandras, C. G., & Lafortune, S. (2008). Introduction to discrete event systems (2nd ed.). Sampath, M., Lafortune, S., & Teneketzis, D. (1998). Active diagnosis of discrete-event
New York, NY: Springer. systems. IEEE Transactions on Automatic Control, 43(7), 908–929.
Chandola, V., Banerjee, A., & Kumar, V. (2009). Anomaly detection: a survey. ACM Sampath, M., Sengupta, R., Lafortune, S., Sinnamohideen, K., & Teneketzis, D. C. (1996).
Computing Surveys, 41(3) pp. 15:1–15: 58. Failure diagnosis using discrete-event models. IEEE Transactions on Control Systems
Dotoli, M., Fanti, M. P., Mangini, A. M., & Ukovich, W. (2009). On-line fault detection in Technology, 4(2), 105–124.
discrete event systems by Petri nets and integer linear programming. Automatica, Sangeetha, J., & Jothilakshmi, S. (2016). Automatic continuous speech recogniser for
45(11), 2665–2672. Dravidian languages using the auto associative neural network. International Journal
ElMaraghy, H. A., Youssef, A. M. A., Marzouk, A. M., & ElMaraghy, W. H. (2017). Energy of Computational Vision and Robotics, 6(1–2), 113–126.
use analysis and local benchmarking of manufacturing lines. Journal of Cleaner Sayed-Mouchaweh, M., & Lughofer, E. (2015). Decentralized fault diagnosis approach
Production, 163(1), 36–48. without a global model for fault diagnosis of discrete event systems. International
Endo, M., Nakajima, H., & Hata, Y. (2012). Simplified Factory Energy Management Journal of Control, 88(11), 2228–2241.
System based on operational condition estimation by sensor data. In: Proc. 8th IEEE Schneider, S., Litz, L., & Danancher, M. (2011). Timed residuals for fault detection and
int conf automation sci and eng (CASE’12), Seoul, South Korea, Aug. 2012 (pp. 14–19). . isolation in discrete event systems. In: Proc. 3rd int workshop on dependable control of
Fan, B., Hu, N., & Cheng, Z. (2013). Fault degradation state recognition for planetary gear discrete syst (DCDS’11), Saarbuecken, Germany, June 2011 (pp. 35–40). .
set based on LVQ neural network. In: Proc. 8th world congr on eng asset manage Wang, G., & Cui, Y. (2013). On line tool wear monitoring based on auto associative neural
(WCEAM’13), Hong Kong, China, Oct. 2013 (pp. 9–18). . network. Journal of Intelligent Manufacturing, 24(6), 1085–1094.
Ghosh, A., Qin, S., Lee, J., & Wang, G. N. (2016). PLAT: an automated fault and beha- Wang, X., Zheng, Y., Zhao, Z., & Wang, J. (2015). Bearing fault diagnosis based on sta-
vioural anomaly detection tool for PLC controlled manufacturing systems. tistical locally linear embedding. Sensors, 15(7), 16225–16247.
Computational Intelligence and Neuroscience, 2016, 1–30 article ID 1652475. Qiu, W., & Kumar, R. (2006). Decentralized failure diagnosis of discrete event systems.
Ghosh, A., Qin, S., Lee, J., & Wang, G. N. (2017a). FBMTP: an automated fault and be- IEEE Transactions on Systems, Man, and Cybernetics A, 36(2), 384–395. http://
havioral anomaly detection and isolation tool for PLC-controlled manufacturing ieeexplore.ieee.org/document/1597408/https://doi.org/10.1109/TSMCA.2005.
systems. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 47(12), 853503.
3397–3417.
Ghosh, A., Qin, S., Wang, G. N., Lee, J., & Jang, H. Y. (2017b). Unified logical model to
identify faults in a PLC controlled manufacturing system. The International Journal of Arup Ghosh was born in Barasat, India on March 22, 1989.
Industrial Engineering: Theory, Applications and Practice, 24(3), 272–283. He received the master’s degree in computer science from
Hashtrudi Zad, S., Kwong, R. H., & Wonham, W. M. (2005). Fault diagnosis in discrete- the University of Trento, Trento, Italy in 2012, and the
event systems: Incorporating timing information. IEEE Transactions on Automatic Ph.D. degree in industrial engineering from the Ajou
Control, 50(7), 1010–1015. University, Suwon, South Korea in 2018. He is currently a
Hernandez-Vargas, M., Cabal-Yepez, E., & Garcia-Perez, A. (2014). Real-time SVD-based Senior Assistant Professor in the School of Information
detection of multiple combined faults in induction motors. Computers & Electrical Technology and Engineering at the Vellore Institute of
Engineering, 40(7), 2193–2203. Technology, Vellore, India. Dr. Ghosh has published more
Hodge, V. J., & Austin, J. (2004). A survey of outlier detection methodologies. Artificial than eight research articles in reputed journals and con-
Intelligence Review, 22(2), 85–126. ference proceedings. His current research interests include
Hu, W., Starr, A. G., & Leung, A. Y. T. (1999). Two diagnostic models for PLC controlled cyber-physical production system, industry 4.0, smart
flexible manufacturing systems. The International Journal of Machine Tools and manufacturing, production system health performance
Manufacture, 39(12), 1979–1991. monitoring, big data analysis, and intelligent process con-
Hu, W., Starr, A. G., & Leung, A. Y. T. (2003). Operational fault diagnosis of manu- trol. Dr. Ghosh was a recipient of the Erasmus Mundus
facturing systems. The Journal of Materials Processing Technology, 133(1–2), 108–117. External Cooperation Window (EMECW) scholarship in 2010, and the Graduate
Jack, H. (2010). Chapter 21. Analog inputs and outputs. Automating manufacturing systems Scholarship from the Ajou University, Suwon, South Korea in 2013. He also received
with PLCs (version 7.0 - April 2010) (pp. 413–434). (7th ed.). Morrisville, NC: Lulu several grants from different reputed South Korean institutions/organizations (as a
Press. Research Assistant) starting from 2013 to 2018.
Jiang, S., & Kumar, R. (2004). Failure diagnosis of discrete-event systems with linear-time
temporal logic specifications. IEEE Transactions on Automatic Control, 49(6), 934–945.
Kramer, M. A. (1991). Nonlinear principal component analysis using autoassociative Gi-Nam Wang received the M.S. degree in industrial en-
neural networks. American Institute of Chemical Engineers Journals, 37(2), 233–243. gineering from the Korea Advanced Institute of Science and
Kramer, M. A. (1992). Autoassociative neural networks. Computers & Chemical Technology, Daejeon, South Korea in 1984, and the Ph.D.
Engineering, 16(4), 313–328. degree in industrial and system engineering from the Texas
Krothapalli, S. R., & Koolagudi, S. G. (2013). Characterization and recognition of emo- A&M University, Texas, USA in 1993. Since 1993, he has
tions from speech using excitation source information. International Journal of Speech been a faculty member with the Ajou University, Suwon,
Technology, 16(2), 181–201. South Korea, where he teaches courses on advanced in-
Kumar, R., & Takai, S. (2009). Inference-based ambiguity management in decentralized formation system design and practice, neural networks, and
decision-making: decentralized diagnosis of discrete-event systems. IEEE Transactions intelligent manufacturing systems. He is currently a Full
on Automation Science and Engineering, 6(3), 479–491. Professor at the Department of Industrial Engineering, Ajou
Laughton, M. A., & Warne, D. F. (2003). Section C – control: chapter 16 programmable University, and also the CEO of UDMTEK Corporation,
controllers. Electrical engineer’s reference book (pp. 16/3–16/52). (16th ed.). Suwon, South Korea. He is the author of sixteen patents,
Burlington, MA: Elsevier Science. and has published more than fifty original research papers
Liu, H., Wang, X., & Lu, C. (2014). Rolling bearing fault diagnosis under variable con- in international journals, book chapters, and international
ditions using hilbert-huang transform and singular value decomposition. conference proceedings. His current research interests include digital manufacturing,
Mathematical Problems in Engineering, 2014(2014), 1–10 Article ID 765621. cyber-physical systems, virtual manufacturing, intelligent process control, and PLC con-
Lunze, J., & Schroder, J. (2004). Sensor and actuator fault diagnosis of systems with trol system design, modelling and simulation.
discrete inputs and outputs. IEEE Transactions on Systems, Man, and Cybernetics Part B:
Cybernetics, 34(2), 1096–1107.
Miranda, V., Castro, A. R. G., & Lima, S. (2012). Diagnosing faults in power transformers Jooyeoun Lee received the master’s degree in business
with autoassociative neural networks and mean shift. IEEE Transactions on Power administration from the Ajou University, Suwon, South
Delivery, 27(3), 1350–1357. Korea in 1993, and the Ph.D. degree in business adminis-
Nazarko, P., & Ziemianski, L. (2016). Damage detection in aluminum and composite tration from the Inha University, Incheon, South Korea in
elements using neural networks for Lamb waves signal processing. Engineering Failure 2004. In past, he worked at Green Business Division,
Analysis, 69(Nov.), 97–107. POSCO ICT as an Executive Vice President, and Strategic
Rao, K. S., Nandi, D., & Koolagudi, S. G. (2014). Film segmentation and indexing using Marketing Division, SK C&C as a Vice President, and
autoassociative neural networks. International Journal of Speech Technology, 17(1), Strategic Solution Division, Oracle as a Senior Director.
65–74. Since 2014, he has been an Associate Professor in the
Roth, M., Lesage, J. J., & Litz, L. (2009). A residual inspired approach for fault locali- Department of Industrial Engineering, Ajou University. He
zation in DES. In: Proc. 2nd IFAC workshop on dependable control of discrete system is also an Ombudsman for Industrial Convergence, Ministry
(DCDS’09), Bari, Italy, June 2009 (pp. 305–310). . of Trade, Industry and Energy (MOTIE), South Korea. His
Roth, M., Lesage, J. J., & Litz, L. (2011). The concept of residuals for fault localization in current research interests include smart grid, smart factory,
discrete event systems. Control Engineering Practice, 19(9), 978–988. virtual or augmented reality, business intelligence, smart
Roth, M., Schneider, S., Lesage, J. J., & Litz, L. (2012). Fault detection and isolation in manufacturing, and ICT convergence technology.
manufacturing systems with an identified discrete event model. International Journal

16