See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/271302599

Fault Diagnosis of Discrete Event Systems Using Hybrid Petri Nets

Article · January 2012

DOI: 10.7763/IJCTE.2012.V4.468

CITATIONS READS

2 50

3 authors, including:

Mohammad Ali Badamchizadeh

University of Tabriz
111 PUBLICATIONS   894 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Virtual Network Embedding View project

Stability of Switched Delay Systems View project

All content following this page was uploaded by Mohammad Ali Badamchizadeh on 08 February 2016.

The user has requested enhancement of the downloaded file.

 International Journal of Computer Theory and Engineering Vol. 4, No. 2, April 2012

Fault Diagnosis of Discrete Event Systems Using Hybrid

Petri Nets
R. Rangarangi Hokmabad, M. A. Badamchizadeh, and S. Khanmohammadi

 cause a distinct change in the state of system components but

Abstract—A new method for fault diagnosis of discrete event do not bring the system to a halt: examples are equipment
systems modeled by Neural Petri Nets (NPNs) is presented in failures (stuck failure of valves, stalling of actuators,
this paper. Assuming that the PN structure and initial marking controller failures, etc.) usual in flight control systems or
are known, faults are modeled by unobservable transitions.
heating and air conditioning systems, and process failures
Neural networks (NNs) have important role to improve the
method. The outputs of them are connected to unobservable (buffer overflow) usual in manufacturing systems [3].
transitions and yield the percentage of faults that may happen Recently, neural networks (NNs) have been applied to the
for prioritize the faults by online computation of the set of fault diagnosis problem because of their good capabilities in
possible fault events. In this method the operator checks the function approximation. Specifically, online approximation
fault that has more value at first. So we reduce the time that method using NNs has been presented for identifying the
spends for repairing the system. Moreover, the graphical
fault functions [4], [5].
representation of the nets allows the diagnoser agent to compute
off-line reduced portions of the net in order to improve the In [6] a new architecture for a fault diagnosis competitive
efficiency of the online computation, without a big increase in neural network is introduced. In this system, the test matrix
terms of memory requirement. and the probability vector of faults are not known a prior. The
neural system starts from a completely vague state and the
Index Terms—Discrete event systems (DES), fault diagnosis, weights of connections, which affect the possibility of
Neural networks (NNs), Petri nets (PNs) detecting the fault in each unit, are modified during the
learning procedure on the base of different tests.
In [7] faults are not explicitly taken into account in the
I. INTRODUCTION
model, and two types of faults have been defined: a place
The diagnostics of industrial processes is a scientific fault that corrupts the net marking, and a transition fault that
discipline aimed at the detection of faults in industrial plants, causes an incorrect update of the marking after events
their isolation, and finally their identification. Its main task is occurrences and In [3] just the maximum and minimum of
the diagnosis of process anomalies and faults in process faults are detected and they are not sure if the specific fault
components, sensors and actuators. Early diagnosis of faults would occurs or not.
that might occur in the supervised process renders it possible In this paper a new technique for the fault diagnosis is used.
to perform important preventing actions. Moreover, it allows Faults are modeled by unobservable transitions. Moreover,
one to avoid heavy economic losses involved in stopped we assume that there may be additional unobservable
production, the replacement of elements and parts [1]. transitions associated with the system legal behaviour and
The operation of large and complex systems requires that that the marking reached after the firing of any transition is
the coordination systems possess fault recovery capabilities.
unknown. The Petri nets (PNs) are connected to NNs and the
In automated manufacturing systems, this ability is included
weights of NNs are training then the outputs of NNs specify
mainly to eliminate unnecessary risks to humans or
the percentage of each fault that may happen. The proposed
hazardous situations into the system as well as to maintain the
production rate. However, introducing this capability to the diagnoser works on-line: it waits for the firing of an
coordination system possesses challenging problems that observable transition and employs an algorithm based on the
have been addressed through several approaches and definition and solution of some integer linear programming
methods. Most of these approaches include stages, such as problems to decide whether the system behaviour is normal
detection, isolation, and confinement of the fault [2]. or exhibits some possible faults. The results characterize the
Discrete event systems (DES) based methodologies for properties that the PN modeling the system fault behaviour
fault diagnosis are applicable not only to systems normally has to fulfill in order to reduce the on-line computational
modeled as DES, but also to systems that traditionally are effort.
treated as continuous-time dynamic systems. In general, DES A good compromise to speed up the online diagnosis is to
approaches to fault diagnosis are suitable for failures that precompute something offline. This is particularly efficient
when PN are used to model the plant. In fact, working on the
Manuscript received February 20, 2010; revised March 25, 2012. net structure some useful information can be computed
Roya Rangharanghi Hokmabad has been graduated from University of
Tabriz (e-mail: [email protected]). offline improving the efficiency of the online diagnosis,
Mohammad Ali Badamchizadeh is with the Department of Control, without a big increase in terms of memory request. In this
Faculty of Electrical and Computer Engineering, University of Tabriz paper it is shown that the programming problems to be solved
(e-mail: [email protected]).
Sohrab Khanmohammadi is with the Control Engineering Department, by the diagnoser can be formulated on reduced portions of the
Faculty of Electrical and Computer Engineering, University of Tabriz net properly computed offline.
(e-mail:[email protected]).

288
 International Journal of Computer Theory and Engineering Vol. 4, No. 2, April 2012

II. BACKGROUNDS AND BASIC ASSUMPTIONS the enabled sequence σ may fire at M yielding M’. We also

A. Basic Petri Nets Notation denote  (t )  q the firing vector associated with a sequence

Petri nets (PNs) are a graphical and mathematical σ, i.e.,  (t )  q if transition t is contained q times in σ. A
modeling tool applicable to many systems. They are a marking M is said reachable from  PN , M  iff there
promising tool for describing and studying information 0
processing systems that are characterized as being concurrent, exists a firing sequence σ such that M [  M . The set of
0
asynchronous, distributed, parallel, nondeterministic, and/or
TABLE I: FORMAL DEFINITION OF A PETRI NET
stochastic. As a graphical tool, PNs can be used as a visual
communication aid similar to flow charts, block diagrams, A Petri net is a 5-tuple, PN = (P, T, F, W, M ) where:
and networks. In addition, tokens are used in these nets to 0
simulate the dynamic and concurrent activities of systems. As P  { p , p ,..., p } is a finite set of places,
1 2 m
a mathematical tool, it is possible to set up state equations,
algebraic equations, and other mathematical models T  {t , t ,..., t } is a finite set of transitions,
1 2 n
governing the behavior of systems. A Petri net (PN) is a Pre : P × T → N (Post : P × T → N)
particular kind of directed graph, together with an initial state C=Post-Pre
called the initial marking, M 0 . The underlying graph N of a F  (P x T) U (T x P) is a set of arcs (flow relation),
W: F {1, 2, 3, ... } is a weight function,
PN is a directed, weighted, bipartite graph consisting of two Mo: p → {0, 1, 2, 3, ... } is the initial marking,
kinds of nodes, called places and transitions, where arcs are P ∩ T =  and P U T ≠  .
either from a place to a transition or from a transition to a A Petri net structure N = (P, T, F, W) without any specific initial marking is
denoted by N.
place. In graphical representation, places are drawn as circles,
transitions as bars or boxes. Arcs are labeled with their A Petri net with the given initial marking is denoted by (N, M ).
0
weights (positive integers), where a k-weighted arc can be
interpreted as the set of k parallel arcs. Labels for unity all markings reachable from M defines the reachability set
0
weight are usually omitted. A marking (state) assigns to each
place a non negative integer. If a marking assigns to place p a of  PN , M  and is denoted by
0
nonnegative integer k, we say that p is marked with k tokens.
R( PN , M )  {M |  : M [  M } .
Pictorially, we place k black dots (tokens) in place p. A 0 0
marking is denoted by M, an m-vector, where m is the total
C. Neural Petri Net
number of places. The pth component of M, denoted by M(p),
is the number of tokens in place p [8].For a complete review Artificial neural networks (ANN) are highly parallel and
on PNs refer to [8]. A formal definition of a PN is given in distributed computation structures that can learn from
Table I. experience and perform inferences. PNs, on the other hand,
provide an effective modeling framework of distributed
B. Marking Projections systems. The basic concepts of PNs are utilized to develop
Definition : A PN [8] is a bipartite graph described by ANN-like multilayered PNs architectures of distributed
PN=(P, T, Pre, Post), where P is a set of places with systems [9].
cardinality m, T is a set of transitions with cardinality n, Pre: The NPN is formally defined as a 6 tuple: NPN = (P, T, Z,
P×T→N and Post: P×T→N are the pre- and post-incidence A,C, M 0 ) where:
matrices, respectively, which specify the arcs connecting P: is a set of places;
places and transitions. Matrix C=Post-Pre is the m × n T: is a set of transitions;
incidence matrix of the net PN. Table I shows a formal Z: is set of arcs, Z  (P×T)  (T×P);
definition of a PN. A: is a pattern of connectivity among places and transitions;
For the pre- and post-sets we use the dot notation, e.g., C: is a set of states of outputs of NNs.
t={p  P: Pre(p,t)>0}. The state of a PN is given by its The resulting NPN is a feed forward network with
current marking, which is a mapping M: P→N, assigning to alternating columns of transitions.
each place of the net a nonnegative number of tokens. A PN In [10], they built ANN-like architectures of distributed
intelligence that can learn from experience. The resulting
system 〈  PN , M  is a net PN with an initial marking
0 Neural Petri Net (NPN) is feedforward network with
M 0 . A transition t f  T is enabled at a marking M if and alternating columns of places ad transition.
The NPN is a pure PN (self –loops are not allowed). This
only if (iff) for each p  .t j , it holds M(p)≥Pre(p, t j ) and we leads to a feedforward architecture. The interaction of ANN
with the environment is through the unobservable transitions
write M [t  to denote that t j  T is enabled at marking and places. Fig.1 shows a simple NPN. t is an unobservable
j 6
M. Let   t b t b ...t b be a sequence of transitions and let transition and models a faulty behavior of a system. The
1 2 k
inputs of ANN { X ...X } are connected to places (a, b…e)
k=|σ| be its length, given by the number of transitions that σ 1 5
contains. If a transition t  T appears in the sequence σ, we and the output of ANN yields the percentage of fault t that
write t   . Moreover, the notation M [  indicates that 6
may happen.
the sequence σ is enabled at M and M [  M  indicates that

289
 International Journal of Computer Theory and Engineering Vol. 4, No. 2, April 2012

either the firing of an observed transition, or the firing of an

unobservable transition that must have fired.
As far as the fault diagnosis is concerned, the g-markings
allow the fault diagnosis agent to store in a compact way all
the needed information about the state space estimation.

IV. FAULT DETECTION

The approach used in this paper is based on the fact that the
firing of an observable transition requires a proper marking of
its input places. If the same event is associated to more than
one transition, the observation of an event could not
necessarily correspond to the firing of a single transition.
Fig. 1. A neural Petri net Taking into account this problem complicates too much
the approach. A similar assumption is not required for
unobservable transitions, which can be assumed to be
III. GENERALIZED MARKING associated to an arbitrary event, since such event is
Let m be the current net marking, C=Post-Pre be the unobservable.
incidence matrix and σ be a firing sequence of transitions If an observable transition fires then if the new marking
 t ...t such that m[t  m [t  m ...[t  m , and this has negative components, the fault diagnoser will start
1 k 1 1 2 2 k k working. The inputs of ANN are connected to the places. If
is denoted as m[  mk . If a sequence σ fires, a new marking the number of tokens in a place was positive, the related input
of ANN would be 0 and if the number of tokens in another
is reached. From the state equation it follows that for the
place was negative the related input of ANN would be 1. The
firing count vector σ it is possible to write
outputs of ANN yield the percentage of faults that may
m  C  m  0 . happen. In this way we prioritize the faults and at first the
Suppose that where ε is a sequence of unobservable operator checks the fault that has more value. So in this
transitions and t  T . Let   m  Ce  m  C (.,t ) . It may method we reduce the time that spends for repairing the
0 t
happen that μ has negative components, since t may not be system. The difference between outputs of ANN and the
enabled under the marking m. The negative components in μ value that operator had have after checking the system be the
mean that the unobservable sequence ε must have fired in error and the weights of ANN is training then these stages
order to explain the firing of t, which is the unique observed repeat again.
event. A marking that may have negative components is
2
called g-markin g .
Suppose that a fault event is associated to the unobservable
transition t f , and that one wants to know if t has occurred
f
prior to the observation of t. Note that it is not necessary to
compute explicitly ε or its firing count vector ε, but simply to
check if ε( t f ) is greater than zero.
The following example shows how the evolution rules for
the g-marking given above can lead to negative marking
components. For a complete review on g-marking refer to [3].
Example : Consider the net N in Fig. 2 and let 0  [0
0 0 1 0 0] T be the initial g-marking of the net and {t 6 ...t8 }
be unobservable transitions. If the firing of t  T is
3 0
observed, then t may fire, since an observable transition is
3
always enabled under any g-marking. The firing of t yields
3
the g-marking 1  mu 0  C(., t3 )  [0 -1 0 2 1 0] T . The
negative marking   1 means that an unobservable
1| p
2

sequence must have fired to explain the firing of t .

3
Throughout this paper, the negative components of a Fig. 2. Example of Petri net model
g-marking represent the tokens that are needed to explain

290
 International Journal of Computer Theory and Engineering Vol. 4, No. 2, April 2012

Example: A Manufacturing System system normal behaviour. In order to achieve this result,
Let us consider the manufacturing system whose PN model g-markings are introduced in this paper. G-markings are net
is shown in Fig. 3, with T  {t , t , t , t , t , t , t , t , t , t } markings that may have negative components and whose
0 1 2 3 4 5 6 7 8 9 10 estimation is always unique. The online computation consists
and T  {t , t , t }. of solving programming problems formulated on net
u 11 12 13
0 structure and based on g-markings.
This manufacturing system consists of three machines,
TABLE II: EXECUTION OF THE FAULT DETECTION ALGORITHM ON THE NET
M , M 2 and M , which process parts conveyed on OF FIG. 3.
1 3
equipments. The system has two operators, F1 and F2 . F1 Action μ The The The
percenta percent percent
could manage M and M . F2 could manage M and M . ge of age of age of
1 2 1 3
fault t11 fault fault
In each equipment the operation is performed in two stages.
may t12 t
Stage1: works with M . 13
1 happen may may
happen happen
Stage2: works with M or M
2 3
• t1 fires when a new equipment is received. Initialization [0 0 0 1 1 1 1 1 0 0 0 33% 33% 33%
0]
• t 2 fires when F1 starts working on equipment with M .
1 [1 0 0 1 1 1 1 1 0 0 0 --- --- ---
t1 fires
• t 3 fires when F1 finishes working on equipment with M . 0]
1
• t 4 fires when F2 starts working on equipment with M . t4 fires [0 0 0 0 1 1 1 0 0 1 0 --- --- ---
1 0]
• t 5 fires when F2 finishes working on equipment with M .
1 t8 fires [0 -1 0 0 1 0 1 -1 0 1 10.35% 15.63 80.2%
0 1] %
• t 6 fires when F1 starts working on equipment with M 2 .
• t 7 fires when F1 finishes working on equipment with M 2 . t9 fires [0 -1 1 0 1 1 1 0 0 1 0 12.32% 40.33 39.92
0] % %
• t 8 fires when F2 starts working on equipment with M .
3 [0 0 1 1 1 1 1 1 0 0 0 --- --- ---
t5 fires
• t 9 fires when F2 finishes working on equipment with M . 0]
3
• t10 fires when the operation on equipment is finished. t2 fires [-1 0 1 0 1 1 0 1 1 0 0 75.32% 14.26 10.23
0] % %
• t11 ,t12 and t13 model the faulty behaviors of M1 , M 2 and
M . t6 fires [-1 -1 1 0 0 1 -1 1 1 0 80.64% 65.45 12.78
3 1 0] % %
• The initial number of tokens in p4 , p5 , p6 , p7 and P8
model the fact that the machines and operators are waiting. With respect to the approaches proposed in the related
Some execution steps of the proposed method are reported in literatures, the proposed method specify the percentage of
Tab. II. In this case, if the estimated g-marking has negative faults in order to provide a reasonably efficient method
components, then the firing of transition that are modeled as suitable use with large systems. However, the algorithm use
faults would test and NNs that are connected to these some off-line calculations based on the structure of the
transitions would work and compute the percentage of faults
considered Petri net (PN) system in order to decrease the
may happen to perform fault detection and identification.
memory capacity. In this way, the proposed fault detection
technique can be more easily applicable.
V. CONCLUSION Further improvements in the efficiency of the proposed
method could be obtained if we assume that after an event
The paper addresses the fault detection problem of
sequence occurrence the reached marking is known or
Discrete Event Systems (DES) and proposes an on-line
univocally determined. In this situation, an incremental
diagnoser in a Neural Petri Net (NPN) framework. A solution approach could be devised. However, identifying the
procedure observes and stores the sequence of system events conditions necessary to univocally determine the reached
and decides on-line whether the system behavior is normal or marking is expected to require a significant amount of effort
some faults may have occurred. To this aim, at each observed to be specified and developed. Hence, this issue will be
event it is provide the possible occurred faults or certifies the tackled in a successive work.

291
 International Journal of Computer Theory and Engineering Vol. 4, No. 2, April 2012

Fig. 3. NPN model of the manufacturing system.

Roya Rangharanghi Hokmabad was born in Tabriz,

Iran in 1982. She received her B.S. degree in Electrical
REFERENCES Engineering from Azad University of Tabriz, Iran in
[1] K. Patan, Artificial Neural Networks for the Modelling and Fault 2005 and her Ms. C. degree in Control engineering form
Diagnosis of Technical Processes, Springer Pub, 2008. University of Tabriz in 2011. Her research interests are
[2] R. Isermann, “Process fault detection based on modeling and fault tolerant, Neural networks, Intelligent control.
estimation methods. A Survey,” Automatica, vol. 20, pp. 387–404,
1984.
[3] F. Basile, P. Chiacchio, and G. Tommasi, “An efficient approach for
online diagnosis of discrete event systems,” IEEE Trans. Automatic
control, vol. 54, no. 4, pp. 748-759, Apr. 2009. Mohammad Ali Badamchizadeh was born in Tabriz,
[4] S. Huang and K. K. Tan, “Fault Detection and Diagnosis Based on Iran, in December 1975. He received the B.S. degree
Modeling and Estimation Methods,” IEEE Trans. On neural networks, in Electrical Engineering from University of Tabriz in
vol. 20, no. 5, pp. 872-881, May 2009. 1998. He received the M.Sc. and Ph.D. degree in
[5] A. T. Vemuri and M. M. Polycarpou, “Neural-network-based robust Control Engineering from University of Tabriz in
fault diagnosis in robotic systems,” IEEE Trans. Neural Netw, vol. 8, 2001 and 2007, respectively. He is now assistant
no. 6, pp. 1410–1420, Nov. 1997. professor in the Faculty of Electrical and Computer
[6] S. Khanmohammadi, I. Hasanzadeh, and H. R. Zarei Poor, “Fault engineering at University of Tabriz. His research
diagnosis competitive neural network with prioritized modification interests include Hybrid dynamical systems, Stability
rule of connection weights,” ELS. Artificial Intelligence in Engineering, of systems, Adaptive Control and intelligent systems.
vol. 14, pp. 127-132, 2000.
[7] Y. Wu and C. N. Hadjicostis, “Algebraic approaches for fault Sohrab Khanmohammadi received his B.S. degree
identification in discrete-event systems,” IEEE Trans. Automat. in Industrial Engineering from Sharif University, Iran
Control, vol. 50, no. 12, pp. 2048–2053, Dec. 2005. in 1977 and M.Sc. degree in Automatic from
[8] T. Murata, “Petri nets: Properties, analysis and applications,” Proc. University Paul Sabatie, France in 1980 and his Ph.D.
IEEE, vol. 77, no. 4, pp. 541–580, Apr. 1989. degree in Automatic from National University,
[9] P. T. T. Binh and N. D Tuyen, “Fault Diagnosis of Power System ENSAE, France in 1983. He is now a professor of
Using Neural Petri Net and Fuzzy Neural Petri Net,” IEEE, 2006. Electrical Engineering at University of Tabriz. His
[10] Richard Zurawski and Mengchi Zhou “Petri Nets and industrial research interests are Fuzzy control, Artificial
application: A tutorial,” IEEE Trans. on industrial electronic, vol. 41, Intelligence applications in Control and Simulation on
no. 6, Dec, 1994. industrial systems and human behavior. Email: [email protected]

292

View publication stats