You are on page 1of 14

ARTICLE IN PRESS

Engineering Applications of Artificial Intelligence 21 (2008) 157–170


www.elsevier.com/locate/engappai

A fuzzy reasoning design for fault detection and diagnosis of a


computer-controlled system
Y. Ting, W.B. Lu, C.H. Chen, G.K. Wang
Department of Mechanical Engineering, Chung Yuan Christian University, 200, Chung Pei Road, Chung Li, Taiwan 32023, ROC
Received 15 March 2005; received in revised form 23 January 2007; accepted 27 April 2007
Available online 5 July 2007

Abstract

A fuzzy reasoning and verification Petri nets (FRVPNs) model is established for an error detection and diagnosis mechanism applied
to a complex fault-tolerant PC-controlled system. The inference accuracy can be improved through the hierarchical design of a two-level
fuzzy rule decision tree and a Petri nets technique to transform the fuzzy rule into the FRVPNs model. Several simulation examples of
the assumed failure events were carried out by using the FRVPNs and the Mamdani fuzzy method with MATLAB tools. The reasoning
performance of the developed FRVPNs was verified by comparing the inference outcome to that of the Mamdani method. Both methods
result in the same conclusions. Thus, the present study demonstrates that the proposed FRVPNs model is able to achieve the purpose of
reasoning, and furthermore, determining of the failure event of the monitored application program.
r 2007 Elsevier Ltd. All rights reserved.

Keywords: Error detection; Error diagnosis; Petri nets; Fuzzy rule; Fuzzy reasoning

1. Introduction and makes predictable estimation on the executing APs. As


been known, while the AP is failed, the OS of computer
The configuration of an error detection and diagnosis sends an error message with illustration of the failure event.
mechanism (EDDM) for a fault-tolerant computer-con- However, most displayed error messages are difficult for the
trolled system has been described in the previous researches user to understand, letting alone for them to know the
(Ting et al., 2002; Ting et al., 2004). The detection damage level, so that one can hardly deal with the failure
mechanism employs the hook process to capture the appropriately. Also, when different APs are executed in the
message in and between the various application programs PC, there may be similar error message but with different
(APs) and the operating system (OS), and detects whether error definition and illustration in different programming
the monitored AP is failed. Establishment of error language (Inprise Corporation, 2002; Richter, 1999). There-
classification and standardization was developed in the fore, it is demanding to investigate how to unify the
previous research (Lu et al., 2003). The diagnosis mechanism illustration of error message and examine the closeness
identifies the failure type and the location of error message, degree of the same failure event but with different error
description, so that a unified error knowledge database
could be established, and then the error symptoms could be
Abbreviations: CPU, central processing unit; DLDM, damage level
decision-making; EDDM, error detection and diagnosis mechanism; EHL,
better inferred via the reasoning algorithm to acquire a final
error hazard level; FRDT, fuzzy rule decision tree; FRVPNs, fuzzy decision-making. The proposed EDDM, which is a fault-
reasoning and verification Petri nets; HT, handling time; MD, member- tolerant computer-controlled system, is designed and aimed
ship degree; MF, membership function; PCPUU, process CPU usage; to satisfy the above needs. Petri nets (PNs) is an ideal
PNs, Petri nets; ProMF, membership function of proposition; RFS, firing candidate for investigating and modeling of systems, and it
strength of rule; RT, response time; TCPUU, total CPU usage; TMU,
total memory usage; WRFSwinning-rule firing strength can represent the inference process as a discrete-event
Corresponding author. Tel.: +886 3 2654319; fax: +886 3 2654096. dynamic system. The advantages of using PNs in rule-based
E-mail address: yung@cycu.edu.tw (Y. Ting). systems include: (1) the graphical formalism, which can

0952-1976/$ - see front matter r 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.engappai.2007.04.007
ARTICLE IN PRESS
158 Y. Ting et al. / Engineering Applications of Artificial Intelligence 21 (2008) 157–170

I RFS
Notation Pk Firing strength of rule
MAXð Þ
i ith input object of an antecedent proposition, R MAX composition operation
i ¼ 1, 2, 3, y PWRFS
k place of winning-rule’s firing strength
j jth fuzzy set of input object of an antecedent I FSC
Rk comparison of firing strength
proposition, j ¼ 1, 2, 3, y I WFR
Pk winning firing rule
p, t, l, s, h, k, m a serial positive integer; p, t, l, s, h, k, I RFS
P½k a set of ‘‘firing strength of rules’’ except the
and m ¼ 1, 2, 3, y firing strength of the winning rule
N a set of non-negative integers I CProO
P linguistic variable of output object of a
Pk kth place in the PNs consequent proposition
Rm mth transition in the PNs I FDR
Rk final decision of winning rule
FRk kth fuzzy rule I DMI
P decision-making identification of winning firing
I AProIO
Pk Input object of antecedent proposition rule
I AProD I CProM
Rk distribution of input object of antecedent Pk marking of consequent proposition
proposition I DMI
Rk identification of consequent proposition
I APro
Pij antecedent proposition of rule I CPro
Pk result of consequent proposition of a rule
I AProMF
Rij calculation of MD of antecedent proposition I CProMD
Rk calculation of MD of consequent proposition
by MF by MF
I
WFij weight function I CProMD
Pk MD of consequent proposition
I AProMD
Pij MD of antecedent proposition RFDM
k final decision-making computation
I RFS
Rk calculation of firing strength of a rule by MIN P FDM
final decision-making result
composition operation

visualize the inference states step by step; (2) the transparent knowledge description and fuzzy reasoning, and developed a
modeling, which has well-established formal mechanisms for weighted fuzzy PNs method (Chen et al., 1990; Chen, 2002).
modeling and structure inconsistency checking; (3) the Looney developed an algorithm of rule-based decision-
analyzing capability, which can express dynamic and making by using fuzzy PNs (Looney, 1998).
structural behaviors of a rule-based system via algebraic In this study, it is attempted to use the PNs, the fuzzy
forms (Scarpelli et al., 1996; Tsang et al., 1999; Yang et al., sets theory, and the fuzzy production rule to establish the
2003). The fuzzy PNs combines the graphical technique of fuzzy reasoning and verification Petri nets (FRVPNs)
PNs, the fuzzy sets theory, and the fuzzy production rule, so model for the diagnosis mechanism. The FRVPNs is
it has the advantage of the graphical power of PNs and the designed based on the fuzzy rule decision tree (FRDT) with
capability of fuzzy to model rule-based decision system the merit of hierarchical structure. It provides different
effectively (Fay, 2001). Thus, fuzzy PNs would outperform level of abstraction, which can be used to represent the sub-
the PNs and improve the efficiency of fuzzy reasoning (Yang model construction and the rule decision for independent
et al., 1997). In general, fuzzy rule-based system consists of level of the FRVPNs. Regarding the efficiency of
knowledge base, database of event facts, rule base, and hierarchical design, it has been discussed in several research
inference engine. In order to deal with the input information articles. For example, Delgado et al. proposed a multi-
from the non-fuzzy system, the fuzzy inference mechanism objective decision making scheme, allowing the evolution-
will carry on fuzzification, rule matching, and defuzzification ary parameters based on a hierarchical genetic fuzzy system
in advance to allow the likely incomplete and imprecise adjustable, so it not only improves the models performance
information to match the antecedent proposition of the (accuracy), but also guarantees the interpretability of the
fuzzy rule and then obtain an inference result. Quite a few resulting fuzzy models (Delgado et al., 2002); Joo et al. also
researches use PNs to construct the fuzzy rule-based system demonstrated a scheme of hierarchical fuzzy system with
(Chen et al., 1990; Koriem, 2000; Chen, 2002). The input constraints on the fuzzy rules approximating with high
and output places of each transition in a PNs can be used to accuracy and fewer fuzzy rules (Joo and Lee, 2005); Wang
represent the knowledge of the antecedent and consequent proposed a hierarchical structure in a fuzzy system
propositions by the fuzzy production rule. Many researches achieving good approximation accuracy (Wang, 1999).
have investigated on the extension of PNs to fuzzy PNs, and According to these works, it is therefore believed that
the modeling as well as the reasoning of fuzzy rule-based improvement of inference accuracy can be achieved by use
system. For instance, Gao et al. used the fuzzy reasoning of the hierarchical design.
PNs model to describe the production rule-based system and Since different types of AP may have different definition
verify its performance as a diagnostic expert system on the of the failure event, it is likely to describe the error
turbine (Gao et al., 2003). Chen used fuzzy PNs to construct symptoms in terms of linguistic variables via the fuzzy set
ARTICLE IN PRESS
Y. Ting et al. / Engineering Applications of Artificial Intelligence 21 (2008) 157–170 159

theory, and construct the inference contexts by use of the handling time (HT), and error hazard level (EHL), etc.
fuzzy production rule. The fuzzy production rule makes are considered for fuzzy reasoning. Fig. 1 demonstrates the
use of the ‘‘IFyTHENy’’ rule to describe the conditions procedure: the EDDM measures these parameters of the
of the antecedent and consequent propositions. The monitored AP, and sends them to the fuzzification module.
membership degrees (MDs) of the propositions on the And then the fuzzification module handles the measured
fuzzy rule are calculated via the membership function parameters of the monitored AP by fuzzification. By using
(MF). Then, the damage level decision-making is deter- the fuzzified parameters, the fuzzy reasoning module will
mined by using the fuzzy reasoning method. In addition, search a corresponding fuzzy rule, read an object knowl-
the rule-checking process and the verification and mod- edge unit from the knowledge database, and then infer an
ification module are included in the FRVPNs. The former outcome of the consequent proposition according to the
one is used to confirm the correctness of the winning rule. input of the antecedent proposition of the fuzzy rule and
The latter one is used to deal with the problem of the contents of the object knowledge unit. Once the
redundancy, conflict, circularity, and incompleteness while parameter fuzzification fails to acquire appropriate match-
new fuzzy rules are added. The rule verification and ing, the fuzzified parameters are sent to the rule self-
modification module will address in another article. learning module. The rule self-learning module collects the
Simulations are carried out with several sets of examples new available information from the user or the information
by using the developed FRVPNs and the fuzzy logic record database, and then carries on self-learning before
toolbox of MATLAB. As compared with the inferred constructing new rules. If it is impossible to acquire a
results, both methods draw the same conclusions. corresponding fuzzy rule from the fuzzy rule base, the
domain expert will be allowed to input the domain
2. Fuzzy diagnosis reasoning structure of EDDM knowledge to the knowledge acquisition module. The rule
verification and modification module will detect whether
The EDDM is proposed for a fault-tolerant computer- the new rules have problems of redundancy, conflict,
controlled system and is a middleware structured in and circularity, and incompleteness by using the input/output
between the user applications of the user mode and the transition function of the place and the inhibit condition.
kernel mode for error detection and diagnosis of the AP. Hence, error of inference and unnecessary expense of
The EDDM consists of the retrieve process, the detection inference due to inaccuracy or redundancy of fuzzy rule
mechanism, the diagnosis mechanism, the information can be reduced; in other words, inference accuracy can thus
record database, and the knowledge base (Ting et al., be improved. The fuzzy inference results will be saved
2004). Several important parameters such as the total to the information record database and the knowledge base
central processing unit (CPU) usage (TCPUU), process of the EDDM, and be sent to the defuzzification module
CPU usage (PCPUU), total memory usage (TMU), for defuzzification and decision-making. Finally, the

Diagnosis Mechanism of EDDM

Reasoning Processor of
Diagnosis Mechanism
Rule Self-Learning
read/write
Module
input

Knowledge Rule Verification


Fuzzy Rule
Acquisition write and Modification read/write Results and
Module Base
Module Illustration

require

Measurement Fuzzification Fuzzy Reasoning Defuzzification


input send send output
Paratemers Module Module Module

read read/write
input write

Information Record Knowledge


Domain
User Database (IRB) Base (KB)
Expert
Databases of EDDM

Fig. 1. Block diagram of fuzzy diagnosis reasoning structure.


ARTICLE IN PRESS
160 Y. Ting et al. / Engineering Applications of Artificial Intelligence 21 (2008) 157–170

defuzzification module delivers an output of damage input space and described by the linguistic variables such as
affection to the monitored AP and the OS. ‘‘Small’’, ‘‘Medium’’, and ‘‘Large’’; Cb is the consequent
proposition defined in a fuzzy region of specified output
3. Representation of fuzzy sets and rules and described by linguistic variables such as ‘‘Low’’,
‘‘Medium’’, and ‘‘High’’; CFkA[0, 1] is called the certainty
3.1. Definition of fuzzy sets and linguistic variables factor (CF); mk is the value of the CF that represents the
strength of the belief in the rule FRk.
The input variables of the antecedent proposition In general, composite fuzzy production rule is classified
considered for the fuzzy rule include the TCPUU, the into several rule types (Gao et al., 2003). The rule type of
PCPUU, the TMU, the HT, and the EHL. The output ‘‘AND’’ connection is used to describe the antecedent
variables of the fuzzy reasoning include the response time proposition in this study. The CF value is designed to
(RT) and the damage level decision-making (DLDM). In reflect the way the experts think (Negnevitsky, 2002; Qu
reference to Negnevitsky (2002), linguistic value, notation and Shirai, 2003). In this article, assuming that the belief
and normalized numerical range of the linguistic variable strength of a fuzzy rule assigned by human expert is
are included in the fuzzy rule and illustrated in Table 1. The CF ¼ 1, which represents that the fuzzy rule is completely
values of linguistic variables are fuzzified to obtain the MD believable, and denoted by
by MF. Values of all the linguistic variables are normalized IF A1 AND A2 AND . . . AND Aa THEN C b ðCF ¼ 1Þ, (2)
in the range of [0, 1], so the MD is bounded in the range of
[0, 1]. The MDs of the concerned variables are all designed
of the trapezoid shape of fuzzy sets, as listed in Table 1. 4. Structure and modeling of FRVPNs

3.2. Fuzzy rules representation In Figs. 2 and 3, an example of a two-level FRVPNs


including the rule-checking process as well as the verifica-
The implication relationship of the antecedent and tion and modification module is presented. The proposed
consequent proposition is used to establish the elements FRVPNs is constructed based on the FRDT with the
for fuzzy rules. Assuming FR is the set of fuzzy rules merits of hierarchical design and PNs technique. It uses
represented by FR ¼ {FR1, FR2, FR3, y, FRs1, FRs}, the rule-checking transitions and places to confirm the
FRkAFR. The kth fuzzy rule is denoted by correctness of the firing rule so as to protect the accuracy of
inference result of the winning rule. The verification and
FRk : IF Aa THEN C b ðCFk ¼ mk Þ, (1) modification module is another important function. It uses
where a, b, k and s are positive integer, and 1pkps; Aa is the input/output transition function of the place and the
the antecedent proposition defined in a fuzzy region of inhibit condition to detect the inconsistent fuzzy rules
(conflict, redundancy, incompleteness, and circularity
Table 1 rules) so that inference accuracy is improved.
Linguistic values and ranges of fuzzy sets for FRVPNs
4.1. Hierarchical structure
Linguistic values Notation Numerical range
(normalized)
FRVPNs is designed based on the FRDT with the
TCPUU, PCPU, TMU hierarchy structure. On the basis of the hierarchical design,
Low L 0.0–0.4 the inference accuracy of FRDT is expected to improve
Medium M 0.2–0.8 (Joo and Lee, 2005; Delgado et al., 2002; Wang, 1999).
High H 0.6–1.0
Fig. 4 shows that there are three inputs considered in the
RT Level_I, including the TCPUU, the PCPUU, and TMU, by
Fast F 0.0–0.4 which the RT can be inferred. In Fig. 5, there are three
Medium M 0.2–0.8
inputs considered in the Level_II, including the RT (the
Slow S 0.6–1.0
output of the Level_I), the HT, and the EHL, by which the
HT DLDM can be inferred in terms of three degrees (Normal,
Short S 0.0–0.4
Slightly High, High) to the AP and the OS.
Medium M 0.2–0.8
Long L 0.6–1.0
4.2. Dynamic decision-making procedure
EHL
Low L 0.0–0.4 The dynamic decision-making procedures of Level_I and
Normal N 0.2–0.8
High H 0.6–1.0
Level_II are shown in Figs. 4 and 5, respectively, and
described as follows:
DLDM
Normal N 0.0–0.4 (1) Level_I:
Slightly high SH 0.2–0.8
Step 1. enter the required new object variables for the
High H 0.6–1.0
fuzzy rule in the Level_I.
ARTICLE IN PRESS
Y. Ting et al. / Engineering Applications of Artificial Intelligence 21 (2008) 157–170 161

I
P1A Pr oIO I
P2A Pr oIO I
P3A Pr oIO
I
R 1A Pr oD I
R A2 Pr oD I
R 3A Pr oD

I
I
P11A Pr o I
P12A Pr o I
P13A Pr o I
P21A Pr o I
P22A Pr o
I
P23A Pr o P31A Pr o I
P32A Pr o I
P33A Pr o
I
I
R 11A Pr oMF I
R 12A Pr oMF I
R 13A Pr oMF I
R A21Pr oMF I
R A22Pr oMF I
R A23Pr oMF I
R 31A Pr oMF I
R 32A Pr oMF R 33A Pr oMF
I I I I I I I I I
WF11 WF12 =1 WF13 =2 WF21 =2 WF22 =1 WF23 WF31 WF32 =2 WF33 =1
I A Pr oMD I A Pr oMD I
I
P A Pr oMD
11
I
P A Pr oMD
12
I
P A Pr oMD
13
I
PA Pr oMD
21
I
P A Pr oMD
22
I
P A Pr oMD
23
P 31 P 32 P33A Pr oMD

I

R 1RFS I
R RFS
2
I
R 3RFS

I
P1RFS I
P2RFS I
P3RFS …

R MAX ()
WRFS
Pk

Rule-che cking
I
R 1FSC I
R FSC
I
R 3FSC
2

I RFS
I
I
P[RFS P[ 2]
I
P[RFS
P1WFR 1] I
P2WFR I
P3WFR 3]

I
P C Pr oO

I I
R FDR I
R 3FDR

R 1FDR 2

Verification and
Modification Algorithm

I
P DMI I
I
P 1
C Pr oM
P3C Pr oM
I
P2C Pr oM
I DMI I
R 1 R DMI
2
I
R 3DMI
I
P1C Pr o I
P2C Pr o P3C Pr o

I
R 1C Pr oMF I
R C2 Pr oMF I
R 3C Pr oMF
I
P1C Pr oMD I
P2C Pr oMD I
P3C Pr oMD

Fig. 2. The FRVPNs model of the Level_I.

Step 2. calculate the MD of the proposition of new Step 5. draw a conclusion of the output sub-object in
object variables via the MF. the Level_I from the selected winning rule.
Step 3. calculate the firing strength by the composition (2) Level_II:
operator MIN. Step 6. import the three object variables, the RT, HT,
Step 4. calculate the maximum firing strength by the and EHL, to the Level_II.
composition operator MAX, and find the winning rule Step 7. calculate the MD of the new object variables by
among the activated rules in the Level_I. the MF.
ARTICLE IN PRESS
162 Y. Ting et al. / Engineering Applications of Artificial Intelligence 21 (2008) 157–170

II
II
P1A Pr oIO P2A Pr oIO
II
R 1A Pr oD II
R A2 Pr oD

II
P11A Pr o II
P12A Pr o II
P13A Pr o II
P21A Pr o II
P22A Pr o II
P23A Pr o
II
R 11A Pr2oMF R 12A Pr2oMF R 13A Pr2oMF R A21Pr oMF II
R A22Pr oMF II
R A23Pr oMF
II
I II WF 21 =
P1C Pr oMD I
P2C Pr oMD I
P3C Pr oMD WF 11 II
WF 12 =2 II
WF 13 =1
2
II
WF 22 II
WF 23 =1
II A Pr oMD II A Pr oMD
II
P A Pr oMD
11
II
P A Pr oMD
12
II
P A Pr oMD
13
P21 P 22
II
P23A Pr oMD

II
II
R 1RFS
II
R RFS
2
R 3RFS …

II
P1RFS II
P2RFS II
P3RFS …

R MAX ()
WRFS
P w

Rule-checking
II
II
R 1FSC R FSC
2
II
R 3FSC …
II RFS II II RFS
II P II
P2WFR P[RFS II
P3WFR P[ 3]
P1WFR [ 1] 2]

II
P C Pr oO

II
R 1FDR II
II
R 3FDR …
R FDR
2

Verification and
Modification Algorithm

II
P DMI
II
II
P1C Pr oM P3C Pr oM
II C Pr oM
P 2
II
II
R 1DMI II
R DMI
2
R 3DMI

II
II
P1C Pr o II
P2C Pr o P3C Pr o

R 1FDM R FDM
2 R 3FDM

P FDM
Fig. 3. The FRVPNs model of the Level_II.

Step 8. calculate the firing strength of each rule in the 4.3. Petri nets for FRVPNs modeling
Level_II by the composition operator MIN.
Step 9. search for the maximum firing strength of the The FRVPNs takes advantage of the PNs to model the
winning rule among the activated rules in the Level_II reasoning structure and represents the dynamic behavior
by the composition operator MAX. of the fuzzy rule. An ordinary PNs structure is a 4-tuple
Step 10. draw a conclusion of the output in the Level_II (P, R, I, O,) defined in Cardoso and Camargo (1999),
from the selected winning rule. Peterson (1981), Girault and Valk (2003), and Zhou
ARTICLE IN PRESS
Y. Ting et al. / Engineering Applications of Artificial Intelligence 21 (2008) 157–170 163

Level
Level_I
_I
MAX Output
Input Membership degree composition sub-
MIN composition operation
objects Calculation operation object

Note:
No
RFS1 = μ Low (TCPUU)
RFS 1
∧ μ Low (PCPUU) ∧ μ Low (TMU)
μLow (TCPUU)
Low (L) RFS2 = μLow (TCPUU)
RFS 2 ∧ μLow (PCPUU) ∧ μMedium (TMU)
TCPU
TCPUU μ (TCPUU)
Medium

Medium (M) RFS3 = μ Low (TCPUU)


RFS 3 ∧ μ Low (PCPUU) ∧ μ High (TMU)
μHigh (TCPUU)

...
..
High (H) Description of Mark A:
RFS 4
The output sub object
of the FRDT Level_I
beco mes the input
μ Low (PCPUU) object of the FR DT
RFS 5
Level_II.
Low (L)

PCPUU
PCP UU μ
Medium (PCPUU) RT A
RFS 6
Medium (M) WFR

μHigh (PCPUU) WRF = MAX (RFS1 , RFS2


RFS 7
High (H) , RFS3 ,...)

RFS 8
μLow (TMU)
Low (L)
RFS 9
TMU μ Medium (TMU)
Medium (M)
...

μHigh (TMU)
RFS 26
High (H)

RFS 27

Step 1 Step 2 Step 3 Step 4 Step 5

Fig. 4. Reasoning the sub-object in the Level_I.

and Venkatesh (1998), but the ordinary PNs cannot Definition 4.1. WF: ASD(P  R)[(R  P)-+N is a weight
satisfy the fuzzy reasoning design for the EDDM. function associated with the weight of the arc or the number
Therefore, the proposed FRVPNs applied 11-tuples, are of the direct arcs, where AS is a set of the arcs, and the current
defined by relation holds only between the places and the transitions.
FRVPNs ¼ (P, R, I, O, WF, M, H, Pro, ProMF, RFS,
WRFS). Definition 4.2. H: P  R-N, H(Pk) ¼ {PkAP: H(Pk, Rm)4
Related definitions are described as below. 0}, is an inhibitor function. An inhibitor arc from a place Pk
ARTICLE IN PRESS
164 Y. Ting et al. / Engineering Applications of Artificial Intelligence 21 (2008) 157–170

FRDT
Level_I
Level _II
Out put
Output
Member ship degr ee MIN composi tion MAX com posi tion Output
Input obj ect s Calculation oper ation objec t
operat ion

Note:
RFS1 = μFast (RT)
RFS 1
μShort (HT) ∧ μShort (HT) ∧ μLow (EHL)

HT Short (S)
RFS2 = μFast (RT)
RFS 2
μ Fast (RT) ∧ μSlow (HT) ∧ μ Normal (EHL)

RT Fast (F)
RFS3 = μFast (RT)
RFS 3
μ Low (EHL) ∧ μMedium (HT) ∧ μ Low (EHL)

.....
EHL Low (L)
RFS 4

μ Medium (HT)
RFS 5
Medium
HT
(M)

μ Medium (RT) DLDM


RFS 6
Medium
RT WRF
(M)

μ Normal (EHL) WRF = MAX (RFS1 , RFS 2 ,


RFS 7
Normal RFS3 ,...)
EHL
(N)

RFS 8
μ Long (HT)
HT Long (L)
RFS 9
μSlow (RT)
RT Slow (S)
...

μ High (EHL)
RFS 26
EHL High (H)

RFS 27

Step 6 Step 7 Step 8 Step 9 Step 10

Fig. 5. Reasoning the object in the Level_II.

to a transition Rm has a ‘‘circle’’ symbol rather than an Definition 4.3. Pro ¼ {Pro11, Pro12, y, Proij} is a finite set
arrowhead at the transition. A transition represented by an of propositions. A proposition ProijAPro is mapped on a
‘‘arrow’’ symbol is enabled when tokens are in all of its place PnAP in the FRVPNs.
normal inputs and no tokens are in all of its inhibitor
inputs. The transition is fired by removing tokens from all Definition 4.4. ProMF is a MF of a proposition. The input
of its normal inputs. object of the proposition can be mapped to a corresponding
ARTICLE IN PRESS
Y. Ting et al. / Engineering Applications of Artificial Intelligence 21 (2008) 157–170 165

MD through the ProMF. The MD of a proposition of a 5. FRVPNs reasoning strategy


fuzzy rule FRk is represented by m(x)A[0, 1]-FR, where
xAX is the input object of the proposition. As illustrated in Fig. 6, the examples of FR1 and FR2
are used to describe the FRVPNs modeling and the
Definition 4.5. a fuzzy composition operation represents a dynamic reasoning behavior. Fig. 6(a) shows the contents
MD of the required proposition calculated by the of the fuzzy rule in the Level_I and Level_II. Fig. 6(b)
composition operator MAX/MIN. The fuzzy composition shows part of the FRVPNs model by means of the PNs
is defined by MAX/MIN-R. technique. Fig. 6(c) shows the dynamic reasoning behavior
of the rules FR1 and FR2. The properties of the
Definition 4.6. RFSk: fuzzy composition-R, represents proposition set of places and the firing transitions are
the firing strength of a fuzzy rule. The larger firing strength described as follows:
indicates the larger belief strength of a rule. Based on the
fuzzy operators AND/OR shown in the antecedent (1) I PAPro
ij and II PAPro
ij : represents the ‘‘antecedent propo-
proposition of a rule FRk, the RFSk is calculated by the sition’’ place of a rule FRk in the Level_I and
composition operator MIN or MAX. Level_II, respectively, where i represents the ith input
object of a fuzzy rule FRk, and j represents the jth
Definition 4.7. WRFS: a winning-rule firing strength, is antecedent proposition of a fuzzy rule FRk.
calculated by the MAX composition operator. The FRDT (2) I RAProMF
ij and II RAProMF
ij : represents the ‘‘MF’’ transi-
structure consists of L-levels, e.g., Level_I and Level_II in tion of an antecedent proposition of a fuzzy rule FRk
this article, and each level includes s-rules, e.g., FR1, FR2, in the Level_I and Level_II, respectively.
y, FRk, y, FRs. Let RFSk be the firing strength of each (3) I PAProMD
ij and II PAProMD
ij : represents the ‘‘MD’’ place
kth rule in the Level_I. Let WRFSk: MAX (RFS1, RFS2, y, of an antecedent proposition of a fuzzy rule FRk in the
RFSk, y, RFSs)-R be the firing strength of the winning Level_I and Level_II, respectively.
rule FRk-FR. The WRFSk is used to compose the firing (4) I RRFS
k and II RRFSk : represents the ‘‘firing strength’’
strengths of the s-rules in the Level_I or Level_II and select transition of a fuzzy rule FRk in the Level_I and
the winning rule FRk with the highest confidence. Level_II, respectively.

Fig. 6. (a) Fuzzy rules in the Level_I and Level_I II. (b) FRVPNs Model. (c) Marking dynamic reasoning behavior.
ARTICLE IN PRESS
166 Y. Ting et al. / Engineering Applications of Artificial Intelligence 21 (2008) 157–170

(5) I PRFS
k and II PRFS
k : represents the ‘‘firing strength’’ Step 3. Perform MIN composition operation and
place of a fuzzy rule FRk in the Level_I and Level_II calculate the firing strength for the activated rules by the
respectively. The place I PRFS
k with a token represents following places and transitions:
the firing strength of a fuzzy rule FRk. 
(1) RRFS ¼ I RRFS ; I RRFS ; . . . ; I RRFS ; . . . ; I RRFS is a set of
(6) RMAX( ): represents the ‘‘firing strength’’ transition of 1 2 k s
‘‘firing strength of rule’’ transitions for each rule in the
a winning rule in the Level_I or Level_II. It is
calculated by the composition operator MAX. Level_I. I RRFS
k uses the fuzzy operator ‘‘AND’’ or ‘‘OR’’
(7) PWRFS : represents the ‘‘winning fuzzy rule’’ place. It to perform MIN or MAX composition operation.
k 
models the firing strength FRSk of the winning rule (2) PRFS ¼ I PRFS
1 ; I PRFS
2 ; . . . ; I PRFS
k ; . . . ; I PRFS
s is a set of
FRk in the Level_I or Level_II. ‘‘firing strength of rule’’ places. While the token exists
(8) I RCProMF
k and II RCProMF
k : represents the ‘‘MF of a in the place I PRFS
k , the transition I RRFS k is disabled by
consequent proposition’’ transition in the Level_I and the inhibitor arc to ensure that the new result of firing
Level_II, respectively. The MD of a consequent strength is transferred into the place I PRFS after the
k
proposition of the winning rule FRk is calculated by previous firing strength in the place have been
the centroid of the aggregate output MF. transferred out .
(9) I PCProMD
k : represents the ‘‘MD of a consequent
proposition’’ place. The place I PCProMD
k with a token Step 4. Perform MAX composition operation on the
represents the MD of a consequent proposition of a firing strengths of the activated rules, and select the
rule FRk in the Level_I. winning rule from the available rules in the Level_I by
(10) PFDM: is a ‘‘final decision-making’’ place. The place the following places and transitions:
PFDM with a token represents the final result in the
Level_II. 
(1) RMAXð Þ ¼ MAX I PRFS
1 ; I PRFS
2 ; . . . ; I PRFS
k ; . . . ; I PRFS
s is
a transition of MAX composition operation. The
5.1. FRVPNs reasoning algorithm for the Level_I transition RMAX( ) is used to calculate the maximum
firing strength among the activated rules.
The FRVPNs model of the Level_I is illustrated in (2) PWRFS
k is a ‘‘WRFSk’’ place, and it represents the firing
Fig. 2, and described as below. strength RFSk of the winning rule FRk.
Step 1. Import the parameters of the proposition of the Step 5. Determine the winning rule for rule-checking
rules by the following places and transitions: purpose by the following places and transitions:
 
(1) I PAProIO ¼ I PAProIO ; I PAProIO ; . . . ; I PAProIO ; . . . ; I PAProIO (1) RFSC ¼ I RFSC ; I RFSC ; . . . ; I RFSC I FSC
k ; . . . ; Rs is a set of
1 2 k s
1 2
is a set of ‘‘input object’’ places. Each place has an ‘‘firing strength comparison’’ transitions. The transition
input variable.  is used to compare the firing strength of an activated
(2) I RAProD ¼ I RAProD
1 ; I RAProD
2 ; . . . ; I RAProD
k ; . . . ; I RAProD
s fuzzy rule with that of the WRFSk of a FRk. If the
is a set of ‘‘input object distribution’’ transitions. It firing strengths are not same, the reasoning process will
represents the transition distribution of linguistic be terminated.
variables for each antecedent proposition of a rule. 
(2) PWFR ¼ I PWFR1 ; I PWFR
2 ; . . . ; I PWFR
k ; . . . ; I PWFR
s is a set
Step 2. Establish the antecedent propositions for each of ‘‘winning firing rule’’ places. While the place I PWFR k
rule and calculate the MD for each antecedent proposition contains a token, the place I PWFR k represents FR k is the
by the following places and transitions: winning firing rule and is selected to fire.
  Step 6. Determine the final decision rule for rule-
(1) PAPro ¼ I PAPro
11 ; I APro I APro
P12 ; P 13 ; . . . ; I APro
P ij is a set of checking purpose by the following places and transitions:
‘‘antecedentproposition’’ places for each rule.
(2) RAProMF ¼ I RAProMF ; I RAProMF ; . . . ; I RAProMF is a set (1) I PRFS
½k is a set of ‘‘firing strength of rule’’ places, except
11 12 ij
of ‘‘MF of an antecedent proposition’’ transitions. the I PRFS
k place.
I
RAProMF represents the MF of the jth fuzzy set (2) I PCProO is a ‘‘linguistic variable of output object of a
ij
consequent proposition’’ place.
corresponding to the ith input object of the antecedent 
proposition, thus the MD of the ith input object of the (3) RFDR ¼ I RFDR 1 ; I RFDR
2 ; . . . ; I RFDR
k ; . . . ; I RFDR
s is a set
of ‘‘final decision of the winning rule’’ transitions. The
antecedent proposition is computed by the jth MF.
transition is used to delete the rules with the firing
(3) IWFij represents the weighted function corresponding
strength less than the rule with the highest firing strength.
to each transition arc of the antecedent propositions.
  (4) I PDMI is a ‘‘decision-making identification’’ place. It
(4) PAProMD ¼ I PAProMD
11 ; I PAProMD
12 ; . . . ; I PAProMD
ij is a set of represents the decision result of the selected winning
‘‘MD of an antecedent proposition’’ places for each rule. rule FRk with the highest firing strength.
ARTICLE IN PRESS
Y. Ting et al. / Engineering Applications of Artificial Intelligence 21 (2008) 157–170 167

Step 7. Construct the consequent propositions of Step 3. Determine the final decision-making by the
the final decision rule by the following places and following places and transitions:
transitions:

 (1) RFDM ¼ RFRDM
1 ; RFRDM
2 ; . . . ; RFRDM
k ; . . . ; RFRDM
s is a
(1) PCProM ¼ I PCProM
1 ; I PCProM
2 ; . . . ; I PCProM
k ; . . . ; I PCProM
s set of ‘‘final decision-making’’ transitions, which
is a set of ‘‘marking of the consequent proposition’’ determines RFDM
k .
places.  (2) PFDM is a ‘‘final decision-making’’ place of entire
(2) RDMI ¼ I RDMI1 ; I RDMI
2 ; . . . ; I RDMI
k ; . . . ; I RDMI
s is a set FRVPNs.
of ‘‘decision-making identification’’ transitions. It
represents the transition distribution of linguistic
variables for a consequent proposition of a fuzzy rule. 6. Simulations

(3) PCPro ¼ I PCPro
1 ; I PCPro
2 ; . . . ; I PCPro
k ; . . . ; I PCPro
s is a set of
‘‘consequent proposition’’ places. It describes the As shown in Figs. 2 and 3, the fuzzy rule base of the
conclusions of the fuzzy rules. reasoning processor in the EDDM is constructed of two
levels in this study. The measurement values [TCPUU,
Step 8. Calculate the MD for each consequent proposi- PCPU, TMU, HT, EML] ¼ [73%, 33%, 65%, 10.5 s,
tion by the following places and transitions: 72%], are assumed to be the inputs variables to the
antecedent propositions of FRVPNs model, which is
(1) RCProMF ¼ I RCProMF1 ; I RCProMF
2 ; . . . I RCProMF
k ;. . . the example_1 with the inference results in the Level_I
I CProMF
Rs Þ is a set of ‘‘MF of a consequent proposi- and Level_II listed in Table 2. With this example, the
tion’’ transitions for each consequent proposition. reasoning process is described in detail through the
following steps.
When the transition I RCProMF
k is fired, the MD is com-
Step 1. Normalize the measurement inputs (e.g.
puted by the kth MF.
[TCPUU, PCPU, TMU, HT, EHL] ¼ [73%, 33%, 65%,
(2) PCProMD ¼ I PCProMD
1 ; I PCProMD
2 ; . . . ; I PCProMD
k ;...;
I CProMD 10.5sec, 72%]) by the fuzzy sets in Section 3.1 and
Ps Þ is a set of ‘‘MD of a consequent proposi- distribute the normalized inputs to the FRVPNs model.
tion’’ places. While the place I PCProMD k contains a The measurement values are normalized to be
token, it receives an MD from the transition I RCProMF k .  
TCPUU PCPUU TMU HT EHL
; ; ; ;
100% 100% 100% 30 100%
5.2. FRVPNs reasoning algorithm for Level_II
¼ ½ð73=100Þ; ð33=100Þ; ð65=100Þ; ð10:5=30Þ; ð72=100Þ
The FRVPNs model of the Level_II is illustrated in ¼ ½0:73; 0:33; 0:65; 0:35; 0:72,
Fig. 4, and described as below. where the maximum HT is defined to be 30 s.
Step 1. Establish the antecedent propositions of the rules Step 2. Assume the Level_I of the reasoning processor
and calculate the MDs for each antecedent proposition as has the following rules:
follows.
FR1: IF ‘‘TCPUU is Medium’’ AND ‘‘PCPUU is Low’’
(1) The output object from Level_I is used as the input AND ‘‘TMU is Low’’ THEN ‘‘RT is Fast’’.
place to Level_II. FR2: IF ‘‘TCPUU is High’’ AND ‘‘PCPUU is Low’’
(2) The new input objects of the antecedent proposition of AND ‘‘TMU is Medium’’ THEN ‘‘RT is Medium’’.
the rules in the Level_II can be modeled by repetition of FR3: IF ‘‘TCPUU is High’’ AND ‘‘PCPUU is Medium’’
the Step1 in Section 5.1. AND ‘‘TMU is High’’ THEN ‘‘RT is Slow’’.
(3) The MDs of the new input objects are calculated by
repetition of the Step2 in Section 5.1.
Then, TCPUU ¼ 0.73, PCPUU ¼ 0.33, and TMU ¼ 0.65
are imported to the places I PAProIO
1 , I PAProIO
2 , and I PAProIO
3 in
Step 2. Describe the decision behaviors of the rules as the Level_I, respectively.
follows. Step 3. The transitions I RAProD , I RAProD , andI RAProD are
1 2 3
concurrently fired, then the tokens are distributed to the
(1) Use Step3 in Section 5.1 to calculate the firing strength place I PAPro I APro I APro I APro I APro I APro I APro
11 , P12 , P13 , P21 , P22 , P23 , P31 ,
of each activated rule. I APro
P32 , and I PAPro
33 , respectively.
(2) Use Step4 in Section 5.1 to find out the maximum firing
Step 4. The transition set I RAProMF is fired, then the
strength from the available activated rules.
(3) Use Step5 and Step6 in Section 5.1 to decide the tokens are transferred to the place I PAProMD11 , I PAProMD
12 ,
I AProMD I AProMD I AProMD I AProMD I AProMD
winning rule among the activated rules. P13 , P21 , P22 , P23 , P31 ,
I AProMD
(4) Use Step7 in Section 5.1 to determine the consequent P32 , and I PAProMD
33 , respectively. The MD of each
propositions of the Level_II. place is calculated by its trapezoidal MF described in
ARTICLE IN PRESS
168 Y. Ting et al. / Engineering Applications of Artificial Intelligence 21 (2008) 157–170

Table 2
Simulation examples of the FRVPNs

Level_I Level_II Matlab RT Matlab DLDM

Rule No. TCPUU PCPUU TMU RT MIN HT EHL DLDM MIN

Example_1 1 M L L F 0 S N N 0.25
2 H L M M 0.35 M N SH 0.4
3 H M H S 0.25 L H H 0
Measurement values 0.73 0.33 0.65 MAX 0.35 0.35 0.72 MAX 0.4 M 0.592 SH 0.416
Example_2 1 M L L F 0.2 S L N 0
2 H L M M 0.35 M N SH 0.3
3 H M H S 0 L L N 0
Measurement values 0.69 0.33 0.36 MAX 0.35 0.26 0.71 MAX 0.3 M 0.432 SH 0.582

Example_3 1 L M L F 0.35 S N N 0
2 H L L F 0 M L N 0.2
3 M L M M 0.4 M H SH 0.25
Measurement values 0.3 0.27 0.28 MAX 0.4 0.36 0.65 MAX 0.25 M 0.391 SH 0.465
Example_4 1 M L H S 0 S N N 0
2 M M L M 0 M N SH 0
3 H H M S 0.6 L H H 0.7
Measurement values 0.72 0.75 0.5 MAX 0.6 0.74 0.86 MAX 0.7 S 0.83 H 0.835
Example_5 1 M L L M 0.25 S N N 0
2 M L M M 0.25 L L N 0
3 H M M S 0.6 L H H 0.75
Measurement values 0.75 0.34 0.32 MAX 0.6 0.78 0.75 MAX 0.75 S 0.71 H 0.838

Note: RT and DLDM are the crisp values calculated by the Mamdani fuzzy method of MATLAB tools.
The italic words are the winning rules.

Section 3.1. In this example, the calculated MDs are: Step 6. According to the result of MAX composition
operation in the above Step5, the final winning rule is
I AProMD
P11 ¼ mLow ðTCPUUÞ ¼ 0,
I AProMD
P12 ¼ mMedium ðTCPUUÞ ¼ 0:35, FR2: IF ‘‘TCPUU is High’’ AND ‘‘PCPUU is Low’’
I AProMD
AND ‘‘TMU is Medium’’ THEN ‘‘RT is Medium’’.
P13 ¼ mHigh ðTCPUUÞ ¼ 0:65,
I AProMD In case that the antecedents of rules are evaluated to be
P21 ¼ mLow ðPCPUUÞ ¼ 0:35,
I AProMD the same level, the max–min inference cannot determine the
P22 ¼ mMedium ðPCPUUÞ ¼ 0:65,
winning rule. Hence, other defuzzification methods are
I AProMD
P23 ¼ mHigh ðPCPUUÞ ¼ 0, considered to use. The center of gravity method (COG) is
I AProMD
P31 ¼ mLow ðTMUÞ ¼ 0, the most popular method (Van Broekhoven and De Baets,
I AProMD 2004; Negnevitsky, 2002; Yager, 1992), which is used to
P32 ¼ mMedium ðTMUÞ ¼ 0:75, solve this problem.
I AProMD
P33 ¼ mHigh ðTMUÞ ¼ 0:25. Step 7. Assume the Level_II of the reasoning processor
has the following rules:
Step 5. The transition set RRFS and the transition RMAX( )
are fired in order, and then the firing strength of each FR4: IF ‘‘HT is Medium’’ AND ‘‘RT is Medium’’ AND
activated rule and the winning rule is calculated by ‘‘EHL is Normal’’ THEN ‘‘DLDM is Slightly High’’.
the MIN and MAX composition operator, respectively. FR5: IF ‘‘HT is Short’’ AND ‘‘RT is Medium’’ AND
It yields ‘‘EHL is Normal’’ THEN ‘‘DLDM is Normal’’.
FR6: IF ‘‘HT is Long’’ AND ‘‘RT is Medium’’ AND
FR1 : MINð0:35; 0:35; 0Þ ¼ 0, ‘‘EHL is High’’ THEN ‘‘DLDM is High’’.
FR2 : MINð0:65; 0:35; 0:75Þ ¼ 0:35,
Then, HT ¼ 0.35 and EHL ¼ 0.72 are imported to the
FR3 : MINð0:65; 0:65; 0:25Þ ¼ 0:25,
places II PAProIO
1 and II PAProIO
2 , respectively. The defuzzifica-
MAX ðFR1 ; FR2 ; FR3 Þ tion of RT is calculated as RT ¼ 0.598 by the centroid of
¼ MAX ð0; 0:35; 0:25Þ ¼ 0:35. the aggregate output MF in the Level_I. This RT value is
ARTICLE IN PRESS
Y. Ting et al. / Engineering Applications of Artificial Intelligence 21 (2008) 157–170 169

then imported to the antecedent propositions in the


Level_II.
Step 8. The transition set II RAProD is fired to distribute
the tokens to the place set II PAPro .
Step 9. As the transition set II RAProMF is fired, the tokens
are then transferred to the place set II PAProMD . The MDs of
antecedent propositions in the Level_II are calculated by
the trapezoidal MF and listed as below.
II AProMD
P11 ¼ mShort ðHTÞ ¼ 0:25
II AProMD
P12 ¼ mMedium ðHTÞ ¼ 0:75
II AProMD
P13 ¼ mLong ðHTÞ ¼ 0
II AProMD
P21 ¼ mLow ðEHLÞ ¼ 0
II AProMD
P22 ¼ mMedium ðEHLÞ ¼ 0:4
II AProMD
P23 ¼ mHigh ðEHLÞ ¼ 0:6
I CProMD
Fig. 8. Final decision of the Level_II.
P23 ¼ mMedium ðRTÞ ¼ 1
Step 10. By using the MIN and MAX composition
operators, the firing strength of each activated rule and the ‘‘DLDM is Slightly High’’. The Mamdani fuzzy method of
winning rule is calculated respectively as the MATLAB tools is also used to compare the inference
results under the same conditions (same inputs, same
FR4 : MINð0:75; 1; 0:4Þ ¼ 0:4 linguistic values, same ranges). As shown in Fig. 7, rules
FR5 : MINð0:25; 1; 0:4Þ ¼ 0:25 FR1, FR2, and FR3 are aggregated and defuzzied to have a
FR6 : MINð0; 1; 0:6Þ ¼ 0 crisp value of RT ¼ 0.592 by calculating the centroid,
which indicates the rule FR2 is the winner. In reference to
MAX ðFR4 ; FR5 ; FR6 Þ ¼ MAX ð0:4; 0:25; 0Þ ¼ 0:4
the consequent proposition of FR2, ‘‘RT is Medium’’ is
Step 11. According to the result of MAX composition thus inferred for Level_I. In Fig. 8, in Level_II, the crisp
operation in the above Step10, the final winning rule is: value of DLDM ¼ 0.416 is calculated, which indicates the
rule FR4 is the winner. In reference to the consequent
FR4: IF ‘‘HT is Medium’’ AND ‘‘RT is Medium’’ AND proposition of FR4, ‘‘DLDM is Slightly High’’ is thus
‘‘EHL is Normal’’ THEN ‘‘DLDM is Slightly High’’. inferred for the monitored AP and the OS. In comparison
with the inferred results, both methods have the same
Following the steps of the FRVPNs Reasoning Algo- reasoning outcomes. This example and other four examples
rithm in Section 5, the final winning rule in Level_I is FR2, by using the two methods are included and listed in
which indicates that the ‘‘RT is Medium’’, and in the Table 2, which shows that the inference outcomes of all the
Level_II the final winning rule is FR4, which indicates the examples are the same.

7. Conclusions

In this article, the fuzzy set theory and the fuzzy


production rule method are used to establish the fuzzy
rules for the failure events of the AP. The proposed
FRVPNs is designed based on the FRDT in association
with the PN technique, and constructed with the rule-
checking process as well as the verification and modifica-
tion module. The computational complexity (Chen et al.,
1990; Gao et al., 2003; Kungas, 2005) is defined by O(ptl)
and determined by the number of p, t and l, where p
and t are the numbers of places and transitions, respec-
tively, and l is the maximum number of transitions in the
longest place-transition path. The FRVPNs reasoning
algorithm apparently has higher computational complexity
than a general FRDT, because it uses more places and
transitions in the rule-checking process and in the
verification and modification module. Even though the
Fig. 7. Final decision of the Level_I. complexity, and higher accuracy could be obtained,
ARTICLE IN PRESS
170 Y. Ting et al. / Engineering Applications of Artificial Intelligence 21 (2008) 157–170

problems of redundancy, conflict, circularity, and incom- Kungas, P., 2005. Petri net reachability checking is polynomial with
pleteness could be reduced. Comparing the FRVPNs with optimal abstraction hierarchies. In: Symposium on Abstraction,
Reformulation and Approximation 3607, 149–164.
the fuzzy logic toolbox of MATLAB by the simulation
Looney, C.G., 1998. Fuzzy Petri nets for rule-based decision-
results of several sets of arbitrarily selected examples, both making. IEEE Transactions on System, Man, and Cybernetics 18
methods conclude the same outcomes of fuzzy reasoning. It (1), 178–183.
verifies that the damage level of the failure event for the Lu, W.B., Ting, Y., Chen, C.H., Chen, Y.C., 2003. The Application of
monitored AP and the OS can be successfully reasoned by Knowledge Representation and Knowledge Acquisition to Fault
the proposed FRVPNs. To sum up, the proposed FRVPNs Diagnosis. In: Proceedings of the 31st International Conference on
Computers & Industrial Engineering, pp. 69–74.
model is capable of reasoning and determining the failure Negnevitsky, M., 2002. Artificial Intelligent–A Guide to Intelligent
event efficiently and accurately, and it is thus suitable to be Systems. Addison-Wesley, New York.
included in the diagnosis mechanism of the EDDM. Peterson, J.L., 1981. Petri Net Theory and the Modeling of Systems.
Regarding the verification and modification algorithm, it Prentice-Hall, Englewood Cliffs, NJ.
is addressed in another article. Qu, W., Shirai, K., 2003. Belief learning in certainty factor model and
its application to text categorization. In: Proceedings of the 2003 Joint
Conference of the Fourth International Conference on Information,
Acknowledgment Communications and Signal Processing and Fourth Pacific-Rim
Conference on Multimedia 12, 1192–1196.
This research is supported by NSC89-TPC-7-033-009 & Richter, J., 1999. Programming Applications for Microsoft Windows, 4th
ed. Microsoft Press, Washington.
NSC88-2212-E-033-018.
Scarpelli, H., Gomide, F., Yager, R.R., 1996. A reasoning algorithm for
high-level fuzzy Petri nets. IEEE Transactions on Fuzzy Systems 4 (3),
References 282–294.
Ting, Y., Lu, W.B., Chen, C.H., Wang, G.K., 2004. Design the reasoning
Cardoso, J., Camargo, H., 1999. Fuzziness in Petri Nets. Physica-Verlag, process of failure detection and diagnosis mechanism (EDDM) by
New York, pp. 4–24. fuzzy rule-based method. The 2004 IEEE International Conference on
Chen, S.M., 2002. Weighted fuzzy reasoning using weighted fuzzy Petri Systems, Man, and Cybernetics. pp. 2344–2349.
nets. IEEE Transactions on Knowledge and Data Engineering 14 (2), Ting, Y., Shan, F.M., Lu, W.B., Chen, C.H., 2002. Implementation
386–397. and evaluation of failsafe computer-controlled systems. Interna-
Chen, S.M., Ke, J.S., Chang, J.F., 1990. Knowledge representation using tional Journal of Computers and Industrial Engineering 42 (2–4),
fuzzy petri nets. IEEE Transaction on Knowledge and Data 401–415.
Engineering 2 (3), 311–319. Tsang, E.C.C., Yeung, D.S., Lee, J.W.T., 1999. Learning capability in
Delgado, M.R., Von Zuben, F., Gomide, F., 2002. Multi-objective fuzzy Petri nets. The 1999 IEEE International Conference on Systems,
decision making: towards improvement of accuracy, interpretability Man, and Cybernetics 3, 355–360.
and design autonomy in hierarchical genetic fuzzy systems. In: Van Broekhoven, E., De Baets, B., 2004. A comparison of three methods
Proceedings of the 2002 IEEE International Conference on Fuzzy for computing the center of gravity defuzzification. 2004 IEEE
Systems 2, 1222–1227. International Conference on Fuzzy Systems, pp. 25–29.
Fay, A., 2001. A fuzzy Petri net approach to decision-making in case of Wang, L.X., 1999. Analysis and design of hierarchical fuzzy systems.
railway track closures. In: IFSA World Congress and 20th NAFIPS IEEE Transactions on Fuzzy Systems 7 (5), 617–624.
International Conference 5, 2858–2863. Yager, R. R., 1992. Fuzzy sets and approximate reasoning in decision
Gao, M., Zhou, M.C., Huang, X., Wu, Z., 2003. Fuzzy reasoning Petri and control. IEEE International Conference on Fuzzy Systems,
nets. IEEE Transactions on System, Man, and Cybernetics, Part A 33 pp. 415–428.
(3), 314–324. Yang, S.J.H., Chu, W.C., Lee, J., Huang, W.T., 1997. A fuzzy Petri nets
Girault, C., Valk, R. (Eds.), 2003. Petri Nets for Systems Engineering– based mechanism for fuzzy rules reasoning. In: The 21st Annual
A Guide to Modeling, Verification, and Applications. Springer, Berlin. International, Computer Software and Applications Conference.
Inprise Corporation, 2002. Borland C++ Builder6 Developer’s Guide. pp. 438–443.
Borland Software Co., USA. Yang, S.J.H., Tsai, J.J.P., Chen, C.C., 2003. Fuzzy rule base systems
Joo, M.G., Lee, J.S., 2005. A class of hierarchical fuzzy systems with verification using high-level Petri nets. IEEE Transactions on Knowl-
constraints on the fuzzy rules. IEEE Transactions on Fuzzy Systems 13 edge and Data Engineering 15 (2), 457–473.
(2), 194–203. Zhou, M.C., Venkatesh, K., 1998. Modeling Simulation and Control of
Koriem, S.M., 2000. A fuzzy Petri net tool for modeling and verification Flexible Manufacturing Systems–A Petri Net Approach. World
of knowledge-based systems. Computer Journal 43 (3), 206–223. Scientific Pub. Co., Singapore.

You might also like