Expert Systems with Applications 38 (2011) 4403–4415

Contents lists available at ScienceDirect

Expert Systems with Applications

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

Failure mode and effects analysis using fuzzy evidential reasoning approach
and grey theory
Hu-Chen Liu, Long Liu ⇑, Qi-Hao Bian, Qin-Lian Lin, Na Dong, Peng-Cheng Xu
College of Mechanical Engineering, Tongji University, 1239 Siping Road, Shanghai 200092, PR China

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

Keywords: Failure mode and effects analysis (FMEA) is a methodology to evaluate a system, design, process or ser-
Failure mode and effects analysis vice for possible ways in which failures (problems, errors, etc.) can occur. The two most important issues
Fuzzy evidential reasoning of FMEA are the acquirement of FMEA team members’ diversity opinions and the determination of risk
Grey theory priorities of the failure modes that have been identified. First, the FMEA team often demonstrates differ-
ent opinions and knowledge from one team member to another and produces different types of assess-
ment information because of its cross-functional and multidisciplinary nature. These different types of
information are very hard to incorporate into the FMEA by the traditional model and fuzzy logic
approach. Second, the traditional FMEA determines the risk priorities of failure modes using the risk pri-
ority numbers (RPNs) by multiplying the scores of the risk factors like the occurrence (O), severity (S) and
detection (D) of each failure mode. The method has been criticized to have several shortcomings. In this
paper, we present an FMEA using the fuzzy evidential reasoning (FER) approach and grey theory to solve
the two problems and improve the effectiveness of the traditional FMEA. As is illustrated by the numer-
ical example, the proposed FMEA can well capture FMEA team members’ diversity opinions and prioritize
failure modes under different types of uncertainties.
Ó 2010 Elsevier Ltd. All rights reserved.

1. Introduction defined as the manner in which a component, subsystem, system,

Failure mode and effects analysis (FMEA) is a widely used engi-
neering technique for defining, identifying and eliminating known
and/or potential failures, problems, errors and so on from system,
design, process, and/or service before they reach the customer
(Stamatis, 1995). It is intended to provide information for making
risk management decisions. When it is used for a criticality analy-
sis, it is also referred to as failure mode, effects and criticality anal-
ysis (FMECA). FMEA as a formal system analysis methodology was
first proposed by NASA in 1960s for their obvious reliability
requirements (Sharma, Kumar, & Kumar, 2005). Since then, it has
been extensively used as a powerful tool for safety and reliability
analysis of products and processes in a wide range of industries
particularly, aerospace, nuclear and automotive industries
(Ebeling, 2000; Gilchrist, 1993; O’Connor, 2000).
FMEA is a structured, bottom-up approach that starts with
known potential failure modes at one level and investigates the ef-
fect on the next subsystem level and, hence, a complete FMEA
analysis of a system often span all the levels in the hierarchy from
bottom to top (see Fig. 1) (Sharma et al., 2005). The failure mode is
⇑ Corresponding author. Tel.: +86 (0)21 2023 1668; fax: +86 (0)21 2023 1661.
E-mail address: liulong@tongji.edu.cn (L. Liu).
process, etc. could potentially fail to meet the design intent. A fail-
ure mode in one component can be the cause of a failure mode in
another component. A failure cause is defined as a weakness that
may result in a failure. For each identified failure mode, their ulti-
mate effects need to be determined, usually by a FMEA team. A fail-
ure effect is defined as the result of a failure mode on the function
of the product/process as perceived by the customer.
A good FMEA can help analysts identify known and potential
failure modes and their causes and effects, help them prioritize
the identified failure modes and can also help them work out cor-
rective actions for the failure modes. Its main objective is to allow
the analysts to identify and prevent known and potential problems
from reaching the customer. To this end, the risks of each identified
failure mode need to be evaluated and prioritized so that appropri-
ate corrective actions can be taken for different failure modes. The
traditional FMEA determines the risk priorities of failure modes
through the risk priority number (RPN), which is the product of
the occurrence (O), severity (S) and detection (D) of a failure. That
is
RPN ¼ S  O  D ð1Þ
where S is the severity of the failure, O is the probability of the fail-
ure occurrence, and D is the probability of the failure being de-
tected. The three risk factors are evaluated using the 10-point

0957-4174/$ - see front matter Ó 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.eswa.2010.09.110
4404 H.-C. Liu et al. / Expert Systems with Applications 38 (2011) 4403–4415

Fig. 1. Hierarchical structure of system.

Table 1 scale described in Tables 1–3. The failure modes with higher RPNs
Traditional ratings for occurrence of a failure (Chin, Chan, & Yang, 2008; Chin et al., are assumed to be more important and will be given higher priori-
2009; Pillay & Wang, 2003; Seyed-Hosseini et al., 2006; Wang et al., 2009).
ties for correction.
Rating Probability of occurrence Possible failure rate
10 Very High: failure is almost inevitable P1/2 1.1. FMEA procedure
9 1/3
8 High: repeated failures 1/8 As outlined by Pillay and Wang (2003), the process for carrying
7 1/20
6 Moderate: occasional failures 1/80
out an FMEA can be divided into several steps as shown in Fig. 2
5 1/400 (Pillay & Wang, 2003; Sharma et al., 2005; Tay & Lim, 2006). These
4 1/2000 steps are briefly explained here:
3 Low: relatively few failures 1/15,000
2 1/150,000
1. Develop a good understanding of what the system is sup-
1 Remote: failure is unlikely 1/1, 500,000
posed to do when it is operating properly.

Table 2
Traditional ratings for severity of a failure (Chin, Chan, & Yang, 2008; Chin et al., 2009; Pillay & Wang, 2003; Seyed-Hosseini et al., 2006; Wang et al., 2009).

Rating Effect Severity of effect

10 Hazardous without Very High severity ranking when a potential failure mode affects safe vehicle operation and/or involves noncompliance with
warning government regulations without warning
9 Hazardous with Very High severity ranking when a potential failure mode affects safe vehicle operation and/or involves noncompliance with
warning government regulations with warning
8 Very High Vehicle/item inoperable, with loss of primary function
7 High Vehicle/item operable, but at reduced level of performance. Customer dissatisfied
6 Moderate Vehicle/item operable, but comfort/convenience item(s) inoperable. Customer experiences discomfort
5 Low Vehicle/item operable, but comfort/convenience item(s) operable at reduced level of performance. Customer experiences some
dissatisfaction
4 Very Low Cosmetic defect in finish, fit and finish/squeak or rattle item that does not conform to specifications. Defect noticed by most
customers
3 Minor Cosmetic defect in finish, fit and finish/squeak or rattle item that does not conform to specifications. Defect noticed by average
customer
2 Very minor Cosmetic defect in finish, fit and finish/squeak or rattle item that does not conform to specifications. Defect noticed by discriminating
customers
1 None No effect

Table 3
Traditional ratings for detection of a failure (Chin, Chan, & Yang, 2008; Chin et al., 2009; Pillay & Wang, 2003; Seyed-Hosseini et al., 2006; Wang et al., 2009).

Rating Detection Criteria

10 Absolutely Design control will not and/or cannot detect a potential cause/mechanism and subsequent failure mode; or there is no design control
impossible
9 Very remote Very remote chance the design control will detect a potential cause/mechanism and subsequent failure mode
8 Remote Remote chance the design control will detect a potential cause/mechanism and subsequent failure mode
7 Very Low Very Low chance the design control will detect a potential cause/mechanism and subsequent failure mode
6 Low Low chance the design control will detect a potential cause/mechanism and subsequent failure mode
5 Moderate Moderate chance the design control will detect a potential cause/mechanism and subsequent failure mode
4 Moderately high Moderately high chance the design control will detect a potential cause/mechanism and subsequent failure mode
3 High High chance the design control will detect a potential cause/mechanism and subsequent failure mode
2 Very High Very High chance the design control will detect a potential cause/mechanism and subsequent failure mode
1 Almost certain Design control will almost certainly detect a potential cause/mechanism and subsequent failure mode
 H.-C. Liu et al. / Expert Systems with Applications 38 (2011) 4403–4415 4405

Fig. 2. FMEA process.

Table 4
Format of an FMEA report.

System FMEA No.

Subsystem Page
Component Prepared by
Core team FMEA Date (org.)
Existing conditions Action results
Component/ Potential Potential Potential Present Severity Occurrence Detection Risk priority Recommend Action S O D RPN
process failure effects of causes of control number actions taken
mode mode mode mechanisms (RPN)

2. Divide the system into sub-systems and/or assemblies in
order to localize the search for components.
3. Use blue prints, schematics and flow charts to identify com-
ponents and relations among components.
4. Develop a complete component list for each assembly.
5. Identify operational and environmental stresses that can
affect the system. Consider how these stresses might affect
the performance of individual components.
6. Determine failure modes of each component and the effects
of failure modes on assemblies, sub-systems, and the entire
system.
7. Categorize the hazard level (severity) of each failure mode
(several qualitative systems have been developed for this
purpose).
8. Estimate the probability. In the absence of solid quantitative
statistical information, this can also be done using qualita-
tive estimates.
9. Calculate the risk priority number (RPN): the RPN is given as
the multiplication of the index representing the probability,
severity and detectability.
10. Determine if action needs to be taken depending on the RPN.
11. Develop recommendations to enhance the system perfor-
mance. These fall into two categories:
 Preventive actions: avoiding a failure situation.
 Compensatory actions: minimizing losses in the event
that a failure occurs.
12. Prepare FMEA report by summarizing the analysis in tabular
form as shown in Table 4 (Puente, Pino, Priore, & de la
Fuente, 2002; Sharma et al., 2005).
1.2. Setbacks of FMEA
The traditional FMEA has been proven to be one of the most
important early preventative actions in system, design, process
or service which will prevent failures and errors from occurring
and reaching the customer. However, the FMEA have been exten-
sively criticized for various reasons (Bowles & Peláez, 1995; Chin,
Wang, Ka Kwai Poon, & Yang, 2009; Pillay & Wang, 2003; Puente
et al., 2002; Seyed-Hosseini, Safaei, & Asgharpour, 2006; Sharma
4406 H.-C. Liu et al. / Expert Systems with Applications 38 (2011) 4403–4415
et al., 2005; Tay & Lim, 2006; Yang, Bonsall, & Wang, 2008; Wang,
Chin, Ka Kwai Poon, & Yang, 2009):
 Different sets of O, S and D ratings may produce exactly the
same value of RPN, but their hidden risk implications may be
totally different. For example, two different failure modes with
values of 2, 3, 2 and 4, 1, 3 for O, S and D, respectively, will have
the same RPN value of 12. However, the hidden risk implica-
tions of the two failure modes may be very different because
of the different severities of the failure consequence. This may
cause a waste of resources and time, or in some cases, a high-
risk failure mode being unnoticed.
 The relative importance among O, S and D is not taken into con-
sideration. The three factors are assumed to have the same
importance. This may not be the case when considering a prac-
tical application of FMEA.
 The mathematical formula for calculating RPN is questionable
and debatable. There is no rationale as to why O, S and D should
be multiplied to produce the RPN.
 The conversion of scores is different for the three factors. For
example, a linear conversion is used for O, but a nonlinear trans-
formation is employed for D.
 RPNs are not continuous with many holes and heavily distrib-
uted at the bottom of the scale from 1 to 1000. This causes prob-
lems in interpreting the meaning of the differences between
different RPNs. For example, is the difference between the
neighboring RPNs of 1 and 2 the same or less than the difference
between 900 and 1000?
 The RPN considers only three factors mainly in terms of safety.
Other important factors such as economical aspects are ignored.
 Small variations in one rating may lead to vastly different
effects on the RPN, depending on the values of the other factors.
For example, if O and D are both10, then a 1-point difference in
severity rating results in a 100-point difference in the RPN; if O
and D are equal to 1, then the same 1-point difference results in
only a 1-point difference in the RPN; if O and D are both 4, then
a 1-point difference produces a 16-point difference in the RPN.
 The three factors are difficult to precisely determine. Much
information in FMEA can be expressed in a linguistic way such
as likely, important or Very High and soon.
1.3. Related work of FMEA
To overcome the above drawbacks, a number of approaches
have been suggested in the literature. For example, Yang et al.
(2008) presented a fuzzy rule-based Bayesian reasoning (FuRBaR)
approach for prioritizing failures in FMEA. The technique is in-
tended to deal with some of the drawbacks concerning the use of
conventional fuzzy logic (i.e. rule-based) methods in FMEA. In
the proposed approach, subjective belief degrees are assigned to
the consequent part of the rules to model the incompleteness
encountered in establishing the knowledge base. A Bayesian rea-
soning mechanism is then used to aggregate all relevant rules for
assessing and prioritizing potential failure modes.
Wang et al. (2009) proposed fuzzy risk priority numbers
(FRPNs) for prioritization of failure modes to deal with the problem
that it is not be realistic in real applications to determine the risk
priorities of failure modes using the risk priority numbers (RPNs)
because they require the risk factors of each failure mode to be pre-
cisely evaluated. In the paper, the authors treat the risk factors O, S
and D as fuzzy variables and evaluate them using fuzzy linguistic
terms and fuzzy grades. The FRPNs are defined as fuzzy weighted
geometric means of the fuzzy grades for O, S and D, and be com-
puted using alpha-level sets and linear programming models. Fi-
nally, the FRPNs are defuzzified using centroid defuzzification
method for ranking purpose.
Chin et al. (2009) presented an FMEA using the evidential rea-
soning (ER) approach to model the diversity and uncertainty of
the assessment information in FMEA. The new model allows FMEA
team members to assess risk factors independently and express
their opinions individually and chooses the traditional numerical
ratings for the assessment of risk factors. It also allows the risk fac-
tors to be aggregated in a rigorous yet nonlinear rather than simple
addition or multiplication manner. However, all individual and
interval assessment grades in the model are assumed to be crisp
and independent of each other in the approach. It did not consid-
erate the occasion in FMEA where an assessment grade may repre-
sent a vague concept or standard and there may be no clear cut
between the meanings of two adjacent grades.
Gilchrist (1993) proposed an expected cost model to improve
the tradition FMEA. It was formulated as EC = CnPfPd, where EC is
the expected cost to the customer, C the cost per failure, n the
items produced per batch or per year, Pf the probability of a failure
and Pd the probability of the failure not to be detected. The author
believed that the expected cost model is more rigorous yet practi-
cal than the RPN and have great benefit of forcing people to think
about quality costs. Ben-Daya and Raouf (1996) proposed an im-
proved FMEA model because they argued that the probabilities Pf
and Pd in the expected cost model were not always independent
and very difficult to estimate at the design stage of a product
and this verity was completely ignored by the expected cost model.
The improved FMEA model addressed gave more importance to the
likelihood of occurrence over the likelihood of detection and com-
bined with the expected cost model to provide a quality improve-
ment scheme for the production phases of a product or service.
Sankar and Prabhu (2001) presented a modified approach for
prioritization of failures in a system FMEA, which uses the ranks
1–1000 called risk priority ranks (RPRs) to represent the increasing
risk of the 1000 possible severity–occurrence–detection combina-
tions. These 1000 possible combinations were tabulated by an ex-
pert in order of increasing risk and can be interpreted as ‘if–then’
rules. The failure having a higher rank was given a higher priority.
Braglia (2000) developed a multi-attribute failure mode
analysis (MAFMA) based on the analytic hierarchy process (AHP)
technique, which considers four different factors O, S, D, and
expected cost as decision attributes, possible causes of failure as
decision alternatives, and the selection of cause of failure as deci-
sion goal. The goal, attributes and alternatives formed a three-level
hierarchy, where the pair wise comparison matrix was used to esti-
mate attribute weights and the local priorities of the causes with
respect to the expected cost attribute, the conventional scores for
O, S and D were normalized as the local priorities of the causes
with respect to O, S, and D, respectively, and the weight composi-
tion technique in the AHP was utilized to synthesize the local pri-
orities into the global priority, based on which the possible causes
of failure were ranked. Braglia, Frosolini, and Montanari (2003a)‘
also presented an alternative multi-attribute decision-making ap-
proach called fuzzy TOPSIS approach for FMECA, which is a fuzzy
version of the technique for order preference by similarity to ideal
solution (TOPSIS). The proposed fuzzy TOPSIS approach allows the
risk factors O, S and D and their relative importance to be assessed
using triangular fuzzy numbers rather than precise crisp numbers.
Chang, Wei, and Lee (1999) used fuzzy sets and gray systems
theory for FMEA, where fuzzy linguistic terms were used to evalu-
ate the degrees of O, S and D, and gray relational analysis was ap-
plied to determine the risk priority of potential causes. To carry out
the gray relational analysis, fuzzy linguistic assessment informa-
tion was defuzzified as crisp values, the lowest level and the
assessment information of the three factors for each potential
cause were defined as a standard series and a comparative series
respectively, whose gray relational coefficients and gray relational
degree were computed. Bigger gray relational degree means
 H.-C. Liu et al. / Expert Systems with Applications 38 (2011) 4403–4415
smaller effect of potential cause. In addition, Chang, Liu, and Wei
(2001) also utilized the gray system theory for FMEA, but the gray
relational degrees were computed using the traditional scores 1–
10 for the three factors rather than fuzzy linguistic assessment
information.
Seyed-Hosseini et al. (2006) proposed a method called decision
making trial and evaluation laboratory (DEMATEL) for reprioritiza-
tion of failure modes in FMEA, which prioritizes alternatives based
on severity of effect or influence and direct and indirect relation-
ships between them. Direct relationships were a set of connections
between alternatives with a set of connection weights representing
severity of influence of one alternative on another. An indirect rela-
tionship was defined as a relationship that could only move in an
indirect path between two alternatives and meant that a failure
mode could be the cause of other failure mode(s). Alternatives hav-
ing more effect on another were assumed to have higher priority
and called dispatcher. Those receiving more influence from an-
other were assumed to have lower priority and called receiver.
Bowles and Peláez (1995) described a fuzzy logic-based ap-
proach for prioritizing failures in a system FMECA, which uses lin-
guistic terms to describe O, S, D and the riskiness of failure. The
relationships between the riskiness and O, S, D were characterized
by a fuzzy if–then rule base which was developed from expert
knowledge and expertise. Crisp ratings for O, S and D were fuzz-
ified to match the premise of each possible if–then rule. All the
rules that have any truth in their premises were fired to contribute
to the fuzzy conclusion. The fuzzy conclusion was then defuzzified
by the weighted mean of maximum method (WMoM) as the rank-
ing value of the risk priority. Similar fuzzy inference method also
appeared in other literatures such as Braglia, Frosolini, and
Montanari (2003b), Pillay and Wang (2003), Puente et al. (2002),
Sharma et al. (2005) and Tay and Lim (2006).
The above literature review shows that much effort has been
paid to the improvement of FMEA by incorporating factor weights,
more factors, expert knowledge and/or fuzziness into the analysis,
but no or little attention has been paid to the diversity and uncer-
tainty of assessment information given by FMEA members. Nor-
mally, FMEA is a group decision behavior and cannot be
performed on an individual basis. Different FMEA team members
may demonstrate different opinions because of their different
expertise and backgrounds (Chin et al., 2009; Naude, Lockett, &
Holmes, 1997). They may provide different types assessment infor-
mation for the same risk factor, some of which may be complete or
incomplete, precise or imprecise, known or unknown and certain
or uncertain. And also the assessment grades in the FMEA may rep-
resent a vague concept or standard and there may be no clear cut
between the meanings of two adjacent grades.
On the other hand, fuzzy logic approaches is not an easy task
which requires experts to make a vast number of judgments and
will be highly costly and time-consuming. Tay and Lim (2006) ar-
gued that not all the rules were actually required in fuzzy RPN
models and, thus, proposed different rules reduction systems to
simplify the fuzzy logic-based FMEA methodology. However, any
inference from an incomplete rule base will be inaccurate and
should be avoided because the reduced rule base will be incom-
plete and contain ignorance information (Wang et al., 2009). Wang
et al. (2009) also argued that using fuzzy if–then rules for FMEA
will result in the problem that the fuzzy if–then rules with the
same consequence but different antecedents are unable to be dis-
tinguished from one another. As a result, the failure modes charac-
terized by these fuzzy if–then rules will be unable to be prioritized
or ranked. Beside, the use of fuzzy if–then rules has no way to
incorporate the relative importance of risk factors into the fuzzy
inference system.
From the above analysis, we can see that there is a clear need to
develop a new approach for FMEA to solve the problems mentioned
4407
above. In this paper we propose a new risk priority model for FMEA
using the fuzzy evidential reasoning (FER) approach and grey
theory. The new model can not only model the diversity and uncer-
tainty of the assessment information in FMEA and incorporate the
relative importance of risk factors into the determination of risk
priority of failure modes, but also take advantage of the benefits
of fuzzy logic and grey theory without the need of asking experts
too much. The paper is organized as follows: In Sections 2 and 3,
we develop the risk priority model using the FER approach and grey
theory. In Section 4, a numerical example is provided to illustrate
the potential applications of the new model in FMEA. The paper is
concluded in Section 5 with a brief summary.
2. Fuzzy evidential reasoning approach
Evidential reasoning (ER) approach, which is the combination of
the Dempster–Shaffer (D–S) theory (Shafer, 1976) with a distrib-
uted modeling framework, was first developed in 1994 for dealing
with multiple attribute decision analysis (MADA) problems charac-
terized by both quantitative and qualitative attributes with various
types of uncertainties (Guo, Yang, Chin, Wang, & Liu, 2009; Yang,
Wang, Xu, & Chin, 2006). The main advantage of the ER approach
is that both precise data and subjective judgments with uncer-
tainty can be consistently modeled under the unified framework.
The ER approach provides a novel procedure for aggregating multi-
ple attributes based on the distributed assessment framework and
the evidence combination rule of the D–S theory.
Extensive research dedicated to the ER approach has been con-
ducted in recent years. Experiences show that a decision maker
may not always be confident enough to provide subjective assess-
ments to individual grades only but at times wishes to be able to
assess beliefs to subsets of adjacent grades. It is to deal with the
problem that the interval grade ER (IER) approach is proposed
(Yang & Sen, 1994). Another extension to the original ER approach
is to take account of vagueness or fuzzy uncertainty, i.e., the assess-
ment grades are no longer clearly distinctive crisp sets, but are de-
fined as dependent fuzzy sets. Yang et al. (2006) proposed the
fuzzy ER approach to extend the original ER individual grades to
fuzzy grades to capture fuzziness caused by the fuzzy evaluation
grades. Recently, Guo et al. (2009) developed a general ER model-
ing framework and an attribute aggregation process, which is re-
ferred to as the fuzzy IER (FIER) algorithm, in order to deal with
both fuzzy and interval grade assessments and provide a more
powerful means to support the solution of complex MADA
problems.
2.1. Assessment of risk factors using belief structures
O, S and D are the three major risk factors identified in FMEA.
Although more risk factors could be included, the main concern
of this paper is not the identification of risk factors, but their
assessment and aggregation. The risk priority model to be devel-
oped in this paper has no limitation on the number of risk factors
and is applicable to any number of risk factors. The three risk fac-
tors can be evaluated numerically or linguistically. Both of them
have been extensively applied and have their merits and demerits.
However, there is a high level of uncertainty involved in FMEA
since FMEA is a group decision behavior and the assessment infor-
mation for risk factors mainly based on experts’ subjective judg-
ments may be complete or incomplete, precise or imprecise, and
certain or uncertain. In addition, most experts are willing to ex-
press their opinions by belief degrees (or possibility measures)
based on a set of evaluation grades, i.e., {Very Low, Low, Moderate,
High, and Very High}. As such, in this paper, we choose linguistic
4408 H.-C. Liu et al. / Expert Systems with Applications 38 (2011) 4403–4415
terms for the assessment of risk factors and the individual evalua-
tion grade set is defined as a fuzzy set HF as follows:
HF ¼ fH11 ; H22 ; H33 ; H44 ; H55 g
¼ fVeryLow; Low; Moderate; High; VeryHighg:
In order to generalize the H b F ¼ fHij ; i ¼ 1; . . . ; 5; j ¼ 1; . . . ; 5g to
fuzzy sets, we assume that a general set of fuzzy individual assess-
ment grades {Hii}, i = 1, . . . , 5 are dependent on each other and only
two adjacent fuzzy individual assessment grades may intersect.
Based on the experts’ opinions, we can approximate all the five
individual assessment grades by trapezoidal fuzzy numbers
(triangular fuzzy numbers are special cases of trapezoidal fuzzy
numbers) for simplifying the discussion and without loss of gener-
ality, and their membership function values can be determined
according to the historical data and the detailed questionnaire an-
swered by all experts, as shown in Fig. 3 and Table 5. Furthermore,
we define the interval fuzzy assessment grades sets Hij for i = 1–4
and j = i + 1 to 5 as trapezoidal fuzzy sets that include fuzzy indi-
vidual grades Hii, H(i+1)(i+1), . . . , Hjj. If the individual assessment
grades are trapezoidal fuzzy sets, every interval grade will be a
trapezoidal fuzzy set as shown in Fig. 4.
As mentioned in the previous section, the assessment grades in
the FMEA may represent a vague concept or standard and there
may be no clear cut between the meanings of two adjacent grades.
In other words, these evaluation grades may not be regarded as
crisp sets. For example, it is difficult to separate the grade Moderate
from the grade Low, especially if evaluations need to be given be-
tween these two grades. Such a problem can be solved by the
FER approach, which allows FMEA team members to provide their
subjective judgments in the following flexible ways:
 A certain grade such as Low, which can be written as {(H22, 1.0)}.
Such an expression is referred to as a belief structure in the FER
approach.
 A distribution such as Low to 0.4 and Moderate to 0.6,which
means that a failure mode is assessed with respect to the risk
factor under consideration to grade Low to the degree of 0.4
and to grade Moderate to the degree of 0.6.Here the degrees of
0.4 and 0.6 represent the confidences (also called belief degrees)
of the FMEA team member in his/her subjective judgments and
the distribution can be equivalently expressed as
{(H22, 0.4), (H33, 0.6)}. When all the confidences are summed to
one, the distribution is said to be complete; otherwise, it is said
to be incomplete. For example, {(H22, 0.4), (H33, 0.5)} is an
incomplete distribution or called incomplete assessment, where
the missing information of 0.1 is referred to as local ignorance
Fig. 3. Fuzzy membership function for linguistic terms.
Table 5
Fuzzy ratings for linguistic terms.
Linguistic terms Fuzzy number
Very Low (0, 0, 1, 2)
Low (1, 2, 3, 4)
Moderate (3, 4, 6, 7)
High (6, 7, 8, 9)
Very High (8, 9, 10, 10)
and could be assigned to any grade between Very Low and Very
High according to the Dempster–Shafer theory of evidence (Sha-
fer, 1976).
 An interval such as Low–Moderate, which means that the grade
of a failure mode with respect to the risk factor under evalua-
tion is between Low and Moderate. This can be written as
{(H23, 1.0)}.
 No judgment, which means the FMEA team member is not will-
ing to or cannot provide an assessment for a failure mode with
respect to the risk factor under consideration. In other words,
the grade by this FMEA team member could be anywhere
between Very Low and Very High and can be expressed as
{(H15, 1.0)}. Such judgments a referred to as total ignorance.
Obviously, the belief structures in the FER approach provide
FMEA team members with an easy-to-use and very flexible way
to express their opinions and can better quantify risk factors than
the traditional RPN methodology. All failure modes with respect to
the three risk factors can be evaluated using belief structures. In
the next subsection, we will see how the belief structures of each
failure mod with respect to every risk factor provided by FMEA
team member individually can be synthesized into a group belief
structure.
2.2. Group belief structures
Suppose there are K members (TM1, . . . , TMK) in a FEMA team
responsible for the assessment of N failure modes (FM1, . . . , FMN)
with respect to L risk factors (RF1, . . . , RFL). Each team member
P
TMk is given a weight kk > 0 (k = 1, . . . , K) satisfying Kk kk ¼ 1 to re-
flect his/her relative importance in the FMEA team. The degree of
importance for each of the team members can be determine by
using direct grade, point allocation, eigenvector method, linear
programming techniques for multidimensional analysis of prefer-
ences (LINMAP), or Delphi method, etc. together with the team
members’ experience and domain knowledge (Chin et al., 2009).
 
~ kl ¼ wkla ; wklb ; wkld is the weight of risk factor RFl given by
Let w
 H.-C. Liu et al. / Expert Systems with Applications 38 (2011) 4403–4415 4409

Fig. 4. Interval fuzzy grades set.

Table 6 where wl is referred to as the crisp number of group weight and is

Fuzzy weights for the relative importance of risk determined by Eq. (8).
factors. n  o
Let Hij ; bkij ðFMn ; RFl Þ ; i ¼ 1; . . . ; 5; j ¼ 1; . . . ; 5 be the belief
Linguistic terms Fuzzy number
structure provided by TMk on the assessment of FMn with respect
Very Low (VL) (0, 0, 0.25)
to RFl, where Hii for i = 1–5 are fuzzy assessment grades defined for
Low (L) (0, 0.25, 0.5)
Moderate (M) (0.25, 0.5, 0.75) risk assessment, Hij for i = 1–4 and j = i + 1 to 5 are the intervals fuz-
High (H) (0.5, 0.75, 1) zy assessment grades between Hii and Hjj, and bkij ðFMn ; RFl Þ are the
Very High (VH) (0.75, 1, 1)
belief degrees to which FMn assessed on RFl to the intervals Hij. All
the grades Hii for i = 1–5 and the intervals Hij for i = 1–4 and j = i + 1
TMk to reflect its relative importance in the determination of risk to 5 together form frame of discernment, which is expressed as
priorities of the failure modes. Since they are not easy to be pre- b F ¼ fHij ; i ¼ 1; . . . ; 5; j ¼ 1; . . . ; 5g
H
cisely determined due to the same reason as O, S and D, the relative
importance weights of the risk factors are assessed using the lin- Or equivalently
guistic terms in Table 6, whose membership functions are visual- 8 9
>
> H11 H12 H13 H14 H15 >
>
ized in Fig. 5 (Wang et al., 2009). The group weight of risk factor >
> >
>
>
> H22 H23 H24 H25 >
>
RFl of the K team members is denoted as < =
! bF ¼
H H33 H34 H35 ð4Þ
X
K X
K X
K X
K >
> >
>
>
> H44 >
H45 >
~l ¼
w ~ kl
kk w ¼ kk wkla ; kk wklb ; kk wkld ; >
> >
>
: ;
k¼1 k¼1 k¼1 k¼1 H55
l ¼ 1; . . . ; L ð2Þ The collective assessment of the K team members for each fail-
The group weights of risk factors are first defuzzified using Eq. (8) ure mode with respect to each risk factor is also a belief structure,
and then normalized using Eq. (3) called group or collective belief structure, which is denoted as
w e n ðlÞ ¼ fðHij ; b ðFMn ; RFl ÞÞ; i ¼ 1; . . . ; 5; j ¼ 1; . . . ; 5g;
X
 l ¼ PL l
w ; l ¼ 1; . . . ; L ð3Þ ij

l¼1 wl n ¼ 1; . . . ; N; l ¼ 1; . . . ; L ð5Þ

Fig. 5. Membership functions of fuzzy weights.

4410 H.-C. Liu et al. / Expert Systems with Applications 38 (2011) 4403–4415

where bij(FMn, RFl) is referred to as group or collective belief degree The values of c and d will remain the same for the defuzzifica-
and is determined by tion of all linguistic terms. The values a0 and b0 are rating values
at the extreme limits of each linguistic term where the member-
X
K
bij ðFMn ; RFl Þ ¼ kk bkij ðFMn ; RFl Þ; ship function is 0 and a1 and b1 are rating values at the extreme
k¼1 limits of each linguistic term where the membership function is
ð6Þ
i ¼ 1; . . . ; 5; j ¼ 1; . . . ; 5; 1 (for a trapezoidal membership function).
n ¼ 1; . . . ; N; l ¼ 1; . . . ; L: Finally, the overall assessment of the failure mode FMn with re-
spect to risk factor RFl is also a crisp number, called overall belief
That is, a group belief degree is the weighted sum of the individ- structure, which can be aggregated by the following equation:
ual belief degrees corresponding to the same grade or interval. In
addition, the group belief structures for the N failure modes with X
5 X
5
X n ðlÞ ¼ hijbij ðFMn ; RFl Þ
respect to the L risk factors form a fuzzy belief decision matrix as ð9Þ
i¼1 j¼1
is shown in Eq. (7), which differs from the traditional decision ma-
trix in that the former consists of fuzzy assessment grades and be- n ¼ 1; . . . ; N; l ¼ 1; . . . ; L
lief structures while the latter is made up of numerical values. e can be defuzzified to crisp
So, the fuzzy belief decision matrix X
2 3 2 3 belief decision matrix X, which is shown as follows:
FM1 e 1 ð1Þ
X e 1 ð2Þ
X  e 1 ðLÞ
X
6 FM 7 6 6Xe ð1Þ e 2 ð2Þ
7
e 2 ðLÞ 7 2 3 2 3
6 2 7 X  X FM1 X 1 ð1Þ X 1 ð2Þ  X 1 ðLÞ
e ¼6
X 7 6 2 7
6 .. 7 ¼ 6 .. 7 ð7Þ 6 FM 7 6 X 2 ð1Þ X 2 ðLÞ 7
4 . 5 4 ...
6 ..
.  . 5
7 6 2 7 6 X 2 ð2Þ  7
X¼6 7 6
6 .. 7 ¼ 6 .. ..
7 ð10Þ
FMN e N ð1Þ
X e N ð2Þ   
X e N ðLÞ
X 4 . 5 4 . .  ... 75
FMN X N ð1Þ X N ð2Þ    X N ðLÞ
2.3. Defuzzification

Based on the fuzzy belief decision matrix, group belief struc- 3. Grey theory
tures on the assessment of each failure mode with respect to the
L risk factors can be aggregated into an overall belief structure Grey theory, proposed by Deng (1989), deals with decisions
using defuzzification method and weighted average method characterized by incomplete information (Shih, Wu, & Huang,
successively. 1996; Wu, Deng, & Wen, 1984), such as operation, mechanism,
Chen and Klien (1997) have proposed an easy defuzzification structure and behavior, which are neither deterministic nor totally
method for obtaining the crisp number of a fuzzy set, which is unknown, but are partially known. It explores system behavior
shown here in Eq. (8) using relation analysis and model construction.
Pn The use of grey theory within the FMEA framework is practica-
i¼0 ðbi  cÞ
hij ¼ Pn P ; i ¼ 1; . . . ; 5; j ¼ 1; . . . ; 5 ð8Þ ble and can be accomplished (Chang et al., 2001; Chang et al.,
i¼0 ðb i  cÞ  ni¼0 ai  d 1999). The flowchart in Fig. 7 shows the proposed grey theory ap-
where hij is the defuzzified crisp number of Hij. As an example, con- proach to rank the failure modes, which are identified in the FMEA
sider the defuzzification of the linguistic term Moderate as shown in process. First, the application of fuzzy sets with grey theory re-
Fig. 6. This linguistic term can be defuzzified to produce a crisp va- quires the defuzzification of the membership functions obtained
lue as seen below in Figs. 3 and 4. The defuzzified values are used to generate the
comparative series, which is represented in the form of a matrix.
½b0  c þ ½b1  c In short, a comparative series is generated based on the crisp belief
h33 ¼
f½b0  c þ ½b1  cg  f½a0  d þ ½a1  dg decision matrix X as shown in Eq. (10). At this step, the standard
½7  0 þ ½6  0 series for the risk factors are generated by determining the optimal
¼ ¼ 0:500
f½7  0 þ ½6  0g  f½3  10 þ ½4  10g level of all risk factors for the failure modes in the FMEA. This stan-
dard series is also represented in the form of a matrix. This could

Fig. 6. Defuzzification of the linguistic term Moderate.

 H.-C. Liu et al. / Expert Systems with Applications 38 (2011) 4403–4415 4411

Fig. 7. Flowchart of proposed approach.

be assumed to be the lowest level of the linguistic terms describing
the risk factors.
Next, the difference between the standard and comparative ser-
ies D0 is obtained and the results are used to determine the grey
relation coefficient.
Using the value of the grey relation coefficient and the group
weights of risk factors, the degree of grey relation of each failure
mode can be calculated. The degree of relation in FMEA denotes
the relationship between the potential causes and the optimal va-
lue of the risk factors. The highest of the value obtained reflect the
smallest priority of the identified failure modes. So, all failure
modes in the FMEA can be ranked according to the ascending order
of the degree of relation (Pillay & Wang, 2003).
3.2. Standard series
Degree of relation can describe the relationship of two series,
thus, an objective series called the standard series shall be estab-
lished, and expressed as X0 = (X0(1), X0(2), . . . , X0(L)). When con-
ducting FMEA, the smaller the score, the less the risk, therefore
the standard series can be the lowest level of all the risk factors
(Chang et al., 2001):
X 0 ¼ ½X 0 ð1Þ; X 0 ð2Þ; . . . ; X 0 ðLÞ ¼ ½H11 ; H11 ; . . . ; H11 
¼ ½h11 ; h11 ; . . . ; h11  ð12Þ
3.3. The difference between comparative series and standard series

3.1. Comparative series The difference between the comparative and standard series,
D0, is calculated and reflected in a form of a matrix as seen below:
An information series with  L components or risk  factors can be 2 3
expressed as, X 0n ¼ X 0n ð1Þ; X 0n ð2Þ; . . . ; X 0n ðLÞ 2 X 0 , where D01 ð1Þ D01 ð2Þ  D01 ðLÞ
X 0n ðlÞ; l ¼ 1; . . . ; L denotes the factors of X 0n . If all information series
6 D02 ð1Þ D02 ð2Þ  D02 ðLÞ 7
6 7
are comparable, the N information series can be described as the D0 ¼ 6
6 .. .. .. 7
7 ð13Þ
4 . .  . 5
following matrix (Deng, 1989):
2 3 2 3 D0N ð1Þ D0N ð2Þ    D0N ðLÞ
X 01 X 01 ð1Þ X 01 ð2Þ  X 01 ðLÞ
6 X 0 7 6 X 0 ð1Þ where, D0n(l) = kX0(l)  Xn(l)k, X0(l) is the standard series and Xn(l) is
6 27 6 2 X 02 ð2Þ  X 02 ðLÞ 7
7
X0 ¼ 6 7 6 7 the comparative series.
6 .. 7 ¼ 6 .. .. .. 7 ð11Þ
4 . 5 4 . .  . 5
3.4. Grey relation coefficient
X 0N X 0N ð1Þ X 0N ð2Þ    X 0N ðLÞ
For the application of this matrix in an FMEA study, the matrix The grey relation coefficient, c(X0(l), Xn(l)), is calculated using
X0 is generated based on the crisp belief decision matrix X, which is Eq. (14) for each risk factor of the failure modes identified in the
determined by Eq. (9). FMEA
4412 H.-C. Liu et al. / Expert Systems with Applications 38 (2011) 4403–4415

minn minl jX 0 ðlÞ  X n ðlÞj þ fmaxn maxl jX 0 ðlÞ  X n ðlÞj
cðX 0 ðlÞ; X n ðlÞÞ ¼
jX 0 ðlÞ  X n ðlÞj þ fmaxn maxl jX 0 ðlÞ  X n ðlÞj
Dmin þ fDmax
¼ n ¼ 1; . . . ; N; l ¼ 1; . . . ; L
D0n ðlÞ þ fDmax
ð14Þ
where X0(l) is the min or max value from the standard series and
Xn(l) is the min or max value from the comparative series. f is an
identifier, f 2 (0, 1); only affecting the relative value of risk without
changing the priority (Chang et al., 2001). Generally, f can be 0.5
(Deng, 1989).
3.5. The degree of relation
This step is to obtain the degree of grey relation based upon the
grey relation coefficient and the group weights of risk factors w  l,
which is determined by Eq. (3). The degree of grey relation is cal-
culated using Eq. (15) for each failure mode
X
L information are shaded and highlighted. The five team members
CðX m ; X n Þ ¼  l cfX m ðlÞ; X n ðlÞg
w ð15Þ
l¼1
For FMEA, the higher the degree of relation obtained from Eq.
(15), the smaller the effect of the failure mode. As a result, all the
failure modes can be prioritized or ranked in terms of the degree
of grey relation of each failure mode.
4. An illustrative example
In this section, we provide a numerical example to illustrate
the potential applications of the proposed FMEA and particularly
the potentials of using the fuzzy evidential reasoning (FER) ap-
proach and grey theory in capturing FMEA team members’ diver-
sity opinions and prioritizing failure modes under different types
of uncertainties. The FMEA example is adapted from Wang et al.
(2009).
A FMEA team consisting of five cross-functional team members
identifies seven potential failure modes in a system and needs to
prioritize them in terms of their failure risks such as probability
of occurrence, severity and detectability so that high risky failure
modes can be corrected with top priorities. Due to the difficulty
in precisely assessing the risk factors and their relative importance
weights, the FMEA team members agree to evaluate them using
the linguistic terms defined in Tables 5 and 6. The assessment
information of the seven failure modes on each risk factor and
the risk factor weights provided by the five team members is pre-
sented in Table 7, where incomplete assessments and ignorance
from different departments are assumed to be of different impor-
tance because of their different domain knowledge and expertise.
To reflect their differences in performing FEMA, the five team
members are assigned the following relative weights: 0.15, 0.20,
30, 0.25 and 0.10, which are shown in the second column of
Table 7.

Table 7
Assessment information on seven failure modes by five FMEA team members.

Table 8
Group assessment of the FMEA team on the seven failure modes and the group weights of three risk factors.

Failure mode Occurrence Severity Detection

1 {(H15, 0.30), (H22, 0.60), (H25, 0.10)} {(H12, 0.15), (H22, 0.45), (H23, 0.25), (H34, 0.15)} {(H11, 0.45), (H12, 0.295), (H15, 0.005), (H22, 0.25)}
2 {(H33, 0.975), (H44, 0.075)} {(H15, 0.10), (H22, 0.30), (H33, 0.60)} {(H11, 0.85), (H13, 0.15)}
3 {(H15, 0.02), (H33, 0.48), (H44, 0.50)} {(H33, 0.85), (H44, 0.15)} {(H11, 1.00)}
4 {(H33, 0.25), (H44, 0.75)} {(H15, 0.015), (H24, 0.135), (H33, 0.40), (H44, 0.45)} {(H11, 0.20), (H22, 0.80)}
5 {(H33, 0.25), (H34, 0.30), (H44, 0.45)} {(H55, 1.00)} {(H11, 0.65), (H22, 0.35)}
6 {(H22, 0.70), (H33, 0.10), (H35, 0.20)} {(H11, 0.55), (H13, 0.20), (H15, 0.05), (H22, 0.20)} {(H13, 0.16), (H15, 0.25), (H22, 0.55), (H44, 0.04)}
7 {(H33, 1.00)} {(H11, 0.405), (H22, 0.595)} {(H12, 0.10), (H33, 0.90)}
Group weights (0.425, 0.675, 0.8625) (0.6125, 0.8625, 1) (0.0625, 0.2875, 0.5375)
 H.-C. Liu et al. / Expert Systems with Applications 38 (2011) 4403–4415 4413

Table 9
Defuzzified crisp number for fuzzy assessment grades.

Table 10
Defuzzified and aggregated assessment information for the seven failure modes and risk priority ranking.

Failure mode Occurrence Severity Detection Grey relation Priority ranking

1 0.379 0.364 0.210 0.734 6
2 0.541 0.438 0.170 0.677 4
3 0.604 0.531 0.130 0.649 3
4 0.656 0.594 0.260 0.588 2
5 0.614 0.870 0.187 0.561 1
6 0.376 0.234 0.377 0.763 7
7 0.500 0.226 0.476 0.718 5
Weights 0.36 0.45 0.19

It is quite obvious that there is almost no existing FMEA method The difference between the comparative and standard series D0
that can be used, without making some kind of assumptions, to is then calculated and expressed as a matrix. Since all entries for
deal with the assessment information in Table 7 because it is dif- the matrix representing the standard series was determined to
ferent from one team member to another and also includes incom- be 0, the difference between the comparative and standard series
plete assessments and ignorance. To carry out a priority analysis, would be equal to the comparative series (considering that
we first use belief structures to express the FMEA team members’ D0n(l) = kX0(l)  Xn(l)k).
individual assessments and synthesize them into group belief Using the values obtained from the difference of the standard
structures by using Eq. (6), as presented in Table 8. The group belief and comparative series, the grey relation coefficient, c(X0(l), Xn(l)),
structures are then defuzzified and aggregated into overall belief is calculated using Eq. (14) for each risk factor of the failure modes
structures using Eqs. (8) and (9) described in Section 2. The results identified in the FMEA. Take the first failure modes in Table 10 for
are presented in Table 10. During this process, all the fuzzy assess- example, the grey relation coefficient is calculated as shown here:
ment grades H b F ¼ fHij ; i ¼ 1; . . . ; 5; j ¼ 1; . . . ; 5g can be defuzzified
0:130 þ 0:5  0:870
using Eq. (8) to produce a crisp number. The result of the defuzzifi- cðX 0 ð1Þ; X 1 ð1ÞÞ ¼ ¼ 0:694
0:379 þ 0:5  0:870
cation is tabulated in Table 9.
Next, the data from the FMEA in Table 10 are analyzed using the 0:130 þ 0:5  0:870
cðX 0 ð2Þ; X 1 ð2ÞÞ ¼ ¼ 0:707
grey theory method. The comparative series is generated based on 0:364 þ 0:5  0:870
the table using Eqs. (9)–(11), as seen in the matrix below. The stan- 0:130 þ 0:5  0:870
cðX 0 ð3Þ; X 1 ð3ÞÞ ¼ ¼ 0:876
dard series is taken to be the lowest level of the linguistic term 0:210 þ 0:5  0:870
describing all three variables, which is Very Low. When the linguis- Similarly, the grey relation coefficient for all the failure modes
tic term Very Low is defuzzified, the crisp number obtained is can be calculated in the same way as shown in the matrix below
0.130, this represents the average value, as such the value 0 (low- 2 3
est possible value) is used to represent the linguistic term Very Low 0:694 0:707 0:876
in the standard series (Pillay & Wang, 2003). A matrix representing 6 0:579 0:647 0:934 7
6 7
the standard series is generated as shown here 6 7
6 0:544 0:585 1:000 7
2 3 6 7
0:379 0:364 0:210 6 7
cðX 0 ðlÞ; X n ðlÞÞ ¼ 6
6 0:518 0:549 0:813 7
7
6 7 6 7
6 0:541 0:438 0:170 7 6 0:539 0:433 0:908 7
6 7 6 7
6 0:604 0:531 0:130 7 6 7
6 7 4 0:697 0:845 0:696 5
6 7
X0 ¼ X ¼ 6
6 0:656 0:594 0:260 77;
6 7 0:604 0:851 0:620
6 0:614 0:870 0:187 7
6 7 On the other side, based upon the information in Table 7, the
6 0:376 0:234 0:377 7
4 5 relative importance weights of all risk factors are first aggregated
0:500 0:226 0:476 using Eq. (2) as shown in the last row of Table 8. The group weights
2 3 2 3
H11 H11 H11 0 0 0 of risk factors are then defuzzified and normalized using Eqs. (8)
6 7 6 7 and (3), respectively. The results are provided in the last row of
6 H11 H11 H11 7 6 0 0 0 7
6 7 6 7 Table 10.
6H H11 7 6 7
6 11 H11 7 60 0 07 Substituting the grey relation coefficient and group weights of
6 7 6 7
X0 ¼ 6
6 H11 H11 H11 7 ¼ 6 0 0 0 7
7 6
7 risk factors into Eq. (15) will give the degree of relation for the first
6 7 6 7 failure mode as seen here:
6 H11 H11 H11 7 6 0 0 0 7
6 7 6 7
6H H11 7 6 7 CðX 0 ; X 1 Þ ¼ ½ð0:694  0:36Þ þ ð0:707  0:45Þ þ ð0:876  0:19Þ
4 11 H11 5 40 0 05
H11 H11 H11 0 0 0 ¼ 0:734
4414 H.-C. Liu et al. / Expert Systems with Applications 38 (2011) 4403–4415
In the same way, the degree of relation is calculated for all the
failure modes identified in the FMEA to produce a ranking that
determines the priority for attention. The results are shown in
the last column of Table 10. The degree of relation of the seven fail-
ure modes give the priority ranking of the seven failure modes as
FM5  FM4  FM3  FM2  FM7  FM1  FM6, which is perfectly
consistent with the ranking achieved by our former intuitive
analysis.
So, the final conclusion for this example is that failure model 5
is given the top priority for correction, followed by failure modes 4,
3, 2, 7, 1, and 6.
5. Conclusions
FMEA is a very important safety and reliability analysis tool
which has been widely used in a wide range of industries. In view
of its difficulty in acquiring precise assessment information on
failure modes, the defects in the traditional RPN modes and the
difficulty in building a complete fuzzy rule base, we proposed a
new FMEA using the fuzzy evidential reasoning (FER) approach
and grey theory in this paper. It can not only capture FMEA team
members’ diversity opinions under different types of uncertain-
ties, but also allow the risk factors and their relative importance
weights to be evaluated in a linguistic manner. More importantly,
the degree of relation rather than a crisp RPN or fuzzy if–then
rules is to be defined for prioritization of failure modes. The core
of the proposed FMEA was the development of the risk priority
model using the fuzzy evidential reasoning (FER) approach and
grey theory, which includes assessing risk factors using belief
structures, synthesizing individual belief structures into group be-
lief structures, aggregating defuzzified group belief structures into
overall belief structure, establishing comparative series standard
series, obtaining the difference between comparative series and
standard series, computing grey relational coefficient and degree
of relation and ranking the failure modes using the degree of
relation.
The potential applications of the proposed FMEA and the de-
tailed computational process of the degree of relation were exam-
ined and illustrated with a numerical example. It was shown that
the proposed FMEA provided a useful, practical, effective and flex-
ible way for risk evaluation in FMEA. In particular, the proposed
FMEA offered a new way for capturing MEA team members’ opin-
ions and prioritizing failure modes in FMEA.
Compared with the traditional RPN and its kinds of variants, the
proposed FMEA has the following advantages:
 It can be used for systems where safety data is unavailable or
unreliable, as it does not force precision.
 The relative importance weights of risk factors are taken into
consideration in the process of prioritization of failure modes,
which makes the proposed FMEA more realistic, more practical
and more flexible.
 There is no need to build any if–then rule base which proves to
be highly subjective, costly and time consuming.
 Risk factors and their relative importance weights are evaluated
in a linguistic manner rather than in precise numerical values.
This enables the experts to express their judgments more real-
istically and makes the assessment easier to be carried out.
 The diversity and uncertainty of FMEA team members’ assess-
ment information can be well reflected and modeled using
belief structures. And it provides an organized method to com-
bine expert knowledge and experience for use in an FMEA
study.
 Failure modes can be fully ranked and well distinguished from
each other unless some of them are assessed to be the same.
 More risk factors can be included if necessary. The proposed
FMEA is not limited to O, S and D, but applicable to any number
of risk factors.
Acknowledgments
Our deepest gratitude goes first and foremost to the authors of
this paper’s references whose work have contributed greatly to the
completion of this thesis.
Second, the authors would like to thank the editor and review-
ers for their constructive suggestions of the paper.
References
Ben-Daya, M., & Raouf, A. (1996). A revised failure mode and effects analysis model.
International Journal of Quality & Reliability Management, 13(1), 43–47.
Bowles, J. B., & Peláez, C. E. (1995). Fuzzy logic prioritization of failures in a system
failure mode, effects and criticality analysis. Reliability Engineering and System
Safety, 50, 203–213.
Braglia, M. (2000). MAFMA: multi-attribute failure mode analysis. International
Journal of Quality & Reliability Management, 17(9), 1017–1033.
Braglia, M., Frosolini, M., & Montanari, R. (2003a). Fuzzy TOPSIS approach for failure
mode, effects and criticality analysis. International Journal of Quality and
Reliability Engineering, 19, 425–443.
Braglia, M., Frosolini, M., & Montanari, R. (2003b). Fuzzy criticality assessment
model for failure modes and effects analysis. International Journal of Quality and
Reliability Management, 20(4), 503–524.
Chen, C. B., & Klien, C. M. (1997). A simple approach to ranking a group of
aggregated fuzzy utilities. IEEE Transactions on Systems Man and Cybernetics, Part
B, 27(1), 26–35.
Chang, C. L., Liu, P. H., & Wei, C. C. (2001). Failure mode and effects analysis using
gray theory. Integrated Manufacturing Systems, 12(3), 211–216.
Chang, C. L., Wei, C. C., & Lee, Y. H. (1999). Failure mode and effects analysis using
fuzzy method and gray theory. Kybernetes, 28, 1072–1080.
Chin, K. S., Chan, A., & Yang, J. B. (2008). Received: Development of a fuzzy FMEA
based product design system. International Journal of Advanced Manufacturing
Technology, 36, 633–649.
Chin, K. S., Wang, Y. M., Ka Kwai Poon, G., & Yang, J. B. (2009). Failure mode and
effects analysis using a group-based evidential reasoning approach. Computers
& Operations Research, 36, 1768–1779.
Deng, J. (1989). Introduction to grey system theory. Journal of Grey Systems, 1(1),
1–24.
Ebeling, C. (2000). An introduction to reliability and maintainability engineering. New
York: Tata McGraw-Hill Company Ltd..
Gilchrist, W. (1993). Modeling failure mode and effect analysis. International Journal
of Quality & Reliability Management, 10(5), 16–23.
Guo, M., Yang, J. B., Chin, K. S., Wang, H. W., & Liu, X. B. (2009). Evidential reasoning
approach for multiattribute decision analysis under both fuzzy and interval
uncertainty. IEEE Transactions on Fuzzy Systems, 17(3), 683–698.
Naude, P., Lockett, G., & Holmes, K. (1997). A case study of strategic engineering
decision making using judgmental modeling and psychological profiling. IEEE
Transactions on Engineering Management, 44(3), 237–244.
O’Connor, P. D. T. (2000). Practical reliability engineering. London: Heyden.
Pillay, A., & Wang, J. (2003). Modified failure mode and effects analysis using
approximate reasoning. Reliability Engineering and System Safety, 79, 69–85.
Puente, J., Pino, R., Priore, P., & de la Fuente, D. (2002). A decision support system for
applying failure mode and effects analysis. International Journal of Quality and
Reliability Management, 19(2), 137–150.
Sankar, N. R., & Prabhu, B. S. (2001). Modified approach for prioritization of failures
in a system failure mode and effects analysis. International Journal of Quality and
Reliability Management, 18(3), 324–335.
Seyed-Hosseini, S. M., Safaei, N., & Asgharpour, M. J. (2006). Reprioritization of
failures in a system failure mode and effects analysis by decision making trial
and evaluation laboratory technique. Reliability Engineering and System Safety,
91, 872–881.
Shafer, G. A. (1976). Mathematical theory of evidence. Princeton, NJ: Princeton
University Press.
Sharma, R. K., Kumar, D., & Kumar, P. (2005). Systematic failure mode effect analysis
(FMEA) using fuzzy linguistic modeling. International Journal of Quality &
Reliability Management, 22(9), 986–1004.
Shih, K. C., Wu, K. W., & Huang, U. P. (1996). Information of grey relation. Taiwan:
Chuan-Hua Co. Ltd..
Stamatis, D. H. (1995). Failure mode and effect analysis—FMEA from theory to
execution. New York: ASQC Quality Press.
Tay, K. M., & Lim, C. P. (2006). Fuzzy FMEA with a guided rules reduction system for
prioritization of failures. International Journal of Quality and Reliability
Management, 23(8), 1047–1066.
Wu, H. H., Deng, J. L., & Wen, K. L. (1984). Introduction to grey analysis. Taiwan: Kao-
Li Co., Ltd.
 H.-C. Liu et al. / Expert Systems with Applications 38 (2011) 4403–4415 4415

Yang, J. B., & Sen, P. (1994). A general multi-level evaluation process for hybrid
MADM with uncertainty. IEEE Transactions on Systems Man and Cybernetics,
24(10), 1458–1473.
Yang, J. B., Wang, Y. M., Xu, D. L., & Chin, K. S. (2006). The evidential reasoning
approach for MADA under both probabilistic and fuzzy uncertainties. European
Journal of Operational Research, 17, 309–343.
Yang, Z. L., Bonsall, S., & Wang, J. (2008). Fuzzy rule-based bayesian reasoning
approach for prioritization of failures in FMEA. IEEE Transactions on Reliability,
57(3), 517–528.
Wang, Y. M., Chin, K. S., Ka Kwai Poon, G., & Yang, J. B. (2009). Risk evaluation in
failure mode and effects analysis using fuzzy weighted geometric mean. Expert
Systems with Applications, 36, 1195–1207.