Beruflich Dokumente
Kultur Dokumente
Research
1. INTRODUCTION
ailure mode, effects and criticality analysis (FMECA) is the most popular systematic assessment of a
process (product) that enables us to determine the location and the mechanism of potential failures, with
the aim of preventing process (product) failures.
FMECA is characterized by a bottom-up approach by which any complex production system is decomposed
into its constituent parts, which are successively analysed to find all the potential failure causes and their
effects. The analyst builds a table with all failure causes and performs a criticality assessment to measure the
risk level for each fault, in terms of criteria such as the chance of failure or the severity of the fault itself.
Correspondence to: Marcello Braglia, Dipartimento di Ingegneria Meccanica, Nucleare e della Produzione, Facolt`a di Ingegneria,
This allows us to prioritize the individuated failure modes with respect to the assessment method adopted.
The US MIL-STD-1629A Procedures for performing a failure mode, effects and criticality analysis describes
a standard criticality assessment methodology which is based on calculation of a criticality number (CN) for
each system failure mode i given by:
CNi = i i p t
where is the failure-effect probability, the failure mode ratio, the failure rate, and t the operating time.
Criticality number technique is mostly used in the aerospace, nuclear, and chemical industries.
A second approach for criticality assessment is based on the risk priority number (RPN) method. The RPN
calculation uses linguistic terms to rank the probability of failure (P ), the severity of its failure effect (S), and
the chance of the failure being undetected (D) on a 1 to 10 numeric scale. Well-known conversion tables
(e.g. Ben-Daya and Raouf1 and Gilchrist2 ) report the typical basis for the linguistic judgement scales used to
estimate the three crisp (in the sense of net, precise) quantities which are used to calculate the RPN value in
the following manner:
RPN = PDS
Due to its simplicity, the RPN method is preferred by the manufacturing industries such as automotive
companies3, domestic appliance firms4 , and tyre companies5.
Unfortunately, when used in real industrial situations, FMECA shows some important weaknesses concerning
the rationality of the approach, especially in terms of criticality analysis (e.g. Ben-Daya and Raouf1 , Gilchrist2 ,
Chang et al.6,7 and Teng and Ho8 ):
FMECA considers only three kinds of attributes, neglecting other important aspects such as the production
quantity and economic aspects;
it is usually difficult to give precise evaluations of intangible quantities such as P , D, and S. In addition,
conversion scores are linear for the chance of failure, but nonlinear for the chance of undetection;
analysts often would like to give different weights to critical criteria;
the RPN cannot measure the effectiveness of proposed corrective measures;
different sets of the three factors can produce exactly the same value of RPN, but the hidden implications
may be totally different;
the mathematical form (i.e. multiplication) adopted for calculating the failure priority is questionable and
strongly sensitive to variations in criticality factor evaluations.
In this paper, a new rational approach to determine the risk priority number and to overcome the limits of
the conventional RPN, as cited above, is proposed. This method is based on a fuzzy version of the technique
for order preference by similarity to ideal solution (TOPSIS)9 . Briefly, TOPSIS is a multi-attribute decisionmaking methodology based on the measurement of the Euclidean distance of an alternative from an ideal
goal. The technique has been specifically adapted to simplify the risk-assessment procedure and to allow a
correct evaluation of pertinent data. To this aim, TOPSIS has been integrated with fuzzy logic theory, to create a
framework able to handle imprecise quantities (such as those deriving from linguistic evaluations or subjective
and qualitative assessments) in a very simple way, giving a final ranking for failure causes that is easy to
interpret, as in the case of the FMECA, but much more consistent.
When considering the applicability of these types (i.e. derived from the control theory) of fuzzy criticality
assessment models in real industrial cases, some doubts remain because of the difficulties involved in defining
the many rules and membership functions required by this methodology. To overcome this problem, Braglia
and Bevilacqua16 propose the use of AHP for obtaining the rules for a particular fuzzy criticality assessment
model. Another characteristic of this model is the use of a triangular approach as crisp inputs in fuzzy models
to evaluate the different opinions of the maintenance staff.
Other FMECA formulations have been proposed in which the evaluation factors are used to classify the
failure into a risk class more than to calculate a risk value. A known example of this type of approach is
the method recommended by the norm DIN V 19 250 Basic safety analyses for process, measuring and control
protection equipment. The risk analysis requirement category is calculated using a risk-o-graph (i.e. a decision
tree) in which the following four parameters are linked: extent of danger (a four-level classification), time of
exposure to risk (two-level), possibility of preventing the risk (two-level), and probability of occurrence of the
risk (three-level). The result of this analysis is a requirement category (RC) based on eight possible levels of
risk, where RC1 means a very low level of risk and RC8 means the highest possible risk.
Puente et al.17 present a criticality assessment approach based on qualitative rules which provide a ranking
of the risks of potential causes of failure. The methodology assigns a risk priority class (RPC) to each cause
of failure, depending on the importance given to the three conventional factors (P , D, and S) related to a
failure mode. In practice, the integer scores (between 1 and 10) of the three factors are translated into their
corresponding qualitative classes (five classes from very low to very high). A set of 125 (= 5 5 5) rules
combining the three inputs of the model permits us to obtain the RPC output variable for each cause of failure
based on a nine-class classification. In order to optimize the risk-discrimination capabilities of the different
causes of failure, a modified version of the technique integrating with fuzzy logic is also proposed by the authors.
Wj = 1
j =1
In terms of criticality assessment application, the failure causes are the alternatives to be ranked. The three
conventional FMECA factors (P , D, and S) related to a failure mode are considered as criteria. But it is evident
that a maintenance manager can potentially consider additional needed criteria during his analysis, such as
expected costs.
The TOPSIS method is based on the following steps18 .
Step 1. Construct the normalized decision matrix R
The first step concerns the normalization of the judgement matrix X = [xij ]. Each element xij is transformed
using the following equation
xij
rij =
n
2
i=1 xij
c 2003 John Wiley & Sons, Ltd.
Copyright
Si+
Si
+ Si
where
0 Ci+ 1
and i = 1, 2, . . . , n
The best alternative is the one which has the shortest distance to the ideal solution. Considering the proposed
calculus, if an alternative has the shortest distance to the ideal solution, then it is assured to have the longest
distance to the non-ideal solution18.
To eliminate the conversion debate by directly evaluating the (crisp) linguistic assessment of three factors, a
fuzzy logic version of TOPSIS can be developed. With the use of fuzzy theory, it is possible to improve the initial
decision system (i.e. the P , D, and S evaluation) by introducing the element of uncertainty, and to optimize the
risk discrimination of different causes of failure.
c 2003 John Wiley & Sons, Ltd.
Copyright
In defining the fuzzy membership functions (FMFs) for the three factors P , D, and S, and for the
corresponding weights of importance, a good compromise between computational complexity and the reachable
precision is needed to avoid an unnecessary waste of resources. As Triantaphyllou and Lin state in their paper,
fuzzy triangular numbers (FTNs) appear to be a valid tool, offering a well balanced compromise between
computational costs and accuracy in the final ranking18.
It is important to underline that, with this approach, each factor is not defined by a set of FMFs covering the
(crisp) interval of existence [1, 10] as in the other FMECA fuzzy models reported in the literature13,14 . Simply, in
our methodology, a generic evaluation of a factor (P , D, or S) for a generic cause of failure is not given with
a crisp value but with a triangular fuzzy number. Similarly for the evaluations of the weights of importance.
In other words, a triangular fuzzy judgement of the factors/weights is only required to the analyst. In this way,
it is not necessary to implement into the model a fuzzy inference engine that, elaborating the inputs, allows us
to obtain the output. This advantage is not negligible because a fuzzy inference engine consists of a knowledge
base supported by a qualitative rule base, generally composed by more than one hundred rules (see, e.g., the 125
reported in Puente et al.17 ). It is evident that this knowledge base is heavy, critical, questionable, and complex
to define.
In general, using more complex forms than FTNs, such as trapezoidal or Gaussian ones, would allow a
more precise description of the problem under analysis. Notwithstanding this, FTNs have been preferred for the
following main considerations.
1. As described by Driankov et al.19 , with respect to fuzzy control theory, more complex functions cause
greater computational complexity, but frequently without significant advantages.
2. The increase of complexity deriving from the adoption of more accurate membership functions could be
particularly evident in fuzzy multi-attribute decision-making theory (MADM). In particular, FTNs require
basic fuzzy mathematical operations (Appendix A) that are easier to use with respect to those needed,
for example, with trapezoid-shaped numbers. In addition, working with triangular shapes facilitates the
ranking of the final outcome of the MADM analysis. Such an operation would be, otherwise, a delicate
and difficult task to carry out, as underlined in Ribeiro20. This aspect is well demonstrated by the fact that
the greater number of fuzzy ranking methods proposed in the literature concern only FTNs.
3. Apart from the mathematical aspects above mentioned, another important matter to be pointed out lies
in the difficulties that usually arise when defining the membership shape. As will be shown below, FTNs
represent the perfect candidate to build with ease a membership function effectively representing the
judgement distribution of multiple experts. This is not true for more complex membership functions, such
as trapezoidal ones.
A generic FTN A is defined by ordered triples A = (a1 , a2 , a3 ) representing, respectively, the lower value,
the modal value, and the upper value of a triangular FMF. Its membership function is described as a bi-linear
relation
x a1
a1 x a2
a2 a1
A (x) = a3 x
a2 x a3
a3 a2
0
otherwise
Obviously, a1 stands for the lower bound, a2 for the modal value, and, finally, a3 for the upper bound (see the
example reported in Figure 1).
To create the fuzzy triangular judgements of three factors and their corresponding weights of importance,
a simple assumption in proposed in this paper. Following this assumption, a maintenance manager has more
confidence that the input parameter lies in the centre of the symmetrical interval than at the edges. In our
case, the concept is linked to the distribution of judgements about the crisp value. The approach draws on a
well-known heuristic procedure for choosing a judgement distribution in the absence of data21 .
Assuming that the experts judgements can be treated as a continuous random variable, and considering
that in practice it is not possible to collect a set of data (i.e. judgements) large enough to permit studying the
c 2003 John Wiley & Sons, Ltd.
Copyright
Membership
0,8
0,6
0,4
0,2
a1=4
0,0
1
a2=5
5
a3=8
6
10
Severity
Figure 1. Example of a severity (S) fuzzy judgement for a generic cause of failure
corresponding probability distribution, a well-known heuristic procedure for choosing a judgement distribution
in the absence of data21 is adopted.
The first step in using the heuristic approach is to identify an interval [a, b] (where a < b and a, b
{1, 2, . . . , 10}) in which it is felt that judgements will lie with probability close to 1. In order to obtain subjective
evaluations of the two extreme judgement evaluations a and b, maintenance staff are asked for their optimistic
and pessimistic estimates, respectively, of the criticality evaluation associated with a given criterion of a given
mode of failure. Once an assessment interval [a, b] has been subjectively identified, the next step is to place
a probability density function (i.e. the fuzzy membership) on [a, b] that is thought to be representative of the
particular evaluation. To this aim, the triangular distribution is used as a rough stochastic model in the absence
of a data distribution.
In the triangular approach, the maintenance experts are also asked for their subjective assessment of the most
likely criterion quantification for a given mode of failure. This most likely value c is the mode of the judgement
triangular distribution and has a height equal to 1.
Such numbers can be used to perform common operations such as addition, multiplication, negation, division,
and so on (Appendix A). A complete set of operations is defined to correctly operate on such entities, allowing
the user to easily obtain a set of fuzzy outputs. This makes it possible to infer the final solution in an easy manner
and avoids the necessity of executing a very large number of calculations. Using these mathematical operations
it is easy to recalculate the TOPSIS approach in a fuzzy manner18 and to obtain, as final output, a fuzzy version
of the relative closeness that represents the triangular fuzzy priority index for each analysed cause of failure.
The major drawback in adopting similar fuzzy approaches arises when having to rank the obtained fuzzy
outputs (i.e. the priority evaluation of the causes of failure). A general and universally accepted method does not
exist. The one proposed by Zhu and Lee22 and used by Triantaphyllou and Lin in their original Fuzzy TOPSIS
method still leaves some doubts when dealing with close fuzzy numbers. Besides, the main remarkable problem
for our type of application is due to the fact that this ranking technique requires a large number of comparisons
between all the alternatives (i.e. causes of failure) analysed. The rank is obtained after a mathematical pairwise
comparison of all fuzzy priority numbers. Considering that in a FMECA for a plant it is easy to find a large
number of causes of failure, this approach is impracticable.
For these reasons, a different ranking technique is proposed for the criticality assessment problem. The fuzzy
outputs are ranked following the technique proposed by Facchinetti et al.23 and specifically valid for FTNs.
c 2003 John Wiley & Sons, Ltd.
Copyright
This approach seems to be a good compromise between complexity and robustness of final results in confused
situations.
The method makes use of a preference function between two FTNs that expresses the degree to which every
alternative is preferred to another. In addition, it considers the propensity for risk of the decision maker thanks
to a convex combination of two different approaches: a pessimistic method and an optimistic method.
Consider two FTNs A = (a1 , a2 , a3 ) and B = (b1 , b2 , b3 ). The first method affirms that a pessimistic
preference function P can be defined in the following manner:
a2 b1
,0 ,1
P (A, B) = min max
a2 a1 + b2 b1
or, symmetrically
P (B, A) = min max
b2 a1
,0 ,1
a2 a1 + b2 b1
As can be noted by the relations reported above, the pessimistic method only refers to the lower value of the
fuzzy triangular numbers and to their mode. Without going deeply into the mathematical aspects of the theory,
this is due to the fact that the pessimistic comparison between two triangular functions is basically realized
by considering only the left half of the two FTNs (i.e. the intervals [a1 , a2 ] and [b1 , b2 ], respectively). In the
same manner, when adopting the optimistic approach described below, the theory refers to the right half of the
triangles. By doing so, it is also easier to understand the meaning of the pessimistic and optimistic approaches.
The former means that for the decision maker it is more interesting/critical to concentrate the analysis and the
corresponding ranking on the lower estimate of the FTN (i.e. a1 and b1 ). Vice versa, for the second one.
The pessimistic preference relation is the following:
A >P B
1
2
In terms of real functions, it is possible to write in an equivalent manner the following relation of (pessimistic)
preference:
A >P B
iff xP (A) =
a1 + a2
b1 + b2
> xP (B) =
2
2
a3 b2
O (A, B) = min max
,0 ,1
a3 a2 + b3 b2
b3 a2
O (B, A) = min max
,0 ,1
a3 a2 + b3 b2
A >O B
A >O B
Finally, the -convex combination of the pessimistic and optimistic preference functions is the characterized by
the following equations:
(A, B) = O (A, B) + (1 )P (A, B), with [0, 1]
A > B iff (A, B) > (B, A)
In terms of real functions we have
A > B
The parameter represents the pessimistic or optimistic attitude of the analyst. In other words, as stated above,
the parameter describes the preference of the decision maker towards a ranking analysis mainly based on the lefthalf values of the FTNs rather than on those on the right half. For more details of this fuzzy ranking procedure
the reader may refer to the paper of Facchinetti et al.23 .
All the movements are obtained by means of electrical motors. The pallets moving the moulds pass through
all the process phases in about 8 minutes, during which the parts are subjected to temperature and pressure
variations and, above all, to the effects of chemical substances. A schematic representation of the system is
reported in Figure 2.
At the loading station an operator performs the manual positioning of the parts on the moulds, fixing the last
ones on a pallet (there are 16 pallets on the machine). Each pallet is then introduced into a pre-heating kiln, at
c 2003 John Wiley & Sons, Ltd.
Copyright
a temperature of 25 C, working with forced air heated by means of electrical resistances. Thermal regulators
provide a constant control of the whole phase. Later on, the pallet is introduced to the foaming zone, where
the most important and delicate procedure takes place. The foaming masks are formed by two plane supports,
crossed by electrical current, supplied with two magnets to hold the metallic part of the door. The foaming
machine controls a maximum of 11 masks at a time, arranged on the two levels of the paternoster (i.e. the masks
conveyor). The movement between the two lines is obtained by means of two elevators formed by mechanical
arms, driven vertically by separate motors, while a set of hydraulic cylinders keeps the masks well joined
together during the process. When a pallet is introduced, the mask is first closed and carried to the foaming
zone, where it is opened again and filled with the foam, at a temperature of about 40 C. Later, the pallet is
driven back to the unloading station. This operation takes about 6 minutes, which is more than sufficient to
allow a correct chemical reaction among the substances which constitute the foam.
Others
10
8
6
4
2
0
0
10
12
14
16
Fuzzy TOPSIS
Figure 3. Correlation analysis of the ranking results obtained with different criticality analysis techniques adopted
investigate the robustness of the obtained results and to evaluate the impact of the uncertainty on the ultimate
outcome.
It is important to state here that the parameter of the fuzzy ranking technique adopted is set equal to 12 .
In this way the real preference functions for two general fuzzy numbers A and B become
x (A) =
a1 + 2a2 + a3
4
and x (B) =
b1 + 2b2 + b3
4
The choice for the neutral value is mainly due to a couple of reasons as follows.
1. To express that the maintenance experts did not show a clear preference for the left extreme with respect
to the right one, for each of the three parameters of the FMECA. In other words, the experts did not weight
in different manners the pessimistic assessments with respect to the optimistic ones.
2. Adopting the value = 12 , the convex preference functions degenerate into the well-known formulas
proposed by Yager24 that in the past have proved to give an effective ranking of fuzzy subsets.
In Table I a comparison of the results concerning a partial example of the first criticality analysis executed with
FMECA and fuzzy TOPSIS approaches is reported.
Considering the reduced additional effort required, the numerical experimental test has also been extended by
considering the results obtained by using the crisp TOPSIS approach. In this case, for each cause of failure the
same linguistic assessments of conventional FMECA have been adopted. Evidently, the three evaluation factors
share the same crisp estimate.
The software application allowed us to calculate the existing correlation among the fuzzy TOPSIS method
results and those deriving from the application of the standard FMECA and crisp TOPSIS techniques (Figure 3).
As one can see, there is a good correlation among all the adopted methods. The correlation coefficient ranges
from a maximum of 0.995 in the case of fuzzy TOPSIS versus crisp TOPSIS, to a minimum of 0.946 in the
case of fuzzy TOPSIS versus standard FMECA. Hence, the TOPSIS method, integrated with the FTN, when
the weights of the performance criteria are kept unchanged, gives a reasonable result. Indeed, apart from some
small differences, due to the uncertainty introduced by using the fuzzy triangular numbers, the analysis confirms
the original standard FMECA outcome, e.g. the robustness of the obtained results.
In a second instance, we modified the failure criteria weights. As stated by the maintenance staff, and also
confirmed by some authors (e.g. Ben-Daya and Raouf1 and Gilchrist2 ), the chance of failure criterion should be
c 2003 John Wiley & Sons, Ltd.
Copyright
Loss of power
Loss of power
Communication
interrupted
Communication
interrupted
Communication
interrupted
Loss of power
Communication
interrupted
(B) Wiring
(D) Breaking of
inout device
(E) Communication
error
(F) Short circuit
(G) Mechanical
shock
Failure mode
Failure cause
2
2
3
2
Machine stop
Machine stop
Machine stop
Issued commands
are not executed
Issued commands
are not executed
Issued commands
are not executed
Loss of control
Failure effect
4
5
2
2
24
20
50
24
32
20
36
RPN
Standard FMECA
P
WP
(1.00, 1.00, 1.00)
D
WD
(1.00, 1.00, 1.00)
S
WS
(1.00, 1.00, 1.00)
Fuzzy TOPSIS
FCN
0.685
0.692
1.000
0.740
0.706
0.828
0.765
Ranking
Table I. Example of FMECA analysis following standard and fuzzy TOPSIS approaches, respectively; system: insulation foam injection line; primary element: control system
considered more important than the others, due to the simple observation that, for instance, a failure cause with a
very high severity value but having a very remote occurrence probability may be less critical than a failure cause
which occurs repeatedly. Shortly, the probabilities are not independent. Therefore, supported by the experts, we
decided to assign as mode values (i.e. a2 ) for the weights of the three failure criteria the following:
WP = 8,
WD = 3,
WS = 3
At the same time, we defined opportune lower and upper bounds, following the indications given by the experts
with respect to their subjective considerations about the performed assessments. Again, the uncertainty of
the judgement concerning the generic weight is captured by adopting the minimum and maximum values,
respectively, from the set composed by the optimistic and pessimistic evaluations proposed by the experts.
After performing the above-described normalization, the FTNs obtained for the three weights are
WP = (0.6, 0.8, 1.0),
On this occasion, the mode of fuzzy triangular judgements (i.e. a2 ) adopted for fuzzy TOPSIS has also been
used to represent the crisp TOPSIS weights for the three evaluation criteria.
Hence, the correlation coefficients decrease sensibly with respect to those obtained in the previous case, as
reported below:
fuzzy TOPSIS versus crisp TOPSIS = 0.974
fuzzy TOPSIS versus standard FMECA = 0.903
As one can see, a good level of correlation is confirmed, though the differences between traditional FMECA
and fuzzy TOPSIS begin to appear clearly. This means that, by changing the weights of importance of the
three factors, the critical final ranking of the failure causes undergoes a number of changes. Unquestionably, the
possibility of weighting differently the three criteria and to introduce into the mathematical model the concept
of uncertainty delivers a great amount of benefits.
We have summarized the obtained results in terms of the fourteen most critical causes of failure in a
normalized stacked-bar chart (Figure 4), showing the outcomes for all the three methods adopted.
A direct comparison between the final ranking of the 14 most critical causes of failure, in the case of fuzzy
TOPSIS and standard FMECA analyses, is finally reported in Table II.
As one can see, the analysis reveals some changes in the final outcome. As expected, the introduction of
weights for the criticality criteria partly modifies the global importance (i.e. degree of criticality) of the various
failure modes and allows a precise estimate of the priority index, discriminating among the results far more
accurately than the traditional RPN.
The discussion is concluded with a fuzzy TOPSIS sensitivity analysis and concerning the weights of the
model. In this way we can test the robustness of the obtained results.
As well underlined in Braglia11 , although the solution reported in Figure 4 reflected a possible scenario where
the chance of failure is the most important criterion, the model solution can potentially change in accordance
with shifts in analyst logic. To explore the response of the model solutions (i.e. the solution robustness) to
potential shifts in the priority of designers strategy, a series of sensitivity analyses of criteria weights can
be performed by changing the triangular fuzzy judgements of weights. As a matter of fact, every criterion is
characterized by an important degree of sensitivity, i.e. the ranking of all causes of failure dramatically changes
over the entire weight range. The problem is to control whether a few changes in the fuzzy judgement evaluations
can lead to significant modifications in the priority final ranking or not.
Considering that weights are now defined as FTNs and not as crisp values, the sensitivity analysis has been
executed as follows. First, we have changed one factor weight at a time. In this way only the main effects
have been considered. In other words, interaction effects of the changes in two or more weights have been
neglected. This simplification has been adopted considering that the introduction of the interaction effects makes
the sensitivity analysis too complex for actual applications. Nevertheless, it is necessary to note that the main
effects are generally the most important aspects in a sensitivity analysis11 .
c 2003 John Wiley & Sons, Ltd.
Copyright
1,000
Ranking
0,800
0,600
0,400
0,200
A)
Lo
w
B)
ba
Br
tte
ea
ry
ki
ng
of
I/O
C
)W
D
)C
iri
ng
ur
re
nt
E)
fe
Fu
ed
se
F)
bu
Ac
rn
tu
ed
at
or
G
br
)F
ok
ee
en
di
ng
H
)S
ca
of
bl
tw
e
ar
e
er
ro
I)
PC r
fa
ul
t
L
)
M
S
)S
ho
of
ck
tw
ar
N
e
)C
l
om oss
m
.E
O
)S
rro
ho
r
rt
P)
ci
rc
Su
ui
rg
t
e
st
re
ss
0,000
Failure Causes
1.000
0.943
0.791
0.755
0.754
0.663
0.571
0.578
0.556
0.510
0.464
0.412
0.403
0.361
FMECA RPN
(N) Communication error
(A) Low battery
(H) Software error
(O) Short circuit
(B) Breaking of I/O
(L) Shock
(C) Wiring
(M) Software loss
(E) Fuse burned out
(G) Feeding cable
(F) Actuator broken
(P) Surge stress
(D) Current feed
(I) PC fault
1.000
0.720
0.640
0.480
0.480
0.400
0.400
0.360
0.320
0.240
0.200
0.160
0.160
0.160
Second, for each fuzzy weight the range of variation has been defined between the following two extremes
(Figure 5).
Lower-bound change (LB): the mode of the FTN a2 is imposed equal to the lower limit a1 .
Upper-bound change (UB): the mode of the FTN a2 is imposed equal to the upper limit a3 .
The fuzzy TOPSIS analysis has been re-executed six times (3 weights 2 FTNs), adopting each time two
weights based on the original fuzzy configurations, and the third weight defined by its upper or lower fuzzy
configuration, alternatively. The results of the final ranking position of the fourteen most critical causes of
failure are reported in Table III.
c 2003 John Wiley & Sons, Ltd.
Copyright
1,0
Original fuzzy
triangular judgement
Membership
0,8
0,6
0,4
0,2
a1=4
0,0
1
a3=8
a2=5
10
Weight
Figure 5. Range of variation of fuzzy weight
Failure cause
(A) Low battery
(B) Breaking of I/O
(C) Wiring
(D) Current feed
(E) Fused burned out
(F) Actuator broken
(G) Feeding cable
(H) Software error
(I) PC fault
(L) Shock
(M) Software loss
(N) Communication error
(O) Short circuit
(P) Surge stress
Original ranking
WP
WD
WS
WP
WD
WS
2
7
10
14
9
12
11
1
13
6
3
5
4
8
2
7
10
14
8
12
11
1
13
6
4
3
5
9
2
7
10
14
9
12
11
1
13
6
3
5
4
8
2
8
10
14
9
12
11
1
13
6
3
5
4
7
2
8
10
14
9
12
11
1
13
6
3
5
4
7
2
7
10
14
9
12
11
1
13
6
3
4
5
8
2
7
10
14
8
12
11
1
13
6
4
3
5
9
As one can observe, the first two most critical failure causes remain the same for every weight configuration
adopted. Also, the less critical causes (i.e. the causes 14, 13, 12, 11, 10) are stable, despite the changes imposed
to the weights values. In the middle, a more variable situation can be found but, in general, no great exchange
in the final ranking position is remarkable. So, we can affirm that the final result of the fuzzy TOPSIS approach
is quite robust and reliable.
c 2003 John Wiley & Sons, Ltd.
Copyright
14
12
10
0
0
10
12
14
16
Figure 6. Scatter diagram of final ranking positions obtained after sensitivity analysis of fuzzy TOPSIS weights
Table IV. Correlation results derived from
sensitivity analysis of fuzzy TOPSIS weights
Correlation
Value
0.982
1.000
0.996
0.996
0.996
0.982
An additional graphical analysis is reported in Figure 6, concerning the scatter diagram obtained by a
comparison between the positions of the six final rankings deriving from sensitivity analyses when compared
with the original one.
It is evident from the figure that a good correlation between the different results exists. The correlation
coefficients range from the perfect correlation (i.e. 1.00) among the final rankings in the case of original weight
configurations versus lower-bound change for the D weight, to a minimum of 0.982 in the case of original weight
configurations versus the lower-bound change for the P and the S weights. All the correlation coefficients are
reported in Table IV. As one can see, the robustness of the final results is confirmed in every situation.
6. CONCLUSIONS
Due to its intrinsic ambiguity and difficult formalization, criticality analysis is a particularly complex task,
usually accomplished by maintenance staff with experience and intuition. While the traditional FMECA
approach still covers a fundamentally important requirement, it is well known that it has several weaknesses.
c 2003 John Wiley & Sons, Ltd.
Copyright
Multi-attribute decision-making approaches and fuzzy logic theory represent two powerful tools to improve the
methodology, mainly due to the capability of:
introducing a potentially larger number of failure criteria;
giving different degrees of importance to the criteria themselves;
making the analysis easier to carry out, due to the possibility of using imprecise data in the form of fuzzy
numbers.
The present paper has successfully demonstrated that the TOPSIS approach represents a multi-attribute decisionmaking tool that can be easily and efficiently integrated with fuzzy triangular numbers and used to support
criticality analyses in real industrial situations. Following this approach, fuzzy logic is only used during the
assessment of factors and relative weights of importance, in place of crisp numbers. In this way, uncertainty
concerning both quantitative data and vague and qualitative information can be used and managed in a consistent
manner during FMECA analyses. Differently from past applications, no inferential phase of fuzzy values is now
required and, as a consequence, the definition of a tedious and critical rule base is now unnecessary. In other
words, with this approach fuzzy logic is integrated in the multi-criteria decision model without needing the
definitions of a rule matrix, as generally required in common fuzzy logic applications for criticality analysis
proposed in the past literature.
The proposed case study has demonstrated the capability of the methodology to manage a criticality analysis
in an intuitive and easy manner. In addition, a sensitivity analysis of the fuzzy judgement weights has confirmed
that the proposed approach gives reasonable and robust final results (i.e. the priority ranking of the failure
causes).
Acknowledgements
The authors wish to express their great appreciation for the referees efforts and for their constructive comments,
which significantly contributed to improve the quality of this paper.
REFERENCES
1. Ben-Daya M, Raouf A. A revised failure mode and effects analysis model. International Journal of Quality and
Reliability Management 1996; 13(1):4347.
2. Gilchrist W. Modelling failure modes and effect analysis. International Journal of Quality and Reliability Management
1993; 10(5):1623.
3. Ford Motor Company. Potential failure mode and effects analysis in design (design FMECA) and for manufacturing
and assembly process (process FMECA) instruction manual. Internal Report Ford Motor Company, Detroit, USA,
September 1988.
4. Zanussi Company. FMEAGuida allanalisi del guasto. Internal Report, Direzione Innovazione Qualit`a del Prodotto,
Zanussi, Italy, 1998.
5. Pirelli Company. FMEA-FMECA: analisi delle modalit`a degli effetti e delle criticit`a dei guasti. Internal Report, Pirelli,
Milan, Italy 1988.
6. Chang C-L, Liu P-H, Wei C-C. Failure mode effects analysis using grey theory. Integrated Manufacturing Systems
2001; 12(3):211216.
7. Chang C-L, Wei C-C, Lee Y-H. Failure mode and effects analysis using fuzzy method and grey theory. Kybernetes
1999; 28(9):10721080.
8. Teng SH, Ho SY. Failure mode and effects analysis: An integrated approach for product design and process control.
International Journal of Quality and Reliability Management 1996; 13(5):826.
9. Hwang CL, Yoon K. Multiple Attribute Decision Making: Methods and Applications. Springer: New York, 1981.
10. Sankar NR, Prabhu BS. Modified approach for prioritization of failures in a system failure mode and effects analysis.
International Journal of Quality and Reliability Management 2000; 18(3):324335.
11. Braglia M. MAFMA: Multi-attribute failure mode analysis. International Journal of Quality and Reliability
Management 2000; 17(9):10171033.
c 2003 John Wiley & Sons, Ltd.
Copyright
12. Bevilacqua M, Braglia M, Gabbrielli R. Monte Carlo simulation approach for a modified FMECA in a power plant.
Quality and Reliability Engineering International 2000; 16:313324.
13. Pelaez CE, Bowles JB. Using fuzzy logic for system criticality analysis. Proceeding of the IEEE Annual Reliability
and Maintainability Symposium, Anaheim, CA, 1994; 449455.
14. Bowles JB, Pelaez CE. Fuzzy logic prioritization of failures in a system failure mode, effects and criticality analysis.
Reliability Engineering and System Safety 1995; 50:203213.
15. Moss TR, Woodhouse J. Criticality analysis revisited. Quality and Reliability Engineering International 1999; 15:117
121.
16. Braglia M, Bevilacqua M. Fuzzy modelling and analytic hierarchy processing as a means to quantify risk levels
associated with failure modes in production systems. Technology, Law and Insurance 2000; 5(34):125134.
17. Puente J, Pino R, Priore P, Fuente D. A decision support system for applying failure mode and effects analysis.
International Journal of Quality and Reliability Management 2001; 19(2):137150.
18. Triantaphyllou E, Lin C-T. Development and evaluation of five fuzzy multi-attribute decision-making methods.
International Journal of Approximate Reasoning 1995; 14:281310.
19. Driankov D, Hellendoorn H, Reinfrank M. An Introduction to Fuzzy Control. Springer: Berlin, 1993.
20. Ribeiro RA. Fuzzy multiple attribute decision making: A review and new preference elicitation techniques. Fuzzy Sets
and Systems 1996; 78:155181.
21. Law AM, Kelton WD. Simulation Modelling and Analysis. McGraw-Hill: New York, 1991.
22. Zhu Q, Lee ES. Comparison and ranking of Fuzzy Numbers, Fuzzy Regression Analysis. Omnitech Press: Warsaw,
1992; 132145.
23. Facchinetti G, Ghiselli Ricci R, Muzzioli S. Note on fuzzy triangular numbers. International Journal of Intelligent
Systems 1998; 13:613622.
24. Yager RR. On choosing between fuzzy subsets. Kybernetes 1980; 9:151154.
APPENDIX A.
Triangular fuzzy memberships can be used to perform common mathematical operations. The basic fuzzy
arithmetic operations on two triangular fuzzy memberships A = (a1 , a2 , a3 ) and B = (b1 , b2 , b3 ) are defined
as follows.
Inverse
A
1 1 1
, ,
a3 a2 a1
Addition
A + B = (a1 + b1 , a2 + b2 , a3 + b3 )
Subtraction
A B = (a1 b3 , a2 b2 , a3 b1 )
Scalar multiplication
k > 0,
k R,
k A = (k a1 , k a2 , k a3 )
Multiplication
A B = (a1 b1 , a2 b2 , a3 b3 )
Division
A
=
B
c 2003 John Wiley & Sons, Ltd.
Copyright
a1 a2 a3
, ,
b3 b2 b1
Authors biographies
Marcello Braglia graduated (with distinction) in 1988 in electronic engineering at the Politecnico di Milano.
In 1990 he obtained a grant in numerical analysis at the Universit`a degli Studi di Brescia. Since 1995 he has
been Researcher in Mechanical Technology and Production Systems at the same university. Since 1998, he has
been a member of the Industrial Plants research staff at the Universit`a di Pisa. Winner on National examination
for Associate Professor position in industrial engineering, July 2001. Winner on National examination for Full
Professor position in industrial engineering, October 2001. His research activities mainly concern equipment
maintenance, reliability, production planning, logistics, and statistical quality control. He is the author of
about a hundred technical papers published in national and international journals and conference proceedings.
In particular, several results have been published in International Journal of Quality & Reliability Management,
Reliability Engineering & System Safety, Quality & Reliability Engineering International, Technology Law
and Insurance, Production Planning & Control, International Journal of Production Research, International
Journal of Physical Distribution and Logistics Management, Logistics Information Management, Journal of
Supply Chain Management, ITOR, RAIRO, Strategic Environmental Management, Technology Management,
Journal of Industrial Technology, OMEGA, and Journal of Industrial Ecology. He is a member of ANIMP
(National Association on Industrial Plants).
Marco Frosolini graduated in 2001 in mechanical engineering at the Universit`a di Pisa. In 2001 he began his
PhD studies in Mechanical Engineering at the same university. His research activities mainly concern equipment
maintenance and reliability. Some of his results have been published in international conferences.
Roberto Montanari graduated (with distinction) in 2000 in mechanical engineering at the Universit`a degli
Studi di Parma. Since 2001, he has been a member of the Industrial Plants research staff at the same university.
His research activities mainly concern equipment maintenance, power plants, and logistics. Some of his results
have been published in national journals and international conferences. He is a member of ANIMP (National
Association on Industrial Plants).