You are on page 1of 12

Research Article

(wileyonlinelibrary.com) DOI: 10.1002/qre.1930 Published online 27 November 2015 in Wiley Online Library

A Multi-attribute Reliability Allocation Method


Considering Uncertain Preferences
Tie Chen,a Songlin Zheng,a,b*† Haitao Liaoc and Jinzhi Fenga,b
In reliability allocation, certain reliability values are assigned to subsystems and components to achieve the required system
reliability. One big challenge in solving such reliability-based design problems is how to handle the uncertain preferences of
a decision maker on multiple attributes of interest. In this paper, we propose a new ordered weighted averaging (OWA)
method based on an analytic hierarchy process to address the decision maker’s uncertain preferences in reliability allocation.
In the proposed OWA operator, a bi-objective mathematical programming model considering both maximal entropy and
minimal variance is transformed into a single-objective mathematical programming model using an ideal-point method.
The maximum entropy minimal variance OWA operator takes full advantage of available information and avoids
overestimating the decision maker’s preferences. A detailed computational procedure is presented to facilitate the
implementation of the proposed method in practice. An illustrative example about the powertrain of fuel cell vehicles is
provided to demonstrate the effectiveness of this method in handling multiple attributes with uncertain preferences in
reliability allocation. Copyright © 2015 John Wiley & Sons, Ltd.

Keywords: reliability allocation; MEMV-OWA operator; uncertainty; AHP; powertrain

1. Introduction
eliability should be considered throughout the life cycle of a product, which has become one of the most common concerns of

R product designers, manufacturers, and users. In particular, reliability allocation is one of the critical aspects in reliability-based
design for a variety of products. The process is to apportion the given reliability index into each subsystem and component with
trade-offs among multiple attributes. To this end, it is important to allocate a reasonable and feasible reliability level to each
subsystem and component. So far, a considerable number of techniques have been developed for reliability allocation. These
techniques belong to two different categories: (i) optimal reliability allocation that focuses on the trade-offs between system
performance and resources and (ii) a weighting method that allocates the system reliability by assigning the weights to multiple
attributes.
In the first category, two objectives often considered are to achieve the highest system reliability under a budgetary constraint and
to minimize the cost of the system while satisfying the required system reliability. Kuo et al.1 provided a comprehensive review on
such reliability optimization problems and their solution methodologies. Numerous reliability optimization problems have been
studied, including traditional reliability redundancy allocation, life percentile optimization, multistate system optimization, and
multi-objective optimization. For example, Coit and Konak2 extended the max–min heuristic method considering multiple weighted
objectives. Yeh et al.3 introduced an artificial bee colony algorithm to solve the reliability redundancy allocation problem, in which the
number of redundant components and their reliability are handled simultaneously in maximizing the system reliability. Sahoo et al.4
formulated four different multi-objective optimization problems with the help of interval-valued reliability and solved the problems
using a genetic algorithm. Mourelatos et al.5 utilized a user-supplied relationship between reliability and cost to maximize the system
reliability and minimize the repair cost for the system. Khalili-Damghani et al.6 solved the multi-objective redundancy allocation
problems by a decision support system and then proposed an efficient ε-constraint method to generate nondominated solutions
on the Pareto front. Liu et al.7 employed an improved genetic algorithm to reallocate the reliability of the fuel cell vehicle (FCV) based
on cost optimization. It is worth pointing out that the feasibility of these methods relies on the analytical relationship between the
reliability of each subsystem and the cost. However, sometimes, it is difficult, if not impossible, to determine the relationship in the
early design stage of a product.

a
School of Mechanical Engineering, University of Shanghai for Science and Technology, Shanghai, 200093, China
b
CMIF Key Lab for Automotive Strength and Reliability Evaluation, Shanghai, 200093, China
c
Department of Industrial Engineering, University of Arkansas, Fayetteville, AR 72701, USA
2233

*Correspondence to: Songlin Zheng, School of Mechanical Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China.

E-mail: songlin_zheng@126.com

Copyright © 2015 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2016, 32 2233–2244
T. CHEN ET AL.

Regarding the weighting method, the Advisory Group of Reliability of Electronic Equipment (AGREE) method8 is based on the
complexity and importance of components to quantify the relationship between component failure and system failure. The
Aeronautical Radio Inc. (ARINC) method9 takes the similarity between the new and old systems and uses a proportional allocation
method to allocate the failure rates to subsystems. Indeed, the AGREE and ARINC methods address reliability allocation slightly
differently, and they do not take other attributes into consideration. As a technical extension, researchers have combined different
attributes in reliability allocation. Falcon et al.10 proposed a reliability allocation methodology called integrated factors method, in
which the criticality, complexity, functionality, and effectiveness were employed. Moreover, Wang et al.11 evaluated seven attributes
in reliability allocation, that is, the frequency of failure, criticality of failure, maintainability, complexity, manufacturing technology,
working condition, and cost. With regard to the multiple attributes decision-making (MADM) problems, many methods have been
proposed to determine the reasonable weights for multiple attributes. Wu et al.12 combined fuzzy comprehensive evaluation with
fuzzy reasoning Petri net to calculate the weights of multiple attributes. Kuo et al.13 proposed an average weighting allocation
method to evaluate each attribute through a standard 1-to-10 scaling. However, because of the discrete ordinal scales and direct
multiplication of rating scores, this method may be misleading sometimes. Another useful alternative is the analytic hierarchy process
(AHP) method that derives the relative weights between multiple attributes according to a hierarchical model, which captures both
qualitative and quantitative aspects of a decision.14–16 However, the evaluation of multiple attributes in this method depends on the
decision maker’s experience and preferences, which are subjective and uncertain in many situations. Therefore, it is necessary to
develop a new approach to aggregate the decision maker’s preferences in such a decision-making process.
The ordered weighted averaging (OWA) operator is of great importance and usefulness in solving MADM problems.17 In recent
years, it has also been widely used in reliability studies. Chang et al.18 used intuitionistic fuzzy set and OWA operators to evaluate
the system reliability of an aircraft propulsion system. Chang et al.19 combined 2-tuple and the OWA operator for prioritization of
failures in a product design failure mode and effect analysis. In particular, reliability allocation can be viewed as one of the MADM
problems, in which the OWA operator can be applied. It is important to process the information aggregation of a decision maker’s
preferences based on the associated weighting vector. In the related research, Chang et al.20 introduced the concept of the maximal
entropy (ME)-OWA operator into reliability allocation. In this method, the orness measure was provided to flexibly compute the
reliability allocation value. When available information was imprecise, incomplete, or uncertain in the early design phase, the orness
measure would be useful in aggregating the decision maker’s preferences under maximum entropy. In fact, the measurement scaling
problem and multi-attribute unequal weighting problem in most traditional reliability allocation methods can be overcome.
Therefore, the proposed ME-OWA method can better help designers make reasonable decisions. Liaw et al.21 combined the
ME-OWA-based decision-making trial and evaluation laboratory method and ARINC method for reliability apportionment, in which
the indirect relationship between subsystems and the predicted failure rate were considered. This method obtains a correct and
discriminating allocation weight, so that the uncertain information of the decision maker’s experience is properly aggregated.
Researchers have proposed different OWA operators and the ways of determining the associated weights. Yager17,22–24 first
introduced the concept of the OWA operator. After that, O’Hagan25 proposed the ME-OWA operator to generate the weighting vector
under a nonlinear constraint of maximum entropy with the prescribed value of orness. Fullér and Majlender26 proposed the minimal
variance (MV)-OWA operator to determine the weighting vector with minimum variance constrained under a given orness level.
The objective of this paper is to propose a new OWA operator, called MEMV-OWA operator, which takes advantage of both the
ME-OWA operator and the MV-OWA operator. The proposed OWA operator can be used to obtain the OWA weighting vector that
minimizes the variance of weighting vector while maximizing the entropy. For clarity, the MEMV-OWA operator not only fully utilizes
the information from a decision maker’s preferences but also provides a potential way to avoid subjective bias in reliability allocation.
Moreover, to overcome the drawback of the average weighting allocation method, we propose an AHP-based method that combines
the MEMV-OWA operator in reliability allocation. To the best of our knowledge, the AHP-based MEMV-OWA reliability allocation
method has not been reported in the related literature.
The remainder of this paper is organized as follows. In Section 2, we introduce the concept of OWA operator and show the method
of using the MEMV-OWA operator for determining the weighting vector. In Section 3, we briefly introduce the average weighting
allocation method first and then present the AHP-based MEMV-OWA reliability allocation method and describe its detailed procedure.
In Section 4, an illustrative example regarding the powertrain of FCV is provided. Finally, conclusions and some final remarks are made
in Section 5.

2. Theory of ordered weighting averaging operator


2.1. Ordered weighting averaging operator definition
Definition 1. OWA operator17
An OWA operator of dimension n is a mapping F: Rn → R, which has an associated weighting vector W = (w1, w2, …, wn)T such that

w 1 þ w 2 þ … þw n ¼ 1; 0≤ w i ≤ 1; i ¼ 1; 2; …; n
with the properties
X
n
F ða1 ;; a2 …; an Þ ¼ w i bi (1)
i¼1

where bi is the ith largest element in the collection of aggregates {a1, a2, …, an }, which are arranged in descending order
2234

(i.e., b1 ≥ b2 ≥ ⋯ ≥ bn) before calculating the scalar product with the weighting vector W.

Copyright © 2015 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2016, 32 2233–2244
T. CHEN ET AL.

Definition 2. Measure of orness17


The measure of orness represents the position of the OWA operator that is closer to the andlike or orlike level. Under a given degree
of orness, the OWA operator makes all the weights as close to each other as possible, so orness is defined as
X
n
ni
ornessðW Þ ¼ wi ¼ α (2)
i¼1
n1
where orness(W) – α ∈ [0, 1]. If orness(W) is closer to zero, then the relationship between the multiple attributes shows a higher andlike
value; that is, the decision maker is maximally noncommittal. If orness(W) is closer to 1, then the relationship between the multiple
attributes shows a higher orlike level; that is, the decision maker is maximally optimistic. In particular, if wi = 1/n, i = 1, 2, …, n, then
orness(W) = 0.5; that is, the decision maker faces a moderate assessment.

Definition 3. Measure of entropy17


The measure of entropy gauges the degree of utilization of information in an uncertain environment. It is also referred to as the
measure of dispersion and is defined as
X
n
dispðW Þ ¼  w i ln w i (3)
i¼1

It is known that the dispersion of W is an actual measure of entropy and has the following properties:
1. When wi = 1 and wj = 0 (j ≠ i), the dispersion of W is minimum, and disp(W) = 0. This shows that only one attribute is considered in
the aggregation process.
2. When wi = 1/n, i = 1, …, n, the dispersion of W is maximum, and disp(W) = ln(n). This indicates that all attributes are considered in
the aggregation process.

Definition 4. Measure of variance24


The measure of variance determines the variability of weighting vector for a given level of orness and is defined as
1X n
1X n
1
D2 ðW Þ ¼ ½w i  E ðw Þ2 ¼ wi 2 2 (4)
n i¼1 n i¼1 n
To avoid overestimating a single attribute, we would like to control the variance of weighting vector in the decision-making process
involving multiple attributes.

2.2. Maximum entropy ordered weighting averaging operator


O’Hagan25 proposed the ME-OWA operator that maximizes the entropy under a constraint of orness value. Mathematically, this
approach is described by the following mathematical programming model:
X
n
maximize :  w i ln w i
i¼1
(5a)
X
n
ni
subject to : w i ¼ α; ð0 ≤ α≤ 1Þ
i¼1
n1

X
n
w i ¼ 1; 0 ≤ w i ≤ 1; i ¼ 1; 2…; n (5b)
i¼1

Fullér and Majlender27 solved this problem using the method of Lagrange multipliers to determine the optimal weighting vector
under the maximal entropy. The associated weighting vector is obtained using Eqs. (6c)–(7b):
j1 nj
ln w j ¼ ln w n þ ln w 1
n1 n1
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi (6a)
wj ¼ w nj j1
n1
1 wn

and
ððn  1Þα  nÞw 1 þ 1
wn ¼ (6b)
ðn  1Þα þ 1  nw 1

then
2235

w 1 ½ðn  1Þα þ 1  nw 1 n ¼ ððn  1ÞαÞn1 ½ððn  1Þα  nÞw 1 þ 1 (6c)

Copyright © 2015 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2016, 32 2233–2244
T. CHEN ET AL.

where the optimal value of w1 should satisfy Eq. (7b). After calculating w1, the optimal value of wn can be determined from Eq. (7a), and
then the other weights can be obtained from Eq. (6c). As an example, Table I shows the weighting vectors of ME-OWA operator
when n = 4.
2.3. Minimal variance ordered weighting averaging operator
Fullér and Majlender26 proposed the MV-OWA operator that minimizes the variance of weighting vector under an orness level
constraint. This approach can be described as
1X n
1
minimize : wi 2  2
n i¼1 n
(7a)
Xn
ni
subject to : w i ¼ α; ð0 ≤α ≤ 1Þ
i¼1
n1
Xn
w i ¼ 1; 0≤ w i ≤ 1; i ¼ 1; 2…; n (7b)
i¼1

which can be solved in an analytical way.


2.4. Maximum entropy minimal variance ordered weighting averaging operator
Based on the concept of the ME-OWA operator and MV-OWA operator, a bi-objective programming model has been built to determine
the weighting vector of the proposed MEMV-OWA operator. That is to say, the uncertain information of the decision maker’s experience
can be utilized as much as possible by maximizing the entropy. Meanwhile, minimizing the variance of the weighting vector is a potential
way to avoid overestimation of the decision maker’s preferences. The constrained mathematical programming model is shown such as
X
n
maximize :  w i ln w i (8a)
i¼1

1X n
1
minimize : wi 2  2
n i¼1 n
(8b)
X
n
ni
subject to : w i ¼ α; ð0 ≤α ≤ 1Þ
i¼1
n1
X
n
w i ¼ 1; 0≤ w i ≤ 1; i ¼ 1; 2…; n (8c)
i¼1

To solve this problem, the ideal-point method (IPM)28,29 can be employed to translate the bi-objective programming model into a
single-objective programming model. IPM evaluates the noninferior alternatives based on their absolute distances from the ideal
point that is defined as a hypothetical alternative that comes as close as possible to the optimal solution. In fact, IPM takes advantage
of overcoming some of the difficulties associated with the interdependence among multiple objectives. In the IPM, there is no need to
treat dependence as being separable, as a noninferior alternative is regarded as an inseparable bundle of objectives.
Xn
1 Xn
Suppose that there are two objectives, (i) minimize f 1 ¼ n1 w i 2  2 and (ii) minimize f 2 ¼ w i ln w i , and each has its optimal
i¼1
n i¼1
values f 01 and f 02 . The IPM is defined as
! !
      1X n
1 X
n
minimize : F  F 0  ¼ f 01  f 1  þ f 02  f 2  ¼ f 01  wi 2  2 þ f 02  w i ln w i (9)
n i¼1 n i¼1

We assume that n ≥ 3 throughout this paper, and let


! ! ! !
1X n
1 Xn Xn
ni Xn
Lðw;λ1 ; λ2 Þ ¼ f 1 
0
wi  2 þ f 2 
2 0
w i ln w i þ λ1 w i  α þ λ2 w i 1 (10)
n i¼1 n i¼1 i¼1
n1 i¼1

Table I. Weighting vector of ordered weighted averaging operator


α

wi 0.5 0.6 0.7 0.8 0.9 1


w1 0.2500 0.3474 0.4613 0.5965 0.7641 1
w2 0.2500 0.2722 0.2756 0.2520 0.1821 0
w3 0.2500 0.2133 0.1647 0.1065 0.0434 0
2236

w4 0.2500 0.1671 0.0984 0.0450 0.0104 0

Copyright © 2015 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2016, 32 2233–2244
T. CHEN ET AL.

which denotes the Lagrange function of the constrained optimization problem, where λ1 and λ2 are real numbers. Then, taking the
partial derivatives of Eq. (12a) with respect to wi, ë1, and ë2 and setting them to zero yield

∂L 2w i ni
¼  ð1 þ ln w i Þ þ λ1 þ λ2 ¼ 0 (11a)
∂w i n n1

∂L Xn
ni
¼ wi  α ¼ 0 (11b)
∂λ1 i¼1
n 1

∂L Xn
¼ w i 1 ¼ 0 (11c)
∂λ2 i¼1

For i = 1, Eq. (12b) becomes


2w 1
  ð1 þ ln w 1 Þ þ λ1 þ λ2 ¼ 0 (12a)
n
and for i = n, Eq. (12b) turns out to be
2w n
  ð1 þ ln w n Þ þ λ2 ¼ 0 (12b)
n
As a result, we have
2w 1 2w n
λ1 ¼ þ ln w 1   ln w n (13a)
n n
2w n
λ2 ¼ þ 1 þ ln w n (13b)
n
Including λ1 and λ2 into Eq. (12b) results in a single-objective model with
   
2 ni 2 i1 2
w i þ ln w i ¼ w 1 þ ln w 1 þ w n þ ln w n (14)
n n1 n n1 n

For example, when n = 4, the weighting vector of the MEMV-OWA operator can be obtained by solving Eqs. (11b) and (18), and the
results are shown in Table II.

3. Methodologies of reliability allocation


3.1. Average weighting allocation method
The average weighting allocation method is commonly used in reliability allocation for complex systems. This method consists of two
major steps: (i) rating multiple attributes by a standard scale and (ii) calculating the allocation weights.
Suppose that Wij is the jth attribute’s rating score of the ith subsystem, m is the number of attributes, and λs is the allocated system
failure rate. The multiple-attribute decision for the ith subsystem can be determined by
m
W i ¼ W i1 W i2 …W ij …W im ¼ ∏ W ij (15)
j¼1

Then, the allocated failure rate of subsystem i is given by


X
n
λi ¼ λs W i Wi (16)
i¼1

However, the average weighting allocation method evaluates multiple attributes by discrete ordinal scales, and their rating scores
are multiplied directly, which is not meaningful and sometimes misleading. For example, two subsystems are evaluated by four

Table II. Weighting vector of maximum entropy minimal variance ordered weighted averaging operator
α

wi 0.5 0.6 0.7 0.8 0.9 1


w1 0.2500 0.3466 0.4580 0.5898 0.7568 1
w2 0.2500 0.2732 0.2800 0.2619 0.1945 0
w3 0.2500 0.2138 0.1659 0.1067 0.0407 0
2237

w4 0.2500 0.1664 0.0961 0.0416 0.0080 0

Copyright © 2015 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2016, 32 2233–2244
T. CHEN ET AL.

attributes, and they have the same rating scores as 6 × 5 × 7 × 9 = 1890 and 7 × 5 × 6 × 9 = 1890. Accordingly, two subsystems will be
allocated the same failure rate according to Eq. (20), which is obviously irrational.

3.2. Analytic hierarchy process-based maximum entropy minimal variance ordered weighting averaging reliability allocation method
3.2.1. Hierarchical model for reliability allocation. To overcome the drawback of the average weighting allocation method, we
consider an AHP-based reliability allocation model for allocating reliability to subsystems. In this model, the top layer is the objective,
that is, the system reliability requirement. The intermediate layer contains multiple attributes, including the design level (D),
manufacturing technology (M), system property (S), and usage maintenance (U). The bottom layer specifies subsystem or component
objects. For example, the hierarchical model for reliability allocation of FCV powertrain is shown in Figure 1.

3.2.2. Comparison matrix. The preferences of the decision maker are often expressed using numerical values when constructing the
comparison matrix. The numerical values range from 1 to 9. Let R = (rpq) be a k × k comparison matrix, where rpq is the rating score for
the pth subsystem comparing with the qth subsystem and k is the number of subsystems. Taking the design level as an example and
comparing the pth subsystem with the qth subsystem, the evaluation criteria result in 1 = same design level, 3 = slightly lower design
level, 5 = lower design level, 7 = much lower design level, 9 = absolutely lower design level, and 2, 4, 6, and 8 = intermediate levels
(Table III). Moreover, it is easy to find that the comparison matrix has the properties: rpp = 1, rpq × rqp = 1.
After constructing the comparison matrix, we can calculate the maximal eigenvalue λmax and the corresponding normalized
eigenvector. Table IV shows the eigenvector that is the weight vector of multiple attributes.
Because of the decision maker’s different psychological states and information inputs, it is difficult to guarantee the consistency of
the comparison matrix. Therefore, it is important to calculate the consistency according to the following equations
λmax  k
CI ¼ (17)
k1
CI
CR ¼ (18)
RI
where CI is the consistency index, CR is the consistency ratio, and RI is the consistency index of a randomly generated matrix as
illustrated in Table V. The comparison matrix is of acceptable consistency, and CR should be less than 0.1. If not, the comparison matrix
should be reviewed, reconsidered, and improved until the matrix with satisfied CR is obtained. An acceptable consistency ratio helps
to ensure that the decision maker determines the priorities of a set of attributes.

Figure 1. Hierarchical model for reliability allocation of fuel cell vehicle powertrain. FCE, fuel cell engine; BMS, battery and management system; EA, electrical appliance;
EDS, electric driving system

Table III. Comparison matrix evaluation scale


rpq Evaluation criteria

1 Subsystem p, q with the same design level


3 Design level of p is slightly lower than that of q
5 Design level of p is lower than that of q
7 Design level of p is much lower than that of q
9 Design level of p is absolutely lower than that of q

Table IV. Weight vector of multiple attributes


Factors

Subsystem Design level Manufacturing technology System property Usage maintenance


Subsystem 1 d1 m1 s1 u1
Subsystem 2 d2 m2 s2 u2
⋯ ⋯ ⋯ ⋯ ⋯
2238

Subsystem k dk mk sk uk

Copyright © 2015 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2016, 32 2233–2244
T. CHEN ET AL.

Table V. The consistency index of randomly generated matrix


n
1 2 3 4 5 6 7 8

RI 0 0 0.58 0.90 1.12 1.24 1.32 1.41

3.2.3. Detailed procedure of analytic hierarchy process-based maximum entropy minimal variance ordered weighting averaging reliability
allocation. The procedure of the proposed OWA reliability allocation method contains eight steps:
Step 1: Determine the system reliability index or system failure rate.
Step 2: Build the hierarchical model for reliability allocation.
Step 3: Construct the comparison matrix R of multiple attributes.
Step 4: Calculate the maximal eigenvalue λmax and the corresponding normalized eigenvector as shown in Table IV.
Step 5: Examine whether or not the comparison matrices meet the consistency requirement. If not, go back to the step 3.
Step 6: Specify α = orness based on Table II, and then obtain the OWA weighting vector W = (w1, w2, …, wn).
Step 7: Calculate the OWA operator F of reliability factors based on Eq. (3) and the selected W = (w1, w2, …, wn).
Step 8: Calculate the failure rate of each subsystem according to

Fj
λj ¼ λs (19)
X
k
Fj
j¼1

where λj is the allocated failure rate of the jth subsystem and Fj is the OWA operator of the jth subsystem.

4. Illustrative example
Vehicle transportation is important in modernity and development. However, the greenhouse gas emissions from petroleum-based
internal combustion engine (ICE) vehicle have caused air pollution and the global climate change. Moreover, more and more ICE
vehicles depending on gasoline/diesel will lead to a critical energy crisis. Therefore, it is urgent to solve these problems to improve
the urban air quality and reduce the petroleum consumption. With the advantages of high energy efficiency, zero tailpipe emissions,
and great driving performance, hydrogen FCV has been regarded as one of the most promising renewable energy vehicles.30,31
Over the past decade, significant progress on the development of FCVs has been achieved. Several governments and automakers
have started their programs and invested in developing and demonstrating FCVs and the associated infrastructures. The USA
launched the National Fuel Cell Electric Vehicle Learning Demonstration from 2004 to 2011. It deployed 183 FCVs traveling 3.6 million
miles and built 25 hydrogen fueling stations.32 Japan started the Japan Hydrogen and Fuel Cell Project that was a 3-year program to
prove their FCV technologies in 2002.31 China also made efforts on FCV demonstration during the 2008 Beijing Olympics and the 2010
Shanghai World Exposition.33 By analyzing the experiences and test data obtained from these FCV demonstration activities, it can be
found that FCV’s reliability, durability, and cost are still critical barriers to commercialization.
Recently, many studies have been conducted on the analysis and modeling of reliability and durability for fuel cell engine (FCE),34–37
auxiliary battery,38,39 electric driving system (EDS),40,41 and direct current/direct current convertor.42,43 In an attempt to enhance their
reliability and durability, these subsystems and components are considered as a whole, called the FCV powertrain. The FCV powertrain
has complex structures including electric, electronic, and mechanical parts. The failure modes of the FCV powertrain have a huge
difference from those of ICE vehicles. Moreover, the high failure rate of the FCV powertrain is a primary reason for the low reliability
of the entire FCV. As a result, it is critical to investigate a reasonable reliability allocation method that considers multiple attributes
and the uncertainties in reliability-based design for the FCV powertrain.

4.1. Reliability test data of fuel cell vehicle powertrain


In this section, the field test data of 28 sample FCVs are utilized. The mileage values of these 28 FCVs are shown in Figure 2. The
accumulative mileage is 833,102.84 km, with a maximum mileage of 34,358.70 km and a minimum mileage of 21,742.70 km. The mean
value is 29,753.67 km, and the standard deviation is 3366.63 km.
The FCV powertrain is mainly composed of FCE, battery and management system (BMS), electrical appliance (EA), and EDS. It can
be modeled as a series system. In other words, if any subsystem fails, the powertrain will break down. Assuming that the lifetime
distribution of powertrain can be described as an exponential distribution, the subsystem’s failure rate can be calculated as
rj
λj ¼ (20)
T
where λj is the failure rate of the jth subsystem, rj is the failure number of the jth subsystem, and T is the accumulative mileage. The
2239

failure rate of each subsystem is given in Table VI.

Copyright © 2015 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2016, 32 2233–2244
T. CHEN ET AL.

Figure 2. Mileage of 28 sample fuel cell vehicles (FCVs)

Table VI. Statistical failure rate of fuel cell vehicle powertrain


Subsystem FCE BMS EA EDS
Failure number 104 51 56 31
Failure rate (104 km1) 1.2484 0.6122 0.6722 0.3721
FCE, fuel cell engine; BMS, battery and management system; EA, electrical appliance; EDS, electric driving system.

4.2. Reliability allocation of fuel cell vehicle powertrain

The failure rate index of powertrain is set to be λs ¼ 2:5104 km1 , and the hierarchical model for reliability allocation of the FCV
powertrain is shown in Figure 1. Senior experts in the field were invited to construct the comparison matrices based on the available
reliability data and their personal experience. Using the method described in Section 3.2.2, the following comparison matrices and
weight vectors for four attributes are obtained:
2 3
1 2 3 5
6 7
6 1=2 1 2 37
6 7
RD ¼ 6 7
6 1=3 1=2 1 27
4 5
1=5 1=3 1=2 1
w D ¼ ð0:4824; 0:2718; 0:1575; 0:0883Þ
2 3
1 4 3 5
6 7
6 1=4 1 1=2 2 7
6 7
RM ¼ 6 7
6 1=3 2 1 3 7
4 5
1=5 1=2 1=3 1
w M ¼ ð0:5423; 0:1397; 0:2333; 0:0847Þ
2 3
1 3 4 4
6 7
6 1=3 1 3 2 7
6 7
RS ¼ 6 7
6 1=4 1=3 1 1=2 7
4 5
1=4 1=2 2 1
w S ¼ ð0:5249; 0:2388; 0:0930; 0:1433Þ
2 3
1 4 2 3
6 7
6 1=4 1 1=3 1=2 7
6 7
RU ¼ 6 7
6 1=2 3 1 2 7
4 5
1=3 2 1=2 1
w U ¼ ð0:4658; 0:0960; 0:2771; 0:1611Þ

After checking the consistency of the comparison matrices, we can see that all the comparison matrices meet the consistency
requirement. The weight vector of multiple attributes, λmax, and consistency index are presented in Table VII.
It can be seen from definition 2 that the prescribed value of orness in reliability allocation is determined based on the decision
2240

maker’s uncertain preferences. If the decision maker is maximally optimistic on the reliability of subsystems, then the orness value
is set to be 1. Conversely, if the decision maker is in a moderate situation, then orness is set to be 0.5. The calculated weighting vector

Copyright © 2015 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2016, 32 2233–2244
T. CHEN ET AL.

Table VII. The weight vector of multiple attributes, λmax, and consistency index
Manufacturing System Usage
Factors Design level technology attribute maintenance
Weight vector of multiple attributes 0.4824 0.5423 0.5249 0.4658
0.2718 0.1397 0.2388 0.096
0.1575 0.2333 0.093 0.2771
0.0883 0.0847 0.1433 0.1611
λmax 4.0145 4.1126 4.0878 4.031
CI 0.0048 0.0375 0.0293 0.0103
CR 0.0054 0.0417 0.0325 0.0115

of MEMV-OWA under the specific orness is presented in Table II, which can be used to aggregate the information of multiple
attributes for subsystems.
Because the decision maker’s preferences on the reliability of subsystems are moderately optimistic, the orness measure is set to
be 0.8, and the weighting vector can be obtained from Table II as W = (w1, w2, w3, w4) = (0.5898, 0.2619, 0.1067, 0.0416). The
MEMV-OWA operator of FCE F1 is then calculated as
X
4
F 1 ðd1 ; m1 ; s1 ; u1 Þ ¼ w i b1i
i¼1
¼ 0:58980:5423 þ 0:26190:5249 þ 0:10670:4824 þ 0:04160:4658 ¼ 0:5282
where b1i is the ith largest element in the vector (d1, m1, s1, u1). In the same way, the OWA operators of BMS, EA, and EDS are obtained
as F2 = 0.2417, F3 = 0.2452, and F4 = 0.1455. Based on Eq. (19), the failure rates of the subsystems are obtained as

λ1 ¼ 1:1377104 km1 ; λ2 ¼ 0:5207104 km1 ; λ3 ¼ 0:5282104 km1 ;


1
λ4 ¼ 0:313410 4
km

4.3. Comparison of different reliability allocation methods


Based on Eqs. (19) and (20) and Table VII, using the average weighting allocation method, the allocated failure rates of subsystems are
obtained as
λ1 ¼ 1:3640104 km1 ; λ2 ¼ 0:8687104 km1 ; λ3 ¼ 0:1943104 km1 ;
1
λ4 ¼ 0:073010 4
km

To compare the effectiveness of different reliability allocation methods, two requirements should be satisfied: (i) the sum of the

X
4
allocated failure rates of subsystems is not larger than the system failure rate index, λ ≤λs and (ii) the allocated failure rate of each
j¼1 j
subsystem should not be higher than the current failure rate. The comparison of the failure rate allocation results is shown in Figure 3.
One can see from Figure 3 that the average weighting method assigns unreasonable failure rates to FCE and BMS, which are much
higher than their failure rates currently demonstrated in the test. Meanwhile, the EDS has been allocated a very low failure rate that
will require a huge investment in technology improvement.
With regard to the proposed AHP-based MEMV-OWA reliability allocation method, the allocation result fulfills both of the two
requirements. Without making the failure rate of any subsystem too low or too high, the new method is quite useful in allocating

2241

Figure 3. Failure rate comparison of FCV powertrain. FCV, fuel cell vehicle; OWA, ordered weighting averaging

Copyright © 2015 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2016, 32 2233–2244
T. CHEN ET AL.

Table VIII. Typical failure modes analysis of fuel cell vehicle powertrain
Subsystem Failure mode Failure reason Improve difficulty
FCE 1. Short-circuit risk of fuel cell stake Design, manufacture, usage Hard
2. Failure of air pump and water pump Manufacture, usage Medium
3. Voltage of single cell is less than the lower limit Manufacture, usage Easy
BMS 1. SOC is extremely low Design Hard
2. SOC is too high and too low Design Medium
EA 1. Sensor failure Design, manufacture, usage Medium
2. CAN communication malfunction Design Medium
EDS 12 V under-voltage(11 V) Usage Easy
FCE, fuel cell engine; BMS, battery and management system; EA, electrical appliance; EDS, electric driving system; SOC, state of charge;
CAN, controller area network.

the failure rate to each subsystem in reliability improvement. Indeed, the failure rates of FCE, BMS, EA, and EDS are reduced by 8.86%,
14.94%, 27.26%, and 18.73%, respectively.

4.4. Analysis of reliability allocation results


Based on the typical failure modes of FCV powertrain as shown in Table VIII, the feasibility of the proposed reliability allocation
method is to be proven, and the direction of technological improvement needs to be studied.
Some failure modes can be considered as easy-to-eliminate failures. For example, ‘12 V under-voltage (11 V)’ may be caused by a
precharge loop not being cut off or a long-term operation without recharging. The failure that ‘voltage of single cell is less than the
lower limit’ is related to the membrane electrode stressed nonuniformly and damages after assembly. By paying attention in the field
and manufacturing, these failures can be effectively controlled.
For medium-eliminated failures, more attention is needed in design, manufacturing, and usage. For example, ‘failure of air pump
and water pump’ is possibly caused by the drive short-axis cracked or material defect, or leakage failure caused by the poor assembly
quality. The failure that ‘state of charge (SOC) is too high and too low’ is caused by the insufficient control accuracy and response
speed of SOC estimation algorithm and imperfect hardware circuit, which make it difficult to guarantee the rational allocation of
energy. ‘Controller area network communication malfunction’ is due to the control system that suffers from serious electromagnetic
interference. It is necessary to reduce the interference by adopting the electromagnetic compatibility design method, such as
grounding, isolation, filtering, and printed circuit board design.
For ‘short-circuit risk of fuel cell stake’ of FCE, it is dangerous but is hard to improve. This failure mode is attributed to the poor
anticorrosion performance of the bipolar plate and bad thermostability of the proton exchange membrane. Indeed, it is critical to
develop high-performance fuel cell materials to improve the reliability and durability of FCE. Moreover, with FCE as a key component
of FCV, its failures have a correlation with other subsystems. Therefore, if the reliability of FCE is improved, other subsystems may be
improved as well.

5. Conclusions
This paper contributes to the research on reliability allocation by proposing a new AHP-based MEMV-OWA operator method that
takes into account the decision maker’s uncertain preferences. We develop bi-objective programming models for obtaining the
weighting vector that minimizes the variance from the equal weights while maximizing the entropy. This method provides a more
unbiased way for aggregating the decision maker’s preferences in reliability allocation. A case study on the FCV powertrain is
provided to demonstrate the use of the method in practice. By comparing the proposed method with the averaging weighted
allocation method, one can see that the weighting vector of multiple attributes aggregated via MEMV-OWA calculation reduces
the variances of the decision maker’s preferences. This is a potential way to make the reliability allocation results more unbiased.
In our future research, we will focus on the method of evaluating multiple attributes that can reduce the subjectivity of a decision
maker’s preferences. In the meantime, we will extend the AHP-based MEMV-OWA method to systems with more complex system
configurations.

Acknowledgements
This work was supported by the National Natural Science Foundation of China (grant no. 51375313), the National High
Technology Research and Development Program of China (2011AA11A265 and 2012AA110701), the U.S. National Science
Foundation (grant no. CMMI-1238304) the Key Project of Shanghai Science and Technology Commission (10JC1411600), Shanghai
2242

Leading Academic Discipline Project (grant no. J50503), and the Innovation Fund Project for Graduate Student of Shanghai (grant
no. JWCXSL1401).

Copyright © 2015 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2016, 32 2233–2244
T. CHEN ET AL.

References
1. Kuo W, Wan R. Recent advances in optimal reliability allocation. Chapter 1. In Computational Intelligence in Reliability Engineering, Gregory L (ed).
Springer Berlin Heidelberg: New York, 2007.
2. Coit DW, Konak A. Multiple weighted objectives heuristic for the redundancy allocation problem. IEEE Transactions on Reliability 2006; 55:551–558.
3. Yeh WC, Hsieh TJ. Solving reliability redundancy allocation problems using an artificial bee colony algorithm. Computers & Operations Research
2011; 38:1465–1473.
4. Sahoo L, Bhunia AK, Kapur PK. Genetic algorithm based multi-objective reliability optimization in interval environment. Computers & Industrial
Engineering 2012; 62:152–160.
5. Mourelatos ZP, Li J, Pandey V, Singh A, Castanier M, Lamb D. A Simulation and Optimization Methodology for Reliability of Vehicle Fleets
(no. TARDEC-21373). Oakland University: Rochester Hills, MI, 2011.
6. Khalili-Damghani K, Abtahi AR, Tavana M. A decision support system for solving multi-objective redundancy allocation problems. Quality, Reliability
Engineering International 2014; 30:1249–1262.
7. Liu XT, Zheng SL, Feng JZ, Chen T. Reliability reallocation for fuel cell vehicles based on genetic algorithm. Quality, Reliability Engineering
International 2014. DOI: 10.1002/qre.1685.
8. Advisory Group of Reliability of Electronic Equipment (AGREE). Reliability of Military Electronic Equipment. Office of the Assistant Secretary of
Defense Research and Engineering: Washington, DC, 1957.
9. Von Alven WH. Reliability Engineering. Prentice Hall: Englewood Cliff, NJ, 1964.
10. Falcone D, Silvestri A, di Bona G. Integrated factors method (IFM): a new reliability allocation technique. Proceedings of Software Engineering and
Application 2002; 4–6. Cambridge, MA.
11. Wang Y, Yam R, Zuo MJ, Tse P. A comprehensive reliability allocation method for design of CNC lathes. Reliability Engineering & System Safety 2006;
72:247–252.
12. Wu JN, Yan SZ, Xie LY, Peng G. Reliability apportionment approach for spacecraft solar array using fuzzy reasoning Petri net and fuzzy
comprehensive evaluation. Acta Astronautica 2012; 76:136–144.
13. Kuo HE. Reliability assurance: application for engineering and management (2nd edn). Chinese Society for Quality 1999; 3:16–23.
14. Satty TL. The Analytic Hierarchy Process. McGraw-Hill: New York, 1980.
15. Cheng EW, Li H. Information priority-setting for better resource allocation using analytic hierarchy process (AHP). Information Management &
Computer Security 2001; 9:61–70.
16. Vaidya OS, Kumar S. Analytic hierarchy process: an overview of applications. European Journal of Operational Research 2006; 169:1–29.
17. Yager RR. On ordered weighted averaging aggregation operators in multicriteria decision making. IEEE Transactions on Systems, Man, and
Cybernetics 1988; 18:183–190.
18. Chang KH, Cheng CH, Chang YC. Reliability assessment of an aircraft propulsion system using IFS and OWA tree. Engineering Optimization 2008;
40:907–921.
19. Chang KH, Wen TC. A novel efficient approach for DFMEA combining 2-tuple and the OWA operator. Expert Systems with Applications 2010;
37:2362–2370.
20. Chang YC, Chang KH, Liaw CS. Innovative reliability allocation using the maximal entropy ordered weighted averaging method. Computers &
Industrial Engineering 2009; 57:1274–1281.
21. Liaw CS, Chang YC, Chang KH, Chang TY. ME-OWA based DEMATEL reliability apportionment method. Expert Systems with Applications 2011;
38:9713–9723.
22. Yager RR. Families of OWA operators. Fuzzy Sets and Systems 1993; 59:125–148.
23. Yager RR. Aggregation operators and fuzzy systems modeling. Fuzzy Sets and Systems 1994; 67:129–145.
24. Yager RR. On the inclusion of variance in decision making under uncertainty. International Journal of Uncertainty, Fuzziness and Knowledge-Based
Systems 1996; 4:401–419.
25. O’Hagan M. Aggregating template or rule antecedents in real-time expert systems with fuzzy set logic. Proceedings of 22nd Annual IEEE Asilomar
Conference in Signals, Systems and Computers 1988; 681–689. Pacific Grove, CA.
26. Fullér R, Majlender P. On obtaining minimal variability OWA operator weights. Fuzzy Sets and Systems 2003; 136:203–215.
27. Fullér R, Majlender P. An analytic approach for obtaining maximal entropy OWA operator weights. Fuzzy Sets and Systems 2001; 124:53–57.
28. Zeleny M. Multiple Criteria Decision Making. McGraw Hill: New York, 1982.
29. Deb K. Multi-objective Optimization Using Evolutionary Algorithms. Wiley: New York, 2001.
30. Ehsani M, Gao Y, Emadi A. Modern Electric, Hybrid Electric, and Fuel Cell Vehicles: Fundamentals, Theory, and Design (2nd edn). CRC press: Boca
Raton, FL, 2009.
31. Haslam GE, Jupesta J, Parayil G. Assessing fuel cell vehicle innovation and the role of policy in Japan, Korea, and China. International Journal of
Hydrogen Energy 2012; 37:14612–14623.
32. Wipke K, Sprik S, Kurtz J, Ramsden T, Ainscough C, Saur G. National Fuel Cell Electric Vehicle Learning Demonstration Final Report. Contract, 303,
275-3000. National Fuel Cell Electric Vehicle Learning Demonstration Final Report. National Renewable Energy Laboratory (NREL): Golden, CO, 2012;
NREL/TP-5600-54860.
33. Dixon RK, Wang X, Wang MQ, Wang J, Zhang Z. Development and demonstration of fuel cell vehicles and supporting infrastructure in China.
Mitigation and Adaptation Strategies for Global Change 2011; 16:775–789.
34. Wu J, Yuan XZ, Martin JJ, Wang H, Zhang J, Shen J, Wu S, Merida W. A review of PEM fuel cell durability: degradation mechanisms and mitigation
strategies. Journal of Power Sources 2008; 184:104–119.
35. Yuan XZ, Li H, Zhang S, Martin J, Wang H. A review of polymer electrolyte membrane fuel cell durability test protocols. Journal of Power Sources
2011; 196:9107–9116.
36. Zhang G, Kandlikar SG. A critical review of cooling techniques in proton exchange membrane fuel cell stacks. International Journal of Hydrogen
Energy 2012; 37:2412–2429.
37. Cerri I, Nagami T, Davies J, Mormiche C, Vecoven A, Hayden B. Innovative catalyst supports to address fuel cell stack durability. International Journal
of Hydrogen Energy 2013; 38:640–645.
38. Schaltz E, Khaligh A, Rasmussen PO. Influence of battery/ultracapacitor energy-storage sizing on battery lifetime in a fuel cell hybrid electric vehicle.
IEEE Transactions on Vehicular Technology 2009; 58:3882–3891.
39. Fadel A, Zhou B. An experimental and analytical comparison study of power management methodologies of fuel cell–battery hybrid vehicles.
Journal of Power Sources 2011; 196:3271–3279.
40. Lu B, Sharma SK. A literature review of IGBT fault diagnostic and protection methods for power inverters. IEEE Transactions on Industry Applications
2009; 45:1770–1777.
2243

41. Rodríguez-Blanco MA, Claudio-Sanchez A, Theilliol D, Vela-Valdés LG, Sibaja-Terán P, Hernández-González L, Aguayo-Alquicira J. A failure-detection
strategy for IGBT based on gate-voltage behavior applied to a motor drive system. IEEE Transactions on Industrial Electronics 2011; 58:1625–1633.

Copyright © 2015 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2016, 32 2233–2244
T. CHEN ET AL.

42. Guilbert D, Gaillard A, N’Diaye A, Djerdir A. Energy efficiency and fault tolerance comparison of DC/DC converters topologies for fuel cell electric
vehicles. Transportation Electrification Conference and Expo (ITEC) 2013; 1–7. IEEE.
43. Hegazy O, Van Mierlo J, Lataire P. Analysis, modeling, and implementation of a multidevice interleaved DC/DC converter for fuel cell hybrid electric
vehicles. IEEE Transactions on Power Electronics 2012; 27:4445–4458.

Authors' biographies
Tie Chen is a PhD candidate at the University of Shanghai for Science and Technology, Shanghai, China. He received his BS degree
(2011) from the University of Shanghai for Science and Technology. His research interests focus on statistical reliability analysis and
reliability-based design in automobile.
Songlin Zheng is a Professor of Automotive Engineering and Director of the Machinery Industry Key Laboratory for Mechanical
Strength and Reliability Evaluation of Auto Chassis Components at the University of Shanghai for Science and Technology, Shanghai,
China. He received his PhD degree in Automotive Engineering in 1999 from Jilin University. He also received MS degrees in
Automobile and Tractor in 1986 from Jilin University. His research areas focus on vehicle load and strength characteristics analysis,
vehicle accelerated testing, and reliability and durability evaluation technique. He is a member of the Mechanical Vibration and
Impact Standardization Technical Committee of China and the Chinese Materials Research Society Fatigue Committee.
Haitao Liao is currently a Professor and Hefley Endowed Chair of Logistics and Entrepreneurship in the Industrial Engineering
Department at The University of Arkansas (UofA), Fayetteville. He is also the Director of Reliability & Intelligent Systems Engineering
(RISE) Laboratory at UofA. He received his PhD degree in Industrial and Systems Engineering from Rutgers University, New Jersey. He
also received MS degrees in Industrial Engineering and Statistics from Rutgers University and a BS degree in Electrical Engineering
from Beijing Institute of Technology. His research interests include modeling of accelerated testing, maintenance models and
optimization, service parts logistics, prognostics, and probabilistic risk assessment. He is a recipient of the National Science
Foundation CAREER Award in 2010, the winner of 2010 and 2013 William A. J. Golomski Award, 2013 IIE QCRE Track Best Paper Award,
and 2015 SRE Stan Ofsthun Best Paper Award. He is a member of IIE, INFORMS, IEEE, and SRE. He is an Associate Editor of IEEE
Transactions on Reliability, and Journal of Quality Technology.
Jinzhi Feng is an Associate Professor of the University of Shanghai for Science and Technology Shanghai, China. She received her PhD
degree (2006) in Automotive Engineering from the Shanghai Jiaotong University, Shanghai, China. Her research area is focused on
structural strength of automobile and reliability design theory.
2244

Copyright © 2015 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2016, 32 2233–2244

You might also like