Reliability Engineering and System Safety 133 (2015) 11–21

Contents lists available at ScienceDirect

Reliability Engineering and System Safety

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

Design of SCADA water resource management control center

by a bi-objective redundancy allocation problem and particle
swarm optimization
Ali Dolatshahi-Zand, Kaveh Khalili-Damghani n
Department of Industrial Engineering, South Tehran Branch, Islamic Azad University, Tehran, Iran

art ic l e i nf o a b s t r a c t

Article history: SCADA1 is an essential system to control critical facilities in big cities. SCADA is utilized in several sectors
Received 14 January 2014 such as water resource management, power plants, electricity distribution centers, traffic control
Received in revised form centers, and gas deputy. The failure of SCADA results in crisis. Hence, the design of SCADA system in
11 July 2014
order to serve a high reliability considering limited budget and other constraints is essential. In this
Accepted 26 July 2014
Available online 12 August 2014
paper, a bi-objective redundancy allocation problem (RAP) is proposed to design Tehran's SCADA water
resource management control center. Reliability maximization and cost minimization are concurrently
Keywords: considered. Since the proposed RAP is a non-linear multi-objective mathematical programming so the
SCADA exact methods cannot efficiently handle it. A multi-objective particle swarm optimization (MOPSO)
Multi-objective redundancy allocation
algorithm is designed to solve it. Several features such as dynamic parameter tuning, efficient constraint
problem
handling and Pareto gridding are inserted in proposed MOPSO. The results of proposed MOPSO are
Meta-heuristics
Multi-objective particle swarm compared with an efficient ε-constraint method. Several non-dominated designs of SCADA system
optimization are generated using both methods. Comparison metrics based on accuracy and diversity of Pareto front
ε-Constraint method are calculated for both methods. The proposed MOPSO algorithm reports better performance. Finally,
Adaptive grid in order to choose the practical design, the TOPSIS algorithm is used to prune the Pareto front.
& 2014 Elsevier Ltd. All rights reserved.

1. Introduction considering several constraints on budget, volume, and weight is

SCADA systems are used to monitor, control, and save the
information about quantity and quality of water at pump stations,
water treatment plant, reservoirs and distribution networks in
water resource management in big cities. These tasks are not
trivial in big cities in which there are lots of stations. RAP has been
utilized in many parts of SCADA systems, such as radio modems
and servers, to increase the reliability.
SCADA control center consists of a set of hardware, software
and communication networks which control the industrial
processes through remote data. Stability and high reliability in
intelligent systems to monitor key installations in major cities and
to restore order, prosperity, and health are very important issues
controlled by SCADA. Failure and downtime of SCADA system can
be catastrophic. Failure of any part of SCADA system can lead to a
disruption in the process control [2,3].Therefore, a proper design
of SCADA system to enhance the reliability of the control center
n
Corresponding author.
E-mail addresses: adolatshahi@yahoo.com (A. Dolatshahi-Zand),
k_khalili@azad.ac.ir, kaveh.khalili@gmail.com (K. Khalili-Damghani).
URL: http://kaveh-khalili.webs.com (K. Khalili-Damghani).
1
Supervisory Control And Data Acquisition: SCADA.
essential. Enhancement of reliability has mainly two options: (1)
using more qualified components which results in higher cost of
the designed system, and (2) using RAP to assign redundant
component using a reasonable cost. Due to these facts, this paper
introduces the RAP as a methodology used to increase reliability of
SCADA systems. The classic form of RAP was proposed by Misra
and Ljubojevic [23]. The RAP was presented to be an NP-hard
problem by Chern [5].
In this paper a multi-objective version of RAP model is
proposed. The proposed RAP model has two objectives called
reliability and cost and several constraints on minimum and
maximum number of allowed components in each sub-system,
cost, weight, and volume of the system. The proposed RAP model
is customized for the design of SCADA in water resource manage-
ment in a real case study. As the problem is hard to solve optimally
using exact algorithms a multi-objective meta-heuristic algorithm,
called the multi-objective particle swarm optimization (MOPSO)
algorithm is proposed. The proposed MOPSO generates a set of
non-dominated solutions on Pareto front. These non-dominated
solutions are associated with different designs of SCADA system
considering both objective functions and all constraints of the
systems. The result of proposed MOPSO is compared with the
result of an efficient algorithm, called epsilon-constraint. Several

http://dx.doi.org/10.1016/j.ress.2014.07.020
0951-8320/& 2014 Elsevier Ltd. All rights reserved.
12 A. Dolatshahi-Zand, K. Khalili-Damghani / Reliability Engineering and System Safety 133 (2015) 11–21

metrics are introduced to compare the accuracy and diversity of
re-generated Pareto front.
The remaining parts of the paper are organized as follows.
In Section 2, a brief literature of related past works is reviewed. In
Section 3, the mathematical formulation of proposed RAP model is
presented. The proposed MOPSO algorithm is also presented in
Section 3. In Section 4, the real case study of SCADA system for
water management resource is presented. The proposed MOPSO
and an efficient exact algorithm are applied on the real case study
in Section 4 and the results are discussed and compared. In Section
5, conclusions and further research directions are introduced.
2. Literature of past works
Swarm intelligence is one of the artificial intelligence techni-
ques that studies the group behavior in decentralized self-
organized agents. Examples for such systems can be found in
nature as ant colony, flock, herd animals, bacteria, and fishes.
However, the principle of particle swarm optimization method
(PSO) is relatively new [9], but since this method is the origin of
biological research and simulating the behavior of living animals,
there are several successful applications of this method in the
literature [13–16]. The mechanism used to search the solution
space in the PSO differs from the evolutionary computations. The
simplicity and the applicability of the PSO method have added to
the popularity of this method for solving a large number of
engineering and management optimization problems [14–16].
In some real cases, the PSO method may suffer from the
problem of premature convergence [28]. In order to overcome
this problem, the MOPSO method makes use of smart techniques
to maintain the level of diversity in the swarm population, thereby
maintaining a good balance between the exploration and exploita-
tion phenomena and preventing premature convergence [14–16].
One of a those smart techniques is gridding. The adaptive grid is used
by the Pareto Archive Evolutionary Strategy (PAES) to maintain
diversity [7].
A series–parallel system is basically characterized through a
predefined number of sub-systems which are connected serially.
Multiple component choices and redundancy levels are available
to connect in parallel for each sub-system. A given component
may have a binary-state or a multi-state in RAP [21,29]. In binary-
state RAP, the problem of a proper structure can be handled by
increasing the reliability of components or supplying parallel
redundant components at some stages.
Konak et al. [20] represented an overview and tutorial describ-
ing genetic algorithms (GA) developed specifically for problems
Due to aforementioned literature of past works and our best
knowledge PSO reported relatively better performance for hand-
ling RAP in different settings [16]. These facts as well as simple
implementation and software coding of PSO in comparison with
other existing procedures persuade us to use PSO in order to
handle a real life problem, called SCADA.
3. Redundancy allocation problem (RAP) and proposed bi-
objective RAP
In this section, the single objective RAP is described. Then the
proposed bi-objective RAP is illustrated and customized for SCADA
systems.
3.1. Single objective RAP
The following notations, indices, parameters, and decision
variables are introduced for RAP:
m: number of sub-systems;
i: index of sub-systems, i¼ 1,2,…,m;
j: index of types of components in each sub-systems, j¼1,2,…,
n;
r ij : reliability of component j in sub-system I;
cij : cost of component j in sub-system i;
wij : weight of component j in sub-system i;
Rs : overall reliability of the series 2 parallel system;
C 0 : allowed cost of system;
W 0 : allowed weight of system;
ai : number of available component choices for sub-system i;
xij : quantity of component j used in sub-system i;
ni : total number of components used in sub-system i;
nmax : maximum number of components in each sub-system;
nmin : minimum number of components in each sub-system;
Rs : reliability of the series–parallel system;
C s : cost of the series–parallel system.
Single objective RAP for a series–parallel system is formulated
as the following:
!
m ai
max Rs ¼ ∏ 1  ∏ ð1 r ij Þxij
i¼1 j¼1
S:T: ð1Þ
m n
∑ ∑ cij xij r C 0 ð2Þ
i¼1j¼1

with multiple objectives. Different applications of meta-heuristic

m n
methods in RAP can be found in literature [22,25,30,31]. Gen and ∑ ∑ wij xij r W 0 ð3Þ
Yun [10] surveyed the GA-based approach for various reliability i¼1j¼1

optimization problems. Li et al. [22] proposed a two-stage

ai
approach for solving multi-objective system reliability optimiza- ∑ xij r nmax ; i ¼ 1; 2; …; m ð4Þ
tion problems. In the first stage, a Multi-Objective Evolutionary j¼1
Algorithm (MOEA) generated non-dominated solutions. Then, a
ai
self-organizing map (SOM) was supplied to cluster similar solu-
∑ xij Z nmin ; i ¼ 1; 2; …; m ð5Þ
tions and finally a DEA represented to select the most efficient j¼1
solutions in each cluster. Khalili-Damghani and Amiri [13] pro-
posed a hybrid approach based on efficient epsilon-constraint, xij A Z þ ; i ¼ 1; 2; …; m; j ¼ 1; 2; …; n ð6Þ
multi-start partial bound enumeration algorithm, and DEA to solve
the binary-state multi-objective reliability redundancy allocation The schematic form of the series–parallel system is shown in Fig. 1.
series–parallel problem. Khalili-Damghani et al. [14] proposed the
decision support system for solving reliability redundancy allocation 3.2. Proposed bi-objective RAP for SCADA systems
problems. Khalili-Damghai et al. [16] proposed a new multi-objective
particle swarm optimization method for solving reliability redun- The following bi-objective RAP models (7)–(13) are proposed
dancy allocation problems. for SCADA system. It is notable that the description of the
 A. Dolatshahi-Zand, K. Khalili-Damghani / Reliability Engineering and System Safety 133 (2015) 11–21 13

maker (DM) may choose preferred solutions from the generated

non-dominated solutions. Hence, an MOPSO meta-heuristic algo-
rithm is proposed to solve models (7)–(13). Following the main
modules of the proposed MOPSO for the design of SCADA system
is described.
…
3.3.1. Structure of a swarm
n1 n2 ni
The structure of swarm should be defined according to number
of sub-system of SCADA system in water resource management as
Sub-System 1 Sub-System 2 Sub-System m well as types of components in each sub-system. The SCADA
Fig. 1. Schematic form of series-parallel system.
control center is assumed to contain three main sub-systems as
follows:

parameters in this model is similar to models (1)–(6).  Radio modem which is used to communicate between stations
! and main control center;
m ai
max Rs ¼ ∏ 1  ∏ ð1  r ij Þxij ð7Þ  FEP which is used to collect data from stations via a proper
i¼1 j¼1
protocol;
m ai
 Server which is used to record information from FEP and to
min C s ¼ ∑ ∑ cij xij send data to SCADA software.
i ¼ 1j ¼ 1

S:T: ð8Þ So the associated scheme of series–parallel configuration for

m n
∑ ∑ cij xij r C 0
i¼1j¼1
m n
∑ ∑ wij xij r W 0
i¼1j¼1
ai
∑ xij r nmax ; i ¼ 1; 2; …; m
j¼1
ai
∑ xij Z nmin ; i ¼ 1; 2; …; m
j¼1
xij A Z þ ; i ¼ 1; 2; …; m; j ¼ 1; 2; …; n
The objective function (7) maximizes the reliability of the series–
parallel system. The objective function (8) minimizes the cost of
the system. Constraint (9) assures that the total cost of the system
is less than or equal to a predefined allowable value. Constraint
(10) assures that the total weight of the system is less than or
equal to a predefined allowable value. A set of constraints (11),
which is written for each sub-system, guarantees that the total
number of components in each sub-system is less than or equal to
maximum allowable number of component. A set of constraints
(12), which is written for each sub-system, guarantees that the
total number of components in each sub-system is greater than
or equal to minimum allowable number of component. A set of
constraints (13) defines the domain of decision variables of
the model.
Models (7)–(13) are properly fitted to SCADA system. Here, the
cost constraint presents the limited budget of SCADA project. The
weight constraint is relaxed through setting a very high upper
bound on total weight of the system. The upper and lower bounds
of number of components are considered in SCADA systems.
Decision variable is the number of parallel extensions and redun-
dant hardware in each sub-system of SCADA system.
SCADA system is presented in Fig. 2.
So, the design of SCADA system is related to determine the
ð9Þ
proper number and types of modems, FEPs, and servers in each
sub-system in a way that a set of non-dominated solutions
considering both objectives (i.e., reliability and cost) and con-
ð10Þ
straints of the system is provided.
The used structure of swarm has two dimensional and is shown
in Fig. 3. Each column of Fig. 3indicates a sub-system in SCADA.
ð11Þ
The rows are reserved for different types of components in each
sub-system. Generally, Xij represents the number of redundant
hardware type j sub-system i. There are five possibilities for the
ð12Þ
types of hardware in each sub-system. So, the structure in Fig. 3
has five rows.
ð13Þ
3.3.2. Initialization of swarm
In the proposed MOPSO, the initial swarm is produced ran-
domly. This procedure is presented as follows:
Step 1. Let k¼ 1. A uniform random value between nmax and nmin
is assigned to kth row of each sub-system.
Step 2. Total number of elements in each sub-system is calcu-
lated, if it was equal to nmax, Stop; otherwise go to Step 3.
Step 3. A value between nmin and nmax  ∑ki ¼ 1 xij is assigned to
(k þ1)th element of each sub-system randomly. Put k¼ kþ 1. Go to
Step 2.
3.3.3. Finding nondominated solutions
The proposed MOPSO algorithm utilizes the non-dominated
sorting mechanism in NSGA-II proposed by Deb et al. [8]. In each
iteration of the algorithm, the dominated particles are marked.
At the end of each iteration, all of the non-dominated particles
are separated and stored in a dynamic memory that is called
repository [6].
1 1 1

3.3. Proposed multi-objective particle swarm optimization (MOPSO) 2 2 2

for SCADA system
…

Due to the NP-hard nature of models (7)–(13), the real scale n3

n1 n2
instances are hard to solve optimally. Moreover, as the proposed
SERVERS
models (7)–(13) are a multi-objective mathematical programming, MODEMS FEP
so generating efficient solutions on Pareto front of the problem is
interesting in practical cases. Under these situations decision Fig. 2. Series–parallel configuration of SCADA system.
14 A. Dolatshahi-Zand, K. Khalili-Damghani / Reliability Engineering and System Safety 133 (2015) 11–21
Modems FEPs Servers
X11 X21 X31
X12 X22 X32
X13 X23 X33
X14 X24 X34
X15 X25 X35
Fig. 3. Structure of a particle.
3.3.4. Gridding solution space
The solution space of the problem is gridded to specific regions.
The gridding is accomplished based on minimum and maximum
values of each objective function in that iteration. All non-
dominated solutions in repository are located in resultant regions.
The number of non-dominated solution, located in a similar
region, is considered as a metrics in order to hold the diversity
of the solutions in next iterations [6,11,18].
3.3.5. Global best selection
A probability is assigned to each particle considering the number
of non-dominated solutions (i.e., particles) in each region. By this
approach the assigned probability values of all non-dominated parti-
cles in a specific region are equal. The lower the number of particles
per area, the higher the probability assigned. A roulette wheel
selection procedure is implemented based on assigned probability
values and a particle is selected as global best particle among the
others [6]. Hence, the accuracy is held while the diversity is also
preserved.
3.3.6. Personal best selection
In each iteration of the algorithm, the best personal experience
of each particle is compared with the current personal experience.
If the current experience has a better quality in the sense of objec-
tive functions, the new experience will be replaced by the old one
and the memory allocated to personal experience with the best
experience of current iteration is updated. Otherwise, no update
is done.
3.3.7. Updating of velocity and position of particles
In each iteration of the algorithm, the velocity and position of
each particle are updated based on global best and personal best
using
vk ðt þ 1Þ ¼ wxk ðtÞ þ r 1 c1 ðPbest k xk ðtÞÞ þ r 2 c2 ðGbest xk ðtÞÞ;
k ¼ 1; 2; …; p
xk ðt þ 1Þ ¼ vk ðt þ 1Þ þ xk ðtÞ; k ¼ 1; 2; …; p
Where xk ðtÞ presents the position of kth particle in iteration t, and
vk ðt þ 1Þ is the velocity of kth particle in iteration tþ 1. Pbestk shows
the best personal experience for the kth particle. Gbest is the best
ever experience for the entire swarm. w is the inertia weight and
determines the tendency of a particle to maintain its previous
exploration direction [24], c1 and c2 are the individual and the
social learning factors, respectively, r 1 ; r 2 A ½0; 1 are the random
numbers, k¼ 1,2,…,p presents the swarm size, and t represents the
iteration number [14–16].
3.3.8. Constraint handling strategy
In the proposed algorithm, a modification method is developed
for infeasible particles in swarm. The modification strategy for the
proposed MOPSO algorithm is as follows:
 If the number of component in the each sub-system of a
solution is greater than nmax, it is set equal to nmax.
 If the number of component in the each sub-system of a
solution is less than nmin, it is set equal to nmin.
 If the number of component in the each sub-system of a
solution is less than nmax and greater than nmin, it has no
change.
 If the number of component in a sub-system is less than zero, it
is set equal to nmin.
 If the number of component in a sub-system is a fraction and
also is less than nmax or is less than nmin, the integer part of it is
calculated. The objective functions are calculated for the
integer part and for the integer part plus unit. Then, the proper
number of component is chosen based on objective function
values. Ties are resolved randomly.
In this modification method, all of the additions and reductions
of components number are done based on the following equa-
tions:
r ij
Ratioij ¼ ; 8 i; j ð16Þ
cij
This ratio is calculated for all component type in all sub-systems.
In each sub-system these ratio values are sorted increasingly. If an
addition is required in the sub-system, the component associated
to the last ratio in the sorted list is selected. If a reduction is
required in the sub-system, the component associated to the first
ratio in the sorted list is selected.
3.3.9. Dynamic characteristics of the proposed algorithm
At initial iterations of the algorithm, the updating of position
and velocity is biased toward Pbestk. While the iterations of the
algorithm go on, the updation of position and velocity is biased
toward Gbest. Eqs. (17)–(19) are developed to insert these proper-
ties in the proposed MOPSO.
 
T
wðt þ 1Þ ¼ wðtÞ þ wðtÞ ð17Þ
Tþ t
 
T
c1 ðt þ 1Þ ¼ c1 ðtÞ c1 ðtÞ ð18Þ
Tþ t
 
T
c2 ðt þ 1Þ ¼ c2 ðtÞ þc1 ðtÞ ð19Þ
T  t þ1
In Eqs. (17)–(19), t is the current iteration, and T is the last iteration
of the algorithm which is a parameter of the proposed algorithm. T
can be determined using a predetermined number of iterations of
the algorithm or based on convergence toward Pareto front of the
ð14Þ problem. In this research we have used the maximum iteration
number or/and maximum CPU time of which each one is met
ð15Þ earlier. All other notations have the same definition in Eqs. (14)
and (15).
The inertia weight is updated in each iteration using Eq. (17). As
the iterations go on the initial inertia weight is decreased. This will
result in reducing the tendency of a particle to maintain its current
position at the final iterations of the algorithm. The individual
learning factor is updated in each iteration using Eq. (18). As the
iterations go on the individual learning factor is decreased. This
will result in reducing the tendency of a particle to follow its Pbestk
at the final iterations of the algorithm. The social learning factor is
updated in each iteration using Eq. (19). As the iterations go on the
 A. Dolatshahi-Zand, K. Khalili-Damghani / Reliability Engineering and System Safety 133 (2015) 11–21 15

individual learning factor is increased. This will result in improv-
ing the tendency of a particle to follow the Gbest at the final
iterations of the algorithm. This properties improve the explora-
tion and exploitation of the algorithm.
3.3.10. Pareto ranking
All non-dominated solutions of all iterations are recorded. After
all iterations of the algorithm are accomplished, these non-
dominated solutions are copied to the dynamic matrix [1]. The
feasible non-dominated particles are chosen and form the estima-
tion of real Pareto front of the problem.
to supply qualified water for Tehran province with more than 15
million populations, 38 cities, and 18,814 km2 area. The optimum
pattern of consumption of water is 130 L/person. These statistics
reveal and illustrate the importance of correct performance of
TWWMC for a metropolis such as Tehran. Moreover, the summar-
ization, conducting, refinery, and recycle of wastewater are
another mission of TWWMC in Tehran. The SCADA system is used
for administration of these jobs in TWWMC. The proposed
approaches are used to present proper design for the process of
dispatching center of SCADA system considering multiple-
objective functions in models (7)–(13).

3.4. Mathematical method for solving multiobjective problems 4.1. Data for redundant component

In Multi-Objective Mathematical Programming (MOMP) the
concept of optimality is replaced with that of Pareto optimality
or efficiency. The Pareto optimal (or efficient, non-dominated,
non-inferior) solutions are the solutions that cannot be improved
in one objective function without deteriorating their performance
in at least one of the rest [19].
The SCADA system in TWWMC has three main sub-systems
(i.e., modems, FEPs, and servers) which work serially. In each sub-
system, five types of redundant component can be utilized. Table 1
presents the reliability and cost of each type in each sub-system.
4.2. Implementation

3.4.1. Efficient epsilon constraint method for MOMP The proposed MOPSO algorithm has been coded in MATLAB
A general form of MOMP problem is shown as software. The efficient epsilon-constraint method has been coded
in LINGO software. The codes were implemented on a PC with MS-
maxðf 1 ðxÞ; f 2 ðxÞ; …; f p ðxÞÞ
Windows 7 Prof., Core i7 CPU, and 4 Gb RAM. Fig. 4 represents the
S:T: pseudo code of the proposed MOPSO algorithm for the design of
X AS ð20Þ SCADA system.

where X is the vector of decision variables, f p ðxÞ is the pth

objective function, and S is the feasible region.
The performance of the proposed MOPSO algorithm is com-
pared with the performance of efficient version of ε-constraint
that was proposed by Mavrotas [19]. In the ε-constraint method
one of the objective functions is optimized while the other
objective functions are treated as constraints [19]. There are
several successful application of ε-constraint method in recent
years [14,17,26,27]. So, the MOMP problem (20) is converted into
model (21) using the proposed method by Mavrotas [19].
s
max f 1 ðxÞ þ εðrs22 þ rs33 þ ⋯ þ rpp Þ
S:T:
X AS
f 2 ðxÞ  s2 ¼ e2
f 3 ðxÞ  s3 ¼ e3
ð21Þ
⋮ options for implementation in the real world. The accuracy of re-
f p ðxÞ  sp ¼ ep
where Si ; i ¼ 2; …; p are the surplus variables used in new
constraints of model (21), and r i ; i ¼ 2; …; p are the range of
objective functions achieved from pay-off tables. By parametrical
variation in the RHS of the constrained objective functions (ei ) the
efficient solutions of the problem are obtained. On the other hand,
the right hand sides of constrained objective functions (i.e. ei ) are
changed in a feasible range from minimum to maximum value for
associated objective function. These minimum and maximum
values are easily calculated based on single objective optimization
problem [13,16,19]. Selection of the most important objective
function is done by the decision maker [4].
4.3. Results
Both methods were applied on the real case of SCADA system
for TWWMC. The results are as follows. The regenerated Pareto
front by the epsilon-constraint method is shown in Fig. 5.
Two hundred non-dominated solutions generated by the
epsilon-constraint method have been plotted in Fig. 5. It can be
concluded from Fig. 5 that the re-generated Pareto front by the
epsilon-constraint method is dense enough and the number of
non-dominated solutions on it is enough to make a decision in real
world. Each solution on Pareto front of Fig. 5 proposes a non-
dominated design for SCADA system. The range of cost function of
the re-generated Pareto front is from 4000 to 13,000 while the
range of reliability function of the re-generated Pareto front is
from 0.71 to 0.99999. These ranges assure the proper diversity of
the re-generated Pareto front. This Pareto front supplies several
generated Pareto front should be compared with another method
in order to be judged.
It is notable that the run time of the LINGO software was set
less than 1 min. The regenerated Pareto front of the proposed
MOPSO algorithm on the same data of SCADA system for TWWMC
is also shown in Fig. 6.
Again the number of generated non-dominated solutions by
the proposed MOPSO algorithm is simply enough to supply proper
practical solutions for a real life design. The ranges of both
Table 1
Reliability and cost of component types in sub-systems.
Sub-system 1 (server) Sub-system 2 (FEP) Sub-system 3 (modem)

Reliability Cost Reliability Cost Reliability Cost

4. Real case study of SACAD system in water resource
management r11 ¼ 0.8 c11 ¼100 r21 ¼ 0.95 c21 ¼500 r31 ¼0.95 c31 ¼ 2000
r12 ¼0.9 c12 ¼ 200 r22 ¼0.9 c22 ¼ 450 r32 ¼ 0.6 c32 ¼1800
The proposed approaches have been customized for the design r13 ¼0.7 c13 ¼ 80 r23 ¼0.8 c23 ¼ 400 r33 ¼0.9 c33 ¼2100
r14 ¼0.75 c14 ¼ 85 r24 ¼0.75 c24 ¼ 350 r34 ¼ 0.95 c34 ¼2175
of SCADA systems in Tehran water and wastewater manage- r15 ¼0.85 c15 ¼ 150 r25 ¼0.97 c25 ¼ 525 r35 ¼0.85 c35 ¼2050
ment company (TWWMC). The TWWMC has the responsibility
16 A. Dolatshahi-Zand, K. Khalili-Damghani / Reliability Engineering and System Safety 133 (2015) 11–21

Start
Initialize the Swarm (Using Procedure in sub-section 3.3.2)
While Termination Conditions not met Do
Find Non-Dominate Solutions
Update the Non-Dominate Archive
Grid Solution Space
Select GlobalBest
Select PersonalBest
Update Velocity and Position of Particles
Constraint Handling (Check Feasibility of Solutions)
Modify Infeasible Solutions
Update Inertia Weight, Social Learning Factor, Individual Learning Factor
Print Archive of Non-Dominate Solutions
End
Fig. 4. Pseudo code of the proposed MOPSO.

1.01
archive size. Moreover, different runs of the epsilon-constraint
method supplies no new solutions on Pareto front of the problem.
0.96 On the other hand, as the epsilon-constraint method is implemented
using mathematical optimization based on partial derivatives so the
generated solutions in different runs of the epsilon-constraint
0.91
method were constant. Epsilon constraint starts a feasible solution
Reliability

and tries to improve it based on partial derivatives of objective

0.86 function. The epsilon constraint method usually reached a unique
region of solution space in which the partial derivatives are near or
0.81 equal to zero, then it stops and saves the solution at this point and
this continues till the maximum archive size is reached.
In contrast, the proposed MOPSO searches the solution space
0.76
more effectively based on several particles. The MOPSO supplies
several solutions in each run while the epsilon constraint tries to
0.71 improve a single solution in each run. Moreover, in the MOPSO the
4000 5000 6000 7000 8000 9000 10000 11000 12000 13000 particles supply each other through the information and data
Cost collected form solution space. This yields to search the solution
Fig. 5. Pareto front by ε-constraint method. space effectively using a set of particles. The MOPSO use random
perturbation in order to improve the diversity of search. The
number of generated solutions by MOPSO was meaningfully larger
1.01 than those generated by epsilon-constraint.
So, in second test, both algorithms are given enough time to
0.96
run. Both methods were set to run 10 min. Smaller step sizes were
supplied for dividing the range of second objective function in the
epsilon-constraint method. Even under these conditions, the
0.91 generated solutions by the epsilon-constraint method were not
different from those represented in Fig. 5. On the other hand
Reliability

0.86 relaxing the computational time limitations and high density of

searching yields to no new solutions in the epsilon-constraint
method.
0.81
But the MOPSO algorithm generated more solutions under
these conditions. The generated solutions were different from
0.76 those represented in Fig. 6. The number of non-dominated solu-
tions increased meaningfully.
The high-density regenerated Pareto front by the proposed
0.71
4000 5000 6000 7000 8000 9000 10000 11000 12000 13000 MOPSO algorithm is shown in Fig. 7. The proposed MOPSO
Cost algorithm was run about five minutes in MATLAB software.
It can be concluded from Fig. 6 that the regenerated Pareto front
Fig. 6. Pareto front by MOPSO algorithm. by the MOPSO algorithm has a proper diversity. The cost and
reliability values are varied in the interval [4460, 14,500] and in
objective functions in re-generated Pareto front by MOPSO are the interval [0.716625, 0.99998995], respectively.
diverse enough to assure the diversity of the algorithm. Each point
in Fig. 6 proposes a non-dominated design for SCADA system. It is 4.4. Real Pareto estimation using reference set
clear that the epsilon-constraint method can handle the problem.
The re-generated Pareto front by the epsilon-constraint method Real Pareto front of an MODM problem contains solutions which
seems to be uniform and qualified. The MOPSO algorithm was run are not dominated by any other solutions while diversity of
for about 46 s in MATLAB software. Total number of iteration was solutions are held. Since the real Pareto front of the benchmark
set equal to 1000. case cannot be achieved exactly, the reference set (RS) is retrieved.
It is notable that although the epsilon-constraint method was RS is used a proper estimation of real Pareto front. Re-generated
able to solve the problem but the number of generated non- Pareto front is a Pareto front which is constructed based on a single
dominated solutions hardly reaches to the maximum number of run of a proposed algorithm. The RS contains the non-dominated
 A. Dolatshahi-Zand, K. Khalili-Damghani / Reliability Engineering and System Safety 133 (2015) 11–21 17

1.01 Table 3 presents the calculated accuracy and diversity metrics

for the proposed MOPSO algorithm and the epsilon constraint
method. As the proposed MOPSO algorithm has a stochastic nature
0.96
so the results may vary in each run. Ten runs are done and the
average result of accuracy and diversity metrics is presented for
0.91 the MOPSO algorithm. The epsilon-constraint method has the
same results in all runs, so the result of a single run is reported.
It can be concluded form Table 2 that the proposed MOPSO
0.86 outperforms the epsilon-constraint method. As it is seen in Table 2
the resultant NSS, ER and SM of proposed MOPSO algorithm have
relative dominance in all runs. The proposed MOPSO algorithm is
0.81
three times faster than the ε-constraint method. The GD for
ε-constraint has a better condition in comparison with GD for
0.76 ε-constraint. According to situation of other measurements, the
MOPSO algorithm has a better relative performance in comparison
with ε-constraint.
0.71
4000 5000 6000 7000 8000 9000 10000 11000 12000 13000
4.6. Comparing reference set and a single run of MOPSO algorithm
Fig. 7. High-density Pareto front by MOPSO algorithm.

Although relative dominance of proposed MOPSO was revealed

solutions for both methods in all runs [14–16]. Both methods were from Table 3, but another experiment, in which the performance of
run 10 times and the non-dominated solutions of all runs were proposed MOPSO was tested against reference set, was also
selected as RS. designed. The reference set was formed based on resultant non-
dominated solutions of 10 runs of the MOPSO algorithm and the
4.5. Comparison metrics epsilon-constraint method. In Fig. 8, there is a comparison
between non-dominated solutions of reference set and the regen-
Following metrics were selected to check the accuracy and the erated non-dominated solutions achieved from a single run of the
diversity of both methods in regenerating the Pareto front [14–16]. proposed MOPSO solutions. As it can be seen, the MOPSO solutions
Number of non-dominated solutions (NNS): This metric repre- approximately are fitted to reference set which is an evidence of
sents the NNS found by each method contained in RS. The higher efficacy and robustness of the proposed MOPSO algorithm. It
this metric, the more the method has converged towards the real means that most of the members of reference set can be re-
Pareto front. generated using a single run of the proposed MOPSO algorithm.
Error ratio (ER): The ER measures the non-convergence of the
methods towards the RS. The definition of the ER is as follows: 4.7. Pruning the non-dominated solutions using TOPSIS method
∑n ei
ER ¼ i ¼ 1 The number of generated non-dominated solutions by the
n
 proposed MOPSO algorithm is too high. This may confuse the
0 If solution i belong to RS
ei ¼ ð22Þ SCADA designers to select and implement a restricted set of
1 Otherwise
solutions. The TOPSIS algorithm proposed by Hwang and Yoon
The closer this metric is to 1, the less the solution has converged
towards the Pareto front. Table 2
Generational distance (GD): This metric calculates the distance Parameters of algorithms.
between the RS and the generated solution set in a single run. GD
MOPSO Epsilon-constraint
is calculated as follows:
∑ni¼ 1 di Parameters Value Parameters Value
GD ¼
N
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Particle no. 20 Archive size 200
m
Archive size 200 Step size of cost 0.01
di ¼ min ∑ ðzik  zpk Þ2 ð23Þ Maximum ieration no. 200 Step size of weight 0.01
p A RS k¼1
Maximum CPU time (s) 1000 Upper bound of cost 1.6  104
where m is the number of objective functions, N is the total Initial C1 2.5 Lower bound of cost 0.4  104
Final C1 0.5 Lower bound of reliability 0.7
number of solution that has been generated by algorithms, RS is
Initial C2 0.5 Upper bound of reliability 1.0
the reference set, di is the minimum distance between i-th Final C2 2.5
solution and RS, and zik is the k-th objective function of i-th Initial W 0.7
solution. Final W 0.4
Spacing metric (SM): The SM measures the uniformity of the
spread of the points in the solution set. In order to calculate SM,
the mean value of all di , should be calculated as
∑ni¼ 1 di Table 3
d¼ ð24Þ Metrics result with the MOPSO algorithm.
N
Then, the SP which is the standard deviation of the closest Method Accuracy Diversity CPU time (s)
distances is calculated using
NSS ER GD SM
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u
u N ðd  di Þ2 MOPSO 208.17 0.33386424 0.6209 2.8806 300
SM ¼ t ∑ Þ ð25Þ ε-Constraint 121 0.51984127 0.2183 2.2432 1000
i ¼ 1 N1
18 A. Dolatshahi-Zand, K. Khalili-Damghani / Reliability Engineering and System Safety 133 (2015) 11–21

anti-ideal solution is shown by A 

A þ ¼ fv1þ ; v2þ ; …; vjþ ; …; vnþ g;
where
vjþ ¼ max vij ; if vij belongs to benefit criteria;
i

vjþ ¼ min vij ; if vij belongs to cost criteria; ð31Þ

i

A  ¼ fv1 ; v2 ; …; vj ; …; vn g;

where
vj ¼ min vij ; if vij belongs to benefit criteria;
i
vj ¼ max vij ; if vij belongs to cost criteria; ð32Þ
i

Step 6. Calculate the separation measure as (33) and (34). If the

* Single Run of MOPSO
O Reference Set
separation from ideal solution is represented by S þ and the
separation from anti-ideal solution is represented by S  then
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
n
Siþ ¼ ∑ ðvij  vjþ Þ2 ; i ¼ 1; 2; …; m ð33Þ
j¼1

sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
n
Fig. 8. Comparison of reference set and MOPSO.
Si ¼ ∑ ðvij  vj Þ2 ; i ¼ 1; 2; …; m ð34Þ
j¼1

[12] is used to prune the non-dominated solutions. In TOPSIS Step 7. Calculate the closeness coefficient as
method, the solutions are ranked based on their distance from Si
CC i ¼ ; i ¼ 1; 2; …; m ð35Þ
both ideal and anti-ideal solutions. The best rank is assigned to the Siþ þ Si
solution which is far from anti-ideal solution and close to ideal
solution simultaneously. For a multiple-attribute decision making The alternative with the highest value of CC i is the most preferred
(MADM) problem with m alternative and n criteria the TOPSIS alternative.
algorithm is as follows [12]: The TOPSIS algorithm was implemented on the solutions (i.e.,
Step 1. Construct the decision matrix. SCADA designs) which have reliability greater than or equal to
0.9999 and budget lower than 9000. The alternatives are shown
2 3 in Table 4.
x11 ⋯ x1n
6 ⋮ ⋱ ⋮ 7 As we have used a preferred part of non-dominated solutions
X¼4 5 ð26Þ
as candidates for the TOPSIS algorithm, so we have alternatively
xm1 ⋯ xmn
set the relative weights of reliability and cost criteria equal to
0.5 in this research. The reliability is assumed as a benefit criterion
In this matrix xij is the score of i-th (i¼1,2,…,m) alternative with
while cost is assumed as a cost criterion. Table 5 presents the final
respect to j-th (j¼1,2,…,n) criterion. It is notable that the alter-
result of the TOPSIS method on candidate designs.
natives are non-dominated solutions generated by the MOPSO
The design number 1 is the most preferred design based on the
algorithm and criteria are objective functions.
TOPSIS method. The structure of this design is shown in Table 6.
Step 2. Construct the normalized decision matrix.
The schematic for this design is shown in Fig. 9.
2 3
r 11 ⋯ r 1n
6 ⋮ 7
Table 4
R¼4 ⋮ ⋱ 5 ð27Þ Candidate designs for SCADA system.
r m1 ⋯ r mn
Candidates Reliability Cost

where rij is calculated using 1 0.999906002888971 8900

2 0.999908502753973 8935
xij 3 0.999916002348978 8950
r ij ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð28Þ
4 0.999921002078982 8951
∑m i ¼ 1 xij
2
5 0.999926001808986 8952
6 0.999931001538989 8953
Step 3. Determine the weights of criteria as 7 0.999936001268993 8954
8 0.999936251194745 8955
9 0.999936751259992 8956
W ¼ ½w1 ; w2 ; …; wn  ð29Þ
10 0.999946250857246 8957
11 0.999951250688497 8958
Step 4. Construct the weighted normalized decision matrix as 12 0.999951750686247 8959
13 0.999956250519748 8960
2 32 3
r 11 ⋯ r 1n w1 ⋯ 0 14 0.999957600469799 8961
6 ⋮ 7 6 ⋮ 7
15 0.999960300369899 8962
V ¼4 ⋮ ⋱ 54 ⋮ ⋱ 5 ð30Þ 16 0.999960750353250 8963
r m1 ⋯ r mn 0 ⋯ wn 17 0.999961650319950 8964
18 0.999962190299970 8965
19 0.999962325294975 8966
Step 5. Determine ideal and anti-ideal solutions as (31) 20 0.999962797777493 8967
and (32), respectively. The ideal solution is shown by A þ and the
 A. Dolatshahi-Zand, K. Khalili-Damghani / Reliability Engineering and System Safety 133 (2015) 11–21 19

It can be concluded from Table 6 and Fig. 9 that, in first sub- sub-systems numbers 1–3 are connected in series configuration in
system of the most preferred design servers types 1 and 2 are order to make the final design of the most preferred design of
used. More formally, two servers of type 1 and three servers of SCADA system for water resource management.
type 2 are set in parallel in first sub-system of the most preferred
design. In second sub-system of the most preferred design only
type 5 of FEP is used. Three FEPs of type 5 are configured in 4.8. Practical implementation
parallel in second sub-system. In third sub-system of the most
preferred design only type 4 of modems are used. Three modems In practice the components of sub-systems of the SCADA
of type 4 are configured in parallel in third sub-system. Finally, control center system are located in different buildings and
connected through a Local Area Network (LAN) or a wireless
LAN. This improves the safety and availability of the whole system
Table 5
considering electricity swing, damages, fire, etc. The components
TOPSIS ranks.
of each sub-system may be implemented as cold or hot standby
Candidates CC i Siþ Si due to the experience of the designer and work load of the center.
In some other cases, two SCADA control center systems with
1 0.992466232799799 0.000012700453022 0.001673103300158 configuration presented in Fig. 8 may be established. This is called
2 0.477587929827110 0.000874093514888 0.000799094309044
high level redundancy in which the new redundant system can be
3 0.253727320120024 0.001248628402518 0.000424524636823
4 0.238807480169208 0.001273590538271 0.000399561135046 parallel, hot or cold standby. Due to cost and reliability considera-
5 0.223889492122116 0.001298554004074 0.000374602061838 tions, the high level redundancy can be formed using lower ranked
6 0.208973967763233 0.001323518724667 0.000349648365580 candidates in Table 4.
7 0.194061689608842 0.001348484630359 0.000324701285948 Alternatively, each component of sub-systems of the SCADA
8 0.179140554145689 0.001373455836643 0.000299736624721
9 0.164220896657059 0.001398426828377 0.000274774646491
control center system may be duplicated. This is called low level
10 0.149335739476021 0.001423391199196 0.000249879050008 redundancy in which the duplicated components can be parallel,
11 0.134445946298072 0.001448360382486 0.000224972873007 hot or cold standby. The final choice is made based on cost-benefit
12 0.119540588024383 0.001473331842807 0.000200035291177 analysis considering the cost and reliability of the alternatives.
13 0.104670624946121 0.001498302178163 0.000175162604641
Due to aforementioned considerations, in the proposed case
14 0.089793552117739 0.001523273597248 0.000150273762027
15 0.074952509950317 0.001548244945645 0.000125447445609 study a high-level redundancy policy has been chosen using
16 0.060121381675541 0.001573216602591 0.000100634224440 alternative Nos. 1 and 2 in Table 4. The systems are replicated.
17 0.045361800814022 0.001598188247615 0.000075941542056 The alternative No. 1 is working and alternative No. 2 is hot
18 0.030752292182337 0.001623159923748 0.000051499619582 standby for it. The schematic design for this configuration is
19 0.016686413964813 0.001648131612497 0.000027968093541
20 0.007533767200201 0.001673103300158 0.000012700453022
shown in Fig. 10.
The data in control center No. 2 will be replicated with site
redundancy protocol in normal condition. Whenever the control
center No. 1 fails, automatically the control center No. 2 will be
Table 6 activated.
Most preferred SCADA system design. There is also another option to exchange data between two
systems shown in Fig. 9. The exchange of data can take place both
Sub-system 1 (server) Sub-system 2 (FEP) Sub-system 3 (modem)
between equal systems and systems with a hierarchical structure.
Type ♯1 2 Type ♯1 0 Type ♯1 0 Information can be transmitted either event-oriented, cyclically or
Type ♯2 3 Type ♯2 0 Type ♯2 0 upon request. In this form, data transfer can take place according
Type ♯3 0 Type ♯3 0 Type ♯3 0 to the Client-Server principle with IEC60870-6/TASE.2 protocol.
Type ♯4 0 Type ♯4 0 Type ♯4 3
Type ♯5 0 Type ♯5 3 Type ♯5 0
Client polls process information from the TASE server and send
control information back to the TASE server.

Sub-System FEPs
Sub-System Modems

Sub-System Servers

Fig. 9. Schematic design of SCADA control center.

20 A. Dolatshahi-Zand, K. Khalili-Damghani / Reliability Engineering and System Safety 133 (2015) 11–21
Another SCADA Center
ICCP TASE.2
Data Replication
Control center 2 Control center 1
Alternative 2 Alternative 1
1 2 N [3] Boyer SA. SCADA: supervisory control and data acquisition. ISA; 2004.
Stations
Fig. 10. High level redundancy in SCADA control center.
5. Conclusion remarks and further research directions
In this paper, the proper configuration of sub-systems of SCADA
water resource management control center was designed. For this
aim a multi-objective redundancy allocation problem (MORAP), in
which the reliability and cost of the system were considered, was
proposed. The mathematical model of MORAP was customized for
SCADA water resource management control center in Tehran. The
SCADA water resource management control center had three main
sub-systems, called modems, servers, and FEPs which formed a
series–parallel configuration. The proposed MORAP was utilized to
determine the proper number of component types in each sub-
system of SCADA water resource management control center.
In order to solve the proposed MORAP for SCADA water
resource management control center a multi-objective particle
swarm optimization (MOPSO) algorithm and a multi-objective
mathematical method, called efficient epsilon-constraint, were
proposed. Several properties such as dynamic parameter tuning,
gridding the solution space, and modification strategy for infea-
sible solutions were inserted in the proposed MOPSO in order to
enhance its performance for the real case study. Both methods
were used to generate the Pareto frontiers of the real case study in
Tehran's SCADA water resource management control center. The
Pareto fronts generated by both methods were compared based on
several metrics of multi-objective programming. The proposed
MOPSO algorithm outperformed the efficient epsilon-constraint
method. Moreover, the computation time was intensively
improved by the proposed MOPSO algorithm. Finally, a TOPSIS
algorithm was supplied in order to rank the non-dominated
solutions on Pareto frontier in the sense if making the most
preferred solutions from all non-dominated solutions.
The SCADA control systems may contain other sub-systems
such as UPS, stabilizers, etc. in other applications as power plants,
electricity dispatching centers, gas turbines, laboratories, and
reactors. So the proposed mathematical modeling and the pro-
posed MOPSO algorithm can easily be used in other applications of
SCADA control systems. The proposed method may be developed
for several stations of SCADA control centers concurrently.
In some real applications of SCADA systems the components
are repairable. This assumption can be taken into account for a
repairable system with multi-state conditions. More intensive
scenarios on standby situations may be considered in future
researches.
Acknowledgment
The authors would like to thank the anonymous reviewers and
the editor for their insightful comments and suggestions.
References
[1] Amiri M, Abtahi A-R, Khalili-Damghani K. Solving a generalized precedence
multi-objective multi-mode time-cost-quality trade-off project scheduling
problem using a modified NSGA-II algorithm. Int J Serv Oper Manag 2013;14
(3):355–72.
[2] Bailey D, Wright E. Practical SCADA for industry. Amsterdam, Netherlands:
Elsevier; 2003.
[4] Branke J, Deb K, Miettinen K, Słowiński R, editors. Multi objective optimiza-
tion: interactive and evolutionary approaches. State-of-the-art survey series of
the lecture notes in computer science, vol. 5252. Berlin: Springer-Verlag; 2008.
p. 470.
[5] Chern MS. On the computational complexity of reliability redundancy alloca-
tion in a series system. Oper Res Lett 1992;11:309–15.
[6] Coello Coello CA, Lechuga MS. MOPSO: a proposal for multiple objective
particle swarm optimizations. In: Proceedings of the 2002 congress on
evolutionary computation. 2002. p. 1051–6.
[7] Coello Coello CA, Pulido GT, Lechuga MS. Handling multiple objectives with
particle swarm optimization. IEEE Trans Evol Comput 2004;8(3):256–79.
[8] Deb K, Pratab A, Agrawal S, Meyarivan T. A fast and elitist multiobjective
genetic algorithm: NSGA-II. IEEE Trans Evol Comput 2002;6(2):182–97.
[9] Eberhart R, Kennedy J. Particle swarm optimization. In: Proceedings of the IEEE
International Conference on Neural Networks. IEEE Press; 1995. p. 1942–8.
[10] Gen M, Yun Y. Soft computing approach for reliability optimization: state-of-
the-art survey. Reliab Eng Syst Saf 2006;91:1008–26.
[11] Hernandez-Diaz AG, Santana-Quintero LV, Coello Coello CA. Improving the
efficiency of ε-dominance based grids. Inf Sci 2011;181:3101–29.
[12] Hwang CL, Yoon KP. Implementation multiple attribute decision making:
methods and applications. New York: Springer-Verlag; 1981.
[13] Khalili-Damghani K, Amiri M. Solving binary-state multi-objective reliability
redundancy allocation series-parallel problem using efficient epsilon-con-
straint, multi-start partial bound enumeration algorithm, and DEA. Reliab
Eng Syst Saf 2012;103:35–44.
[14] Khalili-Damghani K, Abtahi A-R, Tavana M. A decision support system for
solving multi-objective redundancy allocation problems. Qual Reliab Eng Int
2013. http://dx.doi.org/10.1002/qre.1545.
[15] Khalili-Damghani K, Nojavan M, Tavana M. Solving fuzzy multidimensional
multiple-choice knapsack problems: the multi-start partial bound enumera-
tion method versus the efficient epsilon-constraint method. Appl Soft Comput
2013;13:1627–38.
[16] Khalili-Damghani K, Tavana M, Abtahi A-R. A new multi-objective particle
swarm optimization method for solving reliability redundancy allocation
problems. Reliab Eng Syst Saf 2013;111:58–75.
[17] Khalili-Damghani K, Tavana M, Sadi-Nezhad S. An integrated multi-objective
framework for solving multi-period project selection problems. Appl Math
Comput 2012;219:3122–38.
[18] Knowles JD, Corne DW. Approximating the nondominated front using the
Pareto archived evolution strategy. Evol Comput 2000;8(2):149–72.
[19] Mavrotas G. Effective implementation of the ε-constraint method in multi-
objective mathematical programming problems. Appl Math Comput
2009;213:455–65.
[20] Konak A, Coit DW, Smith AE. Multi-objective optimization using genetic
algorithms: a tutorial. Reliab Eng Syst Saf 2006;91:992–1007.
[21] Kuo W, Zuo MJ. Optimal reliability modeling: principles and applications.
Hoboken: John Wiley & Sons, Inc.; 2003.
[22] Li Z, Liao H, Coit DW. A two-stage approach for multi-objective decision
making with applications to system reliability optimization. Reliab Eng Syst
Saf 2009;94:1585–92.
[23] Misra KB, Ljubojevic MD. Optimal reliability design of a system: a new look.
IEEE Trans Reliab 1973;22:255–8.
[24] Shi Y, Eberhart R. Empirical study of particle swarm optimization. In:
Proceedings of the congress on evolutionary computation. Piscataway (NJ):
IEEE Press; 1999, p. 1945–50.
[25] Taboada H, Baheranwala F, Coit DW, Wattanapongsakorn N. Practical solutions
for multi-objective optimization: an application to system reliability design
problems. Reliab Eng Syst Saf 2007;92:314–22.
[26] Tavana M, Abtahi A-R, Khalili-Damghani K. A New multi-objective multi-mode
model for solving preemptive time-cost-quality trade-off project scheduling
problems. Expert Syst Appl 2014;41(4):1830–46.
[27] Tavana M, Khalili-Damghani K, Abtahi A-R. A new variant of fuzzy multi-choice
knapsack for project selection problem. Ann Oper. Res 2013;206(1):449–83.
 A. Dolatshahi-Zand, K. Khalili-Damghani / Reliability Engineering and System Safety 133 (2015) 11–21 21

[28] Tripathi PK, Bandyopadhyay S, Kumar Pal S. Multi-objective particle swarm
optimization with time variant inertia and acceleration coefficients. Inf Sci
2007;177:5033–49.
[29] Zio E. Reliability engineering: old problems and new challenges. Reliab Eng
Syst Saf 2009;94:125–41.
[30] Zio E, Bazzo R. A clustering procedure for reducing the number of represen-
tative solutions in the Pareto front of multi-objective optimization problems.
Eur J Oper Res 2011;210:624–34.
[31] Zio E, Bazzo R. Level diagrams analysis of Pareto front for multiobjective
system redundancy allocation. Reliab Eng Syst Saf 2011;96:569–80.