Journal of Manufacturing Systems 55 (2020) 30–43

Contents lists available at ScienceDirect

Journal of Manufacturing Systems

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

Decomposition-based bi-objective optimization for sustainable robotic T

assembly line balancing problems
Binghai Zhou*, Qiong Wu
School of Mechanical Engineering, Tongji University, Caoan Road 4800, Shanghai, China

ARTICLE INFO ABSTRACT

Keywords: Due to the increasing greenhouse gas emissions and the energy crisis, the manufacturing industry which is one of
Multi-objective optimization the most energy intensive sector is paying close attention to the improvement of environmental performance
Energy consumption models efficiency. Therefore, in this paper the automated assembly line is balanced in a sustainable way which aims to
Robot optimize a green manufacturing objective (the total energy consumption) and a productivity-related objective
Assembly line balancing
(similar working load) simultaneously. A comprehensive total energy consumption of each processing stage was
MOEA/D
analyzed and modeled. To make the model more practical, a sequence-based changeover time and robots with
different efficiencies and energy consuming rates are considered and optimized. To properly solve the problem,
the proposed novel optimal solution takes the well-known MOEA/D as a base and incorporates a well-designed
coding scheme and a problem-specific local search mechanism. Computational experiments are conducted to
evaluated each improving strategies of the algorithm and its superiority over two other high-performing multi-
objective optimization methods. The model allows decision makers to select more sustainable assembly op-
erations based on their decision impacts in both productivity and energy-saving.

1. Introduction in 2013. These industrial robots are used to perform various operations
which are physically demanding, highly repetitive or high-risk, such as
Owing to the continuous increasing energy consumption, global assembly, stamping, die casting, forging and welding, due to their high
energy crisis and climate change, the sustainable development of precision and capability. As a result, the electrical energy consumed by
modern society has become one of the most serious challenges facing robots becomes one of the primary forms of energy consumption in the
human race. Economy, society, and environment are considered as the manufacturing. Fysikopoulos et al. indicated that the energy cost during
three pillars of sustainability. However, the industries have tradition- a car manufacturing process contributes about 9–12 % of the total
ally focused on the first aspect, i.e., improving the quality and efficiency manufacturing cost, and 20 % reduction in energy consumption can
of production, but the environmental or social aspects were with little result in about 2–2.4 % saving in the final manufacturing cost [2].
regard. Seeing into industrial energy consumption, the production Therefore, considering energy efficiency in robotic lines will contribute
processes and manufacturing activities play the major roles, which are significantly to both the economy and the environment, also making the
responsible for approximately 90 % of the total [1]. These facts in- enterprises more competitive and sustainable in the market.
evitably led the researchers and manufactures to pay serious attention Nowadays, it is widely believed that there can be tremendous op-
to improve energy efficiency and control greenhouse gas emissions in portunities in developing novel optimization techniques and strategies
the industrial production process, instead of giving high priority to for sustainable manufacturing, by simultaneously including both eco-
production efficiency and cost. nomic criterions and energy efficiency criterions (e.g., controlling GHG
Differ from the traditional manufacturing, the accelerating utiliza- emissions, decreasing total costs and greening the industry) [3,4].
tion of industrial robots in automatic production lines brings lots of These challenges spread among the optimization of manufacturing ac-
benefits, but meanwhile consumes a significant amount of electrical tivities in various levels such as machining processes (e.g. [5]), sche-
energy. It motivates us to pay close attention to the energy efficiency in duling (e.g. [6]), production planning (e.g. [1]), line balancing, supply
robotic assembly lines. According to the International Federation of chain (e.g. [7]) and so forth. If the robotic assembly line, as one of the
Robotics, by 2018, global sales of industrial robots have accelerated most cost intensive and energy related processes during manufacturing,
reaching an all-time high of more than 380,000 units, twice as much as was balanced in a sustainable way, the benefits could be enormous. In

⁎
Corresponding author.
E-mail addresses: [email protected] (B. Zhou), [email protected] (Q. Wu).

https://doi.org/10.1016/j.jmsy.2020.02.005
Received 26 June 2019; Received in revised form 17 February 2020; Accepted 18 February 2020
0278-6125/ © 2020 The Society of Manufacturing Engineers. Published by Elsevier Ltd. All rights reserved.
B. Zhou and Q. Wu Journal of Manufacturing Systems 55 (2020) 30–43

this paper, we proposed an optimization method of a bi-objective Sus- robot purchasing costs, robot setup costs, sequence dependent setup
tainable Robotic Assembly Line Balancing Problem (SRALBP), which costs, and cycle time for type II robotic mixed-model assembly line
put emphasis on the potential benefits for involving the energy effi- balancing problem [26]. They used NSGA-II and multi-objective par-
ciency criterion into traditional robotic line balancing problem, and ticle swarm optimization (MOPSO) to solve the problem. Zhou and Wu
established a more comprehensive model with energy consumption used an improved immune clonal selection algorithm to solve the bi-
falling into four categories. Specifically, we considered the sequence- objective RALBP considering time and space constraints [27].
based changeover of fixtures and tools, which accounts for a non-ig- However, the studies on RALBP mentioned above so far has focused
norable part of both workstation time and energy consumption when a on traditional objectives, such as minimizing cycle time, minimizing
robot performs various tasks within a workstation. Since the traditional number of workstations and maximizing assembly line efficiency. To
ALBP is an assigning problem, which does not regard the sequencing of the best of authors’ knowledge, researches which take into considera-
tasks assigned to the stations, the proposed mathematical model also tion the sustainability in the robotic assembly line systems is relatively
take the task sequence into consideration. limited. There exist only a few studies concerning energy consumption
The assembly line balancing problem (ALBP) is one of the most or carbon footprint, which will be discussed in this paragraph.
studied problems in the industrial engineering literature. In ALBP, a set Nilakantan et al. investigated the energy consumption in straight ro-
of required tasks are assigned to a serious of workstations in order to botic assembly lines and developed two models to minimize the cycle
produce a product and some objective functions should be optimized time and energy consumption [28]. They utilized particle swarm opti-
subjected to a set of constraints. General Assembly Line Balancing mization (PSO) to solve the problem. Nilakantan et al. minimized the
Problems (GALBPs) regard further specification, such as parallel sta- energy consumption of a U-shaped robotic assembly line [29]. Li et al.
tions (e.g., [8,9]), product diversity (e.g., [10,11]), stationary resources subsequently investigated the reduction of total energy consumption in
(e.g. [12]), equipment selection (e.g. [13]), or space constraints (e.g., two-sided robotic assembly lines and developed a multi-objective re-
[14]), among others. Extensive reviews on ALBP are done by Becker started simulated annealing algorithm to obtain Pareto solutions [30].
and Scholl [15], and Battaïa and Dolgui [16]. Their results indicated that the optimization of line balancing and the
In manual assembly lines, the actual processing times for activities minimization of energy consumption in some situations were con-
vary considerably and can hardly obtain the optimal balance; while the flicting. Zhou and Kang presented a mathematical model with three
performance of robotic assembly lines is rather predictable, which gives objectives of minimizing the cycle time, the sum of energy consump-
rise for improving the system performance through appropriate line tion, and the total cost of robots of assembly lines [31]. They developed
balancing and robot assignment [17]. Assembly lines using robots are a multi-objective hybrid imperialist competitive algorithm with non-
called robotic assembly lines. Since there are different types of robots dominated sorting strategy to solve the problem.
available in the market, which can execute the same task with different One of the limitations of the existing researches on RALBP lies in the
capabilities and efficiencies, the allocation of robots to workstations has lack of practical features of real-life manufacturing systems. It can be
great influence on the performance of assembly lines. First formulated observed that the existing researches on energy-optimized assembly
by Rubinovitz and Bukchin based on ALBPs, the Robotic Assembly Line line problems are highly concentrated on optimizing the energy con-
Balancing Problems (RALBPs) concern not only assigning tasks to sumption during operating and standby process. However, the transport
workstations but also allocating the best fitting robot for each work- of workpieces between workstations and changeover between tasks are
station so as to improve the productivity [18]. RALBPs can be classified also non-negligible processes in the assembly line. On one hand,
into two types. In RALBP-I type, it aims at minimizing the number of transportation and changeover also generate considerable energy con-
workstations and the cycle time is fixed, while RALBP-II type deals with sumption. On the other hand, the changeover time has significant in-
minimizing the cycle time with a given number of workstations [19]. fluence on both workstation time and standby energy consumption. In
A larger number of solution methods have been developed to solve many assembly lines, such as car body welding lines, the operation time
the RALBP, including mathematical programming approaches, heuristic for the tasks is comparatively short. The number of changeovers makes
and meta-heuristic algorithms. Since assembly line balancing problem a big difference to the total workstation time. Therefore, a compre-
is NP-hard, the computational time of an exact method to find an op- hensive energy optimization considering operating, standby, transpor-
timum solution will be much greater than any heuristic method to yield tation and changeover during the balancing of robotic assembly lines
a good near optimum solution [20]. For studies focused on single-ob- urgently needs to be studied.
jective robotic assembly line balancing problems, both optimum and Another limitation lies in the requirement of incorporating problem-
heuristic algorithms can be found. A branch-and-bound algorithm de- specific strategies to guarantee satisfactory performance [32]. Notice
signed by Rubinovitz et al. [18], an integer programming model with a that our robotic assembly line balancing problem SRALBP is a multi-
simulation-based adjustment technique used by Tsai and Yao [21], and objective optimization problem bounded by the industrial related
a cutting plane algorithm proposed by Kim and Park [22] were used to constraints. When solving these constrained multi-objective optimiza-
balance the robotic assembly line. However, the algorithm still requires tion problems (CMOPs), it is important to maintain a balance among
huge amount of computational resources even heuristic rules are in- convergence, diversity and feasibility of a population. And there are
corporated, which is only suitable for small problems. Gao et al. pro- two major aspects to achieve this balance, which are the multi-objective
posed a genetic algorithm which was hybridized with a local search for optimization method and the constraint-handling technique [33].
type-II RALBP [19]. Zhou et al. proposed a genetic algorithm with the In terms of optimization method, multi-objective evolutionary al-
mechanism of simulated annealing for balancing robotic weld assembly gorithms (MOEAs) are widely used to solve multi-objective optimiza-
lines [23]. Nilakantan and Ponnambalam proposed bio-inspired search tion problems in manufacturing systems [34–37, among others], since
algorithms, PSO algorithm and a hybrid cuckoo search-PSO to minimize MOEAs can produce a set of well distributed non-dominated solutions
the cycle time [24]. in a single run. According to the selection strategy used in the evolu-
During the last decades, more complex ALBPs with several con- tionary process, MOEAs can be classified into different types. Most of
flicting objective functions have aroused extensive academic interests. existing MOEAs are dominance-based MOEA, which uses a selection
Pareto Front based approximation methods and aggregative methods, strategy based on Pareto domination. For instance, the most popular
especially meta-heuristic algorithms, have been used to solve the multi- NSGA-II adopts a non-dominated sorting and elitism-preserving
objective problems. Yoosefelahi et al. formulated a multi-objective strategy [38]. Another type of MOEAs is the decomposition-based
mixed integer linear programming model to minimize the cycle time, MOEA, which has attracted much attention in recent years. It decom-
robot setup costs and robot costs [25]. An evolution algorithm was used poses a multi-objective optimization problem into a number of single-
to solve the problem. Rabbani et al. proposed a model to minimize objective optimization problems, and use different weight vectors and

31
B. Zhou and Q. Wu Journal of Manufacturing Systems 55 (2020) 30–43

aggregation functions. One of the major advantages of MOEA/D over changeover may happen between tasks. Since each type of robots have
Pareto dominance based MOEAs is that single objective local search considerably different energy consumption rates and performing cap-
techniques can be readily used in MOEA/D [39]. Therefore, a MOEA/D abilities for the same assembly tasks, total energy consumption and
is adopted in this paper to solve the proposed SRALBP considering production efficiency of a robotic assembly line vary depending on its
energy consumption and changeovers. Meanwhile, a local search me- task sequence, task to workstation assignment and type of robot allotted.
chanism is designed and implemented in the MOEA/D. By exploring the Thus, three kinds of decisions should be made in the sustainable
properties of proposed problem, more focuses on the optimization of robotic assembly line balancing:
energy consumption is placed during the design of the search me-
chanisms to guarantee more satisfactory performance. (1) Assigning assembly tasks to each workstation, ensuring the work
As for the constraint-handling technique, the majority of existing load of each station are similar and cycle time meets the constraint.
methods for ALBPs are feasibility-driven methods, where the feasible (2) Allocating the best fit robot to each workstation considering energy
solutions are always better than infeasible solutions. However, the efficiency and performing capabilities.
method may lead to being trapped in local optima. In this paper, an (3) Sequencing the assembly operations assigned to each workstation
infeasibility-driven strategy is adopted to trade off the feasibility and to make the minimum number of changeovers.
convergence. A small proportion of infeasible solutions is maintained in
the population to take full advantage of the useful information con- 2.1. Assumptions and notations
tained in the infeasible solutions. The improvement in search ability
and convergence of the algorithm is verified through experiments. Before introducing the mathematical model, some basic assump-
The aim of this paper is to effectively apply the enhanced MOEA/D tions considered in this paper are introduced as follows:
and infeasibility-driven strategy to a practical robotic assembly line
balancing problem, optimizing the comprehensive energy optimization (1) The assembly tasks cannot be subdivided.
in a robotic assembly line when balancing the work load and assigning (2) The precedence diagram and changeover relationships between
the best fitting robots to workstations, so that the energy efficiency and activities are known.
sustainability of the assembly line can be improved. In comparison with (3) A task can be allocated only when the cycle time and the pre-
the existing studies, the contributions of this paper are as follows: cedence constraints are satisfied.
(4) Only one robot can be assigned to a station but there is no limita-
(1) In order to optimize the robotic assembly line toward a sustainable tion on the quantity of robots of any type; any task or robot can be
trend, a Type-I SRALBP considering the energy consumption of each assigned to any station.
process was analyzed and modeled, in which the objective is to (5) The time taken to perform a task depends on the assigned robot and
minimize the total energy consumption and number of workstations the time taken for a robot to perform a task is determined.
simultaneously. (6) The energy consumption to perform a task or stand-by depends on
(2) In the proposed model, the energy consumption was more com- the type of robots.
prehensively modeled, divided into the four parts: operating en- (7) Material handling, loading and unloading, and other set-up energy
ergy, standby energy, transport energy and changeover energy. consumption are not neglected in this study. Tool-changing time is
Sequence-dependent changeover is incorporated to achieve a better considered and the changeover time of one robot between every
modelling of the real robotic assembly lines balancing problem. task are equal and constant.
Number of changeovers is optimized to reduce the unproductive (8) The robotic assembly line for a homogenous product is balanced.
time and reduce the number of workstations.
(3) In this paper, an enhanced decomposition-based multi-objective The notations used in this paper are summarized in Table 1.
algorithm (MOEA/D) has been proposed to address the proposed
constrained multi-objective optimization problem. A constraint-
handling technique which reserves informative infeasible in- 2.2. The model of total energy consumption objective function
dividuals during the evolution process and an adaptive PBI method
with regard to the infeasible individual distribution are integrated Basing on the observation of robotic assembly lines, the states of the
to improve its convergence and distribution. whole assembly process include loading, assembly tasks operating by
(4) Taking into account the features of the proposed SRALBP, a pro- robots, tool changing, unloading, transporting between workstations,
blem-specialized local search enhancement is integrated to
strengthen the search capability. The dedicated energy-related local Table 1
Notations and explanations.
search further optimizes the total energy consumption and accel-
erate the search process. Notations Explanation

i task (i = 1,2, 3, …, I )
The rest of this paper is organized as follows. The description of a
s station (s = 1,2, 3, …, UBm )
sustainable robotic assembly line balancing and the model formaliza- CT cycle time
tion is carried out in Section 2. In Section 3, the detailed mechanism of r robot types (r = 1,2, 3, …, R )
the proposed enhanced MOEA/D is discussed. Computational experi- per operation energy consumption of the robot r per time unit
ments are carried out to validate and evaluate the performance of the cer changeover energy consumption of the robot r for one time
ser standby energy consumption of the robot r per time unit
algorithms in Section 4. Finally, conclusions are drawn and the prospect
te transporting, loading and unloading energy consumption of station s
of future research is provided in Section 5. tir processing time of task i by robot type r
cr time for once changeover operation of robot type r
2. Problem description and formulation ps number of changeovers in station s
z ij z ij is 1 if task i and task j are operated by the same tool, and 0
The robotic assembly line consists of a set of workstations and a robot otherwise
Pre (i ) set of immediate predecessors of task i
assigned to each workstation. The workpieces visit workstations succes-
xis xis is 1 if task i is assigned to station s and 0 otherwise
sively as they are moved along the line by some kind of transportation yrs yrs is 1 if robot type r is assigned to station s and 0 otherwise
system with a fixed energy consumption rate. A set of tasks have to be uij uij is 1 if task j is operated immediately after task i and 0 otherwise
executed in an order satisfying the specified precedence constrains and

32
B. Zhou and Q. Wu Journal of Manufacturing Systems 55 (2020) 30–43

energy consumption of executing unloading and loading, and the

transportation system such as a conveyor belt [40]. The transportation
energy consumption can be calculated as follows:
TE = (m 1) te (4)

2.2.4. Model of standby energy consumption

The standby refers to the phase that the robot is waiting to process
the next assembly operation. Standby energy consumption is the energy
consumption of workstation running without assembling and robot
Fig. 1. The energy consumption in a workstation. waiting during changeovers, loading, unloading and idling. The standby
energy consumption in workstation s can be calculated as follows:
and idling. An example of the energy consumption during the whole R I R

process in a workstation can be illustrated by Fig. 1. Hence in this study, SEs = yrs (CT x is yrs tir ) ser
r=1 i=1 r=1 (5)
according to their characteristics, the total energy consumption is di-
vided into four categories: processing, changeover, transportation and
standby. 2.3. The mathematical model

2.2.1. Model of processing energy consumption On the basis of the assumptions and models above, the mathema-
The processing energy consumption is the energy consumption of tical formulation of SRALBP-I with energy saving and changeover
operating an assembly task using a tool by a robot. The consumption consideration is given as follows:
rate depends on the robot type. The processing energy consumption in Min F = ( f1 , f2 ) (6)
workstation s can be calculated as follows:
UBm
I R
f1 (x ) = m = max x is
PEs = x is yrs tir per s=1
i = 1,2, … , I (7)
i =1 r=1 (1)
UBm
f2 (x ) = E = (PES + CES + SES ) + TE
2.2.2. Model of changeover energy consumption s=1 (8)
A tool or fixture change for the assigned robot may be required Subject to
between the executions of two adjacent tasks depending on the pro-
UBm UBm
cessing method. The effect of task sequence in a workstation to chan-
sxjs sx is 0, i, j Pre (i )
geover times is illustrated by Fig. 2. In order to model the changeover s=1 s=1 (9)
energy consumption, the number of changeovers in a workstation is
UBm
introduced.
x is = 1, i
The tasks assigned to station s are sequenced according to the (10)
s=1
precedence relationships and the principal of making the least time of
changeovers. The changeover will happen if task j is operated im- UBm
yrs = 1, r
mediately after task i and the same tool is required by both tasks, i.e., (11)
s=1
x is xjs uij z ij = 1, i, j , s
I R R I I

Notice that the first task in sequence is operated immediately after x is yrs tir + yrs cr ( x is xjs uij z ij ) CT , s
the last task. i =1 r=1 r=1 i=1 j =1 (12)
The number of changeovers ps in station s is given as follows: uij x is xjs , i, j = 1, …, I , s (13)
ps = xis xjs uij z ij , s
i j (2)
x is , yrs, uij = 0 or 1, i , j = 1, …, I , r, s (14)
The Eq. (7) and (8) are the objectives of the model to minimize the
The changeover energy consumption in workstation s can be cal-
number of workstations and total energy consumption. The constrain
culated as follows:
given in Eq. (9) ensures that no precedence relationship is violated, that
R
is, when a task j precedes another task i, it must be assigned to a station
CEs = yrs cer ps
(3) precedes another station which task i is assigned to, or they are assigned
r=1
to a same station. Eq. (10) guarantees that all tasks must be assigned
and every task is assigned to exactly one workstation. Eq. (11) ensures
2.2.3. Model of transportation energy consumption that exactly one robot can be assigned to each workstation. Eq. (12)
The transportation energy consumption is the energy consumed ensures that the total workload time of every workstation, which is
from moving the workpiece along two adjacent workstations, including composed of operating time and changeover time, cannot exceed the

Fig. 2. The effect of task sequence in workstation to change-

over times.

33
B. Zhou and Q. Wu Journal of Manufacturing Systems 55 (2020) 30–43

given cycle time. Eq. (13) ensures task i and task i must be assigned to minSEsr .
r
the same station if task j is operated immediately after task i .
3.2. Constraint-handling techniques
3. Enhanced decomposition-based multi-objective algorithm
Although the coding scheme ensures the precedence constraints, it
Considering the robotic assembly line balancing problem is well- does not take the cycle time into consideration. The solution could be
known as a NP-hard problem, a novel heuristic algorithm, which is built infeasible if the workstation time exceed the given cycle time. Since
upon the framework of decomposition-based multi-objective algorithm infeasible individuals with satisfactory objective values and small cycle
(MOEA/D), is developed in this paper to resolve the proposed con- time violations could possibly generate offspring which have better
strained multi-objective optimization problem (CMOP). Since the ob- objective values, some informative infeasible individuals are main-
jectives conflict with each other, there does not exist a single solution tained in the proposed A-MOEA/D-ID.
that can optimize both objectives simultaneously, and a set of re- The degree of cycle time constraint violation of an individual x is
presentative Pareto solutions are provided by the proposed algorithm. calculated as follows:
The MOEA/Ds apply a linear or nonlinear aggregation method to V (x ) = max{0, maxSTs (x ) CT }
(15)
decompose a multi-objective optimization problem (MOP) into a
s

number of single optimization subproblems (SOPs) and each SOP re- where STs (x ) is the workstation time of x in workstation s .
lates to one solution [41]. A set of uniformly distributed weight vectors
Definition 1. A solution x is called feasible solution if V (x ) = 0 .
is initially generated and the subproblem neighborhoods are used to
Otherwise x is called infeasible solution.
obtain good population diversity and improve the balance between
exploration and exploitation of the algorithm [42]. There are three Then the proposed RALBP with energy consideration is transformed
commonly used decomposition approaches, including the weighted into an unconstrained multi-objective optimization problem with three
sum, Tchebycheff and boundary intersection approaches. In this paper, objectives:
the penalty based boundary intersection (PBI) method is employed and Min F = ( f1 , f2 , f3 ) (16)
improved to obtain better distributed solutions for our practical en-
gineering optimization problems. f1 (x ) = m (17)
In terms of the constraint-handling technique, trades off between
the feasibility and convergence of a population are made simulta- f2 (x )=E (18)
neously in this paper. Based on the analysis of specific features of the
f3 (x )=V (x ) (19)
proposed SRALBP, a small proportion of infeasible solutions is main-
tained in the population to improve the convergence, and each sub-
population uses an adaptive constraint-handling method. 3.3. Weight vectors initialization and adjustment
The details of the proposed adaptive infeasibility-driven MOEA/D
(A-MOEA/D-ID) are explained in the following sections. Since there is no prior knowledge about the Pareto front, a set of
three-dimensional uniform distributed weight vectors w i = (w1i, w2i , w3i)
3.1. Solution representation for each subproblem i is generated to improve the uniformity of the
approximated Pareto front obtained by the algorithm. The simplex
To avoid producing a narrow Pareto Front, a separator-based re- lattice design method is adopted to set initial weight vectors:
presentation scheme expressing both tasks sequence and task-to-station w1i + w2i + w3i = 1 (20)
assignments is adopted for the proposed A-MOEA/D-ID. A potential
solution is expressed by randomly inserting several separators into the
order of task genes satisfying the precedence relationship. Tasks are
wij {0, H1 , H2 , …, HH }, j = 1,2, 3 (21)
encoded by 1,2, 3, …, I , and since separators genes do not represent any the weight vector set is W = (w1, , where N =
w 2, …, w N ) is the CH2 + 2
specific task, they are encoded by number 0 to avoid confusion. Fig. 3 total number of weight vectors.
shows an example of the coding scheme. However, according to Definition 1, the weight vectors dimension
The details of decoding procedure are given as follows: for constraint violation must be equal to zero to make the solution sa-
Step 1. Allocate the tasks to the workstations according to the se- tisfy the cycle time constraint. To guide the individuals towards feasible
parators. region, w3i for each individual is dynamically adjusted and eventually
Step 2. In every workstation, reorder the tasks to make the least becomes zero during the evolution in this paper. As the generation
time of changeovers without violating the precedence constraints. increases, individuals with smaller objective values and constraint
Step 3. For every type of robot r assigned to workstation s , calculate violations are preferred to be selected. The values of weight vectors are
the station time STs r and station energy consumption SEs r . updated as follows:
Step 4. Compare STs r with C , if STs r C , r R 0 . Assign the robot
which performs the allocated tasks with the minSEsr to every station. If w3i
w1i + w2i + =1
r R0 (22)
R0 = , assign the robot which performs the allocated tasks with the

wij {0, H1 , H2 , …, HH }, i = 1,2 (23)

Fig. 3. The separator-based representation scheme.

wij {0, H , 2H , …, HH }, i = 3 (24)

34
B. Zhou and Q. Wu Journal of Manufacturing Systems 55 (2020) 30–43

w
d2 = (F (x ) (z * + d1 )
w (28)

> 0 is a predefined penalty parameter, w = (w1, w2, w3)T is the weight

vector, z * = (z1*, z2*, …, z m* )T is the ideal point in the objective space, i.e.,
for each i = 1,2, 3, z i* = minfi (x ) . Since z i* is generally unknown before
searching, z i* is replaced by fi (x ) with the smallest value during the
searching process.
The PBI decomposition method is illustrated by Fig. 5. It is can be
observed that since d1 and d2 indicate the utopia distance and the per-
pendicular distance respectively between x and z * in direction w , they
Fig. 4. The effective of iteration on weight vectors. are able to represent the convergence and diversity respectively.
Therefore plays a critical role in balancing convergence and diversity
[43]. A small value of can accelerate convergence process, but it may
r( )
gen
=1
1 maxgen
(25) result in a local optimum; while a large value of can improve the
variety of population, but the convergence will be slowed down.
where r is a uniformly distributed random number in (0, 1), gen is the
current generation, maxgen is the maximum generation, is the up-
dating factor. 3.4.2. Adaptive PBI method
As illustrated by Fig. 4, the weight vectors in the early stage guide The penalty factor in the PBI method plays a significant role in the
the search to a wide range, while by decreasing the value of w3i , a set of performance of MOEA/D. For practical engineering optimization pro-
infeasible individuals with better objective values and smaller con- blems with complex POFs, MOEA/D with a fixed PBI penalty factor
straint violations can survive in the evolution. Thus, the proposed al- can difficultly obtain well distributed solutions. To solve the problem,
gorithm can obtain a population of both distributivity and diversity. an adaptive penalty factor scheme, based on infeasible solutions, is
designed in this paper.
3.4. Adaptive PBI method Since some infeasible solutions are preserved during the evolution,
different subproblems with different infeasible solutions quantities
Generally, three aggregation functions, the Penalty Boundary should have different values. When it has a great deal of infeasible
Intersection (PBI), weight sum function and Tchebycheff function are solutions, a smaller penalty factor can help to enhance the convergence
used to decompose the MOP into a number of single-objective optimi- speed; for a small number of infeasible solutions, a larger penalty factor
zation subproblems in MOEA/D. In this study, the decomposition will help to improve population diversity.
method used in A-MOEA/D-ID is the adaptive PBI method. We first define the density of infeasible solutions in the neighbor of
subproblem i :
3.4.1. Basic PBI method ni nmin n _infi
In the PBI method, the optimization problem can be described as =
i
nmax nmin ni (29)
follows:
where ni and n _infi are the number of solutions and infeasible solutions
ming pbi (xw, z *) = d1 + d2 (26)
respectively, nmin and nmax are the minimum and maximum value of n ,
subject to x respectively.
The relationship between the penalty factor of subproblem i and
where the density of infeasible i is as follows:
(F (x ) z *)T w 1
d1 = ( i) =
w (27) 1+ e i (30)

Fig. 5. The PBI method. Fig. 6. An example for crossover operation.

35
B. Zhou and Q. Wu Journal of Manufacturing Systems 55 (2020) 30–43

two cut points of both offspring.

3.5.2. Mutation operation

A mutation operation that reorder a part of the tasks genes and
reassign them to workstations is implemented.
At the reorder stage, two points are chosen randomly and the se-
parators inside the two mutation points are removed, and the task genes
Fig. 7. An example for mutation operation. are reordered into a new sequence with respect to the precedence
constraints. At the reassign stage, a new assignment of those tasks is
carried out by randomly generating new separators between the tasks.
The probability of the separator insertion depends on the current sta-
WT
tion filling rate, fr = CT , where WT is the current workstation time, CT
is the given cycle time.
An illustrative example of this mutation operation is in Fig. 7.
Fig. 8. An example for the energy-based local search.

3.6. The energy-based local search

where and are the controlling factors.
In order to additionally inject diversity of search progress to obtain
3.5. Reproduction operation better distributed Pareto front approximations, as well as incorporate
the problem information to the algorithm, a problem-dependent local
Crossover and mutation operations modified based on the re- search operation is utilized to further improve the energy efficiency.
presentation scheme are used to produce the offspring population. A portion of solutions are selected in each generation after the re-
Parents are selected from some neighboring weight vectors. Offspring production operation to go through the local search with equal prob-
replaces its parent only if it has a better aggregation function value. abilities. In the procedure, the task genes pairs from adjacent work-
stations are exchanged, with a cost of the interchange evaluated, so as
to search the neighborhood of the individual. In terms of the cost, the
3.5.1. Crossover operation
energy consumption is concerned. The search progress can be illu-
To ensure the obtained offspring satisfies the precedence restric-
strated by Fig. 8 and described as follows. For each workstation
tions, partially mapped crossover (PMX), which has been widely used in
s, s = 1, …, S 1, start from the first station:
genetic algorithms for the ALBP is implemented and modified in this
Step 1. Find all the task pairs (x , y ) satisfying x s , y s + 1, and
paper. Two parents are mated to generate two different off-springs, as
the operating sequence of x , y can be exchanged without violating
illustrated in Fig. 6. The crossover operation procedure is described as
precedence constraints.
follows:
Step 2. Calculate the cost, which is the energy consumption increase
Step 1. Two random cut points are selected. Remove the separators
caused by the interchange in workstation s and s + 1.
inside the two cut points.
Step 2. For the first offspring, the genes outside the random points cost x, y = E (31)
are copied directly from the first parent; the task genes inside the two
cut points are copied but in the order they appear in the second parent. Step 3. Exchange the task pair x *, y* with the lowest negative cost
Step 3. For the second offspring, the same mechanism is followed up mincost x , y . If cost x, y 0 for all the task pairs, the interchange will not be
x,y
but with the opposite parents. executed.
Step 4. Insert the separators inside the two cut points randomly into Based on the descriptions for each component of A-MOEA/D-ID

Table 2
Parameter values.
Algorithms Parameter Value

Low Middle High

A-MOEA/D-ID Population size 210 136 91

Size of neighborhood. 15 15 14
Crossover probability 0.8 0.8 0.8
Mutation probability 0.2 0.2 0.2
Maximum generation 30 40 50
Parameter of weight vectors update 2 2 2
Parameter of local search 0.6 0.6 0.6
MO-PSO Particle swarm size 210 136 91
Velocity factor 0.1 0.1 0.1
Inertia factor 0.8 0.8 0.8
Crossover probability 0.8 0.8 0.8
Mutation probability 0.2 0.2 0.2
Maximum generation 30 40 50
Basic MOEA/D Population size 210 136 91
Crossover probability 0.8 0.8 0.8
Mutation probability 0.2 0.2 0.2
Size of neighborhood. 15 15 14
Maximum generation 30 40 50

36
B. Zhou and Q. Wu Journal of Manufacturing Systems 55 (2020) 30–43

given in previous subsections, the pseudo codes of proposed algorithm

are summarized as follows.

4. Computational experiments 4.1. Problem instances and parameters

In order to evaluate the performance of the proposed A-MOEA/D-ID Since no benchmark data set is available for the sustainable RALBP,
in this paper, we implemented our algorithms in MATLAB to solve a set the test problems used in this research are generated based on pre-
of test problem instances with different features. First, we detail how cedence graphs available in http://www.assembly-linebalancing.de/,
the test instances are generated and specify the parameter values. Then, which are benchmark data sets for SALBP-1. The cycle times (CT) in the
we introduce some multi-objective performance indicators used for the data sets for each instance are determined by the method of Hoffmann
computational tests. The analyses on the effects of each proposed al- [44]. Additional factors such as energy consumption rate, required tool
gorithm component are carried out, and the proposed algorithm is for each task, and tool-changing time are generated in the following
compared with the existing benchmark algorithm MO-PSO and the specifications. Two levels, four and six, of robot types were evaluated
basic MOEA/D. Finally, managerial applications are discussed to get with a tool changing time randomly generated from uniform distribu-
managerial insights. tions U [15,25], and an energy consumption rate generated from the

37
B. Zhou and Q. Wu Journal of Manufacturing Systems 55 (2020) 30–43

Table 3
Experimental results of different problems.
Problem I R A0 A1 A2 MO-PSO MOEA/D

BUXEY (CT =36) 29 4 IGD¯ 0.101 1.296 1.435 1.217 3.059

HV¯ 0.059 0.055 0.034 0.042 0.028
N 11 8 9 6 5
T 24.740 20.109 22.668 23.784 19.274
LUTZ1 (CT = 2020 ) 32 4 IGD¯ 10.173 12.367 21.356 13.512 47.517
HV¯ 0.041 0.022 0.037 0.029 0.006
N 6 5 4 4 3
T 38.007 34.719 32.328 34.027 30.255
GUNTHER (CT =36) 35 4 IGD¯ 1.335 2.494 2.081 1.794 3.824
HV¯ 0.025 0.017 0.020 0.019 0.015
N 9 6 8 6 4
T 19.229 15.315 18.336 14.870 13.180
KILBRID (CT = 79 ) 45 4 IGD¯ 4.881 9.803 7.269 7.457 11.630
HV¯ 0.073 0.054 0.054 0.066 0.047
N 10 7 8 6 4
T 34.702 30.578 28.776 31.085 26.802
BUXEY (CT = 36 ) 29 6 IGD¯ 4.350 6.285 6.973 7.236 8.236
HV¯ 0.435 0.274 0.395 0.342 0.153
N 15 13 14 13 11
T 25.345 21.496 24.563 23.597 18.958
LUTZ1 (CT =2020) 32 6 IGD¯ 35.329 48.376 56.362 45.286 63.563
HV¯ 0.063 0.043 0.053 0.051 0.221
N 8 6 7 6 5
T 40.569 36.167 34.163 35.935 33.601
GUNTHER (CT =36) 35 6 IGD¯ 43.214 49.264 53.625 50.628 65.163
HV¯ 0.523 0.369 0.403 0.374 0.204
N 14 11 12 12 9
T 22.626 16.364 20.367 14.943 15.274
KILBRID (CT =79) 45 6 IGD¯ 34.274 40.264 40.747 41.629 53.624
HV¯ 0.094 0.063 0.060 0.058 0.014
N 15 12 12 13 9
T 37.387 33.635 34.627 34.946 28.639
HAHN (CT =2806) 53 4 IGD¯ 10.040 17.844 90.000 32.423 100.577
HV¯ 0.023 0.012 0.005 0.012 0.001
N 11 7 8 7 4
T 34.684 28.480 27.273 30.592 24.061
TONGE (CT =234) 70 4 IGD¯ 16.894 21.208 22.889 17.179 32.683
HV¯ 0.014 0.009 0.008 0.011 0.007
N 7 5 6 5 3
T 25.989 24.407 24.353 21.624 20.245
ARC83 (CT =4206) 83 4 IGD¯ 29.315 135.052 177.993 146.540 514.704
HV¯ 0.610 0.240 0.210 0.250 0.160
N 8 5 6 5 3
T 96.934 61.376 55.530 60.777 46.223
MUKHERJE (CT =222) 94 4 IGD¯ 32.530 53.157 53.722 57.480 63.405
HV¯ 0.019 0.012 0.015 0.011 0.010
N 8 6 6 4 4
T 51.619 44.852 41.600 43.122 40.115
HAHN (CT =2806) 53 6 IGD¯ 43.352 54.253 52.646 59.359 73.626
HV¯ 0.731 0.528 0.529 0.639 0.313
N 14 12 11 11 10
T 38.482 30.492 29.491 31.582 26.326
TONGE (CT =234) 70 6 IGD¯ 84.793 105.842 101.384 99.638 120.269
HV¯ 0.150 0.103 0.110 0.009 0.004
N 9 8 7 6 5
T 29.646 27.367 26.639 26.853 23.468
ARC83 (CT =4206) 83 6 IGD¯ 35.079 56.835 72.624 66.908 103.520
HV¯ 0.720 0.341 0.324 0.410 0.142
N 11 8 9 8 6
T 102.452 82.435 73.526 79.270 55.275
MUKHERJE (CT =222) 94 6 IGD¯ 83.245 124.293 113.834 132.452 183.432
HV¯ 0.120 0.060 0.090 0.080 0.030
N 13 10 12 10 8
T 72.351 63.194 61.295 60.235 48.852
(continued on next page)

38
B. Zhou and Q. Wu Journal of Manufacturing Systems 55 (2020) 30–43

Table 3 (continued)

Problem I R A0 A1 A2 MO-PSO MOEA/D

ARC111 (CT =7916) 111 4 IGD¯ 58.927 78.692 137.328 93.346 202.040
HV¯ 0.022 0.019 0.017 0.017 0.015
N 10 6 11 7 5
T 79.964 66.864 63.548 68.492 49.237
ARC111-2 (CT =5755) 111 4 IGD¯ 584.847 1262.814 1186.692 1294.757 1416.727
HV¯ 0.017 0.013 0.014 0.012 0.010
N 7 5 6 5 3
T 34.514 30.345 30.683 31.668 30.302
BARTHOL2 (CT =133) 148 4 IGD¯ 8.022 10.164 16.318 12.403 43.012
HV¯ 0.190 0.130 0.080 0.110 0.020
N 10 6 7 5 4
T 209.476 122.808 115.916 112.946 77.859
BARTHOL2-2 (CT =513) 148 4 IGD¯ 30.935 45.148 48.886 46.544 67.708
HV¯ 0.028 0.024 0.018 0.021 0.013
N 9 7 8 6 5
T 103.522 87.467 75.113 74.987 72.582
ARC111 111 6 IGD¯ 43.214 62.625 69.463 59.246 82.452
(CT =7916) HV¯ 0.429 0.294 0.231 0.318 0.132
N 15 12 13 11 8
T 84.245 79.482 77.524 76.325 60.352
ARC111-2 111 6 IGD¯ 284.250 524.520 682.460 495.250 729.050
(CT =5755) HV¯ 0.052 0.035 0.033 0.042 0.194
N 10 7 7 8 5
T 42.734 37.238 37.462 36.289 33.623
BARTHOL2 (CT =133) 148 6 IGD¯ 11.341 16.253 18.163 15.362 34.264
HV¯ 0.240 0.120 0.100 0.170 0.050
N 16 13 14 13 8
T 228.345 163.626 158.346 142.569 92.432
BARTHOL2-2 (CT =513) 148 6 IGD¯ 75.727 98.868 100.864 84.858 153.868
HV¯ 0.054 0.028 0.026 0.037 0.011
N 14 10 12 11 9
T 133.787 114.757 110.738 108.468 98.473
Average IGD¯ 65.257 65.257 118.238 130.630 120.104
HV¯ 0.197 0.197 0.119 0.119 0.130
N 11 11 8 9 8
T 67.140 67.140 53.065 51.036 50.791

algorithms and their operators are shown in Table 2.

4.2. Performance indicators

To measure the performance of the multi-objective optimization

algorithms, two widely used metrics are employed: inverted genera-
tional distance [45] and hypervolume [46]. The definitions are as fol-
lows.

(1) Inverted Generational Distance (IGD):

IGD is a metric which represents average distance from the ideal PF

to the approximate PF achieved. It evaluates the performance on con-
vergence and diversity simultaneously. A smaller IGD represents better
performance with respect to both diversity and convergence. Let P * be a
Fig. 9. Obtained Pareto fronts for BUXEY(29,4) by A0, A1, A2, MO-PSO and set of solutions in the ideal PF, and P denote the approximate PF
basic MOEA/D. achieved by the algorithm, then the IGD is calculated as follows:

x P*
d (x , P )
uniform distribution U[0.5,1.5]. Type of tools are randomly generated IGD(P *, P ) =
|P *| (32)
for each task. The variability in activity performance times by different
robot types is set 0.5 on the base of expected time in original instances. where d (x , P ) is the Euclidean distance from P to its nearest member of
The instances scales are divided into three groups according to the the true Pareto front P * .
number of tasks: small (I 50 ), medium (I = 51 to 100 ) and large
(I = 101 to 150 ) datasets. (2) Hypervolume (HV):
We executed each algorithm 10 times with different random seeds,
setting a maximum number of iterations as stopping criterion on a PC HV is Pareto compliant and measures both convergence and di-
with an Intel Core i5 2.8 GHz CPU, 8 GB RAM. The parameters of the versity of the obtained solutions along the PF. A larger HV means a

39
B. Zhou and Q. Wu Journal of Manufacturing Systems 55 (2020) 30–43

Fig. 10. Boxplots of IGD and HV values for A0, A1, A2, MO-PSO and basic MOEA/D.

better dominance relation. It calculates the volume enclosed by the HV(P ) = volume( [f1 (x ), z1r ]×…×[fm (x ), zmr])
(33)
obtained set P and a reference point in the objective space. Let x P

z r = (z1r , z2r , …, z mr)T be the reference point,

where volume is the Lebesgue measure.

40
B. Zhou and Q. Wu Journal of Manufacturing Systems 55 (2020) 30–43

most diversely distributed set of solutions in average among the five

algorithms. Moreover, the results also imply that the solutions found
by A-MOEA/D-ID are closer to the pseudo-optimal Pareto fronts,
which verifies that the solution sorting mechanism of A-MOEA/D-ID
is more effective than that of MO-PSO and basic MOEA/D.
(2) Regarding the HV metric, the solutions achieved by the A-MOEA/D-
ID are significantly better than those achieved by the other four
algorithms, since the higher HV implies the better overall perfor-
mance of the algorithm. Especially as the problem scales up, the
advantage also increases. By comparing the performance of A-
MOEA/D-ID and basic MOEA/D, it is indicated that without the
constraint-handling technique, the individuals can only be dis-
tributed among some narrow and disconnected feasible regions. As
a result, most individuals can easily be trapped in one or a few of
Fig. 11. Pareto fronts obtained by A0 and A2 in BARTHOL2(148,4).
these feasible regions.
(3) In terms of the N indicator values, A-MOEA/D-ID is capable of
finding more non-dominated solutions than MO-PSO and basic
MOEA/D on average. The quantity of obtained non-dominated so-
lutions also increases as the robot types grows, since the search
space becomes larger.
(4) When it comes to the computational time, although the A-MOEA/D-
ID spends more time on its sub-operators, it is able to solve the
realistic problems within an acceptable computational time in view
of its search ability and convergence performance.
(5) These results can also help us to find out if the removed components
are necessary. By comparing the A-MOEA/D-ID with A1, it is ob-
served that the simplified version without self-adaption of the pen-
alty factor has its drawbacks on premature convergence due to the
Fig. 12. Solutions for different scenarios in HAHN(53,4).
lack of diversity. In addition, compared to A2, the problem-depen-
dent local search does help to push forward the Pareto front and
Since we are working with real problems, the true Pareto fronts are improving the diversity according to the average IGD and HV metric.
not known. A pseudo-optimal Pareto set, which is an approximation of
the true Pareto set, obtained by merging all the Pareto set approxima- To get a better intuition about the result of every independent run,
tions generated for each problem instance by any algorithm in any run the IGD and HV performance indicator values of every algorithms
will be considered. Besides, the reference points are specific for each generated in the 10 runs have been represented by boxplots (see
problem instance by maximum values for the two objectives. Fig. 10). In the figure, each box represents the distribution of IGD or HV
values for a certain case by one algorithm (four instances of each pro-
4.3. Experimentation and analysis of results blem scale are shown due to the lack of space, and similar results are
obtained in every instance). It can be observed that the A-MOEA/D-ID
First, a comparison of the proposed A-MOEA/D-ID with two limited outperforms the other four algorithms in most runs on both con-
variants is done to evaluate the performance of its components. Two vergence and diversity.
most significant algorithm components, adaptive penalty factor and Thus, a conclusion can be made that the proposed A-MOEA/D-ID
local search have been selected to be removed from A-MOEA/D-ID can work well on our SRALBP, and A-MOEA/D-ID do outperform the
(A0), thus forming two variants (A1 and A2) of the A-MOEA/D-ID: original MOEA/D in terms of the metrics of IGD, HV, N and computa-
A1: It differs from A0 in the absence of the self-adaption of the tional time. In addition, the enhancing mechanisms show good per-
penalty factor , i.e., using a fixed value in the algorithm. plays a formance and are much helpful to MOEA/D. The application prospect
role in balancing convergence and diversity. for the real-world engineering problems has been proved.
A2: The difference between A0 and A2 lies in the problem specified
local search operator, which is able to induce more diversity into the 4.4. Effect of the local search operator on energy saving
search mechanism.
Secondly, the MO-PSO algorithm and the basic MOEA/D are com- The local search is a critical operator in the proposed algorithm and it
pared with the complete A-MOEA/D-ID. According to the literature influences the solution quality to a considerable extent. Focusing on the
review, PSO algorithm has been widely used to solve assembly line energy consumption, a typical example is shown in Fig. 11, which illus-
balancing problems [24,47–50]. Since studies have shown the super- trates the Pareto front obtained by the complete algorithm A0 and local
iority of MOEA/D over NSGA-II in test instances with complicated PS search removed version A2 in problem BARTHOL2(148,4). It can be evi-
shapes [39,51], NSGA-II is not chosen to be compared with. The same dently observed that the local search procedure has contributed sig-
problem instances and performance indicators are used. nificantly to improving the energy efficiency and the convergence of ob-
Each case is performed 10 times by five algorithms respectively, and tained solutions. Since the global search mechanism including self-
the average of every performance indicators IGD ¯ , HV
¯ , number of so-
adaption of the penalty factor remains the same for both algorithms, the
lutions, N , and computation time T is shown in Table 3. As an ex- diversity and quantity of solutions were almost similar for both algorithms.
ample, the obtained Pareto fronts for BUXEY(29,4) is shown in Fig. 9.
The following observations can be made from the presented results. 4.5. Managerial applications

(1) According to the IGD metric, in most instances of the SRALBP, the A- In practical problems, the production efficiency of the assembly line
MOEA/D-ID clearly outperforms the MOEA/D and MO-PSO in term and the energy efficiency are two crucial but conflict objectives, and it
of the IGDs. It is indicated that the A-MOEA/D-ID can achieve the is important for the decision makers to determine the applicable

41
B. Zhou and Q. Wu Journal of Manufacturing Systems 55 (2020) 30–43

balancing solution among the great number of feasible solutions. Since Declaration of Competing Interest
different decision makers may have different opinions and ambiguity
may exist in the choice of the appropriate relative weights for each None.
objective, all the values for both objectives should be presented. A
preference vector w= (wm, wE )T can be used to denote the preference Acknowledgements
information of decision maker on number of workstations and energy
consumption respectively, where wm + wE = 1. Instead of randomly The authors would like to thank the anonymous referees for their
generating the values of wm within [0,[1]], for better application, in this constructive suggestions and comments. This work was supported by
paper the values of wm and wE are equally divided into four intervals, the National Natural Science Foundation of China [grant numbers
i.e. [0, 0.25), [0.25, 0.5), [0.5, 0.75), [0.75, 1], corresponding to four 7147, 1135].
application scenarios [52]. As an example, a medium-scale problem
HAHN(53,4) is solved by A-MOEA/D-ID and the result is illustrated by References
Fig. 12. When the corresponding interval of preference vector is iden-
tified, the number of feasible solutions for the decision makers can be [1] Salahi N, Jafari MA. Energy-Performance as a driver for optimal production plan-
cut down, helping the manager to make the decision easier. ning. Appl Energy 2016;174:88–100. https://doi.org/10.1016/j.apenergy.2016.04.
085.
Obviously, when greater importance is laid on the objective of [2] Fysikopoulos A, Anagnostakis D, Salonitis K, Chryssolouris G. An empirical study of
minimizing total energy consumption, the right-hand intervals of the the energy consumption in automotive assembly. Procedia Cirp 2012;3:477–82.
Pareto front should be selected. Meanwhile, the objective of workstation https://doi.org/10.1016/j.procir.2012.07.082.
[3] Koho M, Tapaninaho M, Heilala J, Torvinen S. Towards a concept for realizing
number becomes greater relatively. Put in a different way, when the sustainability in the manufacturing industry. J Ind Prod Eng 2015;32:33–43.
tolerance range of workstation number is given by the decision maker, a https://doi.org/10.1080/21681015.2014.1000402.
solution with the lowest total energy consumption can be provided. [4] Kishawy HA, Hegab H, Saad E. Design for sustainable manufacturing: approach,
implementation, and assessment. Sustainability 2018;10:1–15. https://doi.org/10.
Along with the referable applications for real situation production
3390/su10103604.
managements in this section, the proposed problem is believed to [5] Wang H, Zhong RY, Liu G, Mu W, Tian X, Leng D. An optimization model for energy-
provides the managers a method to create a more environmentally efficient machining for sustainable production. J Clean Prod 2019;232:1121–33.
https://doi.org/10.1016/j.jclepro.2019.05.271.
friendly robotic assembly line on the basis of guaranteeing the pro-
[6] Che A, Wu X, Peng J, Yan P. Energy-efficient bi-objective single-machine scheduling
duction efficiency, thus making the enterprises more competitive and with power-down mechanism. Comput Oper Res 2017;85:172–83. https://doi.org/
the industry more sustainable. 10.1016/j.cor.2017.04.004.
[7] Saavedra MMR, de O, Fontes CH, M. Freires FG. Renew Sustain Energy Rev
2018;82:247–59. https://doi.org/10.1016/J.RSER.2017.09.033.
5. Conclusions [8] Öztürk C, Tunalı S, Hnich B, Örnek A. Cyclic scheduling of flexible mixed model
assembly lines with parallel stations. Int J Ind Manuf Syst Eng 2015;36:147–58.
Motivated by the global energy crisis and upward trend of automatic https://doi.org/10.1016/j.jmsy.2015.05.004.
[9] Tiacci L. Mixed-model U-shaped assembly lines: balancing and comparing with
production, this paper models a bi-objective sustainable robotic assembly straight lines with buffers and parallel workstations. Int J Ind Manuf Syst Eng
line balancing problem which deals with both green manufacturing ob- 2017;45:286–305. https://doi.org/10.1016/j.jmsy.2017.07.005.
jective (the total energy consumption) and productivity-related objective [10] Akpinar Ş, Baykasoğlu A. Modeling and solving mixed-model assembly line bal-
ancing problem with setups. Part I: a mixed integer linear programming model. Int J
(number of workstations with a given cycle time). The comprehensive Ind Manuf Syst Eng 2014;33:177–87. https://doi.org/10.1016/j.jmsy.2013.11.004.
and meticulous consideration of energy consumption in different states [11] Lopes TC, Michels AS, Sikora CGS, Magatão L. Balancing and cyclical scheduling of
of robotic assembly line (processing, changeover, transportation and asynchronous mixed-model assembly lines with parallel stations. Int J Ind Manuf
Syst Eng 2019;50:193–200. https://doi.org/10.1016/j.jmsy.2019.01.001.
standby) will significantly contribute to the greening and sustainability [12] Pearce BW, Antani K, Mears L, Funk K, Mayorga ME, Kurz ME. An effective integer
of manufacturing process. An adaptive infeasibility-driven MOEA/D with program for a general assembly line balancing problem with parallel workers and
a problem specified local search operator has been proposed to solve the additional assignment restrictions. Int J Ind Manuf Syst Eng 2019;50:180–92.
https://doi.org/10.1016/j.jmsy.2018.12.011.
problem. An adaptive penalty factor scheme has been incorporated to
[13] Ogan D, Azizoglu M. A branch and bound method for the line balancing problem in
balancing the convergence and diversity of solutions. Aiming at further U-shaped assembly lines with equipment requirements. Int J Ind Manuf Syst Eng
maximizing the energy efficiency, the problem-specific local search 2015;36:46–54. https://doi.org/10.1016/j.jmsy.2015.02.007.
module was designed to further improve the performance of the algo- [14] Chica M, Bautista J, Cordón Ó, Damas S. A multiobjective model and evolutionary
algorithms for robust time and space assembly line balancing under uncertain de-
rithm. Computational experiments are designed and carried out to mand. Omega (United Kingdom) 2016;58:55–68. https://doi.org/10.1016/j.omega.
evaluate the effectiveness of the proposed algorithm. The experimental 2015.04.003.
results show that the improving strategies are of great significance to the [15] Becker C, Scholl A. A survey on problems and methods in generalized assembly line
balancing. Eur J Oper Res 2006;168:694–715. https://doi.org/10.1016/j.ejor.2004.
proposed algorithm and it outperforms MO-PSO and basic MOEA/D on 07.023.
both convergence and diversity. Managerial insights are provided to [16] Battaïa O, Dolgui A. A taxonomy of line balancing problems and their solutionap-
make tradeoffs between productivity goal and energy consumption ob- proaches. Int J Prod Econ 2013;142:259–77. https://doi.org/10.1016/j.ijpe.2012.
10.020.
jective, jointly making the whole robotic assembly line more sustainable. [17] Levitin G, Rubinovitz J, Shnits B. A genetic algorithm for robotic assembly line
Future research may consider other aspects of sustainable manu- balancing. Eur J Oper Res 2006;168:811–25. https://doi.org/10.1016/j.ejor.2004.
facturing such as the greenhouse gas emission and the carbon footprint 07.030.
[18] Rubinovitz J, Bukchin J, Lenz E. RALB – a heuristic algorithm for design and bal-
in the assembly procedure. Multi-objective optimization problems ancing of robotic assembly lines. CIRP Ann Manuf Technol 1993;42:497–500.
concerning productivity, energy and cost can be studied to give more https://doi.org/10.1016/S0007-8506(07)62494-9.
reference to decision makers. Since the attention of this study was paid [19] Gao J, Sun L, Wang L, Gen M. An efficient approach for type II robotic assembly line
balancing problems. Comput Ind Eng 2009;56:1065–80. https://doi.org/10.1016/j.
to single model, future work may include much complex real-world
cie.2008.09.027.
problems such as mixed-model assembly lines and two-sided assembly [20] Gutjahr AL, Nemhauser GL. An algorithm for the line balancing problem. Manage
lines balancing problem. Besides, there are still much can be done to Sci 1964;11:308–15. https://doi.org/10.1287/mnsc.11.2.308.
improve the algorithm. Since most of the existing studies proposed [21] Tsai D-M, Yao M-J. A line-balance-based capacity planning procedure for series-
type robotic assembly line. Int J Prod Res 1993;31:1901–20. https://doi.org/10.
approximate methods and metaheuristics as solution approaches, some 1080/00207549308956831.
effective exact approaches can be proposed to solve small sized in- [22] Kim H, Park S. A strong cutting plane algorithm for the robotic assembly line bal-
stances of the multi-objective ALBPs. Since the proposed enhancing ancing problem. Int J Prod Res 1995;33:2311–23. https://doi.org/10.1080/
00207549508904817.
mechanism is problem-specific, in our future works, the versatility of [23] Zhou B, Wu Q. A novel optimal method of robotic weld assembly line balancing
the algorithm can be exploited and application of A-MOEA/D-ID to problems with changeover times: a case study. Assem Autom 2018;38:376–86.
other engineering optimization problems can be studied. https://doi.org/10.1108/AA-02-2018-026.

42
B. Zhou and Q. Wu Journal of Manufacturing Systems 55 (2020) 30–43

[24] Mukund Nilakantan J, Ponnambalam SG. An efficient PSO for type II robotic as- algorithm: NSGA-II. IEEE Trans Evol Comput 2002;6:182–97. https://doi.org/10.
sembly line balancing problem. IEEE Int Conf Automation Sci Eng (CASE) 1109/4235.996017.
2012:600–5. https://doi.org/10.1109/CoASE.2012.6386398. [39] Peng W, Zhang Q, Li H. Comparison between MOEA/D and NSGA-II on the multi-
[25] Yoosefelahi A, Aminnayeri M, Mosadegh H, Ardakani HD. Type II robotic assembly objective travelling salesman problem. Goh CK, Ong YS, Tan K, editors. Multi-
line balancing problem: an evolution strategies algorithm for a multi-objective Objective Memetic Algorithms, 171. 233 Spring Street, New York, NY 10013,
model. Int J Ind Manuf Syst Eng 2012;31:139–51. https://doi.org/10.1016/j.jmsy. United States: Springer; 2009. p. 309–24.
2011.10.002. [40] Yan P, Wang G, Che A, Li Y. Hybrid discrete differential evolution algorithm for
[26] Rabbani M, Mousavi Z, Farrokhi-Asl H. Multi-objective metaheuristics for solving a biobjective cyclic hoist scheduling with reentrance. Comput Oper Res
type II robotic mixed-model assembly line balancing problem. J Ind Prod Eng 2016;76:155–66. https://doi.org/10.1016/J.COR.2016.06.011.
2016;33:472–84. https://doi.org/10.1080/21681015.2015.1126656. [41] Zhang Q, Li H. MOEA/D: A Multiobjective Evolutionary Algorithm Based on
[27] Zhou B, Wu Q. An improved immune clonal selection algorithm for bi-objective Decomposition. IEEE Trans Evol Comput 2007;11:712–31. https://doi.org/10.
robotic assemble line balancing problems considering time and space constraints. 1109/TEVC.2007.892759.
Eng Comput (Swansea, Wales) 2019;36:1868–92. https://doi.org/10.1108/EC-11- [42] Li H, Zhang Q. Multiobjective optimization problems with complicated pareto sets,
2018-0512. MOEA/D and NSGA-II. IEEE Trans Evol Comput 2009;13:284–302. https://doi.org/
[28] Mukund Nilakantan J, Huang GQ, Ponnambalam SG. An investigation on mini- 10.1109/TEVC.2008.925798.
mizing cycle time and total energy consumption in robotic assembly line systems. J [43] Jiang S, Yang S. An improved multiobjective optimization evolutionary algorithm
Clean Prod 2015;90:311–25. https://doi.org/10.1016/j.jclepro.2014.11.041. based on decomposition for complex pareto fronts. IEEE Trans Cybern
[29] Nilakantan JM, Ponnambalam SG, Huang GQ. Minimizing energy consumption in a 2016;46:421–37. https://doi.org/10.1109/TCYB.2015.2403131.
U-shaped robotic assembly line. 2015 Int. Conf. Adv. Mechatron. Syst. [44] Hoffmann TR. Assembly line balancing: a set of challenging problems. Int J Prod
2015:119–24. https://doi.org/10.1109/ICAMechS.2015.7287140. Res 1990;28:1807–15. https://doi.org/10.1080/00207549008942835.
[30] Li Z, Tang Q, Zhang L. Minimizing energy consumption and cycle time in two-sided [45] Bosman PAN, Thierens D. The balance between proximity and diversity in multi-
robotic assembly line systems using restarted simulated annealing algorithm. J objective evolutionary algorithms. IEEE Trans Evol Comput 2003;7:174–88.
Clean Prod 2016;135:508–22. https://doi.org/10.1016/j.jclepro.2016.06.131. https://doi.org/10.1109/TEVC.2003.810761.
[31] Zhou BH, Kang XY. A multiobjective hybrid imperialist competitive algorithm for [46] Zitzler E, Thiele L. Multiobjective evolutionary algorithms: a comparative case
multirobot cooperative assembly line balancing problems with energy awareness. study and the strength Pareto approach. IEEE Trans Evol Comput 1999;3:257–71.
Proc Inst Mech Eng Part C J Mech Eng Sci 2019;233:2991–3003. https://doi.org/ https://doi.org/10.1109/4235.797969.
10.1177/0954406218803129. [47] Chutima P, Chimklai P. Multi-objective two-sided mixed-model assembly line bal-
[32] Wolpert DH, Macready WG. No free lunch theorems for optimization. IEEE Trans ancing using particle swarm optimisation with negative knowledge. Comput Ind
Evol Comput 1997;1:67–82. https://doi.org/10.1109/4235.585893. Eng 2012;62:39–55. https://doi.org/10.1016/J.CIE.2011.08.015.
[33] Fan Z, Fang Y, Li W, Cai X, Wei C, Goodman E. MOEA/D with angle-based con- [48] Petropoulos DI, Nearchou AC. A particle swarm optimization algorithm for balan-
strained dominance principle for constrained multi-objective optimization pro- cing assembly lines. Assem Autom 2011;31:118–29. https://doi.org/10.1108/
blems. Appl Soft Comput J 2019;74:621–33. https://doi.org/10.1016/j.asoc.2018. 01445151111117700.
10.027. [49] Şahin M, Kellegöz T. A new mixed-integer linear programming formulation and
[34] Du Y, Xing L, Zhang J, Chen Y, He Y. MOEA based memetic algorithms for multi- particle swarm optimization based hybrid heuristic for the problem of resource
objective satellite range scheduling problem. Swarm Evol Comput investment and balancing of the assembly line with multi-manned workstations.
2019;50:100576https://doi.org/10.1016/J.SWEVO.2019.100576. Comput Ind Eng 2019;133:107–20. https://doi.org/10.1016/j.cie.2019.04.056.
[35] Zhang B, Pan Q, Gao L, Zhao Y. MOEA/D for multi-objective hybrid flowshop re- [50] Nilakantan JM, Ponnambalam SG. Robotic U-shaped assembly line balancing using
scheduling problem. 2018. https://doi.org/10.1115/MSEC2018-6581. particle swarm optimization. Eng Optim 2016;48:231–52. https://doi.org/10.
[36] Zhao F, Chen Z, Wang J, Zhang C. An improved MOEA/D for multi-objective job 1080/0305215X.2014.998664.
shop scheduling problem. Int J Comput Integr Manuf 2017;30:616–40. https://doi. [51] Li H, Zhang Q. Multiobjective optimization problems with complicated Pareto sets,
org/10.1080/0951192X.2016.1187301. MOEA/D and NSGA-II. IEEE Trans Evol Comput 2009;13:284–302. https://doi.org/
[37] Manupati VK, Thakkar JJ, Wong KY, Tiwari MK. Near optimal process plan selec- 10.1109/TEVC.2008.925798.
tion for multiple jobs in networked based manufacturing using multi-objective [52] Zhou B, Shen C. Multi-objective optimization of material delivery for mixed model
evolutionary algorithms. Comput Ind Eng 2013;66:63–76. https://doi.org/10. assembly lines with energy consideration. J Clean Prod 2018;192:293–305. https://
1016/J.CIE.2013.06.004. doi.org/10.1016/j.jclepro.2018.04.251.
[38] Deb K, Pratap A, Agarwal S, Meyarivan T. A fast and elitist multiobjective genetic

43