You are on page 1of 27

Computers & Industrial

Engineering
Manuscript Draft

Manuscript Number: CAIE-D-19-01866R1

Title: Eco-friendly multi-skilled worker assignment and assembly line


balancing problem

Article Type: Research Paper

Keywords: Assembly line; Bi-objective optimization; Energy consuming;


Workforce assignment

Abstract: Workforce assignment and energy consumption impact greatly


manufacturing performance. {\color{red}In this work, we study a multi-
skilled worker assignment and assembly line balancing problem with the
due consideration of energy consumption. The problem consists of
scheduling products and assigning workers to workstations appropriately
under given cycle time. Two objectives are minimized simultaneously,
i.e., (1) the total cost including the processing cost and the fixed cost
induced by employing workers, and (2) the energy consumption.} For the
problem, a bi-objective mixed-integer programming model is formulated,
and then $\epsilon$-constraint method is adopted to obtain the Pareto
front for small-scale problems. For large size problems, a processing
time and energy consumption sorted-first rule (PT-EC SFR), a multi-
objective genetic algorithm (NSGA-\uppercase\expandafter{\romannumeral2})
and a multi-objective simulated annealing method (MOSA) are developed to
solve it. Numerical experiments are conducted and computational results
show that the designed PT-EC SFR outperforms the other two algorithms in
terms of computational time and quality of solutions.
*Manuscript
Click here to view linked References

Eco-friendly multi-skilled worker assignment and assembly line


balancing problem

Abstract
Workforce assignment and energy consumption impact greatly manufacturing performance. In this
work, we study a multi-skilled worker assignment and assembly line balancing problem with the
due consideration of energy consumption. The problem consists of scheduling products and assign-
ing workers to workstations appropriately under given cycle time. Two objectives are minimized
simultaneously, i.e., (1) the total cost including the processing cost and the fixed cost induced
by employing workers, and (2) the energy consumption. For the problem, a bi-objective mixed-
integer programming model is formulated, and then -constraint method is adopted to obtain the
Pareto front for small-scale problems. For large size problems, a processing time and energy con-
sumption sorted-first rule (PT-EC SFR), a multi-objective genetic algorithm (NSGA-II) and a
multi-objective simulated annealing method (MOSA) are developed to solve it. Numerical experi-
ments are conducted and computational results show that the designed PT-EC SFR outperforms
the other two algorithms in terms of computational time and quality of solutions.
Keywords: Assembly line; Bi-objective optimization; Energy consuming; Workforce assignment

1. Introduction

In the manufacturing industry, an effective schedule is one of the crucial factors affecting the
productivity of assembly lines. Assembly line is a system containing several stations and products
that are sequently processed by workers at stations. In highly uncertain and fiercely competitive
environments, the flexibility of the production system is important and multi-skilled workers may
act as a potential buffer against varied demands (Nembhard et al., 2014). A multi-skilled worker
is more flexible than a single-skilled one and he/she can be reassigned to process various products.
Generally, multi-skilled workers acquire a set of different skills and proficiency (Bokhorst et al.,
2004). Different proficiency levels further affect the product processing time that can impact
directly energy consumption during production (Norman et al., 2002). The appropriate schedule
of multi-skilled workers can result in high efficiency and low costs. We also consider multiple types
of products in this paper. The type of chosen workers must be compatible with the type of assigned
products.
Furthermore, the world’s energy demand has doubled in the past four decades, and an estima-
tion reveals that it will double again by 2030 (Gong et al., 2018). Of the total energy consumption,
the manufacturing industry accounts for over 33% (Gong et al., 2018). It is necessary and urgent to
save energy consumption of manufacturing industry. Recently, some works on production schedule
with the consideration of energy consumption are made (Wang et al., 2015; Mansouri et al., 2016;
Zhang and Chiong, 2016; Liu et al., 2020). However, in the existing works, few of them are focusing
on saving energy with multi-skilled worker assignment.
In this paper, we address a multi-skilled worker assignment and assembly line balancing problem
considering energy consumption. We optimize the problem globally, i.e., holistic optimization from
the perspective of the entire system. The decisions are made on the schedule of products and
workforce assignment. Two objectives are considered simultaneously. The first objective is to
minimize the total cost including the processing cost and the fixed cost induced by employing
workers. The second one is to minimize energy consumption. The contributions of this paper
include: (1) the trade-off between production cost and energy consumption considering multi-
skilled workforce assignment is studied; (2) a bi-objective mixed-integer programming model is
proposed for the investigated problem; (3) a fast and efficient constructive heuristic approach
(PT-EC SFR) is designed to solve large-scale problems.
The remainder of this paper is organized as following. The relevant literature is reviewed in
Section 2. In Section 3, we describe the investigated problem. A mixed-integer programming
model for the studied problem is formulated in Section 4. In Section 5, three heuristic algorithms
are developed to solve the problem. In Section 6, the three developed algorithms are tested and

Preprint submitted to Computers & Industrial Engineering April 21, 2020


compared via numerical experiments. Finally, this work is concluded and some future research
directions are indicated.

2. Literature review

Assembly line balancing problems have been attracting much attention in recent years. There
are many existing works in the literature, such as heterogeneous worker assignment and so on.
When reviewing the literature, this paper considers four aspects, i.e., multi-type operation, multi-
skilled worker, environmental factor and multiple objective. For each work, we mainly review it
from two perspectives of (1) the investigated problem, i.e., the assembly line balancing problem
(ALBP), and (2) scheduling methods of ALBP. Finally, we conclude the review in Table 1.

2.1. Single-objective optimization


2.1.1. Multi-skilled worker
Multi-skilled workers differ in proficiency level. They are flexible, i.e., they can be assigned
to multiple types of products. Barutçuoğlu and Azizoğlu (2011) study a flexible assembly line
design problem considering multi-skilled worker. The decisions on the assignment of operations
and equipments to workstations are made to minimize the total equipment cost. They propose a
mathematical model and develop a branch and bound algorithm to solve the problem. Blum and
Miralles (2011) study the ALBP with operations and multi-skilled workers assigned to workstations.
An algorithm based on beam search is proposed to minimize the cycle time.
Zaman et al. (2012) study an assembly line balancing problem with multi-skilled worker to
minimize the weighted sum of cycle time and total idle time. They use a genetic algorithm and a
heuristic to solve the problem based on the characteristics.
Vilà and Pereira (2014) investigate the assembly line multi-skilled worker assignment and bal-
ancing problem. An exact enumeration algorithm based on the proposed model is developed to
solve the problem.
Battaı̈a et al. (2015) address an ALBP where the assignment of operations and disabled worker
(multi-skilled worker) is considered. An integer linear programming formulation and constructive
heuristics are proposed to solve the problem. Ritt et al. (2015) study an ALBP with the consider-
ation of disabled or heterogeneous workforce (i.e., multi-skilled worker) to minimize the expected
cycle time. Moreover, the uncertainty of workforce caused by high level of absenteeism is taken
into account. A two-stage mixed-integer programming model and local search are developed to
solve it. Moreira et al. (2015) address an ALBP considering multi-skilled worker. They present
a mathematical model to minimize the number of workstations. Several constructive heuristic
algorithms are proposed to cope with the problem.
Akpinar et al. (2017) address an ALBP with the consideration of setup times of operations
and multi-skilled worker. The problem is to make decisions on assignment and sequence of prod-
ucts to minimize the total number of active workstations. They develop a mixed-integer linear
programming model for the problem, which is solved by a commercial optimizer. Moreover, an
exact algorithm based on Benders decomposition is proposed to solve the problem. Oksuz et al.
(2017) investigate the problem with the U-shaped assembly line considering multi-skilled worker
and worker performance. A linear programming model is formulated to maximize line efficiency.
An artificial bee colony algorithm and a genetic algorithm are proposed to solve it. Moussavi et
al. (2017) consider the assignment of heterogeneous workers to workstations of an assembly line to
minimize the total production time. They present a mixed-integer mathematical model and two
heuristic approaches for it.
Belassiria et al. (2018) investigate an assembly line rebalancing problem considering uncertain
product demand and multi-skilled worker. They propose a mathematical model and a genetic
algorithm embedded with a heuristic procedure to maximize the line efficiency. Defersha et al.
(2018) focus on balancing and sequencing of products in an assembly line considering multi-skilled
worker. They consider three objectives, i.e., minimizing the length, minimizing the number of
workstations and minimizing the cost of workstations and product duplication. They propose a
multi-phased linear programming model and a genetic algorithm is developed with a weighted
sum of the three objectives. Pereira et al. (2018) address a multi-skilled worker ALBP in which
operation assignment and equipment decision are jointly considered. They develop a cost-oriented
formulation with different equipment costs. A genetic algorithm is adopted to solve the problem.
Pereira and Miranda (2018) address an ALBP with uncertain product processing times. A robust
formulation is put forward to minimize the number of stations. For the problem, several lower
bounds and dominance rules are proposed. An enumeration procedure is developed to solve the

2
problem. Yurtkuran et al. (2018) study an ALBP with multi-skilled worker scheduling. The goal
is to minimize the number of workers. A mixed-integer programming model is formulated and an
artificial bee colony algorithm is proposed.
Lai et al. (2019) investigate an ALBP considering the deviation of operation duration from
the initial value. The objective is to minimize the number of workstations. They construct an
algorithm to conduct the stability analysis for the problem.

2.1.2. Multi-type operation


Multi-type operation is considered in some works and each operation type must be compatible
with the type of worker assigned to process the operation.
Dolgui et al. (2018) study an assembly line planning problem with different operations and
identical skill workers to minimize the number of workers. They present a mixed-integer linear
programming model and develop randomized heuristics to solve the problem. Delorme et al. (2019)
study a paced assembly line planning problem to minimize the number of identical workers. They
consider several types of operations, lower and upper bounds of the workforce requirement and
predetermined cycle time. An integer linear programming model, an enumeration algorithm and
a dynamic programming algorithm are developed for the problem.
Multi-skilled workers and multi-type operations are considered in many works. However, ex-
isting works ignore energy consumption induced by different skilled workers.

2.2. Multi-objective optimization


For multi-objective assembly line balancing problems, Nourmohammadi and Zandieh (2011)
study an ALBP considering uncertain demand. They propose a multi-objective differential evolu-
tion algorithm to solve the problem by minimizing the cycle time and the workload smoothness
index of the assembly line simultaneously. Ramezanian and Ezzatpanah (2015) study a bi-objective
workforce assignment problem for an assembly line considering worker abilities. The first objec-
tive is to minimize the total cycle time and the second to minimize the operating costs related
to workers. A model and an evolutionary algorithm named imperialist competitive algorithm are
developed. Babazadeh et al. (2018) consider a straight and U-shaped assembly line balancing
problem. Two objectives, i.e., the number of workstations and the fuzzy cycle time are minimized
simultaneously. An efficient multi-objective genetic algorithm with a new strategy to promote pop-
ulation diversity is proposed. Zhang et al. (2019) investigate an U-shaped assembly line problem
considering environmental factors. They propose a novel multi-objective mixed-integer non-linear
programming model to minimize carbon emission, noise emission and cycle time simultaneously.
A hybrid Pareto grey wolf optimization is designed to solve the multi-objective problem.
In the above researches, the trade-off between production cost and energy in the ALBP has not
been investigated, to the best of our knowledge. In this work, we study a new bi-objective multi-
skilled workforce assignment and assembly balancing problem where (1) multi-type operations
(products) and multi-skilled workers are considered; (2) energy consumption is further considered
by assigning different skilled workers to different products. The two objectives are to minimize the
total cost and energy consumption simultaneously. Moreover, from the perspective of scheduling
methods, we can find that the literature about ALBP is mostly holistic optimization from a global
and integrated way and heuristic algorithms (such as local search, genetic algorithm and so on) are
proposed to solve the problem. In this paper, we also optimize the problem globally and propose
some artificial intelligence algorithms.
We make a comparison between reviewed works and our study in Table 1 to clearly illustrate
our novelty. We can find that we are the first to consider multi-type operations, multi-skilled
workers, environmental impact and multi-objective simultaneously.

3. Problem statement

Before describing the problem, we recall some basic definitions of the assembly line. The cycle
time of the assembly line is the duration between the first product arriving at the first station and
the last product leaving the last station. The duration between a product arriving at a station and
departing from the same station is regarded as regular processing time if the product is processed by
a worker with proficiency 1 on that station. The actual processing time is a non-increasing function
of the number of workers assigned to the station (Delorme et al., 2019). The regular processing
time and actual processing time of product n at station r are respectively denoted by Pnr and
P̂nr . The actual processing time of an operation is the multiplication of the regular processing time
and the proficiency of the workers processing it. Therefore, the actual processing time of operation

3
Table 1: Comparison of researches related to our study

literature multi- multi- considering multi- objectives approaches


type skilled environ- objective
opera- workers mental
tions impact
Nourmohammadi and Zandieh, 2011 X minimize cycle time; minimize smoothness index multi-objective differential
evolution algorithm
Barutçuoğlu and Azizoğlu, 2011 minimize the total equipment cost branch and bound
Blum and Miralles, 2011 X minimize the cycle time algorithm based on beam
search
Zaman et al., 2012 X production rate, cycle time, total idle time genetic algorithm, a heuris-
tic
Vilà and Pereira, 2014 optimal partition of work, workforce assignment branch and bound
Ramezanian and Ezzatpanah, 2015 X X minimize cycle time; minimize cost related to workers imperialist competitive al-
gorithm, genetic algorithm
Moreira et al., 2015 X minimize the number of work stations heuristic methodologies
Battaı̈a et al., 2015 minimize the total number of workers conventional and random-
ized heuristics
Ritt et al., 2015 X minimize the expected cycle time local search heuristics
Oksuz et al., 2017 X maximize of the line efficiency Artificial Bee Colony Algo-
rithm, a Genetic Algorithm
Akpinar et al., 2017 minimize the total number of active workstations Benders decomposition al-
gorithm
Moussavi et al., 2017 X minimize cycle time matheuristic approaches
Yurtkuran et al., 2018 minimize the number of worker discrete Artificial Bee

4
Colony algorithm
Babazadeh et al., 2018 X minimize the number of work stations;minimize the fuzzy cycle time Multi Objective Genetic
Algorithm
Defersha et al., 2018 minimize the length and number of workstations, costs of workstations and product duplications embedded genetic algorith-
m
Pereira et al., 2018 minimize the total cost genetic algorithm with em-
beding hybrid metaheuris-
tic
Dolgui et al., 2018 X minimize the number of workers conventional and random-
ized heuristics
Belassiria et al., 2018 maximize the line efficiency genetic algorithm (GA) hy-
bridised with a heuristic
priority rule-based proce-
dure
Pereira and Miranda, 2018 minimize the number of stations exact method with rules
and an enumeration proce-
dure
Delorme et al., 2019 X minimize the cycle time enumeration algorithm and
a dynamic programming al-
gorithm
Lai et al., 2019 minimize the number of workstations proposed algorithm
Zhang et al., 2019 X X minimize carbon emission; minimize noise emission; minimize cycle time hybrid Pareto grey wolf op-
timization, Pareto grey wolf
optimization
This paper X X X X minimize cost; minimize energy saving constructive heuristic algo-
rithm, NSGA-II and MOSA
Note: ‘X’ indicates that the corresponding item is considered in the corresponding research.
at each station is only related to worker proficiency and the number of workers. The time spent
on switching between producing two kinds of productions is called changeover time. The time is
essential to be short in actual industrial production.
Consider that a product n = 1, 2, ..., n, ..., |N | will visit successively stations 1, 2, ..., r, ..., |R| and
be processed by multi-skilled workers (the set of them is denoted by W ) with different proficiency
levels. Note that different products have different regular processing times and the number of
workers assigned to process a product is limited by an interval [a, b], which is set [1, 3] in this
paper. The proficiency of worker k at station r is denoted by αkr . If a product is processed by
several workers, the proficiency is their mean and the processing efficiency can be improved by this
cumulation assignment. This improvement can be represented by a coefficient β. For example,
a product is assigned to 1, 2 and 3 worker(s) on a station (if allowable), the improvements are
denoted by β1 , β2 and β3 respectively and we have β1 = 1 and 0 < β3 < β2 < 1.
If product n is processed at station r by one worker, the cycle time P̂nr = β1 · Pnr · αkr . When
2 (resp. 3) workers are assigned to process product n at station r, the cycle time is β2 · Pnr · (α1r +
α2r )/2 (resp. β3 ·Pnr ·(α1r +α2r +α3r )/3). Denote regular processing cost cnr as the cost of product
n at station r processed by one worker with proficiency 1. Lower cycle time means higher cost.
The actual processing cost resulted by only one worker k with proficiency αkr to operate product
n is denoted as ĉnr = cnr /(β1 · αkr ). When three workers are assigned to process product n at
station r, the actual processing cost is ĉnr = cnr / [β3 · (α1r + α2r + α3r )/3]. Concerning energy
consumption, the unit power consumption of product n at station r is denoted by Enr , then the
energy consumption of a product is Enr · P̂nr (Liu et al. 2019).
We have the following assumptions:

(1) Each worker is multi-skilled. If his/her proficiency is less than one, then he/she can
operate a product faster than a regular worker (with proficiency 1). Clearly, an unskilled
worker is with proficiency greater than 1 (Lian et al., 2018);
(2) A worker fully occupies a station until completing operations at the station, and then
he/she can move instantly to another station (Ritt et al., 2015);

(3) A worker with high proficiency is paid with a high salary and consumes less energy (Gong
et al., 2018);
(4) If several workers are assigned to a product, the proficiency of these workers is their mean
and can be improved by this cumulated assignment (This improvement can be represented
by a coefficient β̂). For example, two workers (the proficiency levels are a and b respectively)
are assigned to process one product, the proficiency of operating the product is a+b 2 · β̂;

(5) Disturbances do not frequently happen and if a disturbance occurs, the time for computing
a new schedule is short. We also assume that changeover time is short to reduce the impact
on actual production (Gungor and Evans, 2017).
An example with 4 stations, 8 products and 9 workers is illustrated in Fig. 1. The process
sequence of products is 2 → 1 → 3 → 4 → 5 → 6 → 8 → 7. The 8 products are divided into three
types: product types 1, 2, 3 denoted as N1 , N2 and N3 including product sets {1, 2, 3} (red in the
figure), {4, 5, 6} (yellow in the figure) and {7, 8} (purple in the figure) respectively. Similarly, the
9 workers are classified into three types: W1 = {1, 2, 3} (red in the figure), W2 = {4, 5, 6} (yellow
in the figure) and W3 = {7, 8, 9} (purple
S in the
S figure). Moreover, the products of N1 , N2 and N3
can be assigned to workers in W2 W3 , W1 W2 , and W3 , respectively. We can observe in Fig. 1
that after processing products 2 and 1, products 3, 4, 5 and 6 are processing at stations 4, 3, 2 and
1 respectively. As shown in Fig. 2, product 3 finishes operation and product 8 starts processing.
Each available worker is reassigned to process product after completing operation in Fig. 1. In the
Fig. 1, the actual processing time of product 6 that is processed at station 1 by workers 2 and 3
is P̂61 = β2 · P61 · (α21 + α31 )/2, the processing cost ĉ61 = c61 /[β2 · (α21 + α31 )/2], and the energy
consumption is E61 · P̂61 .

4. Problem formulation

In this section, we formulate the problem into a nonlinear model from the perspective of the
entire production system and then we make the linearization.

5
Figure 1: An example of products processing and workers assignment

Figure 2: Products moving in the assembly line

4.1. Notations
Sets and indices:

N : the set of products, indexed by m, n;


W : the set of workers, indexed by k, q;
R: the set of stations, indexed by r;
P : the set of positions, indexed by i, |P | = |N |;

Problem parameters:

Pnr : the regular processing time of product n at station r;


znkr : binary parameter, equal to 1 if the type of product n is compatible with the worker k
at station r, 0 otherwise;
αkr : the proficiency level of worker k at station r;
cnr : the regular processing cost of product n at station r;

6
Enr : the unit processing power of product n at station r;
fk : the fixed cost of worker k;
C: the cycle time;
β1 : the coefficient when assigning one worker to process product;
β2 : the coefficient when assigning two workers to process product;
β3 : the coefficient when assigning three workers to process product;
bigM : a sufficiently large positive number;

Decision variables:

snr : the starting time of product n processed at station r;


hnr : the end time of product n processed at station r;
θ̄nr : binary variable, equal to 1 if the product n is processed by one worker at station r, 0
otherwise;
θ̂nr : binary variable, equal to 1 if the product n is processed by two workers at station r, 0
otherwise;
θ̃nr : binary variable, equal to 1 if the product n is processed by three workers at station r,
0 otherwise;
xnkr : binary variable, equal to 1 if the product n is processed by worker k at station r, 0
otherwise;
ymn : binary variable, equal to 1 if the products m and n are sequently processed at some
station, 0 otherwise;
ρni : binary variable, equal to 1 if the product n is placed at position i, 0 otherwise;
gk : binary variable, equal to 1 if worker k is employed, 0 otherwise;
P̂nr : the cycle time of product n at station r;
ĉnr : the actual cost of product n at station r;

4.2. Mathematical model


XX X
min f1 = ĉnr + (gk · fk ) (1)
n∈N r∈R k∈W
XX
min f2 = (Enr · P̂nr ) (2)
n∈N r∈R

xnkr ≤ znkr n ∈ N, k ∈ W, r ∈ R (3)

X
xnkr = θ̄nr + 2θ̂nr + 3θ̃nr n ∈ N, r ∈ R (4)
k∈W

θ̄nr + θ̂nr + θ̃nr = 1 n ∈ N, r ∈ R (5)

snr ≥ hmr − (1 − ymn ) · bigM m, n ∈ N, m 6= n, r ∈ R (6)

ymn + ynm ≤ 1 m, n ∈ N, m 6= n (7)

sn(r−1) ≥ smr − (1 − ymn ) · bigM m, n ∈ N, m 6= n, r ∈ R (8)

sn(r−1) ≤ smr + (1 − ymn ) · bigM m, n ∈ N, m 6= n, r ∈ R (9)

7
hn|R| ≤ C n∈N (10)

X
ρni = 1 i ∈ P (11)
n∈N
X
ρni = 1 n ∈ N (12)
i∈P

ymn ≥ 1 − (2 − ρmi − ρn(i+1) ) · bigM m, n ∈ N, m 6= n, i ∈ P (13)

ymn ≤ 1 + (2 − ρmi − ρn(i+1) ) · bigM m, n ∈ N, m 6= n, i ∈ P (14)

xmkr + xnk(r−1) ≤ 1 + (1 − ymn ) · bigM m, n ∈ N, m 6= n, k ∈ W, r ∈ R (15)

snr ≥ hn(r−1) n ∈ N, r ∈ R \ {1} (16)

gk ≥ xnkr n ∈ N, k ∈ W, r ∈ R (17)

snr + P̂nr = hnr n ∈ N, r ∈ R (18)


P
k∈W (αkr · xnkr )
P̂nr = Pnr · (β1 · θ̄nr + β2 · θ̂nr + β3 · θ̃nr ) · P n ∈ N, r ∈ R (19)
k∈W xnkr
1
P
 θ̄ θ̂nr
nr θ̃nr  k∈W ( αkr · xnkr )
ĉnr = cnr · + + · P n ∈ N, r ∈ R (20)
β1 β2 β3 k∈W xnkr

snr , hnr , P̂nr , ĉnr ≥ 0 n ∈ N, r ∈ R (21)

θ̄, θ̂, θ̃, xnkr , ymn , ρni , gk ∈ {0, 1} m, n ∈ N, i ∈ P, k ∈ W, r ∈ R (22)


Formulas (1) and (2) are the two objectives respectively, i.e., (1) minimizing the total cost
including the processing cost and the fixed cost induced by employing workers, and (2) minimizing
the energy consumption. Constraints (3) guarantee that the type of product is compatible with
the worker type assigned to process the product. Equations (4-5) ensure that the lower and upper
bounds on the number of workers assigned to each product are respectively 1 and 3. Inequalities
(6-7) state that the relationship on time between the two products processed sequently at station r.
Constraints (8-9) indicate that the product n fully occupies the station r in the time period between
its arrival and departure. Constraints (10) are the limitation of the cycle time. Inequalities (11-14)
are the constraints on the processing order of jobs. ymn = 1 means products m, n are sequently
processed at some station or they are respectively processed at stations r and r − 1 at the same
time. Under the circumstance, constraints (15) ensure that the worker k is arranged at one station.
Constraints (16) state that the start processing time of product n at station r is no less than the
end processing time at station r − 1. Inequalities (17) ensure the worker assigned to process any
product is employed. Equations (18) state that the ending time of the product processing at station
r is equal to the starting time of the one plus the actual processing time. Equations (19-20) are
to calculate the actual processing time and cost respectively when the product n is processed by
worker(s) at station r. Constraints (21-22) are the ranges of variables.

4.3. Linearization
The mathematical model proposed above is non-linear because constraints (19-20) are non-
linear ones. In the following, we linearize the non-linear constraints without introducing any new
variables, and then the model is transformed into a mixed-integer linear programming model.
X
P̂nr ≤ β1 · Pnr · (αkr · xnkr ) + (1 − θ̄nr ) · bigM n ∈ N, k ∈ W, r ∈ R (23)
k∈W
X
P̂nr ≥ β1 · Pnr · (αkr · xnkr ) − (1 − θ̄nr ) · bigM n ∈ N, k ∈ W, r ∈ R (24)
k∈W

8
X
P̂nr ≤ β2 · Pnr · (αkr · xnkr ) + (1 − θ̂nr ) · bigM n ∈ N, k ∈ W, r ∈ R (25)
k∈W
X
P̂nr ≥ β2 · Pnr · (αkr · xnkr ) − (1 − θ̂nr ) · bigM n ∈ N, k ∈ W, r ∈ R (26)
k∈W
X
P̂nr ≤ β3 · Pnr · (αkr · xnkr ) + (1 − θ̃nr ) · bigM n ∈ N, k ∈ W, r ∈ R (27)
k∈W
X
P̂nr ≥ β3 · Pnr · (αkr · xnkr ) − (1 − θ̃nr ) · bigM n ∈ N, k ∈ W, r ∈ R (28)
k∈W

cnr X 1 
ĉnr ≤ · · xnkr + (1 − θ̄nr ) · bigM n ∈ N, k ∈ W, r ∈ R (29)
β1 αkr
k∈W

cnr X 1 
ĉnr ≥ · · xnkr − (1 − θ̄nr ) · bigM n ∈ N, k ∈ W, r ∈ R (30)
β1 αkr
k∈W

cnr X 1 
ĉnr ≤ · · xnkr + (1 − θ̂nr ) · bigM n ∈ N, k ∈ W, r ∈ R (31)
β2 αkr
k∈W

cnr X 1 
ĉnr ≥ · · xnkr − (1 − θ̂nr ) · bigM n ∈ N, k ∈ W, r ∈ R (32)
β2 αkr
k∈W

cnr X 1 
ĉnr ≤ · · xnkr + (1 − θ̃nr ) · bigM n ∈ N, k ∈ W, r ∈ R (33)
β3 αkr
k∈W

cnr X 1 
ĉnr ≥ · · xnkr − (1 − θ̃nr ) · bigM n ∈ N, k ∈ W, r ∈ R (34)
β3 αkr
k∈W

bigM is a sufficiently large positive number. We assume that the productP n is processed by
one worker at station r, i.e., θ̄nr = 1. From constraints (23-24), P̂nr = β1 · Pnr · k∈W (αkr · xnkr ).
Constraints (25-26) and (27-28) are corresponding to θ̂nr = 1, θ̃nr = 1, respectively. In other words,
constraints (19) can be replaced by constraints (23-28). Similarly, constraints (20) are equivalent
to constraints (29-34).

5. Solving approaches

In the following section, we get the Pareto front of this problem by -constraint method, which
is only applied when the size of the problem is small. The optimization of the assembly line problem
with worker assignment is known to be NP-hard (Vilà and Pereira, 2014) and it is extremely time-
consuming under large-scale problems by -constraint method. Heuristic algorithms are appropriate
to solve large-scale problems (Chen et al., 2017; Chen et al., 2019b). We develop three heuristic
algorithms to solve the investigated problem under large-scale instances.
Firstly, some notations are recalled in the following description. In this paper, Pareto front
consists of a serious of non-dominated solutions (f1 , f2 ), where f1 denotes the value of first ob-
jective, and f2 is the second objective. For two solutions (f11 , f21 ), (f12 , f22 ) from Pareto front, if
f12 ≤ f11 and f22 ≤ f21 , where at least one inequality is strict, we call that (f12 , f22 ) dominates (f11 ,
f21 ). Otherwise, the two are non-dominated solutions.
We first adopt -constraint (Wu et al., 2018) to get the Pareto solutions under small-scale
problems and then we introduce three algorithms to solve the investigated problem when the s-
cale grows. The first two methods are classic heuristics including NSGA-II and MOSA. Another
proposed algorithm is based on the sorted multiplication of regular processing time and energy
consuming called processing time and energy consuming sorted-first rule (PT-EC SFR). The com-
putational results reveal that PT-EC SFR is better than NSGA-II and MOSA from perspectives
of solution quality and calculation time.

5.1. -constraint method


-constraint method is one of the most widely (Mavrotas, 2009; Balaman, 2016; Xing et al.,
2019) and well-organized techniques to handle the multi-objective structure of complex problem
(Haimes et al., 1971). The core idea of the method is to minimize only one objective function by

9
limiting other objective(s). In this paper, we adopt -constraint method by dividing the interval
of  uniformly into a number of subintervals (we set the length of subinterval 20 in this paper, i.e.,
4=20). The following concepts are needed when using -constraint method in this paper.

• Ideal point: let f I =(f1I , f2I ) with f1I = min{f1 (X)} and f2I = min{f2 (X)}, X ∈ Z V ;
• Nadir point: let f N =(f1N , f2N ) with f1N = min{f1 (X) : f2 (X) = f2I } and f2N = min{f2 (X) :
f1 (X) = f1I }, X ∈ Z V ;

• Extreme point: f E ={f1I , f2N } and f E = {f1N , f2I } are two extreme points on the Pareto
front.

In this paper, the two objectives are the total cost and energy consumption. The total cost is
set to be the first objective and the other is energy consumption. The detailed -constraint method
is showed in Algorithm 1.

Algorithm 1 Pseudo-code of the -constraint method


1: Compute the Ideal point fI = (f1I ,f2I ) and Nadir point fN = (f1N ,f2N ).
2: Set z0 ={(f1N ,f2I )} and  = f1N - 4.
3: while  ≥ f1I do
4: Solve the -constraint problem by adding f1 ≤  as an additional constraint and the f2 as the single
objective function to optimality.
5: Add the optimal solution value (f1∗ ,f2∗ ) to z0 .
6: Set  = f1∗ - 4.
7: end while
8: Obtain the Pareto front z by removing dominated points from z0 .

5.2. Non-dominated solution genetic algorithm(NSGA-II)


NSGA-II is one of the most popular GA-based methods and employed in a variety of applications
(Babazadeh et al., 2018). As an evolutionary algorithm, it integrates a fast non-dominated sorting
procedure and an elitist-preserving approach. The approach is widely used to solve multi-objective
problems effectively (Bandyopadhyay et al., 2013; Brownlee and Wright, 2015; Wang et al., 2017)
and the general procedures are shown in Fig. 3.

5.2.1. Initialization
The chromosome shown in Fig. 4 is divided into two parts. The first is to denote the processing
sequence and the second is the arrangement of workers. For example, there are 3 products and 2
stations. The length of one chromosome is 3+3*2*3=21, i.e., 21 positions in the chromosome. The
first three positions are respectively 3, 1, and 2 if the process sequence is 3 → 1 → 2. The remaining
18 positions denote the workers arranged to process the product. The first three positions denote
workers 2 and 3 are arranged to process product 1 at station 1. 0 in the position denotes no
arranged worker.
We generate the product process sequence randomly. And then, we choose compatible worker(s)
for each product. For example, under the status of Fig. 1, products 3, 4, 5, 6 are processing
at stations 4, 3, 2 and 1 respectively. We first filter the compatible workers for each type of
products and then generate an integer in {1, 2, 3} to denote the number of workers for each product
respectively. Finally, we choose worker(s) to process the products from compatible workers. Under
the next status shown in Fig. 2, we generate the assignment of workers similarly. As for the
solution exceeding the given cycle time, we give a penalty on the two objectives.

5.2.2. Genetic operation


We set two kinds of genetic operations. The first operation is to choose one chromosome
randomly in which we choose two products randomly and exchange their processing sequence. And
then, the worker arrangements of the two products are changed accordingly (reselect compatible
worker(s) for the two products from remaining workers). The second operation is to reselect
compatible worker(s) for each product.

10
Figure 3: The procedure of NSGA-II

Figure 4: The presentation of one chromosome

5.3. Multi-objective simulated annealing method


Kirkpatrick et al. (1987) first proposed simulated annealing (SA), which uses the concept of
the annealing process to seek optimal solution from the feasible domain, to address the complexity
of single objective problems. The algorithm iterates from a set of initial solutions with a high
temperature, Tmax , and finishes with a final temperature, Tf inally . Under each current temper-
ature, there are M arkov length iterations. The new solutions generated in each iteration are
received with probability, whose value is related to the current temperature. An annealing rate
tempcool ∈ (0, 1) is used to determine the speed of annealing. The smaller value of tempcool means
annealing with a lower pace which can approach the optimal solutions gradually.
However, there are two objectives in this paper. A modified SA algorithm is developed to
find appropriate Pareto solutions rather than getting one used in a single objective environment.
Based on the framework of Bandyopadhyay et al. (2008), the multi-objective simulated annealing
(MOSA) method is proposed to solve our problem and the steps are shown in Fig. 5.
The initialization operations are similar to the ones of NSGA-II. As for perturbation, we choose
one solution in each iteration under some temperature and modify the process sequence through
exchanging two positions randomly. And then, we reselect the compatible workers from remaining
workers to process the changed products.

5.4. Constructive heuristic method


We propose a method to solve the problem. The key step of the rule is to choose worker(s) to
process product n on station r according to some probability which is related to the multiplication
of regular processing time and energy consuming of the product n at station r. We call the method

11
Figure 5: The procedure of MOSA

PT-EC SFR (processing time and energy consuming sorted-first rule), and the detailed steps are
given in Algorithm 2. In this problem, the algorithm outperforms the two other algorithms in
terms of solution quality and running time.

6. Computational experiments

In this section, we first tune the parameter values in the NSGA-II and MOSA to find the
best parameter combinations. Then we make a comparison among the three algorithms using two
indicators. Moreover, we discuss the changeover time after one disturbance under a single-objective
in the Appendix. All these algorithms are coded in Python 3.7, and are conducted on PC with
3.90 GHz Intel Core (TM) CPU processor and 4GB RAM memory.

6.1. Tuning
We generate the regular processing time (Pnr ) in [5, 25] (Ritt et al., 2015), regular processing
cost (cnr ) in [20, 50] (Manavizadeh et al., 2013), unit processing power in [0.80, 1.20] (Zhang et
al., 2019), the proficiency (αkr ) in [0.60, 1.30] (Manavizadeh et al., 2013) and fixed cost (fk ) in
[50, 200] randomly (Ramezanian and Ezzatpanah, 2015). Concerning other parameters, we set
β1 = 1, β2 = 0.6, β3 =P 0.4, and C grows with the number of processing products. znkr is generated
randomly and we set k znkr ≥ |W | − 2 for more feasible solutions. When tuning parameters, we
set {|N |, |W |, |R|} = {5, 6, 2} and C = 80.
In NSGA-II, the pop is in {150, 200, 250, 300} and gen is in {300, 400, 500, 600, 700, 800}. The
results are given in Table 2. The ID (The smaller ID is, the better are the solutions) is one
indicator to measure the performance of NSGA-II with Pareto solutions got by -constraint method
as reference set and the detail information of this indicator is introduced in the following section.
From the table, we can find ID is minimal when pop = 300 and gen = 500. That is to say NSGA-II
performs best under the circumstance.
Similarity, we set Tmax = {400, 500, 600, 700, 800}, Tf inally = {0.01, 0.1, 1.00, 3.00, 500} and
tempcool = {0.90, 0.95, 0.97, 0.99}. From Table 3, the best combination of {Tmax , Tf inally , tempcool }
is {400, 0.01, 0.99}.

12
Algorithm 2 Pseudo-code of the PT-EC SFR
1: Calculate the multiplication Λnkr of regular processing time and energy consumption of each
worker when processing product n on station r.
2: Let SETnr denote the compatible worker for product n processing on station r.
3: Let x denote the set of product process sequence and Iter the number of iterations.
4: Calculate the sum of Λnkr in the k dimension as SU Mnr and then calculate the probabilities
P ROnkr = Λnkr /SU Mnr where k ∈ SETnr .
5: for j in Iter do
6: for i in x do
7: if worker k is not chosen then
8: Generate an integer N U M randomly to denote the number of workers processing prod-
uct n on station r;
9: Choose N U M workers from SETnr according to the P ROnkr (the P ROnkr denote the
selected probability of worker k to process product n on station r);
10: end if
11: non-dominated sort for all solutions and eliminate the solutions whose ranks are not one.
12: end for
13: non-dominated sort for current solutions and get the final solutions.
14: end for

Table 2: Computational results for parameter tuning of NSGA-II

{pop, gen} ID {pop, gen} ID {pop, gen} ID


{150 ,300} 0.1785 {200 ,500} 0.1529 {250 ,700} 0.1521
{150 ,400} 0.1761 {200 ,600} 0.1587 {250 ,800} 0.1462
{150 ,500} 0.1432 {200 ,700} 0.1374 {300 ,300} 0.1625
{150 ,600} 0.1517 {200 ,800} 0.1372 {300 ,400} 0.1562
{150 ,700} 0.1527 {250 ,300} 0.1561 {300 ,500} 0.1316
{150 ,800} 0.1451 {250 ,400} 0.1472 {300 ,600} 0.1412
{200 ,300} 0.1392 {250 ,500} 0.1563 {300 ,700} 0.1514
{200 ,400} 0.1431 {250 ,600} 0.1434 {300 ,800} 0.1592
Note: Bold format means the targeted parameter combination.

6.2. Performance Measurement


It is necessary to evaluate the solving methods (Chen et al., 2019a; Guo et al., 2019). In this
section, we adopt two indicators to make a comparison among the three algorithms. The solution
reference set is comprised of solutions solved by the three algorithms.

6.2.1. Introduction of indicators


Since the optimal solutions on the Pareto front cannot dominate each other, the performance
evaluation of multi-objective optimization cannot be assessed by a unique numerical value as that
of single-objective optimization. Some techniques are proposed to compare the multi-objective
solutions. In this paper, we adopted two indicators, which have been widely used in previous
studies.
The first one is Distance from Reference Set Indicator ID . It is used to measure the ‘distance’
between the approximation Pareto front and the true Pareto front (Coello and Cortés, 2005). A
smaller ID means the approach performs better. For an approximation non-dominated solution
set A and a reference solution set Ω, ID is calculated by the following expression:
1 X
ID (A, Ω) = min d(x, y)
|Ω| x∈A
y∈Ω

where d(x, y) presents the distance between solutions x ∈ A and y ∈ Ω. M is the number of
objectives (M = 2 in this paper), and fimin , fimax , i ∈ [1, M ] are minimum and maximum values
of the ith objective in set Ω, respectively.
v
uM 
uX fi (x) − fi (y) 2
d(x, y) = t
i=1
fimax − fimin

13
Table 3: Computational results for parameter tuning of MOSA

{Tm , Tf , tempc } ID {Tm , Tf , tempc } ID {Tm , Tf , tempc } ID {Tm , Tf , tempc } ID


{400, 0.01, 0.90} 0.2741 {500, 0.10, 0.95} 0.3123 {600, 1.00, 0.97} 0.2568 {700, 3.00, 0.99} 0.2531
{400, 0.01, 0.95} 0.3022 {500, 0.10, 0.97} 0.2810 {600, 1.00, 0.99} 0.2761 {700, 5.00, 0.90} 0.3285
{400, 0.01, 0.97} 0.2729 {500, 0.10, 0.99} 0.3290 {600, 3.00, 0.90} 0.3080 {700, 5.00, 0.95} 0.2727
{400, 0.01, 0.99} 0.2402 {500, 1.00, 0.90} 0.2811 {600, 3.00, 0.95} 0.2926 {700, 5.00, 0.97} 0.3242
{400, 0.10, 0.90} 0.3286 {500, 1.00, 0.95} 0.3145 {600, 3.00, 0.97} 0.2795 {700, 5.00, 0.99} 0.2626
{400, 0.10, 0.95} 0.2923 {500, 1.00, 0.97} 0.2532 {600, 3.00, 0.99} 0.2542 {800, 0.01, 0.90} 0.2917
{400, 0.10, 0.97} 0.2754 {500, 1.00, 0.99} 0.2536 {600, 5.00, 0.90} 0.2621 {800, 0.01, 0.95} 0.2982
{400, 0.10, 0.99} 0.2986 {500, 3.00, 0.90} 0.2751 {600, 5.00, 0.95} 0.2845 {800, 0.01, 0.97} 0.3123
{400, 1.00, 0.90} 0.2816 {500, 3.00, 0.95} 0.2907 {600, 5.00, 0.97} 0.3119 {800, 0.01, 0.99} 0.3125
{400, 1.00, 0.95} 0.3262 {500, 3.00, 0.97} 0.2996 {600, 5.00, 0.99} 0.2772 {800, 0.10, 0.90} 0.2723
{400, 1.00, 0.97} 0.2665 {500, 3.00, 0.99} 0.2901 {700, 0.01, 0.90} 0.2612 {800, 0.10, 0.95} 0.2932
{400, 1.00, 0.99} 0.3341 {500, 5.00, 0.90} 0.2926 {700, 0.01, 0.95} 0.3079 {800, 0.10, 0.97} 0.3021
{400, 3.00, 0.90} 0.2891 {500, 5.00, 0.95} 0.2790 {700, 0.01, 0.97} 0.3224 {800, 0.10, 0.99} 0.2941
{400, 3.00, 0.95} 0.2849 {500, 5.00, 0.97} 0.3083 {700, 0.01, 0.99} 0.2956 {800, 1.00, 0.90} 0.2981
{400, 3.00, 0.97} 0.2521 {500, 5.00, 0.99} 0.3126 {700, 0.10, 0.90} 0.3133 {800, 1.00, 0.95} 0.3001
{400, 3.00, 0.99} 0.3058 {600, 0.01, 0.90} 0.2823 {700, 0.10, 0.95} 0.3152 {800, 1.00, 0.97} 0.3166
{400, 5.00, 0.90} 0.2922 {600, 0.01, 0.95} 0.3233 {700, 0.10, 0.97} 0.2888 {800, 1.00, 0.99} 0.2466
{400, 5.00, 0.95} 0.2715 {600, 0.01, 0.97} 0.2712 {700, 0.10, 0.99} 0.2805 {800, 3.00, 0.90} 0.2677
{400, 5.00, 0.97} 0.3012 {600, 0.01, 0.99} 0.3254 {700, 1.00, 0.90} 0.2911 {800, 3.00, 0.95} 0.3026
{400, 5.00, 0.99} 0.3238 {600, 0.10, 0.90} 0.2878 {700, 1.00, 0.95} 0.3155 {800, 3.00, 0.97} 0.3090
{500, 0.01, 0.90} 0.3126 {600, 0.10, 0.95} 0.2972 {700, 1.00, 0.97} 0.2777 {800, 3.00, 0.99} 0.2784
{500, 0.01, 0.95} 0.3116 {600, 0.10, 0.97} 0.2746 {700, 1.00, 0.99} 0.2689 {800, 5.00, 0.90} 0.2746
{500, 0.01, 0.97} 0.2896 {600, 0.10, 0.99} 0.2679 {700, 3.00, 0.90} 0.3012 {800, 5.00, 0.95} 0.2677
{500, 0.01, 0.99} 0.2669 {600, 1.00, 0.90} 0.2774 {700, 3.00, 0.95} 0.2785 {800, 5.00, 0.97} 0.3101
{500, 0.10, 0.90} 0.3051 {600, 1.00, 0.95} 0.2958 {700, 3.00, 0.97} 0.2769 {800, 5.00, 0.99} 0.3068
Note: Bold format means the targeted parameter combination.

The second indicator is called Average e-dominance (AD), which can measure the average
distance from a solution set A to the reference set Ω (Ziter et al., 2003). The indicator AD is
defined below.
1 X
AD(A, Ω) = e(x, y)
|Ω|
y∈Ω

where e(x, y) presents that a point (f1 (x), f2 (x)) ∈ A is e-dominated a point (f2 (x), f2 (y)) ∈ Ω.
e(x, y) < 1 means that (f1 (y), f2 (y)) is dominated by (f1 (x), f2 (x)). e(x, y) = 1 indicates that
at least one dimension of the two points are overlapping. That is to say larger AD means the
approach is better. e(x, y) is calculated by:
 
f1 (x) f2 (x)
e(x, y) = min max ,
x∈A f1 (y) f2 (y)
Exactly, the reference set Ω in the indicator expression refers to the true Pareto front. However,
it is extremely difficult to obtain the true Pareto front, especially for large-scale instances. In this
situation, Ω is formed by the total solutions obtained by the three algorithms.

6.2.2. Comparison of proposed algorithms


In the comparison, we generate 20 instances and the results are listed in Table 4. Column
‘Instance’ denotes different parameter combinations ({|N |, |W |, |R|}). From the table, we can find
the results of the two indicators are consistent. The PT-EC SFR is the best of the three algorithms
in terms of computational time and the quality of solutions.
From the perspective of the solution, NSGA-II and PT-EC SFR are better than MOSA for
each parameter combination under the two indicators. In terms of average ID , the two algorithms
(NSGA-II and PT-EC SFR) are better than MOSA about 73.42% and 74.21% respectively. Under
the indicator AD, the percents are respectively 0.92% and 1.10%. The mean values of the ID
indicator are 0.0294 and 0.0303 for the two algorithms (NSGA-II and PT-EC SFR), which states
that the PT-EC SFR is about 2.97% better than NSGA-II. Concerning the indicator AD, we can
find the PT-EC SFR is better than NSGA-II from the mean values, which are 1.0033 and 1.0015
severally.

14
Table 4: Computational results for 20 instance combinations

NSGA-II MOSA PT-EC SFR


Instance
ID (A1 , Ω) AD(A1 , Ω) Time(s) ID (A2 , Ω) AD(A2 , Ω) Time(s) ID (A3 , Ω) AD(A3 , Ω) Time(s)
{5, 6, 2} 0.0375 1.0172 299.59 0.1385 0.9825 931.72 0.0237 1.0071 184.43
{5, 9, 3} 0.0570 1.0004 394.83 0.1466 1.0016 1579.97 0.0478 1.0035 168.11
{5, 12, 4} 0.0434 1.0003 493.31 0.1561 0.9997 1768.92 0.0393 1.0054 153.90
{5, 15, 5} 0.0353 1.0004 623.36 0.0399 0.9994 878.60 0.0340 1.0022 184.11
{10, 6, 2} 0.0344 1.0007 568.15 0.1764 0.9995 1741.39 0.0243 1.0026 155.73
{10, 9, 3} 0.0302 1.0221 788.09 0.1001 0.9871 2350.28 0.0312 1.0102 157.83
{10, 12, 4} 0.0287 1.0005 800.62 0.0918 0.9994 1488.03 0.0191 1.0042 178.68
{10, 15, 5} 0.0509 1.0004 734.93 0.1287 0.9998 2519.77 0.0485 1.0056 235.32
{15, 6, 2} 0.0218 1.0006 765.20 0.1047 1.0011 2681.24 0.0290 1.0050 171.81
{15, 9, 3} 0.0236 1.0007 554.82 0.1055 0.9999 1063.73 0.0288 1.0034 168.92
{15, 12, 4} 0.0223 1.0007 609.50 0.0968 0.9046 1705.89 0.0251 1.0054 205.55
{15, 15, 5} 0.0257 1.0006 770.52 0.1137 1.0000 1386.75 0.0335 1.0022 285.23
{20, 6, 2} 0.0222 1.0005 1018.63 0.1066 1.0000 1488.01 0.0240 1.0052 155.87
{20, 9, 3} 0.0228 1.0009 937.45 0.1229 0.9994 1467.19 0.0270 1.0027 182.22
{20, 12, 4} 0.0284 1.0006 1019.59 0.0963 0.9997 1132.56 0.0262 1.0034 258.39
{20, 15, 5} 0.0193 0.9796 959.06 0.0969 0.9778 1131.04 0.0338 0.9837 343.01
{25, 6, 2} 0.0313 1.0011 1193.48 0.0966 0.9978 1453.14 0.0235 1.0039 164.97
{25, 9, 3} 0.0245 1.0007 765.34 0.1270 0.9992 1618.93 0.0245 1.0026 192.08
{25, 12, 4} 0.0199 1.0006 634.96 0.1307 0.9991 1813.94 0.0191 1.0044 283.13
{25, 15, 5} 0.0264 1.0009 892.74 0.1048 1.0009 1235.60 0.0252 1.0026 398.78
Average 0.0303 1.0015 741.21 0.1140 0.9924 1571.83 0.0294 1.0033 211.40
Note: Bold format means the best algorithm.

From the perspective of running time, the mean running times of the three algorithms are
741.21s, 1571.83s and 211.40s respectively. The MOSA is the slowest one and PT-EC SFR is the
quickest one. The PT-EC SFR has a greater advantage in computational time than NSGA-II,
which is about 71.48%. For NSGA-II and PT-EC SFR, they are much better than MOSA.
For vividly demonstrating the comparison of the three algorithms, we draw some insight pictures
under the combinations of {|N | = 10, |W | = 9, |R| = 3} and {|N | = 25, |W | = 15, |R| = 5}. The
pictures are shown in Fig. 6, and Fig. 7. We also get the observation that our proposed algorithm
(i.e., PT-EC SFR) is better than the two heuristic algorithms (i.e., NSGA-II and MOSA).

7. Conclusion

In this paper, we study a workforce assignment and assembly line balancing problem with a
given cycle time. Simultaneously, we take environmental protection into account described by
energy consumption. To the best of our knowledge, we are the first to study the trade-off of
the total cost and environment protection with two objectives in this topic concurrently. The
first objective is to minimize the total cost and minimizing the volume of energy consumption is
regarded as the second objective.
To find a trade-off relation between the two objectives, we establish a bi-objective mathematical
model from the perspective of the entire production system. An -constraint method is introduced
to obtain Pareto front for small scale instances. Two classic heuristic algorithms, i.e., NSGA-II and
MOSA, are proposed to solve large-scale problems. Moreover, we propose an algorithm (PT-EC
SFR) based on sort rule. Computational experiments are conducted to illustrate the efficiency of
the three algorithms. We can conclude that the PT-EC SFR outperforms NSGA-II and MOSA in
this problem in terms of the quality of solutions and computing time.
Future research directions may cover the following aspects:
(1) In practical, the processing time of product may be indeterministic due to uncertain
machine maintenance or the damage to product during transportation. It may be necessary
and interesting to consider the uncertainty of processing time.

(2) We cannot get the true Pareto solutions by heuristic. It may be indispensable to develop
exact algorithms to solve the problem in actual production.

15
Figure 6: The comparison of the three algorithms (|N | = 10, |W | = 9, |R| = 3)

Reference
[1] Akpinar S, Elmi A, Bektaş T (2017), Combinatorial benders cuts for assembly line balancing
problems with setups, European Journal of Operational Research, 259(2): 527-537.
[2] Babazadeh H, Alavidoost M H, Fazel M H F, Sayyari S T (2018), An enhanced NSGA-II
algorithm for fuzzy bi-objective assembly line balancing problems, Computers & Industrial
Engineering, 123: 189-208.
[3] Balaman Ş Y (2016), Investment planning and strategic management of sustainable systems
for clean power generation: An -constraint based multi-objective modelling approach, Journal
of Cleaner Production, 137: 1179-1190.
[4] Bandyopadhyay S, Bhattacharya R (2013), Solving multi-objective parallel machine scheduling
problem by a modified NSGA-II, Applied Mathematical Modelling, 37(10-11): 6718-6729.
[5] Bandyopadhyay S, Saha S, Maulik U, Deb K (2008), A simulated annealing-based multiobjec-
tive optimization algorithm: amosa, IEEE Transactions on Evolutionary Computation, 12(3):
269-283.
[6] Barutçuoğlu Ş, Azizoğlu M (2011), Flexible assembly line design problem with fixed number
of workstations, International Journal of Production Research, 49(12): 3691-3714.
[7] Battaı̈a O, Delorme X, Dolgui A, Hagemann J, Horlemann A, Kovalev S, Malyutin S (2015),
Workforce minimization for a mixed-model assembly line in the automotive industry, Inter-
national Journal of Production Economics, 170: 489-500.
[8] Belassiria I, Mazouzi M, ELfezazi S, Cherrfi A, Elmaskaoui Z (2018), An integrated model
for assembly line re-balancing problem, International Journal of Production Research, 56(16):
5324-5344.
[9] Blum C, Miralles C (2011), On solving the assembly line worker assignment and balancing
problem via beam search, Computers & Operations Research, 38(1): 328-339.
[10] Bokhorst J A C, Slomp J, Gaalman G J C (2004), On the who-rule in dual resource constrained
(DRC) manufacturing systems, International Journal of Production Research, 42(23): 5049-
5074.
[11] Brownlee A E I, Wright J A (2015), Constrained, mixed-integer and multi-objective optimi-
sation of building designs by NSGA-II with fitness approximation, Applied Soft Computing,
33: 114-126.

16
Figure 7: The comparison of the three algorithms (|N | = 25, |W | = 15, |R| = 5)

[12] Chen C, Ding Y, Wang Z, Zhao J F, Guo B, Zhang D Q (2019a), VTracer: when online
vehicle trajectory compression meets mobile edge computing, IEEE Systems Journal, DOI:
10.1109/JSYST.2019.2935458.
[13] Chen C, Ding Y, Xie X F, Zhang S, Wang Z, Feng L (2019b), Trajcompressor: an online
map-matching-based trajectory compression framework leveraging vehicle heading direction
and change, IEEE Transactions on Intelligent Transportation Systems, DOI: 10.1109/TIT-
S.2019.2910591.
[14] Chen C, Zhang D Q, Ma X J, Guo B, Wang L Y, Wang Y S, Sha E (2017), Crowddeliver:
planning city-wide package delivery paths leveraging the crowd of taxis, IEEE Transactions
on Intelligent Transportation Systems, 18(6): 1478-1496.
[15] Coello C A C, Cortés N C (2005), Solving multiobjective optimization problems using an
artificial immune system, Genetic Programming & Evolvable Machines, 6(2): 163-190.
[16] Defersha F M, Mohebalizadehgashti F (2018), Simultaneous balancing, sequencing, and work-
station planning for a mixed model manual assembly line using hybrid genetic algorithm,
Computers & Industrial Engineering, 119: 370-387.
[17] Delorme X, Dolgui A, Kovalev S, Kovalyov M Y (2019), Minimizing the number of workers
in a paced mixed-model assembly line, European Journal of Operational Research, 272(1):
188-194.

[18] Dolgui A, Kovalev S, Kovalyov M Y, Malyutin S, Soukhal A (2018), Optimal workforce


assignment to operations of a paced assembly line, European Journal of Operational Research,
264(1): 200-211.
[19] Gong G, Deng Q, Gong X, Liu W, Ren Q (2018), A new double flexible job-shop scheduling
problem integrating processing time, green production, and human factor indicators, Journal
of Cleaner Production, 174: 560-576.
[20] Gungor Z E, Evans S (2017), Understanding the hidden cost and identifying the root causes
of changeover impacts, Journal of Cleaner Production, 167: 1138-1147.
[21] Guo S M, Chen C, Wang J Y, Liu Y X, Xu K, Yu Z W, Zhang D Q, Chiu D M (2019),
ROD-revenue: seeking strategies analysis and revenue prediction in ride-on-demand service
using multi-source urban data, IEEE Transactions on Mobile Computing, DOI: 10.1109/TM-
C.2019.2921959.

17
[22] Haimes Y, Lasdon L, Wismer D (1971), On a bicriterion formulation of the problems of
integrated system identification and system optimization, IEEE Transactions on Systems,
Man, and Cybernetics, 1(3): 296-297.
[23] Kirkpatrick S, Gelatt J, Vecchi M P (1987), Optimization by simulated annealing, Readings
in Computer Vision, 220(4598): 606-615.
[24] Lai T C, Sotskov Y N, Dolgui A (2019), The stability radius of an optimal line balance with
maximum efficiency for a simple assembly line, European Journal of Operational Research,
274(2): 466-481.
[25] Lian J, Liu C G, Li W J, Yin Y (2018), A multi-skilled worker assignment problem in seru pro-
duction systems considering the worker heterogeneity, Computers & Industrial Engineering,
118: 366-382.
[26] Liu M, Yang X N, Chu F, Zhang J T, Chu C B (2020), Energy-oriented bi-objective optimiza-
tion for the tempered glass scheduling, Omega, 90: 12-32.

[27] Liu Z C, Guo S S, Wang L (2019), Integrated green scheduling optimization of flexible job shop
and crane transportation considering comprehensive energy consumption, Journal of Cleaner
Production, 211: 765-786.
[28] Manavizadeh N, Hosseini N S, Rabbani M, Jolai F (2013), A Simulated Annealing algorithm
for a mixed model assembly U-line balancing type-I problem considering human efficiency and
just-in-time approach, Computers & Industrial Engineering, 64(2): 669-685.
[29] Mansouri S A, Aktas E, Besikci U (2016), Green scheduling of a two-machine flowshop: trade-
off between makespan and energy consumption, European Journal of Operational Research,
248(3): 772-788.
[30] Mavrotas G (2009), Effective implementation of the -constraint method in multi-objective
mathematical programming problems, Applied Mathematics and Computation, 213(2): 455-
465.
[31] Moreira M C O, Miralles C, Costa A M (2015), Model and heuristics for the assembly line
worker integration and balancing problem, Computers & Operations Research, 54: 64-73.

[32] Moussavi S E, Mahdjoub M, Grunder O (2017), Productivity improvement through a sequenc-


ing generalised assignment in an assembly line system, International Journal of Production
Research, 55(24): 7509-7523.
[33] Nembhard D A, Bentefouet F (2014), Selection policies for a multifunctional workforce, In-
ternational Journal of Production Research, 52(16): 4785-4802.

[34] Norman B A, Tharmmaphornphilas W, Needy K L, Bidanda B, Warner, R C (2002), Worker


assignment in cellular manufacturing considering technical and human skills, International
Journal of Production Research, 40(6): 1479C1492.
[35] Nourmohammadi A, Zandieh M (2011), Assembly line balancing by a new multi-objective dif-
ferential evolution algorithm based on TOPSIS, International Journal of Production Research,
49(10): 2833-2855.
[36] Oksuz M K, Buyukozkan K, Satoglu S I (2017), U-shaped assembly line worker assignment and
balancing problem: a mathematical model and two meta-heuristics, Computers & Industrial
Engineering, 112: 246-263.

[37] Pereira J, Miranda E Á (2018), An exact approach for the robust assembly line balancing
problem, Omega, 78: 85-98.
[38] Pereira J, Ritt M, Vásquez Ó C (2018), A memetic algorithm for the cost-oriented robotic
assembly line balancing problem, Computers & Operations Research, 99: 249-261.
[39] Ramezanian R, Ezzatpanah A (2015), Modeling and solving multi-objective mixed-model
assembly line balancing and worker assignment problem, Computers & Industrial Engineering,
87: 74-80.

18
[40] Ritt M, Costa A M, Miralles C (2015), The assembly line worker assignment and balancing
problem with stochastic worker availability, International Journal of Production Research,
54(3): 907-922.
[41] Vilà M, Pereira J (2014), A branch-and-bound algorithm for assembly line worker assignment
and balancing problems, Computers & Operations Research, 44: 105-114.
[42] Wang H F, Fu Y P, Huang M, Huang G Q, Wang J W (2017), A NSGA-II based memetic
algorithm for multiobjective parallel flowshop scheduling problem, Computers & Industrial
Engineering, 113: 185-194.
[43] Wang S, Lu X, Li X X, Li W D (2015), A systematic approach of process planning and
scheduling optimization for sustainable machining, Journal of Cleaner Production, 87: 914-
929.
[44] Wu P, Che A D, Chu F, Zhou M C (2015), An improved exact-constraint and cut-and-solve
combined method for bi-objective robust lane reservation, IEEE Transactions on Intelligent
Transportation Systems, 16(3): 1479-1492.
[45] Xing X K, Yan Y M, Zhang H R, Long Y, Wang Y F, Liang Y T (2019), Optimal design
of distributed energy systems for industrial parks under gas shortage based on augmented
-constraint method, Journal of Cleaner Production, 218: 782-795.
[46] Yurtkuran A, Yagmahan B, Emel E (2018), A novel artificial bee colony algorithm for the
workforce scheduling and balancing problem in sub-assembly lines with limited buffers, Ap-
plied Soft Computing, 73: 767-782.
[47] Zaman T, Paul S K, Azeem A (2012), Sustainable operator assignment in an assembly line
using genetic algorithm, International Journal of Production Research, 50(18): 5077-5084.
[48] Zhang R, Chiong R (2016), Solving the energy-efficient job shop scheduling problem: a multi-
objective genetic algorithm with enhanced local search for minimizing the total weighted
tardiness and total energy consumption, Journal of Cleaner Production, 112(4): 3361-3375.
[49] Zhang Z K, Tang Q H, Zhang L P (2019), Mathematical model and grey wolf optimization
for low-carbon and low-noise U-shaped robotic assembly line balancing problem, Journal of
Cleaner Production, 215: 744-756.
[50] Zitzler E, Thiele L, Laumanns M, Fonseca C M, Fonseca V G (2003), Performance assessment
of multiobjective optimizers: an analysis and review, IEEE Transactions on Evolutionary
Computation, 7(2): 117-132.

Appendix
We discuss the computational time of the algorithm (PT-EC SFR) after one disturbance, which
is regarded as rescheduling. We consider the same model except for the objective, which is the
weighted sum of the two objectives in this discussion. We assume that the weights are both 0.5.
The model is as follows.

min f = 0.5 · f1 + 0.5 · f2 (35)

s.t. (3) − (18), (21) − (34)

XX X
f1 = ĉnr + (gk · fk ) (36)
n∈N r∈R k∈W
XX
f2 = (Enr · P̂nr ) (37)
n∈N r∈R

We test the same instances with experiments above and the results are listed in Table 5.
Column ‘Instance’ denotes different parameter combinations ({|N |, |W |, |R|}). We can find that
the computational time increases with the scale of instance growing. The average computational
time is 67.72s, i.e., the changeover time or rescheduling time is short. It is applicable for real
production. A manufacturer can adjust the production planning by applying the algorithm almost
without loss of efficiency after one disturbance.

19
Table 5: Computational results for 20 instance combinations under single objective

Instance Objective Time(s)


{5, 6, 2} 960.29 9.16
{5, 9, 3} 1609.39 13.59
{5, 12, 4} 21664.08 20.46
{5, 15, 5} 22243.71 29.44
{10, 6, 2} 21640.33 15.04
{10, 9, 3} 22418.48 28.20
{10, 12, 4} 22905.30 47.76
{10, 15, 5} 24127.33 86.78
{15, 6, 2} 21882.53 21.18
{15, 9, 3} 22589.90 39.04
{15, 12, 4} 24199.98 71.36
{15, 15, 5} 25481.33 140.75
{20, 6, 2} 22761.27 26.93
{20, 9, 3} 23544.91 52.00
{20, 12, 4} 24979.66 100.80
{20, 15, 5} 26042.05 190.00
{25, 6, 2} 22714.64 30.57
{25, 9, 3} 24417.59 63.24
{25, 12, 4} 25779.56 118.28
{25, 15, 5} 28145.45 249.75
Average 21505.39 67.72

20
*Detailed Response to Reviewers

<Computers & Industrial Engineering>

<Eco-friendly multi-skilled worker assignment and assembly line balanc-


ing problem>

Dear Editor,
We would like to thank you and the anonymous reviewers for the thorough
reviewing and constructive comments on our manuscript. According to the
comments, we have revised and polished the manuscript thoroughly. The
major correction parts are highlighted in RED in the revised manuscript.
We hope that the revised version can meet with approval.

The detailed corrections are listed below point by point:

Detailed response to the comments of Reviewer #2

Comment 1. This article is not well written that it’s a bit difficult to read
and understand. It is recommended to edit it.
Response: Thanks for pointing this out! We have carefully revised the en-
tire paper to make it easy to understand. For example: (1) we have revised
Abstract and Introduction (please see pages 1 and 2); (2) we have explained
-constraint method to make it clear (please see page 9); (3) we have modified
the conclusion (please see page 16).

Comment 2. The problem under study has been well-researched in the lit-
erature. The objective function in order to analyze technical performance is

1
quite clever. Taking risks is very good, but most used in the world of theory
and the assumption must be improved.

Response: Thanks for pointing this out! We have improved the assump-
tions in the paper as suggested (please see page 5). We have also clarified
the utilization of mixed-integer programming and linearization. Please see
pages 8 and 9.

Comment 3. After years of working on these issues by different people, we


need real-world problems examined. It is recommended to reduce restrictions
after issue to a close issue becomes a reality in order to simulate in the real
world to a real problem.

Response: Thanks for the suggestion! We have re-conducted all compu-


tational experiments, applying parameters based on real-world applications,
which are stated in the literature (Manavizadeh et al., 2013; Ramezanian
and Ezzatpanah, 2015; Ritt et al., 2015; Zhang et al., 2019). Please see page
12.

Detailed response to the comments of Reviewer #4

Reviewer’s comments: This paper studied a bi-objective multi-skilled work-


er assignment and assembly line balancing problem with energy consumption
considerably. This paper is well organized, and the analysis of the results is
impressive. The presentation is satisfactory. Minor revision is suggested for

2
this paper.

Comment 1. In the manufacturing execution, stochastic disturbances (such


as machine breakdown or rush orders) will affect the executability/performability
of the periodical schedule resulting from this type of sophisticated optimization
algorithm. How frequently is the proposed algorithm invoked for optimizing?
Please present more discussion on the computing time of the algorithm and
the changeover time of the assembly line.

Response: Thank you for your comments! We assume that disturbances do


not frequently happen. Moreover, changeover time is assumed to be short
and can be omitted in line with Gungor and Evans (2017). We have refined
the assumptions (please see page 5). Besides, we have also added computa-
tional experiments to demonstrate short rescheduling time after disturbances
(please see page 20).

Comment 2. In my opinion, there exist three scheduling classes for pro-


duction. 1) Holistic optimization from a global and integrated way. For
instance, ‘Dynamic scheduling in RFID-driven discrete manufacturing sys-
tem by using multi-layer network metrics as heuristic information’. This
kind of schedule usually uses artificial intelligence algorithms and data an-
alytics, supposing they could get all the manufacturing data promptly. 2)
Decentralized self-organizing way. For instance, ‘Makerchain: A Blockchain
with Chemical Signature for Self-Organizing Process in Social Manufactur-
ing’. This kind of schedule usually builds a multi-agent system based on
blockchain and smart contracts, supposing the manufacturing data are dis-

3
tributed indifferent nodes. 3) Hybrid models. For instance, ‘ManuChain:
Combining Permissioned Blockchain With a Holistic Optimization Model as
Bi-Level Intelligence for Smart Manufacturing’. This kind of scheduling com-
bines both the lower-level crowd self-organizing intelligence and upper-level
holistic optimization intelligence. The authors may present more discussion
on the scheduling methods.

Response: Thank you for pointing this out! We are very sorry for our
careless writing. We investigate a multi-skilled worker assignment and as-
sembly line balancing problem and formulate the model from the perspective
of holistic production system. When solving the problem, we adopt two
classic algorithms (i.e., a multi-objective genetic algorithm and a simulated
annealing method) and propose a constructive algorithm based on the prob-
lem characteristics. In other words, we optimize the problem from a global
and integrated way. Moreover, we have further added the discussion of the
scheduling methods in the literature review as suggested. We have also ad-
justed the entire paper accordingly.

Comment 3. References format should be consistent with the journal tem-


plate.

Response: Thanks for your detailed comments. We are sorry for our negli-
gence. We have revised the reference format.

Dear Editor and anonymous referees, thanks for your constructive comments
again! We have made a great effort to improve the organization and presen-

4
tation of this work. The revised manuscript has now been resubmitted to
the journal. We are looking forward to your positive response!

Sincerely,
Dr. Ming Liu

Reference

[1] Gungor Z E, Evans S (2017), Understanding the hidden cost and identi-
fying the root causes of changeover impacts, Journal of Cleaner Production,
167: 1138-1147.

[2] Manavizadeh N, Hosseini N S, Rabbani M, Jolai F (2013), A Simulated


Annealing algorithm for a mixed model assembly U-line balancing type-I
problem considering human efficiency and just-in-time approach, Computers
& Industrial Engineering, 64(2): 669-685.

[3] Ramezanian R, Ezzatpanah A (2015), Modeling and solving multi-objective


mixed-model assembly line balancing and worker assignment problem, Com-
puters & Industrial Engineering, 87: 74-80.

[4] Ritt M, Costa A M, Miralles C (2015), The assembly line worker assign-
ment and balancing problem with stochastic worker availability, International
Journal of Production Research, 54(3): 907-922.

[5] Zhang Z K, Tang Q H, Zhang L P (2019), Mathematical model and grey


wolf optimization for low-carbon and low-noise U-shaped robotic assembly
line balancing problem, Journal of Cleaner Production, 215: 744-756.

5
*Highlights (for review)

Highlights:
In this work, the contributions include: (1) the trade-off between production cost
and energy consumption considering multi-skilled workforce assignment is studied;
(2) a bi-objective mixed-integer programming model is proposed for the investigated
problem; (3) a fast and efficient constructive heuristic approach (PT-EC SFR) is
designed to solve large-scale problems.

You might also like