Journal of Cleaner Production 174 (2018) 560e576

Contents lists available at ScienceDirect

Journal of Cleaner Production

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

A new double flexible job-shop scheduling problem integrating

processing time, green production, and human factor indicators
Guiliang Gong, Qianwang Deng*, Xuran Gong, Wei Liu, Qinghua Ren
State Key Laboratory of Advanced Design and Manufacturing for Vehicle Body, Hunan University, Changsha, China

a r t i c l e i n f o a b s t r a c t

Article history: In this paper, we propose an original double flexible job-shop scheduling problem (DFJSP), in which both
Received 14 January 2017 workers and machines are flexible. Because of the characteristics of double flexibility, DFJSP conforms to
Received in revised form practical production better than the flexible job-shop scheduling problem (FJSP). In addition, a multi-
29 July 2017
objective optimization mathematic model according to the DFJSP is proposed, which is concerned
Accepted 17 October 2017
with the processing time indicator that is usually optimized by most existing studies; green production
Available online 27 October 2017
indicators, namely, factors regarding environmental protection; and human factor indicators, which are
actual indispensable elements that exist in the production system. Furthermore, ten benchmarks of
Keywords:
Double flexible job shop scheduling
DFJSP are presented and solved using a newly proposed hybrid genetic algorithm (NHGA). With the
problem proposed NHGA, a new well-designed three-layer chromosome encoding method and some effective
Green production indicators crossover and mutation operators are developed. To obtain the best combination of key parameters in
Human factors NHGA, the Taguchi design of experiment method is used for their evaluation. Finally, comparisons be-
Multi-objective optimization tween NHGA and NSGA-II show that the proposed NHGA has advantages in terms of the solving accuracy
Hybrid genetic algorithm and efficiency of the DFJSP, particularly at a large scale. It would be beneficial to apply our proposed
model to the multi-objective optimization of scheduling problems, especially those considering human
factor and green production-related indicators.
© 2017 Elsevier Ltd. All rights reserved.

1. Introduction the flexibility of the machines. However, to the best of our knowl-
edge, there have been no studies that consider the flexibility of the
Effective scheduling is one of the crucial factors affecting the machines and workers simultaneously, both of which widely exist
efficiency and productivity of job-shop production. In a practical in practical production and greatly affect the production efficiency
production system, the operation of a job can usually be processed and productivity.
by a machine chosen from a set of machines, i.e., machine flexibility, Fig. 1 shows an example of a practical job-shop scheduling
and a machine can also usually be operated by a worker chosen model including the flexibility of the machines and workers. There
from a set of workers, i.e., worker flexibility. Ever since the two- and are three operations, twelve machines, and three workers. Opera-
three-stage flow-shop scheduling problems were first proposed tion 1 can be processed using machines 1 through 4, operation 2
and optimized by Johnson (1954), the job-shop scheduling problem can be processed using machines 5 through 8, and operation 3 can
(JSP) has been comprehensively studied by researches in recent be processed using machines 9 through 12. Worker 1 can operate
decades. With the JSP, each operation can be processed once and all of the machines except machines 9 through 12, worker 2 can
only once on a single machine. The flexible job shop scheduling operate all of the machines except machines 1 through 4, and
problem (FJSP), also called JSP with machine flexibility (Dauze
re- worker 3 can operate all of the machines except machines 5
Pe
re

s and Paulli, 1997; Deng et al., 2017), is a generalization of through 8.
the classical JSP proposed by Brucker and Schlie (1991). In the FJSP, Facing the above proposed practical job-shop scheduling prob-
each operation can be processed once and only once on a machine lem, the production managers must consider the following
chosen from a previously identified machine set, that is, it considers situations:

C Owing to the complexity of the machines, the workers who

* Corresponding author. can operate them will be different. For example, supposing
E-mail address: [email protected] (Q. Deng).

https://doi.org/10.1016/j.jclepro.2017.10.188
0959-6526/© 2017 Elsevier Ltd. All rights reserved.
 G. Gong et al. / Journal of Cleaner Production 174 (2018) 560e576 561

best of our knowledge, there have been no studies integrating both

green production related indicators and human related factors into
the FJSP, which usually considers the processing time related
factors.
In this paper, to solve practical job-shop scheduling problems
using the flexibility of the machines and workers simultaneously,
and sufficiently consider the processing time related factors, green
production related factors, and human related factors, the
following researches were conducted. First, a new model of the
double flexible job-shop scheduling problem (DFJSP) considering
the flexibility of the machines and workers simultaneously is pro-
posed to achieve a better production performance. In addition, a
new mathematic model is constructed to describe the proposed
DFJSP. Furthermore, a new hybrid genetic algorithm (NHGA) is
constructed to solve the proposed DFJSP through the use of new
well-designed three-layer chromosome encoding, chromosome
decoding, and genetic operators, in which the processing time in-
dicator, human factor indicator, and green production indicator are
simultaneously considered. For the processing time indicator, this
model considers the processing time of each operation. For the
human factor indicator, this model mainly considers the skill level
and cost of the workers. For the green production indicator, this
model mainly considers the energy consumption, noise, recycling
of the tool chips, and safety. Finally, ten benchmarks of the DFJSP
are presented and solved using NHGA and NSGA-II to obtain Pareto-
optimal solutions based on the optimization objectives of the
Fig. 1. An example of a practical job-shop scheduling model including flexibility of the
machines and workers.
makespan, total worker costs, and total influence of the green
production indicators.
The contributions of this paper can be summarized as follows:
there is a very complex machine and a set of workers, and
that some or all of these workers can operate the machine, C We first propose the DFJSP model, which considers the ma-
who will then be assigned to the machine through the chine and worker flexibility simultaneously. Because of the
optimization objectives such as the makespan and total characteristic of double flexibility, this model is more suit-
worker cost? able for practical production compared with the FJSP.
C Because of the processing difficulty of the operations, the C This is the first time the processing time indicator, human
workers who can process such operations will also be factor indicator, and green production factor indicator have
different. For example, supposing there are a very difficult been considered simultaneously in the DFJSP model.
operation and a set of workers, some or all of which can C NHGA is constructed to solve the proposed DFJSP, in which a
process the operation, who will process the operation based new well-designed three-layer chromosome encoding
on the optimization objectives such as the makespan and method is developed. In addition, ten benchmarks are put
total worker cost? forward to show the model of the DFJSP and apply NHGA.
C Owing to the varying skill levels or other factors of the C Managerial implications of green production scheduling for
workers, the operations and machines that could be assigned manufacturing companies are provided.
to them may also differ. For example, supposing there are a
worker and a set of operations and machines, and that some The rest of this paper is organized as follows. A review of pre-
or all of these operations can be processed by the worker vious studies related to green production and human factor in-
using certain machines, which operation/machine will be dicators is described in the next section. In Section 3, the
processed/operated by the worker based on the optimization mathematical model of the proposed DFJSP problem is described. In
objectives such as the makespan and total worker cost? Section 4, NSGA-II is first briefly introduced, and details of the
chromosome encoding, chromosome decoding, and genetic oper-
Furthermore, according to an analysis by the IEA (2008), the ators that are adapted to the proposed NHGA are then presented.
world's energy demand has doubled over the last 40 years, and it The experiment results and a discussion based on the use of NHGA
has been estimated that it will double again by 2030; in addition, are provided in Section 5. Finally, some concluding remarks and
the manufacturing sector is accountable for over 38% of CO2 areas of future studies are described in Section 6.
emissions, and is responsible for 33% of total consumption. There is
therefore an urgency to reduce energy consumption by means of 2. Literature review
energy-efficient measures for both economic and environmental
reasons (May et al., 2013). In recent years, some researchers have 2.1. Human factor indicator related studies
studied production scheduling based on the minimization objec-
tives of green production related indicators (He et al., 2005; Subai Human factors, also called ergonomics, can be defined as “the
et al., 2006; Drake et al., 2006; Mouzon et al., 2007; Mouzon and theoretical and fundamental understanding of human behavior and
Yildirim, 2008; He and Liu, 2010; Fang et al., 2011; Fang and Lin, performance in purposeful interacting sociotechnical systems, and
2013; Dai et al., 2013; Liu and Huang, 2014; Wang et al., 2015; the application of that understanding to the design of interactions
Mansouri et al., 2016; Zhang and Chiong, 2016). However, to the in the context of real settings” (Wilson, 2000). As one of the key
elements in a production system, human factors cannot be ignored
562 G. Gong et al. / Journal of Cleaner Production 174 (2018) 560e576

because they have the potential to improve the production per- occupational health and safety (Othman et al., 2012). A mixed-
formance (Neumann and Medbo, 2009; Udo and Ebiefung, 1999). integer linear programming (MILP) model that describes the fa-
During the last decades, ergonomics has not been taken into full tigue and recovery in dual-resource constrained systems was
consideration in the production system development process, constructed by Jaber and Neumann (2010), who found that short
which makes the majority of managerial decisions difficult to rest breaks after each task, as well as short cycle times and faster
change (Helander, 2000; Jensen, 2002; Neumann and Medbo, recovery rates, can improve the system performance. Eklund (1997)
2009). Some studies have focused on the reasons for ignoring er- discovered that worker fatigue can greatly impact the system per-
gonomics in the production system. Seven common reasons for not formance in terms of quality, and Oxenburgh et al. (2004) also
considering ergonomics early in the production system develop- pointed out that it can significantly affect human productivity.
ment process were discussed by Helander (2000) and Bidanda et al. Corominas et al. (2010) presented a problem of assigning and
(2005), and it was pointed out that the difficulty in quantifying scheduling a set of tasks to a set of workers when the worker's
human issues is the major reason for its omission. In addition, some performance regarding the task depends on the experience of the
researchers have held other reasons for this, such as ergonomics worker and the other tasks involved. The application of a learning
being too abstract to be useful; the fact that the technical system curve of workers in different industrial settings has been well
should be designed prior to considering the ergonomics; and the documented in the literature (Corominas et al., 2010; Anzanello and
idea that people are adaptive, and there is therefore no need for Fogliatto, 2011; Jaber et al., 2013). Attia et al. (2014) mentioned that
dealing with the issue (Othman et al., 2012). However, these should flexible work assignments and the suitable assignment of workers
not be the reasons for ignoring human factors, which are actually to shifts can both have an effect on worker productivity.
indispensable elements in the production system. As discussed above, the literature review described herein
Recently, with the development of technology and the wide demonstrates the idea that human factors should be considered in
study of production scheduling problems, many researchers have production scheduling because of the potential to improve the
made an effort to study a production system incorporating human production efficiency. However, to solve the production scheduling
factors. Many researches have concluded that human factors should when incorporating human factors, which corresponds more to
be considered in production scheduling and operation manage- actual production systems, a considerable amount of research is
ment activities (Boudreau et al., 2003; Gino and Pisano, 2008; still needed in this area.
Neumann and Dul, 2010; Ryan et al., 2011; Neumann and Village,
2012). These studies are extensions of the “sociotechnical sys- 2.2. Green production indicator related studies
tems” movement of the 1970s (Cherns, 1976; Clegg, 2000), which is
an outgrowth of general systems theory (Skyttner, 2001) in which The first two papers considering green objectives in the area of
the operation systems are made up of both technological and hu- manufacturing scheduling were written by He et al. (2005) and
man elements, both of which must work well together for Subai et al. (2006). Drake et al. (2006) proposed a framework to
improved results. Hunter (1986) studied how a person's cognitive characterize the energy consumption of machines and their sub-
capability, cognitive aptitude, job knowledge, and job performance components (e.g., a spindle, coolant pump, controller, heater, and
affect the production performance. Based on a General Motors car hydraulic system in an injection molding machine, as well as the
assembly plant-door AGV system, Ragotte (1990) studied the hu- motors used in these components). To minimize the energy con-
man operator variability affecting the throughput of an AGV sys- sumption and total completion time, several dispatching rules and
tem. Using a simulation to investigate the concept of a multi-level a multi-objective mathematical programming formulation were
flexibility workforce, Felan and Fry (2001) found that it is better to put forward by Mouzon et al. (2007) for scheduling jobs on a single
have a combination of workers with high flexibility and workers CNC machine. This work showed that energy can be saved by
with no flexibility than employing all workers with equal flexibility. turning the machines off during idle times, but did not consider
Bidanda et al. (2005) highlighted the importance of the interactions energy saving during actual machine operation. In a subsequent
between the production system and certain key human factors, and work, a metaheuristic algorithm that minimizes the two conflicting
pointed out that the skills, communication, training, and assign- objectives of total energy consumption and total tardiness on a
ment strategies of workers, as well as the reward/compensation single machine using multi-objective optimization was developed
system, autonomy, teamwork aspects, and conflict management all by Mouzon and Yildirim (2008). This work also found that energy
require special attention for companies implementing cellular can be saved by turning machines off during idle times, but did not
manufacturing. Da Silva et al. (2006) constructed an aggregate consider the energy saving that can be achieved during machine
production planning model based on multiple-criteria mixed operation.
integer linear programming, which includes worker training, legal He and Liu (2010) constructed a new model based on their
restrictions on workload, and workforce size. A mixed integer previous paper (He et al., 2005), which considers other environ-
programming model was developed by Gel (2007) to determine the mental impacts such as the production of solid, liquid, gaseous, and
amount of hiring, firing, and cross-training for each GCA level to other kinds of waste. A multi-objective mixed integer linear pro-
minimize the total costs, which include training costs, salary costs, gramming formulation including the completion time and energy
firing costs, and missed production costs over multiple time pe- considerations with varying operation speeds on a single machine
riods. The author concluded that the differences among workers was proposed by Fang et al. (2011), in which the operation speed
should be considered when planning and managing the workforce. can be changed to affect the peak load and energy consumption. In
To deal with a simultaneous dynamic cell formation and the worker a similar work, Fang et al. (2013) studied a permutation flowshop
assignment problem, Aryanezhad and Deljoo (2009) constructed a problem with peak power consumption constraints using a mixed
mathematical model, and discussed the importance of incorpo- integer programming formulation.
rating human issues into a traditional dynamic cell formation. A Diaz et al. (2011) studied how the processing speed affects the
conceptual framework to help ergonomists in their research, edu- machine time, which dominates the energy demand and specific
cation, and practice to understand how to support the strategic energy consumption of a machine tool. In the computing field,
objectives of a company was developed by Dul and Neumann similar to parallel-machine scheduling problems, Fang and Lin
(2009). This framework helps ergonomics experts focus on ergo- (2013) found that the computing energy consumed increases
nomics from the point of view of business performance rather than with a higher execution speed of the processors.
 G. Gong et al. / Journal of Cleaner Production 174 (2018) 560e576 563

Dai et al. (2013) presented the scheduling problem of a flexible noise, recycling of the tool chip, and safety of the machine operated
flow shop, and proposed an energy-efficient model. This model was by the corresponding workers on the corresponding machines. For
solved using an improved genetic-simulated annealing algorithm, example, job 1 has three operations: operation O11 can be processed
and it was found that a significant potential improvement in energy using machines M1, M2, and M4 (where machine M1 can be operated
efficiency can be achieved. Liu and Huang (2014) studied power by workers W1, and W2; machine M2 can be operated by workers
consumption and carbon footprints using a multi-objective genetic W1, W2, and W3; and machine M4 can be operated by worker W1);
algorithm. They found that the makespan and energy consumption operation O12 can be processed using machines M1, M3, and M5
may be apparently conflicting, and there could be significant po- (where machine M1 can be operated by workers W1 and W2; ma-
tential for minimizing the energy consumption through the use of chine M3 can be operated by workers W1, W2, and W3; and machine
an energy-saving decision method. Wang et al. (2015) proposed a M5 can be operated by workers W1 and W2); and operation O13 can
systematic approach of process planning and scheduling optimi- be processed by machines M3, M4, and M5 (where machine M3 can
zation for sustainable machining including the objectives of energy be operated by worker W1; machine M4 can be operated by workers
efficiency of the milling process, productivity, and a constraint in W1, W2, and W3; and machine M5 can be operated by workers W1
the surface quality. Mansouri et al. (2016) studied the trade-off and W2). Other jobs can be interpreted similarly.
between makespan and energy consumption, and developed a In this paper, we aim to achieve the following three objectives:
constructive heuristic for a fast trade-off analysis between the
makespan and energy consumption. To solve the energy-efficient C Minimization of the maximum total worker cost.
JSP, Zhang and Chiong (2016) proposed a multi-objective genetic C Minimization of completion time of the machines
algorithm, through which some effective local search methods (makespan).
were designed. C Maximization of the green-production related indicator
To summarize, although a few researchers have recently made objective.
certain contributions regarding production scheduling with mini-
mization objectives related to green production related indicators, Some assumptions are put forward:
such contributions are mainly theoretical, and have yet to satisfy
the actual scheduling requirements. There are still many arduous C Each machine can process only one operation at a time on
tasks to deal with, such as constructing new models closer to any job.
practical production scheduling using green production indicators, C Each worker can operate only one machine at a time for any
and developing new algorithms that solve the production sched- operation.
uling problems more efficiently. C Each operation can be performed only once on one machine,
and its sequence is respected for every job.
3. Mathematical model of the proposed DFJSP C Each operation can be processed only once by one worker,
and its sequence is respected for every job.
3.1. Problem description C The operations of different jobs do not have precedence
constraints.
The problem of DFJSP is described as follows. There is a set of n C A temporary interruption of an operation is not allowed after
independent jobs J ¼ {J1, J2, …, Jn}, a set of m machines M ¼ { M1, M2, it has started.
…, Mm }, and a set of l workers W ¼ {W1, W2, …, Wl}. A job Ji has a C An operation of any job cannot be processed until its pre-
sequence of r operations {Oi1, Oi2, …, Oir} to be processed one after ceding operations are completed.
another according to the precedence constraint. Each operation Oij, C The processing time corresponding to the jobs, operations,
namely, the jth operation of Ji, must be executed on a given machine machines, and workers are given in advance.
chosen from the given machine set M. Each machine Mi must be
operated by a given worker chosen from the given worker set W. For The standards of quantification of the human indicator factors,
the processing time indicator, this model considers the processing time indicator factors, and green indicator factors in Table 1 are
time of each operation. For the human factor indicator, this model illustrated as follows:
mainly considers the skill level and cost of the workers. For the
green-production indicator, this model mainly considers the en- C Skill level of the workers: According to the worker's technical
ergy consumption, noise, recycling of the tool chip, and safety. The level, the worker ranks are divided into three classes, in
workers are divided into three types according to their skill level. which W1, W2, and W3 represent the best skilled worker, an
The processing cost of the workers, processing time, energy con- average skilled worker, and the worst skilled worker,
sumption, noise, recycling of the tool chip, and safety of each respectively.
operation conducted by a worker on a machine are all ascertained. C Processing cost of workers: This represents the unit cost of a
The scheduling consists of three sub-problems: (1) sequencing the worker for each operation on a corresponding machine.
operations, (2) assigning the machines to the corresponding oper- C Processing time: This represents the unit processing time for
ations, and (3) assigning the workers to the corresponding each operation on a corresponding machine by a corre-
machines. sponding worker.
For an explicit illustration, an example of the proposed DFJSP is C Energy consuming: This represents the unit energy con-
shown in Table 1. There are three jobs, five machines, and workers sumption, which is determined based on the corresponding
with three types of skill level. Jobs 1, 2, and 3 have three, two, and processing time.
two operations, respectively. Column 1 denotes the jobs. Column 2 C Noise: This represents the working noise of each machine in
shows the operations of each job. Column 3 indicates the machines decibels, the data of which are obtained from practical pro-
that process the corresponding operations. Column 4 lists the type duction testing.
and processing cost of the workers for operations on corresponding C Recycling of tool chip: This presents the recovery rate of a tool
machines. Column 5 shows the processing time of the operations chip, where the data are obtained from practical production
processed by the corresponding workers on the corresponding testing.
machines. Finally, column 6 indicates the energy consumption,
564 G. Gong et al. / Journal of Cleaner Production 174 (2018) 560e576

Table 1
Data of the proposed 3  5  3 DFJSP.

Job Operation Machine Human factor indicator Time indicator Green indicator

Skill level of workers Processing cost of workers Processing time Energy consuming Noise Recycling of tool chip Safety

J1 O11 M1 W1 100 9 4.05 80.0 0.7 0.7

W2 85 10 5.2
M2 W1 120 6 3.25 79.8 0.8 0.8
W2 100 7 3.62
W3 70 9 3.85
M4 W1 90 14 4.32 84.1 0.6 0.6
O12 M1 W1 110 5 3.46 78.1 0.7 0.7
W2 70 8 5.2
M3 W1 120 6 2.36 79.3 0.9 0.9
W2 90 7 2.63
W3 60 9 3.01
M5 W1 130 7 3.56 80.2 0.7 0.7
W2 70 9 4.01
O13 M3 W1 80 6 3.44 79.4 0.7 0.9
M4 W1 110 7 3.55 82.5 0.7 0.6
W2 100 8 3.85
W3 100 10 4.25
M5 W1 190 7 4.32 78.5 0.7 0.7
W2 120 12 6.35
J2 O21 M1 W1 150 5 3.28 82.6 0.7 0.7
W2 100 6 3.62
W3 50 8 4.59
M4 W1 110 7 3.55 76.3 0.8 0.6
W2 80 8 4.26
M5 W1 75 6 3.77 80.8 0.7 0.7
O22 M1 W1 120 4 4.21 81.2 0.8 0.7
W2 100 6 5.82
M2 W1 120 3 5.23 79.5 0.7 0.8
M5 W1 180 6 3.21 80.5 0.7 0.6
W2 120 7 3.65
J3 O31 M1 W1 120 5 3.85 82.5 0.8 0.7
W2 80 8 5.36
M2 W1 120 6 4.35 81.6 0.6 0.8
W2 80 9 6.12
M4 W1 140 5 3.44 79.8 0.7 0.6
W3 80 6 4.26
O32 M2 W1 200 4 2.35 80.6 0.7 0.8
M4 W2 210 3 3.35 78.8 0.9 0.6

C Safety: This represents the safety factor of each machine, C Cij: Completion time of operation Oij
where the data are obtained from the previous records on C Ck: The completion time of machine k
the machines or the experience of the production managers. C Lijks: The cost of Oij on machine k by worker s
C EIijks: The value of green indicator of Oij on machine k by
worker s
C Eijks: Energy consumption of Oij on machine k by worker s
3.2. Problem formulation C Nijk: Noise of Oij on machine k
C Rijk: Recycling of tool chip of Oij on machine k
The notions used in this paper are as follows: C Sijk: Safety of Oij on machine k
C xijk: A binary variable defined as
C n: Total number of jobs
C m: Total number of machines


C l: Total number of workers 1; if machine k is selected for operation Oij
C ri: Total number of operations of job i xijk ¼
0; otherwise
C i, h: Index of jobs, i, h ¼ 1,2, …,n
C j, g: Index of operations, j, g ¼ 1,2, …,r
C k: Index of machines, k ¼ 1,2 …,m C xijks: A binary variable defined as

1; if worker s is selected to operate machine k for operation Oij
xijks ¼
0; otherwise

C s: Index of workers, k ¼ 1,2 …,l

C Oij: The jth operation of job i The mathematical model can be given as follows.
C Pijks: Processing time of Oij on machine k by worker s The objective functions.
 G. Gong et al. / Journal of Cleaner Production 174 (2018) 560e576 565

Human factor indicator related objective. formulated as follows. First, Pt and Ut represent the current parent
and offspring population, respectively. The sizes of Pt and Ut are
(1) Minimization of the maximal total worker cost, i.e., both N. Then, a new population Rt ¼ Pt∪Ut (of size 2N) is formed by
combining Pt with Ut. Furthermore, an operator called non-
n X
X ri X l 
m X  dominated sorting is executed, which defines Rt as different non-
LC ¼ Lijks  xijks (1) dominated levels (rank1, rank2, and so on), to choose the best N
i¼1 j¼1 k¼1 s¼1
members as a new population called Ptþ1 for the next generation. In
Processing time indicator related objective. NSGA-II, the crowding distance is computed using a special oper-
ator, and the solutions with larger crowding distance values are
(2) Minimization of makespan, i.e., then selected. For more details of NSGA-II, refer to (Deb et al., 2002).

CM ¼ max ðCk Þ (2) 4.2. Proposed NHGA

1km

Green production indicator related objectives. 4.2.1. Framework

In this paper, NHGA is proposed for the DFJSP, in which a new
(3) Maximization of the green production indicator related well-designed three-layer chromosome encoding method is used
objective, i.e., for exploring the solution space in a more precise manner. A job-
based crossover (JBX) operator is also proposed for an operation
n X
X ri X l 
m X  sequence (OS) vector, and a random probability crossover (RPX)
G¼ EIijks  xijks (3) operator is proposed for the machine assignment (MA) and worker
i¼1 j¼1 k¼1 s¼1
assignment (WA) vectors. In addition, three mutation methods are
Here, EIijks is obtained by integrating the four green indicators proposed for OS, MA, and WA vectors. For the selection procedure,
with Equation (4). we use the method of (mþl)-ES (Xiao et al., 2015), where m denotes
the number of individuals in a parent population, and l denotes the
 . 0  . 0
0 0 size of the child population generated from the parent population
EIijks ¼ Eijks þ Nijk þ 1 Rijk þ 1 Sijk ; (4)
through a crossover and mutation. By evaluating the combined
population using adaptive ensemble fitness ranking, elite in-
where x’ denotes the unified value of x, and xmin and xmax are the dividuals can be selected and constitute a new parent population.
minimum and maximum of the x-related indicator, respectively. The framework of NHGA is shown in Fig. 2.
The computation equation of dimensionless unity is defined
through Equation (5). 4.2.2. Chromosome representation
0 Based on the particularity of DFJSP, a three-layer chromosome
x ¼ ðx  xminÞ=ðxmax  xminÞ; (5)
representation method is proposed, which consists of OS, MA, and
Subject to WA vectors. To initialize the population, a mapping relation be-
tween operations and integers is defined such that each operation
Cij  Ciðj1Þ  Pijks xijks ; (6) can be denoted through an integer ranging from 1 to Nop, where Nop
is the number of all operations. An example of this is shown in
where ci; j; i ¼ 1; …; n; j ¼ 2; …; ri ; Fig. 3, where job 1 has three operations, and jobs 2 and 3 both have
two operations, and thus the mapping relation between each
X
m X
l
operation and each integer can be constructed.
xijks ¼ 1; (7)
After the mapping relation is confirmed, a new chromosome can
k¼1 s¼1
be generated. In Fig. 4, an example using three jobs is shown, in
where ci; j; i ¼ 1; …; n; j ¼ 2; …; ri ; and which the OS vector is generated randomly under the condition of
the process constraints, namely, operation 1 has to be scheduled
h  i
before operations 2 and 3; operation 2 has to be scheduled before
Chg  Cij  thgk xhgk xijk  0 ∨
h  i (8) operation 3; operating 4 has to be scheduled before operation 5;
Cij  Chg  tijk xhgk xijk  0 ; and operation 6 has to be scheduled before operation 7. The MS
vector consists of three parts because three jobs are applied. The
where length of this vector is equal to that of the OS vector. Each part
represents the machines selected for the corresponding job oper-
ci; j; h; g; k; i; h ¼ 1; …; n; j ¼ 1; …; ri ; g ¼ 1; …; rh ; k ¼ 1; …; m.
ations. For example, the first part contains three numbers: 1, 3, and
Equations (1)e(3) are used for the objectives of LC minimization,
5. Number 1 means that machine 1 is selected for operation 1 of job
CM, and G, respectively. Inequality (6) ensures the operation pre-
1; number 3 means that machine 3 is selected for operation 2 of job
cedence constraint. Inequality (7) ensures that only one machine
1; and number 5 means that machine 5 is selected for operation 3 of
can be selected for one operation, and only one worker can be
job 1. The WS vector also consists of three parts because three jobs
chosen to operate the machine. Inequality (8) ensures that each
are applied. The length of this vector is also equal to that of the OS
machine processes only one operation at each time.
vector. Each part presents the workers selected for the corre-
sponding job operations. For example, the first part contains three
4. Proposed algorithm numbers: 2, 2, and 1. The first 2 indicates that worker 2 is selected
for operation 1 of job 1; the second 2 means that worker 2 is
4.1. Brief introduction of NSGA-II selected for operation 2 of job 1; and 1 means that worker 1 is
selected for operation 3 of job 1. An example of this is shown in
NSGA-II is a population-based multi-objective evolutionary al- Fig. 4, where the operation, machine, and worker sequence are as
gorithm that is widely used in the optimization of multi-objective follows: (O11, M1, W2), (O21, M4, W2), (O12, M3, W2), (O31, M2, W2),
problems. The core procedure of NSGA-II can be briefly (O22, M2, W1), (O13, M5, W1), and (O32, M4, W2), and we can obtain
566 G. Gong et al. / Journal of Cleaner Production 174 (2018) 560e576

left to right, if max (ASij, t_start) þ Pijks  t_end, the earliest start
Parameter time Sij ¼ max (ASij, t_start); otherwise, Sij ¼ max (ASij, LMk),
settings where LMk denotes the end time of the last operation on ma-
chine k.
Initialize the first population P (L=0)
Step 6: Calculate the completion time of every operation. Here,
comprised of individuals Cij ¼ Sij þ Pijks.
Step 7: Generate the set of start and completion times for every
operation of every job.

Generate offspring population

A Gantt chart of the schedule based on the chromosome in Fig. 4
Generate offspring P'2 (L) with P (L) comprised of
individuals with mutation is shown in Fig. 5. The data of this chromosome were obtained from
population P'1 (L) with
P(L) comprised of method Table 1.
individuals with crossover
method If < , ( - ) new
individuals are generated OS 1 4 2 6 5 3 7

A combined population with P (L), P'1 (L) and P'2 (L) is O11 O21 O12 O31 O22 O13 O32
L = L+1 generated
Job1 Job2 Job3
Adaptive ensemble fitness ranking
MS 1 3 5 4 2 2 4
Select individuals and generate offspring population
P(L)
O11 O12 O13 O21 O22 O31 O32
NO
The termination
condition is met? WS 2 2 1 2 1 2 2
YES Job1 Job2 Job3
Output the best individual
Fig. 4. Three-layer encoding example of a chromosome of 3  5  3 DFJSP.

Fig. 2. Framework of NHGA for DFJSP.

O11 O12 O13 O21 O22 O31 O32

1 2 3 4 5 6 7

Fig. 3. Mapping relation between operations and integers.

the values of the objectives of the work by referring to Table 1.

When the chromosome representation is decoded, each oper- Fig. 5. Decoding example of a chromosome of 3  5  3 DFJSP.
ation starts as soon as possible following the precedence and ma-
chine constraints (Kacem et al., 2002). A schedule generated using 4.2.3. Crossover operators
this decoding method can be ensured to be an active schedule The goal of a crossover operator is to obtain better solutions
(Cheng et al., 1996). The decoding procedure is implemented as based on the father population by exchanging better genes be-
follows: tween two chromosomes. In this paper, to guarantee the feasibility
of the solutions and obtain a high-quality offspring population, we
Step 1: Identify the workers of all operations based on the use two different crossover operators during the crossover pro-
worker assignment vector. cedure. The overall procedure of the crossover operator is shown in
Step 2: Identify the set of machines used to process every job. Algorithm 1. A JBX operator, whose procedure is shown in
Step 3: Identify the set of operations for every machine. Algorithm 2, is proposed for the OS. Fig. 6 shows an example of a
Step 4: Determine the allowable start time of every operation. JBX operator, in which job 1 is selected for jobSet, and all operations
Here, ASij ¼ Ci(j-1), where ASij denotes the allowable start time of belonging to job 1, namely, operations 1, 2, and 3, are consequently
operation Oij, and Ci(j-1) is the completion time of operation Oi(j-1) chosen for the positioned operations. To prevent the generation of
for the same job, here if j ¼ 1 then Ci(j-1) ¼ 0. an infeasible solution, we combine the MA vector with the WA
Step 5: Calculate the idle time of the machine of operation Oij, vector. A random probability crossover (RPX) operator, described in
and obtain the idle areas [t_start, t_end], where t_start is their Algorithm 3, is used for the combined vector, an example of which
start time, and t_end is their end time. Scanning these areas from is shown in Fig. 7.
 G. Gong et al. / Journal of Cleaner Production 174 (2018) 560e576 567

Fig. 7. RPX crossover for machine and worker sequences.

PM 4 5 6 7 1 2 3

OM 4 1 5 6 7 2 3
Fig. 8. Mutation for operation sequence.

PM 2 1 4 4 3 2 4

P1 4 5 6 7 1 2 3 OM 2 1 4 4 2 2 4
Job1 Job2 Job3
O1 4 6 5 7 1 2 3
Fig. 9. Mutation for machine sequence.

P2 1 4 2 3 6 5 7 PW 3 1 2 2 3 1 1
O2 1 4 2 3 5 6 7
OW 3 1 2 1 3 1 1
P1 Job1 Job2 Job3
4 5 6 7 1 2 3
Fig. 6. JBX crossover for operation sequence. Fig. 10. Mutation for worker sequence.
 Algorithm 3: excuteMWCross(P1, P2)
1: get the MA&WA array [PA1; PA2]
2: randomly generate a binary vector unit (Nop)
3: set index = 1;
4: For index =1 to Nop Do
5: If unit(index) = 1 Then
6: C1(index) = PA1(index)
7: C2(index) = PA2(index)
8: Else
9: C1(index) = PA2(index)
10: C2(index) = PA1(index)
11: End If
12: End For
13: Return [C1, C2]

4.2.4. Mutation operators

A mutation operator is generally used for jumping out of the
local optimal solution and enhancing the population diversity as a
4.2.5. Adaptive ensemble fitness ranking
result. In this paper, we construct three different mutation methods
The ensemble fitness ranking (EFR) is a many-objective opti-
for the OS, MA, and WA vectors. The overall procedure of a mutation
mization method proposed by Yuan et al. (2014). In comparison
operator is shown in Algorithm 4, and the mutation methods used
with NSGA-II, EFR adopts a more general selection method that
for the OS, MA, and WA vectors are described in Algorithms 5, 6,
incorporates a ranking method and fitness functions. Based on EFR,
and 7, respectively. Three examples are given in Figs. 8e10.
an adaptive ensemble fitness ranking (AEFR) method is proposed in
this paper.
For solution x in population R waiting for a ranking, each
objective (LC, CM, and G in this paper) should be investigated and
ranked. Assuming that there are altogether M objectives, as a result,
M ranking positions are obtained for x, as denoted by R(x) ¼ (r1(x),
r2(x), …, rM(x))T. Through the overall consideration of M ranking
positions R(x), the global rank Rg(x) can be obtained. In this paper,
Rg(x) is a weighted sum of R(x), and follows Equation (9).

Rg ðxÞ ¼ w  RðxÞ (9)

Ranking Rg in ascending order, all solutions in R can be

Algorithm 5: excuteProcessMutation(P (L))

1: P’21 (L)
2: For i = 1 to Do
3: If rand<MR Then
4: randomly select one element (Se) in the operation sequence of individual i
5: find the allowable range of motion ([start, end]) of the element
6: randomly select a position in [start, end]
7: generate Child by moving Se to the selected position
8: P’21 (L) = [P’21 (L); Child]
9: End If
10: End For
11: Return P’21 (L)
 G. Gong et al. / Journal of Cleaner Production 174 (2018) 560e576 569

Algorithm 6: excuteMachineMutaion(P (L))

1: P’22 (L)
2: For i = 1 to Do
3: If rand<MR Then
4: randomly select one operation (OP) in OS of individual i
5: get the machine (M) used to process OP in MA
6: generate Child by replacing M with another machine among the available
machines
7: randomly select one worker among the available workers
8: P’22 (L) = [P’22 (L); Child]
9: End If
10: End For
11: Return P’22 (L)

Algorithm 7: excuteWorkerMutation(P (L))

1: P’23 (L)
2: For i = 1 to Do
3: If rand<MR Then
4: randomly select one operation (OP) in OS of individual i
5: get the machine (M) and worker (W) used to process OP
6: generate Child by replacing W with another worker among the available workers
7: P’23 (L) = [P’23 (L); Child]
8: End If
9: End For
10: Return P’23 (L)

partitioned into several solution sets {F1, F2, …}. All solutions in the experiment simulation was run independently ten times for each of
same set have an identical Rg. In the ith iteration, we select N so- the compared algorithms.
lutions to constitute Ptþ1, following Algorithm 8.

5.1. DFJSP benchmark construction

5. Experiment studies and discussions
The DFJSP was first proposed, and accordingly, we constructed
To test the performance of the proposed NHGA in solving the ten DFJSP benchmarks, called DFJSPs 01e10, to verify the effec-
DFJSP, the algorithm is coded in MATLAB R2014a and implemented tiveness of the algorithm and future development. We constructed
on a computer configured using an Intel 2.67 GHz Core i3 CPU and the benchmarks of DFJSP based on the benchmarks proposed by
8 GB of RAM. We report our experiment results and compare them Brandimarte (1993), which are the most popular benchmarks for
with the results obtained using NSGA-II (Deb et al., 2002). The FJSP. The method for constructing the DFJSP benchmarks is shown
570 G. Gong et al. / Journal of Cleaner Production 174 (2018) 560e576

Algorithm 8: select(R)
1: Pt+1
2: i 1
3: While |Pt+1| + |Fi| N Do
4: Pt+1 Pt+1 Fi
5: i i +1
6: End While
7: Pt+1 Pt+1 Fi[1 : (N |Pt+1|)]

Algorithm 9: generateNewBenchmark(TB)
1: set variance of normal distribution
2: get the operation Oij and its available machine set Mij
3: For k = 1: size(Mij)
4: get the kth machine Mk for Oij and its available worker set Wk; get the processing
time Pijk in TB
5: For s = 1: size(Wk)
6: generate the processing time Pijks with a normal distribution
7: While Pijks
8: regenerate Pijks
9: End While
10: generate the cost of worker s
11: generate the energy consumption of worker s
12: End For
13: generate the values of the other three green indicators of machine k
14: End For

in Algorithm 9, in which TB denotes the data of the Brandimarte the population size, PS; crossover rate, CR; and mutation rate, MR.
benchmarks. The processing time, Pijks, is generated using a normal To obtain the best combination of these parameters, we use the
distribution method, the mean value of which is determined based Taguchi design of experiment method (Montgomery, 2008) for the
on the processing time of Oij on Mk (denoted by Pijk in Algorithm 9) investigation with instance DFJSP 04. Combinations of the different
in these FJSP benchmarks. The cost and value of the green in- values of these parameters are listed in Table 2.
dicators are generated randomly. It should be noted that, with a For each combination of parameters, the NHGA is run ten times
decrease in the processing time, the cost increases and the energy independently, and we choose the mean value of the best results of
consumption decreases. The benchmark data of DFJSPs 01-10 can the ten runs as the evaluation parameter of each combination. The
be downloaded from https://pan.baidu.com/s/1mhHfv6K. parameter combinations and their results are shown in Table 3.
Analogous to the traditional FJSP, DFJSP also has both partial and Clearly, the smaller the mean value is, the better the combination
total flexibility. In this paper, we consider total worker flexibility, achieved.
namely, any machine can be operated by any worker. According to the orthogonal table, the trend of each factor level
is illustrated in Fig. 11. The response value of each parameter is then
determined to analyze the significance rank of each parameter. The
results are listed in Table 4.
From Table 4 and Fig. 11, it can be seen that the population size,
5.2. Parameter settings
PS, is the most significant among these three parameters. A large PS
value makes the NHGA convergence very slow and time
Three critical parameters are used in our proposed NHGA, i.e.,
 G. Gong et al. / Journal of Cleaner Production 174 (2018) 560e576 571

Table 2 consuming, whereas a small value may cause a premature

Combinations of parameter values. convergence. A high value of the crossover probability, CR, requires
Parameters Factor level significant computational effort, whereas a small value cannot
1 2 3 4
excavate the potential of the best solution well. An appropriate
mutation probability, MR, can expand the solution space and
PS 50 100 150 200
maintain a good solution. According to the above analysis, we chose
CR 0.2 0.5 0.7 0.9
MR 0.05 0.1 0.15 0.2 PS ¼ 150, CR ¼ 0.5, and MR ¼ 0.15 for the following simulations of
the DFJSP benchmarks.

5.3. Experiment results

Table 3 5.3.1. Experiment 1

Results of parameter analysis.
The data of Experiment 1 were taken from Table 1. Fifty Pareto-
Experiment number Three parameters Mean value optimal solutions with three optimization objectives, i.e., make-
PS CR MR span, maximal total labor cost, and green production indicator
related objective values for DFJSP, were obtained using the pro-
1 1 1 1 87.140
2 1 2 2 87.700
posed NHGA. We listed twenty of these solutions in Table 5
3 1 3 3 87.700 (n  m  w indicates that the problem includes n jobs, m ma-
4 1 4 4 90.560 chines, and w workers; the OS, MS, and WS vectors indicate the
5 2 1 2 89.390 operation sequence, machine assignment, and worker assignment
6 2 2 1 87.800
solutions corresponding to all operations; and CM, LC, and G are the
7 2 3 4 88.670
8 2 4 3 87.090 optimization objectives of the makespan, maximal total labor cost,
9 3 1 3 86.293 and green-production indicator related objective values, respec-
10 3 2 4 85.167 tively). The Pareto-optimal solution distribution space of the pro-
11 3 3 1 87.620 posed problem is shown in Fig. 11. The 50 Pareto-optimal solutions
12 3 4 2 88.293
13 4 1 4 85.360
provide 50 schemes for production managers to choose from in this
14 4 2 3 87.435 proposed scheduling problem according to their emphasis. Take the
15 4 3 2 89.365 first solution, i.e., ([6 1 4 2 7 3 5], [2 1 3 4 2 4 2], [1 1 1 2 1 3 1], LC ¼ 17,
16 4 4 1 88.135 CM ¼ 790, and G ¼ 11.6157), as an example. The OS vector [6 1 4 2 7 3
5] means that the operation sequence is O31, O11, O21, O12, O32, O13,

Fig. 11. Factor level trend of key parameters.

572 G. Gong et al. / Journal of Cleaner Production 174 (2018) 560e576

Table 4
Response value.

Level PS CR MR

1 88.275 87.046 87.674

2 88.738 87.026 88.205
3 86.843 88.339 87.130
4 87.324 88.520 87.439

The bold 86.843 is the best value for PS, with this we chose the level 3 for PS;the bold
87.026 is the best value for CR, with this we chose the level 2 for CR; and the bold
87.130 is the best value of MR,with this we chose the level 3 for MR.

Fig. 12. Optimization solution 1 of the proposed 3  5  3 DFJSP (LC ¼ 17, CM ¼ 790,
and O22. The MS vector [2 1 3 4 2 4 2] means that machines 2, 1, 3, 4, and G ¼ 11.6157).
2, 4, and 2 are chosen to process operations O11, O12, O13, O21, O22,
O31, and O32, respectively. The WS vector [1 1 1 2 1 3 1] indicates that
workers with skill levels of 1, 1, 1, 2, 1, 3, and 1 are chosen to operate
the corresponding machines 2, 1, 3, 4, 2, 4, and 2, respectively. Here,
LC ¼ 17 means that the objective of makespan is 17, CM ¼ 790 in-
dicates that the objective of the total cost is 790, and G ¼ 11.6157
means that the green-production indicator related objective value
is 11.6157. Other solutions can be interpreted in a similar manner.
Two Gantt charts of two solutions, indicated in bold in Table 5, are
shown in Figs. 12 and 13.

5.3.2. Experiment 2
DFJSPs 01e10 were constructed using Algorithm 9, which is
based on the data of Brandimarte (1993). For each instance, the
algorithms were run ten times. The experiment results are shown
in Table 6 (only ten solutions are listed randomly for each bench- Fig. 13. Optimization solution 2 of the proposed 3  5  3 DFJSP (LC ¼ 17, CM ¼ 765,
mark solved using the corresponding algorithm), where the first and G ¼ 12.3284).
column denotes the proposed DFJSP benchmarks; the second and
third columns denote the results of NHGA and NSGA-II, respec-
tively; CM, LC, and G denote the optimization objectives of make-
span, maximal total labor cost, and green-production indicator set is also non-dominated with solutions from NSGA-II. For DFJSP
related objective values; and T denotes the average computation 02 and DFJSP 04, NHGA is biased toward CM and G, which result in a
time (in seconds) of each benchmark using the corresponding large value of LC, although the solution of the optimal solution set is
algorithm. also non-dominated with solutions from NSGA-II. For DFJSP 03,
From the values of the three optimization objectives shown in DFJSP 05, DFJSP 06, DFJSP 07, DFJSP 09, and DFJSP 10, NHGA is
Table 6, for DFJSP 01, NHGA is biased toward LC and G, which result biased toward G, which results in a large value of LC and CM;
in a large value of CM, although the solution to the optimal solution however, the solution of the optimal solution set is also non-
dominated with solutions from NSGA-II. For DFJSP 08, the NHGA
is biased toward LC, which results in a large value of CM and G;
however, the solution of the optimal solution set is also non-
Table 5
Experiment results. dominated with solutions from NSGA-II.
Based on the values of the average computation time, NHGA is
Problem Optimal sequence Optimal value
significantly less than NSGA-II, which means the proposed algo-
nmw OS vector MS vector WS vector LC CM G rithm is more effective in solving DFJSP problems.
353 6 1 4 2 7 3 5 2 1 3 4 2 4 2 1 1 1 2 1 3 1 17 790 11.6157 To evaluate the performance of the proposed NHGA, the C ~
6 1 2 3 4 5 7 2 3 3 4 1 1 2 1 1 1 1 1 1 1 18 870 9.6308 matric (Wisittipanich and Kachitvichyanukul, 2013) is used to
1 4 6 2 3 5 7 2 3 3 1 1 1 2 3 3 1 3 2 2 1 24 640 11.6838
compare the set of non-dominated solutions obtained from NHGA
6 1 2 3 4 5 7 2 3 3 4 1 1 2 1 1 1 1 1 1 1 18 870 9.6308
4 6 1 7 2 5 3 2 1 3 1 1 4 2 3 2 1 3 2 3 1 23 650 13.3491 ~ ðA; B Þ measures the
with those obtained from NSGA-Ⅱ. Here, C
6 1 2 3 4 7 5 2 3 3 5 1 1 2 3 3 1 1 2 2 1 24 665 11.2481 fractions of members of B that are dominated by members of A.
4 6 1 7 2 5 3 2 1 3 1 1 4 2 3 2 1 3 2 3 1 23 650 13.3491 ~ matric function.
Equation (10) shows the C
1 6 4 2 3 5 7 2 1 3 1 1 4 2 3 1 1 3 2 3 1 20 690 12.9141
4 1 6 2 7 5 3 2 1 3 1 1 4 2 3 1 1 3 2 3 1 20 690 12.9141
1 6 4 7 2 3 5 2 3 3 1 1 4 2 1 1 1 3 1 3 1 18 770 11.0976 ~ ðA; BÞ ¼ jfb2Bjda2A :
C
a dominates bgj
; (10)
6 1 2 3 4 7 5 2 3 3 4 1 1 2 3 1 1 1 1 1 1 21 820 9.7808 jBj
1 4 2 3 6 5 7 2 3 3 1 1 1 2 1 3 1 3 1 2 1 21 710 11.1313
6 1 2 3 4 7 5 2 1 3 4 1 1 2 1 1 1 2 1 1 1 17 830 11.0723
~ ðA; B Þ ¼ 1
where |B | is the number of solutions in B. Therefore, C
1 6 4 2 3 7 5 2 3 3 4 1 1 2 1 2 1 1 1 2 1 19 800 10.0758
6 1 4 2 3 5 7 2 3 3 4 1 1 2 1 2 1 2 1 2 1 19 770 10.2533 indicates that each solution in B is dominated by some solutions in
6 1 2 3 4 7 5 2 3 3 5 1 1 2 3 2 1 1 2 2 1 22 695 11.1531
A. On the other hand, C~ ðA; B Þ ¼ 0 shows that all solutions in B
4 6 1 7 2 3 5 2 3 3 4 1 1 2 3 2 1 1 2 2 1 22 730 10.6283
4 1 2 3 6 5 7 2 1 3 5 1 4 2 1 1 1 1 2 3 1 17 765 12.3284 are not dominated by any solution in A. The lower the ratio
6 4 1 2 3 5 7 2 1 3 4 1 1 2 1 1 1 1 1 1 1 17 860 10.8948 ~ ðA; B Þ is, the better the solution set in B that is obtained.
C
1 6 7 2 4 3 5 2 1 3 5 1 4 2 1 1 1 1 1 3 1 17 785 11.9259 Comparison results of NHGA and NSGA-Ⅱ are shown in Table 7.
The bold two solutions were chosen to draw the two Gantt charts. By comparing the data in Table 7, it can be concluded that the
 G. Gong et al. / Journal of Cleaner Production 174 (2018) 560e576 573

Table 6
Results of experiment 2.

Problem (n  m  w) NHGA NSGA-II

LC CM G T (s) LC CM G T (s)

DFJSP 01(10  6  3) 53 4976 78.518 19.21 42 6660 74.645 66.43

53 4950 78.740 66 4732 86.320
54 4938 78.851 82 5883 74.182
53 4969 78.629 67 4719 83.169
53 4950 78.740 69 6892 72.205
54 4938 78.851 66 4733 85.519
54 4931 78.962 50 4951 82.592
55 4964 78.629 66 4733 85.519
56 4938 78.851 82 5923 74.071
55 4938 78.851 57 4792 82.102
DFJSP 02(10  6  3) 48 4982 76.964 19.72 43 6759 71.887 70.54
48 4984 76.964 49 4707 81.146
47 4993 76.964 36 6372 79.281
47 5008 76.821 43 6759 71.887
48 5007 76.821 46 4798 82.305
48 5007 76.821 58 5537 74.846
47 5032 76.678 56 5202 76.184
49 4993 76.964 47 6574 72.115
48 5008 76.821 58 4950 77.824
50 4989 76.964 48 6325 72.720
DFJSP 03(15  8  3) 234 14059 208.19 49.14 203 13918 217.99 103.94
237 14064 208.17 224 14877 211.36
230 14125 208.16 243 13099 220.83
237 14052 208.20 261 13819 214.18
229 14143 208.16 217 13406 227.32
226 13896 208.94 220 15252 211.54
233 14084 208.23 224 14742 211.51
232 14137 208.16 214 13652 225.42
224 13813 209.64 252 13597 216.19
227 14240 208.01 261 13819 214.18
DFJSP 04(15  8  3) 94 7679 136.08 29.34 91 7784 140.48 80.53
95 7638 136.35 72 8636 137.97
90 7666 136.59 76 10156 131.90
92 7654 136.68 93 10068 130.39
95 7652 136.36 120 8658 133.09
94 7681 136.17 82 8074 141.68
94 7710 135.99 114 7887 138.05
96 7658 136.27 120 8658 133.09
92 7673 136.59 100 7897 138.96
96 7734 135.74 111 7932 137.36
DFJSP 05(15  4  3) 203 10047 176.21 37.72 183 11299 182.17 81.78
207 10051 175.94 216 12342 170.48
205 10010 176.33 187 12418 172.44
208 10060 175.81 221 9188 191.49
204 10053 176.08 253 10105 177.31
201 10077 176.08 244 10394 176.57
211 9998 176.19 249 10280 176.67
205 10072 176.08 201 9611 184.33
207 10018 176.46 211 9250 187.61
205 10103 175.83 220 9301 187.28
DFJSP 06(10  15  3) 103 14338 176.05 51.28 96 15016 192.11 108.33
102 14381 176.09 98 15188 185.58
110 14434 175.69 136 14234 186.48
102 14465 175.82 121 12823 201.60
106 14406 175.96 99 15010 186.57
103 14469 175.78 131 12886 196.17
114 14409 175.69 100 14666 185.73
107 14319 176.18 116 12882 200.01
104 14415 176.09 129 12991 197.51
105 14476 175.69 106 13874 190.02
DFJSP 07(20  5  3) 205 9605 137.62 32.95 170 9926 143.47 85.63
205 9600 137.64 204 8595 142.70
212 9675 137.35 196 10581 138.23
205 9761 137.50 242 9325 139.23
209 9554 137.65 202 8614 145.22
214 9657 137.30 217 10354 137.36
211 9587 137.62 206 10178 138.35
208 9566 137.72 236 9403 138.68
206 9718 137.62 186 8746 143.82
197 9747 137.81 199 8623 145.00
DFJSP08(20  10  3) 554 20593 400.89 74.61 518 25451 400.12 121.24
545 20589 401.38 574 22131 401.27
558 20527 401.10 637 25025 391.44
(continued on next page)
574 G. Gong et al. / Journal of Cleaner Production 174 (2018) 560e576

Table 6 (continued )

Problem (n  m  w) NHGA NSGA-II

LC CM G T (s) LC CM G T (s)

556 20588 400.94 590 26368 391.89

555 20589 400.99 525 23397 403.97
554 20609 400.89 637 24975 391.50
557 20536 401.20 590 26263 392.15
545 20575 401.52 629 23574 394.07
545 20590 401.47 561 25979 394.12
553 20606 401.09 553 25815 394.69
DFJSP 09(20  10  3) 383 22959 354.51 80.24 358 23650 371.53 133.49
383 23240 354.17 358 23619 371.58
383 23240 354.17 407 23130 357.74
387 23124 354.24 411 21589 366.47
381 23508 353.69 436 22004 363.87
379 22651 356.65 359 23892 370.81
378 22737 356.57 416 22295 359.51
378 22737 356.57 389 21831 369.13
386 22871 354.92 394 21741 367.80
395 22898 354.51 431 22828 358.75
DFJSP 10(20  15  3) 312 22199 338.40 80.52 345 22366 346.92 135.44
317 22127 338.48 291 23426 360.71
315 22173 338.42 352 21070 359.66
320 22112 338.51 294 23444 359.80
319 22136 338.49 293 23312 360.82
315 22213 338.40 355 21615 352.32
310 22388 338.00 346 21316 352.58
315 22 338.35 312 23135 349.51
312 22170 338.57 295 23237 360.68
318 22302 338.02 341 22263 348.70

optimization results of DFJSPs 01-10 obtained from NHGA are (3) A new hybrid genetic algorithm (NHGA) was proposed to
better than those obtained from NSGA-II in terms of solution solve the proposed DFJSP, in which a new well-designed
quality, particularly for large-scale problems. Figs. 14 and 15 show three-layer chromosome encoding method and some effec-
the Gantt chart of the instances DFJSP 03 and DFJSP 10, respectively. tive crossover and mutation operators were developed. In
the selection part, an adaptive ensemble fitness ranking
6. Conclusions and future studies method is designed.
(4) Ten benchmarks of DFJSP were presented and solved using
(1) A new model of the double flexible job-shop scheduling both the proposed NHGA and NSGA-II. Based on a compari-
problem (DFJSP), which originally considers the machine and son of their results, we can see that the proposed NHGA has
worker flexibility simultaneously, was proposed. Because of advantages in solving accuracy and efficiency for the DFJSP.
the characteristics of double flexibility, DFJSP is more suitable
for practical production compared with the flexible job-shop In the future, we will study the DFJSP problem, while consid-
scheduling problem (FJSP). ering the processing time indicator, human factor indicator, and
(2) For the manufacturing and scheduling fields, this marks the green production factor indicator simultaneously in greater depth.
first time that the time indicator, human factor indicator, and Meanwhile, the redistribution of job operations and machine
green production factor indicator are considered simulta- breakdowns may also be taken into consideration.
neously. In addition, a new mathematical model for the
proposed DFJSP was constructed.

Table 7
Mutual dominance of each set of solutions.

Problem (n  m  w) ~
C(NHGA, NSGA-II) ~
C(NSGA-II, NHGA)

DFJSP 01 (10  6  3) 0.0712 0.0000

DFJSP 02 (10  6  3) 0.0677 0.0000
DFJSP 03 (15  8  3) 0.2922 0.0000
DFJSP 04 (15  8  3) 0.1822 0.0000
DFJSP 05 (15  4  3) 0.2921 0.0000
DFJSP 06 (10  15  3) 0.1546 0.0000
DFJSP 07 (20  5  3) 0.5448 0.0000
DFJSP 08 (20  10  3) 0.5168 0.0000
DFJSP 09 (20  10  3) 0.5836 0.0000
DFJSP 10 (20  15  3) 0.6475 0.0000
DFJSP 9 0.5900 0.0000
DFJSP 10 0.6262 0.0000
 G. Gong et al. / Journal of Cleaner Production 174 (2018) 560e576 575

M8 1,1,1 15,1,3 1,3,1 4,4,3 8,3,3 9,2,3 11,2,3 5,2,2 9,4,3 5,4,3 2,4,3 6,7,3 6,8,2 12,9,3 1,9,2 1,10,2 9,10,3

8,6,3 13,10,3

M7 2,1,2 15,3,3 1,4,3 6,2,3 3,6,3 9,3,1 6,4,3 13,5,2 13,6,3 7,5,1 9,5,3 14,7,2 5,6,3
10,9,3

3,2,2 13,4,2 12,3,2 15,5,2 11,3,2 7,6,2 15,10,2 2,8,3

M6 8,1,2 5,1,2 10,2,2 15,2,2 1,5,3 10,4,3 4,5,1 12,5,3 8,4,2 11,5,312,6,2 13,8,3 10,7,1 2,6,3 9,9,2 2,9,3 4,10,3 7,10,2

3,5,2 10,8,3 5,8,3 2,10,2

M5 3,1,2 4,2,2 7,1,3 3,4,3 14,4,2 14,5,3 5,3,3 6,6,3 7,7,3

12,8,3 13,9,2 11,7,39,7,3 6,9,15,7,311,9,3 10,10,3 5,9,2 5,10,1

13,2,3 1,6,31,7,2 4,7,3 3,8,2 12,7,3 14,8,3 11,8,3

M4 4,1,3 1,2,2 13,3,3 12,2,1 6,1,2 10,3,3 15,4,3 11,1,3 7,2,3 6,3,3 14,2,311,4,2 10,5,3 6,5,1 13,7,3 14,6,3 15,8,2 2,5,2 9,6,3 2,7,3 8,7,3 7,8,3 8,9,3 7,9,3

10,6,1

M3 3,3,3 14,1,3 4,8,2 15,7,3 15,9,2

6,10,1 8,10,1

M2 13,1,3 4,3,3 3,7,2 8,5,3 5,5,2 11,10,3

2,3,2 7,3,2 4,6,214,3,3 9,8,3 12,10,3 8,8,3 3,10,3

217
M1 10,1,2 12,1,1 8,2,1 9,1,2 2,2,3 12,4,1 7,4,2 15,6,2 1,8,3 11,6,1 4,9,3 14,9,3 3,9,3 14,10,1

0 10 20 30 40 50 60 70 80 90 100 110 120 130 140 150 160 170 180 190 200 210

Fig. 14. Gantt chart of a solution to DFJSP 03.

M15

M14

M13 20,3,3 6,10,3 19,12,2

M12

M11

M10 17,2,2 18,2,3 5,2,3 7,4,3 2,2,2 18,4,3 5,4,2 20,5,3 13,4,3 4,5,3 13,6,3 3,6,3 8,5,2 5,9,1 15,8,3 4,8,3
3,9,3 14,8,3 8,10,3 5,13,2 16,14,3 11,10,2 19,13,2
1,10,2 16,11,1 2,12,2
M9 10,1,3 6,1,3 4,2,3 10,4,3 20,2,2 6,3,3 7,5,3 12,3,3 15,3,3 11,4,3 9,4,3 13,5,3 8,6,3 3,8,3 16,8,3 17,7,3 14,7,3 15,9,3 20,10,3 13,10,3 16,13,3 4,11,2 10,12,2

M8 18,3,2 1,4,3 15,5,3 12,6,2 1,7,1 8,4,3 17,6,2 16,7,2 11,9,3 20,8,3 12,10,3 13,11,3

M7 4,1,3 7,2,1 2,1,3 1,2,2 3,3,3 5,3,3 17,4,3 1,5,3 14,3,2 2,6,3 6,7,3 20,7,1 16,9,3 13,8,3
17,10,3 15,10,3
3,10,2 12,11,2 14,10,2
13,1,2 6,9,3
M6 16,1,3 9,1,3 5,1,3 12,1,2 8,2,2 2,3,33,4,3 17,5,2 10,6,2 14,4,3 4,6,3 6,6,3 13,7,3 12,9,2 20,9,3 19,9,3 8,11,2 9,10,2
1,1,2 17,9,3 10,9,2
M5 10,2,3 16,2,3 11,2,3 12,2,3 15,2,3 4,3,3 11,5,3 3,5,3 8,3,3 19,6,3 9,5,3 20,6,3 14,6,1 17,8,3 10,8,3 7,10,2 16,12,3 17,11,3 2,13,3 6,11,3
15,12,2
M4 17,1,3 18,1,3 15,1,2 17,3,2 9,2,3 11,3,3 9,3,3 14,2,2 15,4,3 1,6,3 7,7,3 15,7,3 12,7,3 1,9,3 9,6,2 9,7,1 8,8,3 8,9,3 2,9,2 2,10,2 7,11,3 19,11,220,11,3 18,11,3 9,9,3 10,11,3
13,3,1
M3 11,1,3 20,1,3 3,1,3 8,1,2 19,5,3 10,5,3 5,5,2 2,5,2 12,5,3 5,7,3 14,5,3 2,7,2 1,8,1 3,7,2 19,7,3 8,7,3 18,7,2 2,8,3 9,8,3 13,9,3 2,11,2 7,12,3 8,12,3 20,12,3 10,10,1 5,14,117,13,1

M2 14,1,3 7,1,3 19,1,3 19,2,3

19,3,216,3,3 1,3,3 13,2,3 6,4,3 2,4,1 18,5,2 4,4,2 18,6,2 11,6,3 6,5,2 5,8,3 11,7,3 7,8,2 11,8,312,8,3 10,7,3 16,10,3
18,8,2
6,8,3 1,11,3 19,10,3 14,9,3 3,11,2
4,9,3 4,10,1 8,13,2
10,3,3 20,4,2 16,5,1 18,10,3 15,11,3 17,12,2
M1 6,2,2 3,2,3 7,3,2 19,4,2 16,4,3 7,6,3 12,4,3 5,6,3 16,6,2 15,6,1 4,7,3 19,8,3 7,9,3 5,10,3 5,11,25,12,2 18,9,2 1,12,3 7,13,3 7,14,120,13,39,11,2 306

0 10 20 30 40 50 60 70 80 90 100 110 120 130 140 150 160 170 180 190 200 210 220 230 240 250 260 270 280 290 300

Fig. 15. Gantt chart of a solution to DFJSP 10.

Acknowledgements Ann. Oper. Res. 41 (3), 157e183.

Brucker, P., Schlie, R., 1991. Job-shop scheduling with multi-purpose machines.
Comput. 45 (4), 369e375.
The authors want to thank the editor and anonymous referees Cheng, R., Gen, M., Tsujimura, Y., 1996. A tutorial survey of job-shop scheduling
for their valuable comments and suggestions. This work was sup- problems using genetic algorithmsdi: representation. Comput. Ind. Eng. 30 (4),
ported by the National Natural Science Foundation of China (grant 983e997.
Cherns, A., 1976. The principles of sociotechnical design. Hum. Relatsh. 29 (8),
number 71473077), and the National Key Technology R&D Program 783e792.
of China (2015BAF01B00). Clegg, C.W., 2000. Sociotechnical principles for system design. Appl. Ergon. 31 (5),
463e477.
Corominas, A., Olivella, J., Pastor, R., 2010. A model for assignment of a set of tasks
References when work performance depends on experience of all tasks involved. Int. J.
Prod. Econ. 126 (2), 335e340. https://doi.org/10.1016/j.ijpe.2010.04.012.
Anzanello, M.J., Fogliatto, F.S., 2011. Learning curve models and applications: liter- Da Silva, C.G., Figueira, J., Lisboa, J., Barman, S., 2006. An interactive decision support
ature review and research directions. Int. J. Ind. Ergon. 41 (5), 573e583. system for an aggregate production planning model based on multiple criteria
Aryanezhad, M.B., Deljoo, V., 2009. Dynamic cell formation and the worker mixed integer linear programming. Int. J. M. S Sci. 34 (2), 167e177. https://
assignment problem: a new model. Int. J. Adv. Manuf. Tech. 41 (3), 329e342. doi.org/10.1016/j.omega.2004.08.007.
Attia, E.A., Duquenne, P., Le-Lann, J.M., 2014. Considering skills evolutions in multi- Dai, M., Tang, D., Giret, A., Salido, M.A., Li, W.D., 2013. Energy-efficient scheduling
skilled workforce allocation with flexible working hours. Int. J. Prod. Res. 52 for a flexible flow shop using an improved genetic-simulated annealing algo-
(15), 4548e4573. rithm. Rob. Comput. Integr. Manuf. 29 (5), 418e429.
Bidanda, B., Ariyawongrat, P., Needy, L.S., Norman, B.A., Tharmmaphornphilas, W., Dauze
re-Pere

s, S., Paulli, J., 1997. An integrated approach for modeling and solving
2005. Human related issues in manufacturing cell design, implementation, and the general multiprocessor job-shop scheduling problem using tabu search.
operation: a review and survey. Comput. Ind. Eng. 48 (3), 507e523. Ann. Oper. Res. 70, 281e306.
Boudreau, J., Hopp, W., Mcclain, J.O., Thomas, L.J., 2003. On the interface between Deb, K., Pratap, A., Agarwal, S., Meyarivan, T., 2002. A fast and elitist multiobjective
operations and human resources management. M Som 5 (3), 179e202. genetic algorithm: nsga-ii. IEEE Trans. Evol. Comput. 6 (2), 182e197.
Brandimarte, P., 1993. Routing and scheduling in a flexible job shop by tabu search. Deng, Q., Gong, G., Gong, X., Zhang, L., Liu, W., Ren, Q., 2017. A bee evolutionary
576 G. Gong et al. / Journal of Cleaner Production 174 (2018) 560e576

guiding non-dominated sorting genetic algorithm ii for multi-objective flexible Mansouri, S.A., Aktas, E., Besikci, U., 2016. Green scheduling of a two-machine
job-shop scheduling. Comput. Intell. Neurosci. 2017, 5232518. flowshop: trade-off between makespan and energy consumption. Eur. J. Oper.
Diaz, N., Redelsheimer, E., Dornfeld, D., 2011. Energy consumption characterization Res. 248 (3), 772e788.
and reduction strategies for milling machine tool use. In: Glocalized Solutions Montgomery, D.C., 2008. Design and Analysis of Experiments. John Wiley and Sons.
for Sustainability in Manufacturing. Springer Berlin Heidelberg, pp. 263e267. Mouzon, G., Yildirim, M.B., 2008. A framework to minimise total energy con-
Drake, R., Yildirim, M.B., Twomey, J.M., Whitman, L.E., Ahmad, J.S., Lodhia, P., 2006. sumption and total tardiness on a single machine. Int. J. Sustain. Eng. 1 (2),
Data Collection Framework on Energy Consumption in Manufacturing. 105e116.
Dul, J., Neumann, W.P., 2009. Ergonomics contributions to company strategies. Appl. Mouzon, G., Yildirim, M.B., Twomey, J., 2007. Operational methods for minimization
Ergon. 40 (4), 745e752. https://doi.org/10.1016/j.apergo.2008.07.001. of energy consumption of manufacturing equipment. Int. J. Prod. Res. 45
Eklund, J., 1997. Ergonomics, quality and continuous improvement e conceptual (18e19), 4247e4271.
and empirical relationships in an industrial context. Ergon 40 (10), 982e1001. Neumann, W.P., Dul, J., 2010. Human factors: spanning the gap between om and
Fang, K.T., Lin, B.M., 2013. Parallel-machine scheduling to minimize tardiness hrm. Int. J. Oper. Prod. Mana 30 (9), 923e950.
penalty and power cost. Comput. Ind. Eng. 64 (1), 224e234. Neumann, W.P., Medbo, P., 2009. Integrating human factors into discrete event
Fang, K., Uhan, N., Zhao, F., Sutherland, J.W., 2011. A new approach to scheduling in simulations of parallel flow strategies. Prod. Plan. Control 20 (1), 3e16. https://
manufacturing for power consumption and carbon footprint reduction. doi.org/10.1080/09537280802601444.
J. Manuf. Syst. 30 (4), 234e240. Neumann, W.P., Village, J., 2012. Ergonomics action research II: a framework for
Fang, K., Uhan, N.A., Zhao, F., Sutherland, J.W., 2013. Flow shop scheduling with peak integrating HF into work system design. Ergon 55 (10), 1140e1156.
power consumption constraints. Ann. Oper. Res. 206 (1), 115e145. Othman, M., Gouw, G.J., Bhuiyan, N., 2012. Workforce scheduling: a new model
Felan, J.T., Fry, T.D., 2001. Multi-level heterogeneous worker flexibility in a Dual incorporating human factors. J. Ind. Eng. Mana 5 (2), 259e284.
Resource Constrained (DRC) job-shop. Int. J. Prod. Res. 39 (14), 3041e3059. Oxenburgh, M., Marlow, P., Oxenburgh, A., Rapport, N., Oxenburgh, P.M., 2004.
https://doi.org/10.1080/00207540110047702. Increasing Productivity and Profit through Health and Safety: the Financial
Gel, E.S., 2007. Modelling inherent worker differences for workforce planning. Int. J. Returns from a Working Environment, vol. 57. CRC Press, pp. 1e12 (4).
Prod. Res. 45 (3), 525e553. Ragotte, M.J., 1990. The Effect of Human Operator Variability on the Throughput of
Gino, F., Pisano, G., 2008. Toward a theory of behavioral operations. M Som 10 (4), an AGV System. A Case Study: General Motors Car Assembly Plant - Door AGV
676e691. System. M.A.Sc Thesis. Department of Management Sciences, University of
He, Y., Liu, F., 2010. Methods for integrating energy consumption and environmental Waterloo, Canada.
impact considerations into the production operation of machining processes. Ryan, B., Qu, R., Schock, A., Parry, T., 2011. Integrating human factors and operational
Chin. J. Mech. Eng-En. 428 (4). research in a multidisciplinary investigation of road maintenance. Ergon 54 (5),
He, Y., Liu, F., Cao, H.J., Li, C.B., 2005. A bi-objective model for job-shop scheduling 436e452.
problem to minimize both energy consumption and makespan. J. Cent. South Skyttner, L., 2001. General Systems Theory e Ideas and Applications. World Sci-
Univ. 12 (2), 167e171. entific, London.
Helander, M.G., 2000. Seven common reasons to not implement ergonomics. Int. J. Subai, C., Baptiste, P., Niel, E., 2006. Scheduling issues for environmentally
Ind. Ergon. 25 (1), 97e101. responsible manufacturing: the case of hoist scheduling in an electroplating
Hunter, J.E., 1986. Cognitive ability, cognitive aptitude, job knowledge, and job line. Int. J. Prod. Econ. 99 (1), 74e87.
performance. J. Vocat. Behav. 29 (3), 340e362. Udo, G.G., Ebiefung, A.A., 1999. Human factors affecting the success of advanced
IEA, 2008. Worldwide Trends in Energy Use and Efficiency: Key Insights from IEA manufacturing systems. Comput. Ind. Eng. 37 (1e2), 297e300. https://doi.org/
Indicator Analysis. Accessed June 29, 2013. http://www.iea.org/Textbase/ 10.1016/S0360-8352 (99)00078-9.
Papers/2008/indicators_2008.pdf. Wang, S., Lu, X., Li, X.X., Li, W.D., 2015. A systematic approach of process planning
Jaber, M.Y., Neumann, W.P., 2010. Modeling worker fatigue and recovery in dual- and scheduling optimization for sustainable machining. J. Clean. Prod. 87,
resource constrained systems. Comput. Ind. Eng. 59 (1), 75e84. https:// 914e929.
doi.org/10.1016/j.cie.2010.03.001. Wilson, J.R., 2000. Fundamentals of ergonomics in theory and practice. Appl. Ergon.
Jaber, M.Y., Givi, Z.S., Neumann, W.P., 2013. Incorporating human fatigue and re- 31 (6), 557e567.
covery into the learningeforgetting process. Appl. Math. Modell. 37 (12), Wisittipanich, W., Kachitvichyanukul, V., 2013. An efficient pso algorithm for
7287e7299. finding pareto-frontier in multi-objective job shop scheduling problems. Ind.
Jensen, P.L., 2002. Human factors and ergonomics in the planning of production. Int. Eng. Manag. Syst. 12 (2), 151e160. https://doi.org/10.7232/iems.2013.12.2.151.
J. Ind. Ergon. 29 (3), 121e131. Xiao, S., Sun, S., Guo, H., Jin, M., Yang, H., 2015. Hybrid estimation of distribution
Johnson, S.M., 1954. Optimal two-and three-stage production schedules with setup algorithm for solving the stochastic job shop scheduling problem. J. Mech. Eng.
times included. Nav. Res. Logist. Q. 1 (1), 61e68. 51 (20), 27e35 (In Chinese).
Kacem, I., Hammadi, S., Borne, P., 2002. Pareto-optimality approach for flexible job- Yuan, Y., Xu, H., Wang, B., 2014. Evolutionary Many-objective Optimization Using
shop scheduling problems: hybridization of evolutionary algorithms and fuzzy Ensemble Fitness Ranking, pp. 669e676.
logic. Math. Comput. Simul. 60 (60), 245e276. Zhang, R., Chiong, R., 2016. Solving the energy-efficient job shop scheduling
Liu, C.H., Huang, D.H., 2014. Reduction of power consumption and carbon footprints problem: a multi-objective genetic algorithm with enhanced local search for
by applying multi-objective optimisation via genetic algorithms. Int. J. Prod. minimizing the total weighted tardiness and total energy consumption. J. Clean.
Res. 52 (2), 337e352. Prod. 112, 3361e3375.