IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. XX, NO.

XX, XX 2020 1

Energy-efficient Scheduling of Distributed Flow

Shop with Heterogeneous Factories: A Real-world
Case from Automobile Industry in China
Chao Lu, Liang Gao, Senior Member, IEEE, Jin Yi, and Xinyu Li, Member, IEEE

Abstract—Distributed flow shop scheduling of a camshaft at distinct sites in order to optimize one or some criteria,
machining is an important optimization problem in the auto- which is regarded as a distributed shop scheduling problem.
mobile industry. The previous studies on distributed flow shop Meanwhile, as the more and more serious global warming and
scheduling problem mainly emphasized homogeneous factories
(shop types are identical from factory to factory) and economic frequent occurrence of extreme weather, green manufacturing
criterion (e.g., makespan and tardiness). Nevertheless, hetero- has received widespread concern. Production scheduling has
geneous factories (shop types are varied in different factories) played a crucial role in implementing the target of green smart
and environment criterion (e.g., energy consumption and carbon manufacturing system [2], [3]. Consequently, green scheduling
emission) are inevitable because of the requirement of practical of distributed shop problem has been studied in recent years.
production and life. In this paper, we address this energy-
efficient scheduling of distributed flow shop with heterogeneous In most of the existing studies on distributed shop scheduling
factories for the first time, where contains permutation and problem, factories are deemed to be homogeneous where shop
hybrid flow shops. First, a new mathematical model of this types are identical from factory to factory. Nevertheless, in
problem with objectives of minimization makespan and total real-world manufacturing environment, factories are usually
energy consumption (TEC) is formulated. Then, a hybrid multi- heterogeneous where shop types are varied in different facto-
objective optimization algorithm, which integrates the iterated
greedy (IG) and an efficient local search, is designed to provide ries. Therefore, the research along distributed shop scheduling
a set of trade-off solutions for this problem. Furthermore, the problem with heterogeneous factories seems more significant
parameter setting of the proposed algorithm is calibrated by for some manufacturing enterprises.
using a Taguchi approach of design-of-experiment. Finally, to This work investigates one energy-efficient distributed flow
verify the effectiveness of the proposed algorithm, it is compared shop scheduling problem with heterogeneous factories (DF-
against other well-known multi-objective optimization algorithms
including MOEA/D, NSGA-II, MMOIG, SPEA2, AdaW, and SPHF). To the best of our knowledge, DFSPHF has not
MO-LR in an automobile plant of China. Experimental results yet been reported so far. Most previous studies focus on
demonstrate that the proposed algorithm outperforms these six the distributed shop scheduling problem with homogeneous
state-of-the-art multi-objective optimization algorithms in this factories, which covers the distributed scheduling of permu-
real-world instance. tation flow shop [4]–[10], hybrid flow shop [11], [12], and
Index Terms—distributed shop scheduling, heterogeneous fac- job shop [13]–[15]. Recently, some efforts have been devoted
tories, multi-objective optimization, energy-efficient scheduling, to variants of distributed permutation flow shop scheduling
iterated greedy algorithm. problem (DPFSP) such as assembly DPFSP [16], no-wait
DPFSP [17], blocking DPFSP [18], and limited buffer DPFSP
I. I NTRODUCTION [19]. Meanwhile, many intelligent optimization algorithms
have been presented to address such a scheduling problem.
W ITH the development of economic globalization, tradi-
tionally centralized manufacturing is very hard to meet
the requirement of the market. Whereas, distributed manufac-
For example, these optimization algorithms include genetic
algorithm [16], artificial bee colony algorithm [20], fruit fly
turing has become a common modern production mode due optimization algorithm [21], iterated greedy algorithm [1] and
to its lower management risk and higher production profit so on. Among them, iterated greedy algorithm has received
[1]. In this distributed manufacturing, all production tasks are approval from most researchers because of its simplicity of
assigned among a set of collaborative factories (or shops) algorithm structure and ease of implementation. In comparison
with traditional distributed production scheduling problems
Chao Lu is with the School of Computer Science, China University of with identical factories or shops, DFSPHF is more complicated
Geosciences, Wuhan, Hubei, 430074 China e-mail: [email protected]. than classic flow shop scheduling problem and DPFSP. Fur-
Liang Gao and Xinyu Li are with State Key Laboratory of Digital Manufac-
turing Equipment & Technology, Huazhong University of Science and Tech- thermore, DFSPHF is also NP-hard since flow shop scheduling
nology, Wuhan, Hubei, 430074 China. E-mail: [email protected], problem is known to be NP-hard [22]. In addition, DFSPHF is
[email protected]. much closer to the realistic production and living environment,
Jin Yi is with Department of Mechanical Engineering, Chongqing Univer-
sity, Chongqing, China. E-mail: [email protected]. which is commonly encountered in the charging infrastructure
Manuscript received XXX XX, 2020. This work was supported [23] and computing systems [24]. Therefore, it is very impor-
by the National Science Foundation of China (NSFC) under Grant tant to study DFSPHF in terms of both theory and practice.
No. 51805495,51885502 and the fundamental research funds for the
central universities, China University of Geosciences (Wuhan) Grant In this paper, DFSPHF is extracted from a realistic camshaft
CUGGC03.(Corresponding author: Jin Yi.) manufacturing process in an automobile plant of China. In
Digital Object Identifier 10.1109/TII.2020.3043734

1551-3203 © 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. XX, NO. XX, XX 2020 2

this company, manufacturing a camshaft should pass through Factory 1

Stage 1 Stage 2 Stage S
a mixed production line with permutation flow shop scheduling
problem (PFSP) and hybrid flow shop scheduling problem M1,1 M2,1 MS,1

....

....

....

....
(HFSP) at different factories. DFSPHF aims to minimize both MS,m
M1,m1 M2,m2
makespan and total energy consumption (T EC). Obviously, Fh
s

DFSPHF is a multi-objective optimization problem (MOP)

....

....

....
in the realistic production. Consequently, multi-objective op- Stage 1
Factory Fh
Stage 2 Stage S
timization algorithms are suitable for handling such a compli- M1,1 M2,1 MS,1

cated MOP and can be classified into two categories: priori Jobs
Assign
jobs to Product

....

....

....
....
factories
and posteriori method [25]. Concerning the priori method, a
M1,m1 M2,m2 MS,ms
linear weighted sum method can be used to solve a MOP by
assigning an appropriate weighted value to each objective. This
Factory Fh +1
method is very straightforward but has several weaknesses. In Stage 1 Stage 2 Stage 3

contrast, a Pareto-based posteriori method is very effective in M1,1 M2,1 Ċ MS,1

solving MOPs because of its strong search capability [17].

....
....

....
Fp
Factory F
Recently, the iterated greedy algorithm has shown a powerful
M1,1 M2,1 Ċ MS,1
performance for solving the distributed shop scheduling prob-
lems [1]. However, no free lunch theorem has indicated that it
is of great importance to utilize problem-specific knowledge Fig. 1. A layout of this DFSPHF.
for designing a high-performance algorithm. Therefore, when
solving this complex DFSPHF, we refine valuable knowledge
by analyzing problem property and then integrate this knowl- handle multiple jobs at a time. All heterogeneous factories can
edge into this algorithm as a general local search heuristic. handle all jobs. Once one job is allocated to any one of these
That is, a hybrid multi-objective iterated greedy algorithm factories, all operations of this job must be handled in this
(HMOIG) is presented to tackle this DFSPHF. assigned factory and cannot be transferred to another factory.
The main contributions of this work are summarized as The setup times and transport times are included in the job
follows: processing times. The storage capacity of buffer area is infinite
(1) Introducing the concept of green manufacturing as a between adjacent the machines. In general, different machines
new performance criterion into DFSPHF; have different power. When one job is processed on the
(2) Formulating a novel mathematical model of energy- machine with higher power, its processing speed is faster and
efficient DFSPHF; its processing time is shorter as well. In this case, makespan
(3) Developing a novel multi-objective iterated greedy may be decreased but the energy consumption will increase.
algorithm (i.e., HMOIG) to solve DFSPHF; Thus, there exists a conflicting relationship between makespan
(4) Verifying the good behavior of the proposed algo- and T EC. In this paper, DFSPHF aims to reasonably assign
rithm by comparison experiments. jobs to factories and determine job sequence in each factory
The rest of this article is organized as follows. The mathe- for optimizing makespan and T EC criteria. In addition, Fig.
matical model for DFSPHF is established in Section II. Section 1 depicts a layout of this DFSPHF.
III describes a new optimization approach for this DFSPHF.
The experiment study is provided in Section IV. Section V B. Problem modelling
concludes our work.
Before modelling this DFSPHF, notations of this problem
should be given below:
II. P ROBLEM STATEMENT AND MODELLING
n: the number of jobs.
A. Problem statement s: the number of stages (or number of operations of each job).
DFSPHF is described as follows: All jobs are allocated F : the number of factories.
on all F heterogeneous factories, which contains a mixed Fp : factories with the permutation flow shop.
production line consisting of factories with the identical PFSP Fh : factories with the hybrid flow shop.
(denoted by Fp ) and factories with the identical HFSP (de- j: index of jobs, j={1,2,...,n}.
noted by Fh ). In Fp factories, partial jobs assigned in these k: index of stages, k={1,2,...,s}.
factories are processed on all machines in the same direction. f : index of factories, f ={1,2,...,F }.
In Fh factories, partial jobs assigned in these factories are i: index of machines.
processed through all stages in the same direction. Each stage mk,f : the number of machines at stage k in the factory f .
has at least one machine, and there exist multiple identical Oj,k : the operation of job j at stage k.
parallel machines in at least one stage for Fh factories. Each pj,k : the processing time of Oj,k .
job has a set of operations and each operation can be handled Cf : the maximum completion time of jobs assigned in the
on any one of identical parallel machines in Fh factories or factory f .
a given machine in Fp factories. One job cannot be handled Cj,k : the completion time of Oj,k .
on more than one machine at a time, and one machine cannot Ci,k,f : the completion time of operations on machine i at stage
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. XX, NO. XX, XX 2020 3

k in factory f . Idj,i,k,f denotes the idle time of operation Oj,k on

Cmax : the makespan of a schedule. machine i in the factory f . Then insert Idj,i,k,f into
ti,k,f : the idle time of machine i at stage k in factory f . a set W ti,k,f .
idle
Pi,k,f : the idle energy consumption of machine i per unit time Step 3: Sort this set W ti,k,f in a non-ascending order and
at stage k in factory f . choose the former T members from W ti,k,f as a new
run
Pi,k,f : the work energy consumption of machine i per unit set W T i,k,f,T .
time at stage k in factory f . Step 4: For all machines at stage k, use the following strat-
EC w : energy consumption during the work phase. egy. Shut down the machine i if Idj,i,k,f meets
EC s : energy consumption during the idle phase. the following conditions:Idj,i,k,f ∈ W T i,k,f,T ; and
of f /on
Ei,k,f : energy consumption required to use the turn off/on Idj,i,k,f > Tb . Otherwise, keep the machine idle.
strategy. In summary, a mathematical model of energy-efficient DF-
T : the maximum allowable number of switching off each SPHF is as follows:
machine. 
Tb : the breakeven duration. min f1 = Cmax
Objectives: (3)
Π: a feasible solution representation. min f2 = T EC
L: an infinite large positive number. Subject to:
xj,f : equals to 1 if job j is allocated to factory f and otherwise
F
equals to 0. X
yj,i,k,f : equals to 1 if job j is processed on machine i at stage xj,f = 1 ∀j = 1, · · · , n. (4)
f =1
k in factory f and otherwise equals to 0.
zj,j ′ ,k,f : equals to 1 if job j precedes j ′ at stage k in factory n
X
f and otherwise equals to 0. xj,f ≥ 1 ∀f = 1, · · · , F. (5)
The objectives of DFSPHF include makespan and T EC, j=1

which are elaborated as follows: mk,f

X
1) Makespan criterion: Makespan is usually considered yj,i,k,f = xj,f ∀f = 1, · · · , F ;
as the economic criterion in scheduling problems. That is, i=1
(6)
makespan can reflect the production benefit of an enterprise k = 1, · · · , s; j = 1, · · · , n.
to some extent. Thus, makespan Cmax objective of DFSPHF
can be defined below: Cj,k ≥ pj,k ∀j = 1, · · · , n; k = 1, · · · , s. (7)

min f1 = Cmax = max {Cj,k } (1)

1≤j≤n,1≤k≤s Cj,k+1 −pj,k+1 ≥ Cj,k ∀j = 1, · · · , n;
(8)
2) T EC criterion: T EC during the manufacturing pro- k = 1, · · · , s − 1.
cess could be seen as one key environment indicator. T EC zj,j ′ ,k,f + zj ′ ,j,k,f ≤ 1 ∀j, j ′ = 1, · · · , n;
criterion reflects the carbon emission of an enterprise at (9)
manufacturing system layer. In this study, the second objective k = 1, · · · , s; f = 1, · · · , F.
is to minimize T EC through the following formula:
zj,j ′ ,k,f + zj ′ ,j,k,f ≥ yj,i,k,f + yj ′ ,i,k,f − 1∀i = 1, · · · , mk,f ;
F X s m k,f Z Ci,k,f
j, j ′ = 1, · · · , n; k = 1, · · · , s; f = 1, · · · , F.
X X
min f2 = T EC = θi (t) dt
f =1 k=1 i=1 t=0
(2) (10)
= ECw + ECs
Cj ′ ,k − pj ′ ,k − Cj,k + L · (3 − yj,i,k,f − yj ′ ,i,k,f − zj,j ′ ,k,f )
where θi (t) represents the instantaneous
≥ 0 ∀i = 1, · · · , mk,f ; j, j ′ = 1, · · · , n;
power of
PF Ps Pmk,f Pn machine i at time t, EC w =
run k = 1, · · · , s; f = 1, · · · , F.
f =1 k=1 i=1 j=1 Pi,k,f ·h pj,k · yj,i,k,f , ECs =
PF Ps Pmk,f Pn idle
(11)
f =1 k=1 i=1 yj,i,k,f Pi,k,f · ti,k,f · (1 − Hj,i,k,f )
i j=1
of f /on xj,f ∈ {0, 1}, ∀j = 1, · · · , n; f = 1, · · · , F. (12)
+Ei,k,f · Hj,i,k,f , Hj,i,k,f is a binary value, which equals
to 1 if machine i to process job j at stage k is turned off in
the factory f and otherwise equals to 0. Because frequently
yj,i,k,f ∈ {0, 1}, ∀j = 1, · · · , n; k = 1, · · · , s;
using this turn off/on strategy will damage to machine tools, (13)
we present a new energy conservation strategy inspired by f = 1, · · · , F ; i = 1, · · · , mk,f .
the literature [26]. The specific procedure of this strategy is
zj,j ′ ,k,f ∈ {0, 1}, ∀j, j ′ = 1, · · · , n;
below: (14)
k = 1, · · · , s; f = 1, · · · , F.
Step 1: Conduct the following steps for all jobs on each
machine i at each stage k in the factory f . Eq. (3) gives the optimization objectives including
Step 2: Scan each job j on the machine i at the stage k in the makespan and T EC. Constraint (4) guarantees that each job
factory f , and calculate the idle time Idj,i,k,f , where must be assigned to one factory. Constraint (5) ensures that
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. XX, NO. XX, XX 2020 4

Algorithm 1 Framework of the proposed HMOIG

F1 F2 F3
1: Input: Parameters like P opSize, Maximum number of function
evaluations (M axN F Es), destruction(ds), crossover and muta-
tion probability(Pc and Pm ). 2 5 7 4 6 1 3 8
2: Output: Non-dominated solutions.
3: Encoding and decoding Hybrid Flow Shop Permutation Flow Shop
4: P0 ← Initialize population (P opSize)
5: t←0
6: while t ≤ M axN F Es do Fig. 2. Encoding representation.
7: St ← Selection (Pt )
8: St′ ← Search Operator (St ,ds,Pc ,Pm )
9: Qt ← Local Search (St′ )
10: Ct ← Qt ∪ Pt A. Encoding and decoding
11: {F1 , F2 , · · · } ← Fast Non-dominated Sort (Ct ) DFSPHF requires to address two decision sub-problems:
12: Pt+1 ← ∅ (1) assignment of jobs to factories; (2) job sequence at the
13: i←1
14: while |Pt+1 | + |Fi | ≤ P opSize do first stage in each factory. Thus, an encoding representation
15: Crowding Distance Assignment (Fi ) of each solution should include two layers of information.
16: Pt+1 ← Pt+1 ∪ Fi Here, each solution is encoded as F parts, where each part
17: i++ denotes one factory and each part contains a string of numbers
18: end while representing a permutation of partial jobs in this
19: Pt+1 ← Elitist Strategy (Pt+1 )
factory. Let
20: Sort (Fi ) Π = π 1 , · · · , π f , · · · , π fh |π fh+1 , · · · , π F be a feasible
21: t=t+1 solution of DFSPHF, where π f is a permutation of partial
22: end while jobs at the first stage in the factory f , the former fh factories
contain the identical HFSP, and the rest factories contain the
identical PFSP. To explain this encoding representation, Fig. 2
each factory contains at least one job. Constraint (6) defines shows a solution to one problem with 8 jobs and 3 factories.
that each job should pass through all stages and be handled In this solution, jobs 2, 5 and 7 are assigned to factory 1,
on only one machine at each stage. Constraint (7) defines that namely, the partial job sequence π 1 is [2-5-7] in the factory 1
the completion time of each operation is greater or equal to its (i.e., F1 ). Similarly, job 4 and 6 are assigned to the factory 2
processing time. Constraint (8) ensures that one operation can (i.e., F2 ) and so on. Meanwhile, F1 and F2 are Fh factories,
start to be processed after its previous operation is completed. and F3 is a Fp factory.
Constraints (9)-(11) guarantee that one job cannot be handled The decoding mechanism for this DFSPHF is as follows:
on multiple machines at a time and one machine cannot handle First, all jobs are allocated to all factories according to en-
multiple jobs at a time. Constraints (12)-(14) impose the binary coding representation. Then, for jobs assigned in Fh factories,
decision variables. the job sequence denotes the job processing order at the first
stage in the assigned factory. In the following stage, the job
processing order is confirmed based on the completion times
III. P ROPOSED MULTI - OBJECTIVE OPTIMIZATION
of jobs in the previous stage. That is, jobs are allocated on the
ALGORITHM
first available machine according to First Come First Served
The proposed hybrid multi-objective iterated algorithm (FCFS) strategy [26]. For jobs assigned in Fp factories, the
(HMOIG) is a Pareto-based optimization algorithm and its job sequence denotes the job processing order at all stages,
framework is depicted in Algorithm 1. The motivation of because each stage has only one machine.
designing this HMOIG is as follows: First, iterated greedy
shows popularity due to its simplicity of algorithm structure B. Initialization
and excellent performance. Second, the iterated greedy based
local search algorithm can balance the global and local search. As for initialization, we propose a cooperative initialization
These merits drive us to propose a HMOIG for this distributed mechanism inspired by the previous work [4]. This cooper-
flow shop with heterogeneous factories. HMOIG involves ative initialization strategy assembles three different rules to
three innovations: (1) A new energy conservation strategy is generate one high-quality population. The procedure of this
embedded into the proposal. This energy conservation will not initialization strategy is stated below:
only reduce energy consumption but protect against machine Step 1: For half of all solutions, assign all jobs among all
damage (It is mentioned in section II(B)); (2) A cooperative factories as evenly as possible so that machine load
initialization heuristic is proposed. This initialization heuristic is balanced.
can not only generate three high-quality initial solutions but Step 2: For jobs in the Fp , perform the following strategy.
guarantee the diversity of population; (3) A knowledge-based 1) Sort all jobs in non-ascending order of total
local search strategy is designed to improve the exploitation processing times;
capacity. This local search can guide candidate solutions 2) Test and evaluate each job in all possible posi-
toward the potential promising regions by utilizing problem tions on each factory of Fp
properties. In this section, the main improvement procedures 3) Insert it to the position with current minimum
of HMOIG are elaborated. completion time.
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. XX, NO. XX, XX 2020 5

Step 3: The rest of the solutions are randomly generated to Algorithm 2 Proposed local search
keep the diversity of population. 1: Input: a trial schedule Π and an archive Q
2: Output: an archive Q.
3: set1 ← Insertion on critical factory (Π).
C. Search operator 4: set2 ← Swap on critical factory (Π).
5: set3 ← Insertion on non-critical factory (Π).
This search operator consists of two stages of operation: 1) 6: set4 ← Insertion on T EC (Π).
genetic operator; 2) destruction and reconstruction. The first 7: for Π′ ∈ {set1 , set2 , set3 , set4 , Π} do
stage operator is a genetic search operator, which includes 8: if Π′ ≺ Π then
crossover and mutation. According to characteristics of DF- 9: Π ← Π′
10: insert Π into Q
SPHF, partially matching crossover (PMX) and swap mutation
11: end if
are employed [27]. The procedure of PMX is as follows: 12: end for
Step 1: Select two positions at random on two parents, and
then determine the interval between two positions.
The interval is also called a matching interval. can obtain Cmax (Π) = Cf c (Π) according to the previous
Step 2: Create two temporary offsprings by swapping the definition. By the reverse proving, if job j has been assigned
matching interval of two parents. to the critical factory f c , moving job j will generate a new
Step 3: Identify the conflicting elements or jobs of temporary schedule Π′ and result in the following two conditions:
offsprings, and find mapping relation of jobs. (1) if Cmax (Π′ ) = max {Cf (Π′ )|f = 1, · · · , F } ≤
Step 4: Make sure that the job permutation of offsprings is Cf c (Π), then Cmax (Π′ ) = Cf c (Π); (2) if
feasible based on the mapping relation. Cmax (Π′ ) = max {Cf (Π′ |f = 1, · · · , F } >Cf c (Π),
Swap mutation operator is as follows: Two positions are the Cmax (Π′ )>Cmax (Π) = Cf c (Π). On the whole,
randomly selected and two corresponding jobs in two positions Cmax (Π) ≤ Cmax (Π′ ). Therefore, we can conclude that
are then exchanged. makespan objective can only be reduced by moving jobs in
The second stage operator is destruction and construction. the critical factory.
This destruction is to randomly remove ds jobs from the Therefore, the local search operator is proposed according
factories, where ds is the number of the deleted jobs. This to the problem properties.
destruction will cause two independent permutations consist- (1) In insertion on critical factory, extract one job from
ing of the deleted jobs (πd ) and the remaining jobs (πr ). the critical factory, and insert this job into other
This destruction and construction operator in this work is as possible positions in this critical factory.
follows: Extract ds jobs from some factories and reinsert the (2) In swap on critical factory, extract two jobs from the
removed ds jobs back into the destroyed solution. For more critical factory, and exchange two positions of jobs.
details, please refer to [1]. (3) In insertion on non-critical factory, extract one job
from the critical factory and insert it into other
possible positions among the non-critical factories.
D. Local search
(4) In insertion for T EC, extract one job with the idle
Local search is an effective heuristic to improve the qual- time interval and insert it into other all possible
ity of solutions in the field of scheduling [28]. In general, positions among all factories.
makespan depends on the critical factory (i.e., the factory The newly generated offspring Qt is merged with the parent
with the maximum completion time is called the critical population Pt to constitute a new combination population
factory), whereas T EC is sensitive to the idle time interval Qt ∪ Pt at the current iteration t. Then, a fast non-dominated
of a schedule. Therefore, the local search can be carefully sorting scheme is conducted on the merged population. Finally,
designed based on properties of this DFSPHF. We provide the best population with size P S (P S is the population
two properties before proposing this local search. size) is maintained from the combined population based on
Property 1: Energy consumption ECw during the work Pareto dominance relationship and crowding distance [29].
phase cannot be changed by moving job sequence . The pseudocode of the proposed local search is presented in
Proof: Job processing time pj,k of each operation is fixed, Algorithm 2.
run
and the work power of machine Pi,k,f is not influenced
by job sequence and job assignment among factories. Thus, IV. E XPERIMENT STUDY
regardless of changes within a schedule, energy consumption
To demonstrate the effectiveness of the proposed algorithm
ECw during the work phase remains constant.
in addressing the realistic DFSPHF problem, comparison
Property 2: Makespan criterion can only be reduced by
experiments are conducted on the same computer environment
moving jobs in the critical factory, if there is only one critical
with Intel Core i7, 2.9 GHz, 16 GB RAM, and Windows 10
factory.
OS. Meanwhile, all algorithms are implemented in Java.
Proof: Each factory has a corresponding maximum
completion time Cf for a schedule Π, therefore,
makespan of a schedule Π can be denoted as A. Description of a real-world case and instances
Cmax (Π) = max {Cf (Π)|f = 1, · · · , F }. Let the factory This realistic DFSPHF problem is from an automobile plant
f c ∈ {1, · · · , F } be only one critical factory. Obviously, we in China. As we know, camshaft is one of the core components
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. XX, NO. XX, XX 2020 6

TABLE II
T HE PROCESSING TIME OF JOBS ON EACH STAGE ( UNIT: MIN ).

Job NO. Stage 1 Stage 2 Stage 3 Stage 4 Stage 5 Stage 6 Stage 7 Stage 8 Stage 9
1 6 8 6 8 16 7 8 4 4
2 10 14 12 9 21 15 16 6 7
3 12 15 14 13 26 16 17 8 8
4 15 20 18 21 28 21 24 11 10
5 22 24 21 25 30 23 22 5 6
6 30 32 26 30 32 27 28 8 9
7 24 22 18 28 30 25 26 25 27
Fig. 3. Camshaft product. 8 10 17 13 18 26 10 9 10 11
9 23 30 26 34 18 15 15 9 9
10 30 26 20 23 30 20 16 16 15
11 25 30 33 32 22 15 20 8 12
TABLE I 12 20 32 25 28 20 28 19 13 15
D ATA OF ALL OPERATIONS . 13 35 25 20 30 22 20 15 15 17
14 15 28 10 34 25 14 18 18 10
15 14 10 15 17 26 15 20 20 20
Stage Processing operation Equipment NO. of machine Work/idle power (kW) 16 20 23 26 29 32 22 16 12 13
O1 Milling end face, drilling enter hole and marking Machining center 1 24/5 17 16 18 16 15 19 14 15 8 8
O2 Rough turning CNC machine 2 40/8
18 7 8 7 9 21 8 8 6 4
O3 Finish turning CNC machine 2 43/8
O4 Keyseat and milling cam Milling machine 1 25/7 19 8 9 7 9 20 8 9 5 6
O5 Quenching, tempering and straightening Machining center 1 20/6 20 9 6 8 5 13 6 8 5 3
O6 Grinding spindle and cam Grinding machine 3 30/8
O7 Flaw detection and brushing Machining center 1 15/8
O8 Polishing Polishing center 2 40/8
O9 Cleaning, heat and pressing Machining center 1 15/5

C. Parameter calibration
The parameter configuration can impact the performance of
of the Engine as shown in Fig. 3. In this real-world case, the algorithm in solving this problem. The proposed HMOIG
a camshaft (product model is X7) is manufactured by using contains four important parameters: Population size (P S),
a carbon steel wedged camshaft method in two factories. destruction (ds), crossover rate (Pc ) and mutation rate (Pm ).
This manufacturing process mainly has nine operations or A Taguchi approach of design-of-experiment (DOE) [31]
stages like milling, turning, quenching, grinding spindle, flaw is employed to measure the optimal combination of these
detection, polishing and so on. All jobs can be allocated parameters. The level of each parameter is as follows: P S =
and processed on one factory (Fh ) with a hybrid flow shop {40, 60, 80, 100}, ds = {2, 3, 4, 5}, P c = {0.7, 0.8, 0.9, 1.0}
or on one factory (Fp ) with a permutation flow shop. The and P m = {0.1, 0.2, 0.3, 0.4}. An orthogonal array L16 (44 ) is
job processing time of each operation is the same from adopted in this calibration experiment. HMOIG is conducted
factory to factory. Energy consumption of the turn off/on 30 times for each parameter configuration on this case. Fig. 4
of f /on
strategy is Ei,k,f = 0.67 kWh and the maximum number also shows the main effects plot of four parameters for three
of switching off machine is T = 1. The other related data performance metrics. Note that the lower these metric values
are provided in Tables I-II. Notably, Table I only presents the are, the better the performance is. Based on the comprehensive
number of parallel machines at each stage in Fh , where Ok observation, the best configuration of parametric values is set
denotes the operation of the stage k. Each stage has only as P S = 100, ds = 2, Pc = 0.7, and Pm = 0.2.
one machine in Fp . To further verify the effectiveness of
the proposed algorithm for solving large-scale instances, we
also design large-scale instances based on the above realistic D. Effectiveness of each improvement part of HMOIG
problem. All instances have n jobs, s stages and 2 factories To verify the effectiveness of each improvement part
(one is Fh and the other one is Fp ), n = {150, 200, 250}, of HMOIG, we compare three variants of HMOIG, where
s = {10, 15, 20, 25, 30}. The processing time pj,k ∈ [10, 100] HMOIG1 represents the HMOIG without the cooperative ini-
uniformly discrete distribution, the number of identical parallel tialization heuristic, HMOIG2 denotes the HMOIG without the
machines mk,f ∈ [1, 5] uniformly discrete distribution for the proposed energy conservation strategy, and HMOIG3 denotes
run
factory Fh , the power of energy consumption Pi,k,f ∈ [15, 40] the HMOIG without the proposed local search. Table III lists
idle
uniformly distribution, Pi,k,f ∈ [5, 10]. The other parameter statistical metric values over 30 times on 15 large instances,
sets are the same with those of the above realistic case. There- where the symbol ”+” denotes the proposal is significantly
fore, there are 15 large-scale instances in total. Each instance superior to the second best algorithm at the confidence level
is remarked by ins n s, which represents the problem with α = 0.05. The best mean values are marked in bold.
n jobs and s stages. The results from Table III indicate HMOIG substantially
outperforms its three variants in terms of three metrics.
This confirms that each improvement part of HMOIG can
B. Performance metrics contribute to the performance of the algorithm in solving the
DFSPHF.
To assess the behavior of multi-objective optimization al-
gorithms for solving this type of problem, three different
performance metrics like Spread [29], Generation Distance
(GD), and Inverse Generation Distance (IGD) [30] are utilized E. Results and discussion
in this paper. Spread can assess the diversity of solutions. To testify the effectiveness of the proposed HMOIG,
GD can evaluate the convergence of a certain algorithm. IGD HMOIG is compared to the current main stream multi-
can measure the comprehensive performance of one algorithm. objective evolutionary algorithms (MOEAs) like MOEA/D
Regarding these metrics, a lower value is better. [32], MMOIG [27], NSGA-II [29], SPEA2 [33], AdaW [34],
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. XX, NO. XX, XX 2020 7

0DLQ(IIHFWV3ORWIRU,*' 0DLQ(IIHFWV3ORWIRU6SUHDG 0DLQ(IIHFWV3ORWIRU*'

36 GV 3F 3P 36 GV 3F 3P 36 GV 3F 3P
  




0HDQRI6SUHDG
0HDQRI,*'

 

0HDQRI*'



 


 

  
               
                                       
       
       

(a) IGD (b) Spread (c) GD

Fig. 4. Main effects plot of three metrics: (a) IGD, (b) Spread and (c) GD.

TABLE III TABLE IV

S TATISTICAL RESULTS OF ALL METRICS OF HMOIG AND ITS VARIANTS S TATISTICAL VALUES OF THREE METRICS AMONG ALL ALGORITHMS .
IN ALL INSTANCES .
IGD Spread GD
IGD
MOEAs
mean std mean std mean std
HMOIG1 HMOIG2 HMOIG3 HMOIG
Instance
Mean std Mean std Mean std Mean std
SPEA2 4.15E-3 6.1E-4 1.25 7.6E-2 5.16E-3 2.5E-3
Ins 150 10 2.14E-03 3.00E-04 5.34E-03 6.10E-04
2.15E-03 4.30E-04 1.27E-03(+) 3.90E-04 MOEA/D 5.03E-3 1.1E-3 1.22 6.6E-2 4.28E-3 1.9E-3
Ins 150 15 4.25E-03 5.40E-04 9.05E-03 9.10E-04
4.44E-03 6.60E-04 2.18E-03(+) 4.20E-04 MMOIG 6.33E-3 5.4E-4 7.14E-1 7.0E-2 1.70E-2 2.4E-3
Ins 150 20 4.19E-03 5.20E-04 8.19E-03 1.20E-03
4.23E-03 5.70E-04 1.92E-03(+) 4.50E-04 NSGA-II 3.64E-3 5.9E-4 1.24 7.1E-2 4.24E-3 1.4E-3
Ins 150 25 4.22E-03 5.30E-04 8.36E-03 7.80E-04
4.38E-03 5.00E-04 2.33E-03(+) 5.50E-04
Ins 150 30 4.11E-03 5.40E-04 8.35E-03 8.70E-04
4.21E-03 4.60E-04 1.96E-03(+) 4.80E-04 AdaW 3.56E-3 7.1E-4 1.24 8.6E-2 4.75E-3 1.0E-3
Ins 200 10 3.40E-03 4.60E-04 7.68E-03 4.90E-04
3.40E-03 4.00E-04 1.70E-03(+) 4.40E-04 MO-LR 5.16E-3 1.1E-3 1.26 8.2E-2 5.69E-3 1.0E-3
Ins 200 15 4.48E-03 5.80E-04 9.34E-03 1.00E-03
4.75E-03 3.80E-04 2.15E-03(+) 5.20E-04 HMOIG 2.39E-3 6.5E-4 1.12 6.5E-2 2.11E-3 4.2E-4
Ins 200 20 3.91E-03 3.70E-04 8.30E-03 9.20E-04
3.90E-03 4.40E-04 1.62E-03(+) 3.60E-04
Ins 200 25 4.85E-03 5.10E-04 9.21E-03 8.50E-04
4.71E-03 5.00E-04 2.03E-03(+) 4.90E-04
Ins 200 30 4.90E-03 5.70E-04 9.51E-03 7.50E-04
5.10E-03 4.60E-04 2.18E-03(+) 4.60E-04
Ins 250 10 2.45E-03 2.70E-04 6.30E-03 5.20E-04
2.52E-03 3.10E-04 1.12E-03(+) 2.80E-04
Ins 250 15 3.39E-03 4.40E-04 6.95E-03 7.70E-04
3.48E-03 4.40E-04 1.38E-03(+) 3.10E-04 TABLE V
Ins 250 20 3.04E-03 4.00E-04 5.85E-03 6.30E-04
3.11E-03 2.80E-04 1.26E-03(+) 2.70E-04
Ins 250 25 3.33E-03 3.10E-04 6.63E-03 8.00E-04
3.60E-03 4.20E-04 1.52E-03(+) 3.00E-04 OVERALL RANKS THROUGH THE F RIEDMAN TEST OF THREE METRICS
Ins 250 30 3.38E-03 3.10E-04 5.67E-03 5.80E-04
3.28E-03 3.70E-04 1.58E-03(+) 3.60E-04 AMONG ALGORITHMS ( A LEVEL OF SIGNIFICANT α = 0.05).
GD
HMOIG1 HMOIG2 HMOIG3 HMOIG
Instance
Mean std Mean std Mean std Mean std
Ins 150 10 3.09E-03 4.20E-04 7.70E-03 9.60E-04 3.09E-03 6.00E-04 1.81E-03(+) 5.60E-04 IGD Spread GD
MOEAs
Ins 150 15 5.51E-03 7.30E-04 1.15E-02 1.40E-03 5.74E-03 8.40E-04 2.68E-03(+) 6.70E-04 Rank p-value Rank p-value Rank p-value
Ins 150 20 5.45E-03 6.70E-04 1.04E-02 1.40E-03 5.48E-03 6.80E-04 2.40E-03(+) 6.40E-04 SPEA2 4.0 5.0 4.0
Ins 150 25 5.38E-03 6.60E-04 1.07E-02 1.20E-03 5.60E-03 6.50E-04 2.61E-03(+) 6.90E-04
Ins 150 30 5.49E-03 7.10E-04 1.09E-02 1.30E-03 5.61E-03 5.90E-04 2.50E-03(+) 7.40E-04 MOEA/D 5.0 6.0 5.0
Ins 200 10 4.51E-03 6.20E-04 9.94E-03 8.50E-04 4.52E-03 5.00E-04 2.22E-03(+) 6.00E-04 MMOIG 7.0 3.95E-24 1.0 1.24E-21 7.0 3.25E-22
Ins 200 15 5.98E-03 7.30E-04 1.17E-02 1.30E-03 6.35E-03 5.10E-04 2.92E-03(+) 7.00E-04 NSGA-II 3.0 3.0 3.0
Ins 200 20 5.31E-03 4.50E-04 1.13E-02 1.40E-03 5.40E-03 5.60E-04 2.28E-03(+) 5.30E-04
Ins 200 25 6.36E-03 6.50E-04 1.16E-02 1.20E-03 6.28E-03 6.60E-04 2.83E-03(+) 6.20E-04
AdaW 2.0 4.0 2.0
Ins 200 30 6.22E-03 7.10E-04 1.24E-02 1.10E-03 6.47E-03 5.50E-04 2.85E-03(+) 5.70E-04 MO-LR 6.0 7.0 6.0
Ins 250 10 3.92E-03 4.10E-04 9.76E-03 9.80E-04 4.02E-03 5.00E-04 1.74E-03(+) 4.50E-04 HMOIG 1.0 2.0 1.0
Ins 250 15 4.70E-03 5.90E-04 9.58E-03 1.10E-03 4.85E-03 5.80E-04 1.92E-03(+) 4.50E-04
Ins 250 20 4.31E-03 5.40E-04 8.18E-03 9.30E-04 4.36E-03 4.00E-04 1.86E-03(+) 4.30E-04
Ins 250 25 4.48E-03 3.90E-04 8.84E-03 1.10E-03 4.85E-03 5.10E-04 2.04E-03(+) 3.90E-04
Ins 250 30 4.79E-03 4.20E-04 7.98E-03 8.50E-04 4.63E-03 5.10E-04 2.24E-03(+) 4.90E-04
Spread
Instance
HMOIG1
Mean std
HMOIG2
Mean std
HMOIG3
Mean std Mean
HMOIG
std
standard deviation values) in this real-world case. The optimal
Ins 150 10 3.23E-01 3.30E-02 6.93E-01 8.70E-02 3.13E-01 3.30E-02 2.29E-01(+) 2.60E-02
Ins 150 15 4.63E-01 5.40E-02 1.10E+00 8.90E-02 4.60E-01 3.70E-02 3.47E-01(+) 3.30E-02 results are marked with bold. Table V records Friedman rank
Ins 150 20 4.13E-01 4.60E-02 9.68E-01 1.10E-01 4.02E-01 5.10E-02 3.17E-01(+) 4.00E-02
Ins 150 25 4.09E-01 3.10E-02 9.31E-01 1.20E-01 4.17E-01 3.60E-02 3.34E-01(+) 2.80E-02 test results among all algorithms in this real-world case, where
Ins 150 30 4.11E-01 5.00E-02 9.80E-01 1.10E-01 4.18E-01 3.80E-02 3.10E-01(+) 3.40E-02
Ins 200 10 3.26E-01 2.90E-02 1.04E+00 1.10E-01 3.29E-01 3.00E-02 2.34E-01(+) 3.10E-02 the confidence level α = 0.05.
Ins 200 15 3.63E-01 2.50E-02 1.03E+00 1.10E-01 3.75E-01 3.40E-02 2.61E-01(+) 3.90E-02
Ins
Ins
200
200
20
25
3.43E-01
3.55E-01
2.50E-02
3.00E-02
9.19E-01
9.54E-01
1.00E-01
1.10E-01
3.47E-01
3.70E-01
3.70E-02
3.70E-02
2.31E-01(+)
2.45E-01(+)
2.00E-02
2.70E-02
As observed in Table IV, HMOIG achieves lower value than
Ins
Ins
200
250
30
10
3.55E-01
3.16E-01
3.20E-02
3.30E-02
8.64E-01
9.68E-01
1.00E-01
1.10E-01
3.54E-01
3.36E-01
2.70E-02
3.40E-02
2.43E-01(+)
2.32E-01(+)
2.20E-02
2.10E-02
its competitors for the GD and IGD metrics. Regarding the
Ins
Ins
250
250
15
20
3.46E-01
3.17E-01
3.70E-02
2.50E-02
8.15E-01
7.19E-01
9.10E-02
8.30E-02
3.21E-01
3.19E-01
2.90E-02
2.10E-02
2.30E-01(+)
2.37E-01(+)
2.20E-02
2.40E-02
Spread, HMOIG is inferior to MMOIG but superior to the
Ins
Ins
250
250
25
30
3.42E-01
3.07E-01
2.50E-02
2.70E-02
7.58E-01
6.76E-01
7.70E-02
7.20E-02
3.44E-01
3.15E-01
2.80E-02
2.60E-02
2.45E-01(+)
2.19E-01(+)
2.30E-02
2.10E-02
other algorithms. As observed in Table V, HMOIG ranks the
first for IGD and GD metrics, but ranks the second for Spread
metric. Also, since the p-values for algorithms are less than
and MO-LR [35]. The parameters of these compared algo- the significance level of 0.05, it can be concluded that at least
rithms are calibrated as the same in section IV(C). Due to one of seven algorithms has a significant difference effect for
space limitation, we only provide the final parameter settings three metrics. To further determine the significant difference
of its compared algorithms as follows: Pc and Pm are set between a control method HMOIG and other algorithms, a
to 0.9 and 0.3 for NSGA-II and MO-LR, 0.9 and 0.2 for post-hoc test of Dunnett method is performed on these results
SPEA2, MOEA/D and AdaW. The neighborhood size is 20 by different algorithms. Fig. 5 shows a multiple comparison
for AdaW and MOEA/D. The archive size is 60 for SPEA2. interval plot for three metrics. If an interval does not contain
We set ds = 2 for MMOIG. P S is 100 for all algorithms. a zero, the corresponding mean is significantly different from
To conduct a fair comparison, these algorithms all use the the control mean. The confidence interval for the difference
common termination criterion (i.e., M axN F Es = 25, 000). between HMOIG and other algorithms contains only positive
Meanwhile, they also use the above-mentioned encoding and numbers with regard to IGD and GD metrics. For Spread met-
decoding mechanism, and energy-saving strategy. For all the ric, the confidence interval for the difference between HMOIG
algorithms, we implemented them by Java. This comparison and the MMOIG except for other algorithms contains only
experiment is conducted 30 independent runs in this real-world a negative number. Thus, these results indicate that HMOIG
case and instances. Table IV lists statistical results (mean and significantly outperforms the other algorithms for IGD and
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. XX, NO. XX, XX 2020 8

'XQQHWW6LPXOWDQHRXV&,VIRU,*' 'XQQHWW6LPXOWDQHRXV&,VIRU6SUHDG 'XQQHWW6LPXOWDQHRXV&,VIRU*'

63($+02,* 63($+02,* 63($+02,*

02($'+02,* 02($'+02,* 02($'+02,*

002,*+02,* 002,*+02,* 002,*+02,*

16*$,,+02,* 16*$,,+02,* 16*$,,+02,*

$GD:+02,* $GD:+02,* $GD:+02,*

02/5+02,* 02/5+02,* 02/5+02,*

                      

(a) IGD (b) Spread (c) GD

Fig. 5. Multiple comparison: (a) IGD, (b) Spread and (c) GD.

GD metrics and is also competitive to its rivals for Spread  63($
metric. Furthermore, Table VI records statistical results (mean 
02($'
002,*
$
and standard deviation values) in 15 large-scale instances. The 
16*$,,
$GD:
results from Table VI show that HMOIG is overwhelmingly 
02/5
&3/(;

7(& N:K
better than its rivals for solving such instances. The success  +02,*

of HMOIG lies in its algorithm design. First, a cooperative 

initialization strategy is proposed to generate one high-quality  %

population. That is to say that it spends a little time improving 

the quality of initial solutions during the search progress. Sec- 


ond, this search operator consists of two stages of operation,      

which can enhance the search diversity to some extent. Third, 0DNHVSDQ PLQ

we utilize problem-specific knowledge and then regard this

knowledge as a general local search heuristic. This knowledge- Fig. 6. Pareto front approximations by different algorithms on this real-world
case.
based local search can guide trial solutions towards Pareto
optimal solutions. By contrast, the other MOEAs does not IDFWRU\+)63

VWDJHPDF          
have this local search. VWDJHPDF
VWDJHPDF
  
 






VWDJHPDF    
VWDJHPDF      
VWDJHPDF          
VWDJHPDF          
VWDJHPDF    
VWDJHPDF   
VWDJHPDF   
VWDJHPDF         
VWDJHPDF     
VWDJHPDF     

To visualize the behavior of different algorithms in this VWDJHPDF

  


 
 


 


   



real-world case, Fig. 6 presents their own Pareto fronts (PFs) IDFWRU\3)63

with the best IGD. We find the extreme objectives by a VWDJH         
VWDJH         

software CPLEX, which overlaps with partial results obtained VWDJH

VWDJH
 
 
 
 


 




 

by HMOIG. It also verifies the correctness of mathemati- VWDJH

VWDJH












 
 
 
 
VWDJH         
cal model by comparing results obtained by our proposal VWDJH
VWDJH




 
 










and CPLEX. Regarding the convergence and distribution of       



non-dominated solutions, HMOIG can obtain better solutions

than its rivals, because approximations found by HMOIG Fig. 7. Gantt chart of solution A with the best makespan.
are much closer toward the optimal Pareto front and more
evenly distributed. Meanwhile, we can see from Fig. 6 that IDFWRU\+)63

VWDJHPDF   

there is an obvious conflicting relation between makespan VWDJHPDF

VWDJHPDF
VWDJHPDF




 
VWDJHPDF 
and T EC. Furthermore, Table VII shows two extreme ob- VWDJHPDF
VWDJHPDF
VWDJHPDF
 




 

jectives and the corresponding two non-dominated solutions VWDJHPDF

VWDJHPDF
VWDJHPDF


 


VWDJHPDF  
by HMOIG (namely, solution A and B in Fig. 6). Fig. VWDJHPDF
VWDJHPDF

  

         
7 presents a Gantt chart of the solution A with the best
makespan (f1 = 341 min and f2 = 1634.56 kWh). Fig. 8 IDFWRU\3)63

VWDJH              
 

shows a Gantt chart of the solution B with the best T EC VWDJH

VWDJH
  
  
 

 

 
   
 

   
   
 

(f1 = 544 min and f2 = 1495.68 kWh). Based on the VWDJH

VWDJH
  
 






 
 




 
 
 
     
VWDJH                 
practical productive experiments (First Come First Served), VWDJH
VWDJH
 


 
 
 


  
 
     
  
 

the solution Π = (1, 2, · · · , 9, 10|11, 12, · · · , 19, 20) and the VWDJH           


     
corresponding result is (390 min, 1688.33 kWh). The proposed
HMOIG can improve makespan by 14.37% compared to Fig. 8. Gantt chart of solution B with the best T EC.
the practical production. The HMOIG can reduce T EC by
12.88% compared to the practical schedule. Thus, the proposed
HMOIG can solve this problem well.
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. XX, NO. XX, XX 2020 9

TABLE VI
S TATISTICAL RESULTS OF ALL METRICS OF ALL ALGORITHMS IN ALL INSTANCES .
IGD
SPEA2 MOEA/D MMOIG NSGAII AdaW MO-LR HMOIG
instance
mean std mean std mean std mean std mean std mean std mean std
ins 150 10 6.04E-03 4.90E-04 5.36E-03 1.00E-03 1.04E-02 3.10E-04 5.70E-03 4.00E-04 5.80E-03 5.00E-04 5.35E-03 9.80E-04 1.77E-03 (+) 4.50E-04
ins 150 15 6.32E-03 6.00E-04 6.52E-03 1.50E-03 1.31E-02 7.50E-04 6.10E-03 6.90E-04 6.02E-03 6.20E-04 6.49E-03 9.90E-04 2.17E-03 (+) 4.10E-04
ins 150 20 4.81E-03 3.90E-04 4.36E-03 5.60E-04 9.92E-03 4.60E-04 4.67E-03 3.90E-04 4.72E-03 4.50E-04 4.60E-03 6.60E-04 1.51E-03 (+) 3.20E-04
ins 150 25 5.47E-03 3.80E-04 5.05E-03 9.10E-04 1.02E-02 9.60E-04 5.20E-03 4.70E-04 5.17E-03 5.10E-04 4.96E-03 5.60E-04 2.17E-03 (+) 4.50E-04
ins 150 30 5.17E-03 3.90E-04 5.12E-03 5.60E-04 1.01E-02 7.70E-04 5.07E-03 4.80E-04 5.12E-03 4.60E-04 5.02E-03 6.20E-04 2.20E-03 (+) 3.40E-04
ins 200 10 4.73E-03 4.20E-04 4.44E-03 9.10E-04 9.04E-03 4.90E-04 4.50E-03 4.40E-04 4.53E-03 4.70E-04 4.52E-03 9.50E-04 1.29E-03 (+) 2.90E-04
ins 200 15 4.53E-03 3.60E-04 4.00E-03 9.40E-04 8.44E-03 4.90E-04 4.32E-03 4.20E-04 4.36E-03 3.90E-04 4.08E-03 7.50E-04 1.48E-03 (+) 2.10E-04
ins 200 20 3.68E-03 2.60E-04 3.35E-03 5.70E-04 7.45E-03 5.40E-04 3.42E-03 3.10E-04 3.48E-03 3.50E-04 3.39E-03 4.30E-04 1.43E-03 (+) 2.40E-04
ins 200 25 4.65E-03 3.20E-04 5.04E-03 8.90E-04 9.21E-03 5.00E-04 4.42E-03 2.60E-04 4.46E-03 2.60E-04 4.86E-03 9.00E-04 1.96E-03 (+) 2.50E-04
ins 200 30 4.24E-03 3.30E-04 4.06E-03 5.10E-04 7.82E-03 1.20E-03 4.19E-03 3.60E-04 4.10E-03 3.30E-04 4.59E-03 8.60E-04 1.88E-03 (+) 3.30E-04
ins 250 10 5.45E-03 4.60E-04 5.95E-03 7.10E-04 1.08E-02 5.50E-04 5.24E-03 7.60E-04 5.22E-03 6.80E-04 6.17E-03 1.40E-03 1.92E-03 (+) 2.90E-04
ins 250 15 5.18E-03 4.40E-04 4.15E-03 5.90E-04 8.62E-03 4.90E-04 4.95E-03 4.40E-04 4.80E-03 5.00E-04 4.25E-03 8.10E-04 1.72E-03 (+) 4.30E-04
ins 250 20 3.73E-03 3.60E-04 3.62E-03 1.00E-03 7.54E-03 7.60E-04 3.66E-03 3.30E-04 3.67E-03 3.10E-04 3.46E-03 5.60E-04 1.43E-03 (+) 2.70E-04
ins 250 25 3.44E-03 2.30E-04 3.48E-03 5.90E-04 7.61E-03 8.10E-04 3.34E-03 2.60E-04 3.28E-03 3.20E-04 3.54E-03 7.80E-04 1.53E-03 (+) 3.20E-04
ins 250 30 3.56E-03 2.70E-04 3.59E-03 6.90E-04 7.37E-03 8.50E-04 3.45E-03 3.00E-04 3.51E-03 3.50E-04 3.62E-03 6.50E-04 1.65E-03 (+) 2.40E-04
GD
SPEA2 MOEA/D MMOIG NSGAII AdaW MO-LR HMOIG
instance
mean std mean std mean std mean std mean std mean std mean std
ins 150 10 8.64E-03 6.10E-04 5.39E-03 8.00E-04 1.63E-02 7.00E-04 8.40E-03 6.50E-04 8.50E-03 8.30E-04 5.43E-03 8.30E-04 2.68E-03 (+) 7.00E-04
ins 150 15 8.51E-03 1.00E-03 5.49E-03 8.90E-04 1.90E-02 1.10E-03 8.42E-03 1.00E-03 8.25E-03 1.10E-03 5.75E-03 7.50E-04 2.97E-03 (+) 5.80E-04
ins 150 20 8.56E-03 8.00E-04 5.48E-03 8.80E-04 1.78E-02 7.30E-04 8.45E-03 8.40E-04 8.47E-03 9.20E-04 5.34E-03 1.00E-03 2.61E-03 (+) 6.10E-04
ins 150 25 8.64E-03 7.70E-04 5.31E-03 9.20E-04 1.53E-02 7.00E-04 8.29E-03 8.40E-04 8.24E-03 8.70E-04 5.13E-03 7.60E-04 3.05E-03 (+) 6.10E-04
ins 150 30 8.52E-03 8.60E-04 5.43E-03 6.60E-04 1.50E-02 6.40E-04 8.57E-03 8.10E-04 8.63E-03 7.90E-04 5.61E-03 7.10E-04 3.05E-03 (+) 5.00E-04
ins 200 10 7.46E-03 8.00E-04 4.89E-03 6.80E-04 1.52E-02 7.10E-04 7.52E-03 6.80E-04 7.44E-03 8.50E-04 4.97E-03 8.00E-04 2.17E-03 (+) 4.70E-04
ins 200 15 7.27E-03 6.70E-04 4.04E-03 5.80E-04 1.35E-02 3.70E-04 7.38E-03 7.30E-04 7.34E-03 6.50E-04 4.55E-03 6.00E-04 2.38E-03 (+) 3.50E-04
ins 200 20 6.61E-03 4.70E-04 3.79E-03 5.30E-04 1.18E-02 3.80E-04 6.25E-03 6.00E-04 6.28E-03 7.50E-04 3.89E-03 5.60E-04 2.27E-03 (+) 3.60E-04
ins 200 25 6.78E-03 6.50E-04 4.17E-03 5.90E-04 1.32E-02 3.90E-04 6.52E-03 5.60E-04 6.65E-03 3.80E-04 3.93E-03 6.30E-04 2.62E-03 (+) 3.70E-04
ins 200 30 6.91E-03 5.30E-04 3.94E-03 5.50E-04 1.10E-02 4.10E-04 7.13E-03 5.30E-04 6.96E-03 4.90E-04 3.94E-03 5.90E-04 2.70E-03 (+) 4.40E-04
ins 250 10 8.30E-03 9.70E-04 5.91E-03 8.30E-04 1.90E-02 1.40E-03 8.13E-03 1.30E-03 8.28E-03 1.20E-03 5.76E-03 8.90E-04 2.83E-03 (+) 5.20E-04
ins 250 15 9.38E-03 1.10E-03 5.30E-03 9.60E-04 1.90E-02 1.30E-03 9.18E-03 1.00E-03 8.81E-03 9.60E-04 5.27E-03 1.10E-03 3.00E-03 (+) 6.70E-04
ins 250 20 6.44E-03 7.30E-04 3.34E-03 6.10E-04 1.15E-02 5.90E-04 6.56E-03 6.50E-04 6.57E-03 6.30E-04 3.54E-03 6.00E-04 2.44E-03 (+) 4.50E-04
ins 250 25 6.45E-03 5.60E-04 3.71E-03 6.00E-04 1.15E-02 4.30E-04 6.46E-03 5.30E-04 6.27E-03 5.80E-04 3.80E-03 6.50E-04 2.51E-03 (+) 4.40E-04
ins 250 30 6.68E-03 6.50E-04 3.95E-03 7.10E-04 1.27E-02 4.90E-04 6.54E-03 6.80E-04 6.66E-03 7.50E-04 3.94E-03 6.80E-04 2.78E-03 (+) 4.00E-04
Spread
SPEA2 MOEA/D MMOIG NSGAII AdaW MO-LR HMOIG
instance
mean std mean std mean std mean std mean std mean std mean std
ins 150 10 1.08E+00 1.00E-01 1.35E+00 8.70E-02 7.19E-01 3.80E-02 9.35E-01 9.60E-02 9.40E-01 1.00E-01 1.33E+00 7.20E-02 3.30E-01 (+) 2.70E-02
ins 150 15 1.15E+00 9.60E-02 1.34E+00 9.00E-02 7.38E-01 4.50E-02 1.05E+00 1.20E-01 1.08E+00 1.00E-01 1.28E+00 6.70E-02 3.47E-01 (+) 3.30E-02
ins 150 20 9.65E-01 9.80E-02 1.32E+00 7.20E-02 6.95E-01 5.40E-02 8.65E-01 8.00E-02 8.87E-01 1.00E-01 1.31E+00 9.10E-02 3.26E-01 (+) 3.10E-02
ins 150 25 1.02E+00 9.70E-02 1.33E+00 6.80E-02 7.10E-01 5.00E-02 9.28E-01 1.10E-01 9.05E-01 9.80E-02 1.34E+00 7.00E-02 3.48E-01 (+) 3.40E-02
ins 150 30 9.66E-01 9.90E-02 1.35E+00 6.60E-02 6.44E-01 4.80E-02 8.44E-01 9.20E-02 8.53E-01 7.90E-02 1.35E+00 5.80E-02 3.33E-01 (+) 3.30E-02
ins 200 10 1.04E+00 9.40E-02 1.15E+00 9.50E-02 7.48E-01 4.20E-02 9.47E-01 9.60E-02 9.04E-01 1.10E-01 1.16E+00 9.50E-02 3.11E-01 (+) 3.20E-02
ins 200 15 9.34E-01 1.10E-01 1.21E+00 1.00E-01 6.87E-01 6.80E-02 7.70E-01 7.70E-02 7.94E-01 7.50E-02 1.20E+00 8.60E-02 3.11E-01 (+) 2.80E-02
ins 200 20 8.33E-01 9.10E-02 1.32E+00 9.10E-02 6.01E-01 6.00E-02 7.08E-01 6.50E-02 7.17E-01 7.70E-02 1.32E+00 7.10E-02 3.03E-01 (+) 2.80E-02
ins 200 25 8.83E-01 8.10E-02 1.34E+00 7.50E-02 6.09E-01 6.00E-02 7.65E-01 6.30E-02 7.56E-01 6.60E-02 1.33E+00 7.90E-02 3.29E-01 (+) 3.00E-02
ins 200 30 8.04E-01 9.40E-02 1.36E+00 6.60E-02 6.15E-01 5.20E-02 7.04E-01 7.00E-02 7.18E-01 7.40E-02 1.39E+00 6.30E-02 2.96E-01 (+) 2.90E-02
ins 250 10 1.13E+00 8.50E-02 1.16E+00 6.40E-02 7.69E-01 4.80E-02 1.00E+00 8.30E-02 1.02E+00 9.10E-02 1.17E+00 6.70E-02 4.12E-01 (+) 3.80E-02
ins 250 15 1.22E+00 9.30E-02 1.41E+00 8.20E-02 8.40E-01 6.30E-02 1.10E+00 1.10E-01 1.16E+00 1.10E-01 1.40E+00 6.80E-02 3.48E-01 (+) 4.40E-02
ins 250 20 8.95E-01 7.70E-02 1.34E+00 8.70E-02 6.32E-01 6.60E-02 7.49E-01 6.30E-02 7.48E-01 5.90E-02 1.34E+00 8.20E-02 3.10E-01 (+) 2.10E-02
ins 250 25 8.40E-01 7.90E-02 1.39E+00 7.20E-02 5.77E-01 4.60E-02 7.26E-01 6.30E-02 7.33E-01 8.50E-02 1.38E+00 7.10E-02 3.03E-01 (+) 2.70E-02
ins 250 30 8.70E-01 6.80E-02 1.42E+00 5.40E-02 6.06E-01 5.60E-02 7.97E-01 7.20E-02 7.75E-01 7.10E-02 1.41E+00 7.40E-02 3.21E-01 (+) 3.00E-02

TABLE VII of distributed shop scheduling problems with heterogeneous

E XTREME SOLUTIONS FOUND BY HMOIG ON THE REAL - WORLD CASE . factories are also of interest. Third, developing a more effective
No. solution: Π f1 (min) f2 (kWh) algorithm by analyzing the properties for DFSPHF is another
A
F1 :7,14,1,2,13,12,10,9,17,19
341 1634.56 future research direction.
F2 :3, 15, 4, 8, 5, 6, 16, 11, 18, 20
F1 :4,8,18
B 544 1495.68
F2 :6,15,5,16,11,10,7,2,3,12,9,13,17,14,19,1,20
R EFERENCES
[1] R. Ruiz, Q. Pan, and B. Naderi, “Iterated greedy methods for
the distributed permutation flowshop scheduling problem,” Omega-
V. C ONCLUSION international Journal of Management Science, vol. 83, pp. 213–222,
2019.
Green manufacturing and cooperation manufacturing among [2] X. Gong, Y. Liu, N. Lohse, T. De Pessemier, L. Martens, and W. Joseph,
multiple factories have become a modern production mode “Energy-and labor-aware production scheduling for industrial demand
in large enterprises. Although green distributed scheduling response using adaptive multiobjective memetic algorithm,” IEEE Trans-
actions on Industrial Informatics, vol. 15, no. 2, pp. 942–953, 2018.
problem with heterogeneous factories has widely appeared in [3] Y. Fang, C. Peng, P. Lou, Z. Zhou, J. Hu, and J. Yan, “Digital-twin-based
many industries, no research on this problem is reported. Thus, job shop scheduling toward smart manufacturing,” IEEE Transactions on
this work is the first attempt to investigate an energy-efficient Industrial Informatics, vol. 15, no. 12, pp. 6425–6435, 2019.
[4] J.-J. Wang and L. Wang, “A knowledge-based cooperative algorithm for
DFSPHF. The major contribution of this work is to extend the energy-efficient scheduling of distributed flow-shop,” IEEE Transactions
research field of distributed shop scheduling problems. At the on Systems, Man, and Cybernetics: Systems, no. 99, pp. 1–15, 2018.
same time, the HMOIG algorithm is proposed to solve a real- [Online]. Available: doi:10.1109/TSMC.2017.2788879
[5] J. Deng and L. Wang, “A competitive memetic algorithm for multi-
world case from an automobile plant in China. Experiment objective distributed permutation flow shop scheduling problem,” Swarm
is conducted to assess the performance of the HMOIG algo- and Evolutionary Computation, vol. 32, pp. 121–131, 2017.
rithm. Experimental results show that HMOIG is significantly [6] Y. Fu, G. Tian, A. M. Fathollahi-Fard, A. Ahmadi, and C. Zhang,
“Stochastic multi-objective modelling and optimization of an energy-
better than other algorithms in terms of obtaining compromise conscious distributed permutation flow shop scheduling problem with
solutions with good spread and convergence performance. the total tardiness constraint,” Journal of Cleaner Production, vol. 226,
Concerning future research, first, it is interesting to take pp. 515–525, 2019.
[7] J. Gao and R. Chen, “A hybrid genetic algorithm for the distributed
into account realistic constraints in DFSPHF, such as transport permutation flowshop scheduling problem,” International Journal of
time and sequence-dependent setup time. Second, other types Computational Intelligence Systems, vol. 4, no. 4, pp. 497–508, 2011.
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. XX, NO. XX, XX 2020 10

[8] S. Lin, K. Ying, and C. Huang, “Minimising makespan in distributed [30] E. Zitzler, K. Deb, and L. Thiele, “Comparison of multiobjective
permutation flowshops using a modified iterated greedy algorithm,” evolutionary algorithms: Empirical results,” Evolutionary Computation,
International Journal of Production Research, vol. 51, no. 16, pp. 5029– vol. 8, no. 2, pp. 173–195, 2000.
5038, 2013. [31] R. C. V. Nostrand, “Design of experiments using the taguchi approach:
[9] H. Bargaoui, O. B. Driss, and K. Ghedira, “A novel chemical reaction 16 steps to product and process improvement,” Technometrics, vol. 44,
optimization for the distributed permutation flowshop scheduling prob- p. 289, 2002.
lem with makespan criterion,” Computers & Industrial Engineering, vol. [32] H. Li and Q. Zhang, “Multiobjective optimization problems with com-
111, pp. 239–250, 2017. plicated pareto sets, MOEA/D and NSGA-II,” IEEE Transactions on
[10] V. Fernandezviagas, P. Perezgonzalez, and J. M. Framinan, “The dis- Evolutionary Computation, vol. 13, no. 2, pp. 284–302, 2008.
tributed permutation flow shop to minimise the total flowtime,” Com- [33] E. Zitzler, M. Laumanns, and L. Thiele, “SPEA2: Improving the strength
puters & Industrial Engineering, vol. 118, pp. 464–477, 2018. pareto evolutionary algorithm,” TIK-report, vol. 103, 2001.
[11] K.-C. Ying and S.-W. Lin, “Minimizing makespan for the distributed [34] M. Li and X. Yao, “What Weights Work for You? Adapting Weights
hybrid flowshop scheduling problem with multiprocessor tasks,” Expert for Any Pareto Front Shape in Decomposition-Based Evolutionary
Systems with Applications, vol. 92, pp. 132–141, 2018. Multiobjective Optimisation,” Evolutionary Computation, vol. 28, no. 2,
[12] W. Shao, Z. Shao, and D. Pi, “Modeling and multi-neighborhood iterated pp. 227–253, JUN 2020.
greedy algorithm for distributed hybrid flow shop scheduling problem,” [35] S. Wang, M. Gong, Y. Wu, and M. Zhang, “Multi-objective optimization
Knowledge Based Systems, p. 105527, 2020. for location-based and preferences-aware recommendation,” Information
Sciences, vol. 513, pp. 614–626, MAR 2020.
[13] B. Naderi and A. Azab, “Modeling and heuristics for scheduling of
distributed job shops,” Expert Systems with Applications, vol. 41, no. 17,
pp. 7754–7763, 2014. Chao Lu received the Ph.D. degree in industrial en-
[14] C.-Y. Hsu, B.-R. Kao, K. R. Lai et al., “Agent-based fuzzy constraint- gineering from the Huazhong University of Science
directed negotiation mechanism for distributed job shop scheduling,” and Technology, Wuhan, China, 2017.
Engineering Applications of Artificial Intelligence, vol. 53, pp. 140–154, He is an associate professor of computer sciences
2016. with the School of Computer Science, China Uni-
[15] I. Chaouch, O. B. Driss, and K. Ghedira, “A novel dynamic assignment versity of Geosciences, Wuhan. His research inter-
rule for the distributed job shop scheduling problem using a hybrid ant- ests include multiobjective evolutionary algorithm,
based algorithm,” Applied Intelligence, vol. 49, no. 5, pp. 1903–1924, scheduling, etc.
2019.
[16] X. Zhang, X.-T. Li, and M.-H. Yin, “An enhanced genetic algorithm
for the distributed assembly permutation flowshop scheduling problem,”
International Journal Of Bio-inspired Computation, vol. 15, no. 2, pp. Liang Gao (M’08- SM’20) received the B.Sc. de-
113–124, 2020. gree in mechatronic engineering from Xidian Uni-
[17] W. Shao, D. Pi, and Z. Shao, “Local Search Methods for a Distributed versity, Xi’an, China, in 1996, and the Ph.D. degree
Assembly No-Idle Flow Shop Scheduling Problem,” IEEE Systems in mechatronic engineering from the Huazhong Uni-
Journal, vol. 13, no. 2, pp. 1945–1956, JUN 2019. versity of Science and Technology (HUST), Wuhan,
[18] G. Zhang, K. Xing, and F. Cao, “Discrete differential evolution algorithm China, in 2002.
for distributed blocking flowshop scheduling with makespan criterion,” He is a Professor of the Department of Indus-
Engineering Applications of Artificial Intelligence, vol. 76, pp. 96–107, trial and Manufacturing Systems Engineering, the
NOV 2018. Deputy Director of State Key Laboratory of Digital
[19] G. Zhang and K. Xing, “Differential evolution metaheuristics for dis- Manufacturing Equipment and Technology. He was
tributed limited-buffer flowshop scheduling with makespan criterion,” supported by the National Science Fund for Dis-
Computers & Operations Research, vol. 108, pp. 33–43, AUG 2019. tinguished Young Scholars of China in 2018. His research interests include
[20] D. Gong, Y. Han, and J. Sun, “A novel hybrid multi-objective artificial operations research and optimization, big data and machine learning etc. He
bee colony algorithm for blocking lot-streaming flow shop scheduling has published over 200 papers indexed by SCIE, authored 7 monographs.
problems,” Knowledge-Based Systems, vol. 148, pp. 115 – 130, 2018.
[21] Y. Han, D. Gong, J. Li, and Y. Zhang, “Solving the blocking flow
shop scheduling problem with makespan using a modified fruit fly Jin Yi received his B.S and Ph.D. degrees in indus-
optimisation algorithm,” International Journal of Production Research, trial engineering from the Huazhong University of
vol. 54, no. 22, pp. 6782–6797, 2016. Science and Technology (HUST), Wuhan, China, in
[22] M. R. Garey, D. S. Johnson, and R. Sethi, “The complexity of flowshop 2012 and 2017. He was a research fellow at the De-
and jobshop scheduling,” Mathematics of Operations Research, vol. 1, partment of Industrial System Engineering, National
no. 2, pp. 117–129, 1976. University of Singapore from 2017 to 2020.
[23] S. Wang, Z. Y. Dong, F. Luo, K. Meng, and Y. Zhang, “Stochas- He is currently an assistant professor at the de-
tic collaborative planning of electric vehicle charging stations and partment of Mechanical Engineering, Chongqing
power distribution system,” IEEE Transactions on Industrial Informatics, University, China. His research interests include
vol. 14, no. 1, pp. 321–331, 2017. intelligent algorithm, surrogate-assisted global opti-
[24] H. Yuan, J. Bi, M. Zhou, Q. Liu, and A. C. Ammari, “Biobjective task mization etc.
scheduling for distributed green data centers,” IEEE Transactions on
Automation Science and Engineering, 2020. Xinyu Li (M’19) received his Ph.D. degree in
[25] Y. Yuan and H. Xu, “Multiobjective flexible job shop scheduling using industrial engineering from Huazhong University of
memetic algorithms,” IEEE Transactions on Automation Science and Science and Technology (HUST), China, 2009.
Engineering, vol. 12, no. 1, pp. 336–353, 2013. He is a Professor of the Department of Industrial
[26] C. Lu, L. Gao, Q. Pan, X. Li, and J. Zheng, “A multi-objective cellular & Manufacturing Systems Engineering, State Key
grey wolf optimizer for hybrid flowshop scheduling problem considering Laboratory of Digital Manufacturing Equipment &
noise pollution,” Applied Soft Computing, vol. 75, pp. 728–749, 2019. Technology, School of Mechanical Science & En-
[27] J.-Y. Ding, S. Song, and C. Wu, “Carbon-efficient scheduling of flow gineering, HUST. He had published more than 90
shops by multi-objective optimization,” European Journal of Opera- refereed papers. His research interests include intel-
tional Research, vol. 248, no. 3, pp. 758–771, 2016. ligent algorithm, scheduling and machine learning
[28] R. Zhang, S. Song, and C. Wu, “Robust scheduling of hot rolling etc.
production by local search enhanced ant colony optimization algorithm,”
IEEE Transactions on Industrial Informatics, vol. 16, no. 4, pp. 2809–
2819, 2019.
[29] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, “A fast and elitist
multiobjective genetic algorithm: NSGA-II,” IEEE Transactions on
Evolutionary Computation, vol. 6, no. 2, pp. 182–197, 2002.