Download as pdf or txt
Download as pdf or txt
You are on page 1of 19

International Journal of Production Research

ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/tprs20

Optimised redesign of reverse logistics network


with multi-level capacity choices for household
appliances

Ada Che, Jieyu Lei & Zhibin Jiang

To cite this article: Ada Che, Jieyu Lei & Zhibin Jiang (2021): Optimised redesign of reverse
logistics network with multi-level capacity choices for household appliances, International Journal of
Production Research, DOI: 10.1080/00207543.2021.1967499

To link to this article: https://doi.org/10.1080/00207543.2021.1967499

Published online: 01 Sep 2021.

Submit your article to this journal

Article views: 95

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at


https://www.tandfonline.com/action/journalInformation?journalCode=tprs20
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH
https://doi.org/10.1080/00207543.2021.1967499

Optimised redesign of reverse logistics network with multi-level capacity choices


for household appliances
Ada Chea , Jieyu Leia and Zhibin Jianga,b
a School of Management, Northwestern Polytechnical University, Xi’an, People’s Republic of China; b Antai College of Economics and
Management, Shanghai Jiao Tong University, Shanghai, People’s Republic of China

ABSTRACT ARTICLE HISTORY


Reverse logistics networks reintroduce end-of-life products to remanufacturing, which is significant Received 25 November 2020
for sustainable development and environmental protection. In this paper, we investigate the current Accepted 14 July 2021
reverse logistics network of household appliances in China and redesign a new network that intro- KEYWORDS
duces necessary facilities: disassembly centres and three types of remanufacturing plants to improve Reverse logistics network
recycling rates. A mixed-integer linear programming model with multi-level capacity choices is redesign; facility location;
proposed to determine the locations and capacities for disassembly centres and remanufacturing capacity choices;
plants. An efficient decomposition-and-expansion heuristic is developed to solve the model. The mixed-integer programming;
redesigned network is evaluated using a real case. The results indicate that the recycling rate is decomposition-and-
largely improved by the new network in which almost all useful materials can be recycled, whereas, expansion
in the traditional network, only metals are efficiently recycled. A sensitivity analysis of model param- heuristic
eters is conducted using the Taguchi method to identify their effects on recycling decisions. The
proposed algorithm is further evaluated using a set of randomly generated instances. The results
show that the algorithm can yield high-quality solutions within a short time. The best configuration
of the algorithm is suggested via sensitivity analysis of parameters using the generated instances.

1. Introduction
collection centres and smelters, as shown in Figure 1(a).
Rapid technological innovation leads to many house- Collection centres collect various end-of-life household
hold appliances being quickly replaced by new prod- appliances from cities to obtain valuable metal materials;
ucts. The increasing number of end-of-life household smelters reprocess materials that can be reused in man-
appliances has attracted considerable attention in reverse ufacturing. Although most metals can be recycled by the
logistics network design. According to the Beijing Con- network, some materials such as plastic and glass are still
sumer Association, about 7 billion household appliances poorly treated because of insufficient disassembly at the
were used in China in 2018. With the purchase of new collection centres and the shortage of remanufacturing
products, over 20 million end-of-life household appli- plants.
ances need to be recycled every year. These end-of- To improve the recycling performance, we redesign
life products are rich in valuable materials that can the reverse logistics network, as shown in Figure 1(b),
be widely reused in manufacturing (Park et al. 2006; which introduces disassembly centres, plastic plants, and
Li et al. 2021). An efficient reverse logistics network glass plants into the traditional network. In the new net-
plays a significant role in improving the recycling rate work, collection centres collect end-of-life various house-
while reducing the negative influence on the environ- hold appliances from cities and send them to disassem-
ment (Pishvaee, Farahani, and Dullaert 2010; Lu et al. bly centres. Disassembly centres sort these products and
2021). Five household products are on the recycling allocate them in different disassembly lines. In general,
list for their large quantities and high values, includ- three disassembly lines are included in a disassembly cen-
ing TVs, computers, air conditioners, refrigerators, and tre: television/computer, refrigerator/air conditioner, and
washing machines. In recent years, more electronic prod- washing machine disassembly lines. End-of-life house-
ucts are being considered for addition to the recycling hold products are separated into metals, plastics, and
list. glass pieces via these disassembly lines; then, they are
Traditional reverse logistics networks for household delivered to smelters, plastic plants, and glass plants for
appliances in China consists of two-echelon facilities: further treatments.

CONTACT Zhibin Jiang zbjiang@sjtu.edu.cn


© 2021 Informa UK Limited, trading as Taylor & Francis Group
2 A. CHE ET AL.

Figure 1. Traditional (a) and redesigned (b) reverse logistics network of household appliances in China.

Disassembly centres and various remanufacturing on RLND and RLNR (RLND/R) to highlight our
plants play an important role in improving the recy- contributions.
cling rate (Tang, Zhou, and Caudill 2001). Disassem- The RLND/R have some characteristics that differ-
bly centres break end-of-life products into materials. entiate them from the forward logistics network design
Researchers focus on various disassembly problems such problem (Fleischmann et al. 2001; Jayaraman, Patter-
as disassembly sequence, disassembly line balancing, and son, and Rolland 2003; Min, Jeung Ko, and Seong Ko
lot-sizing disassembly problem (Guo et al. 2018; Zhao 2006). Firstly, multiple products and multiple materials
et al. 2018; Guo, Zhou, Abusorrah, et al. 2020; Guo, are simultaneously considered in the reverse flow. Sec-
Zhou, Liu, et al. 2020; Guo et al. 2021). Remanufacturing ondly, products with different quality or components are
plants reprocess materials and reuse them in produc- generally handled in different manners in the reverse
tion. Researchers pay attention to various problems on logistics network. Thirdly, more types of facilities are
reprocessing planning, remanufacturing scheduling, and required for recycling products and materials, such as
marketing strategy (Polotski, Kenne, and Gharbi 2017; collection centres, inspection centres, disassembly cen-
Lechner and Reimann 2020; Shi et al. 2021). We consider tres, repair centres, remanufacturing plants, and landfills.
both disassembly centres and remanufacturing plants in Finally, the uncertain supply of products makes it harder
the redesigned network. The key issue is to determine to decide locations and capacities for facilities in the
their best locations and capacities. During redesigning RLND/R. Due to these characteristics, researchers sug-
the network, some plants have been established in the tra- gest that more attention should be paid to the reverse
ditional network or existed in many places. For example, problems.
in the Shaanxi province of China, 21 smelters, 13 plas- Several research aspects have been discussed in
tic pants, and 8 glass plants were established to meet the the existing RLND/R literature, such as multi-echelon
recycling demand. Another challenge in the reverse logis- reverse logistics networks (e.g. Jayaraman, Patterson, and
tics network redesign problem is to determine if these Rolland 2003; Min, Jeung Ko, and Seong Ko 2006; Eskan-
existing remanufacturing plants should be used in the darpour et al. 2014; Tosarkani, Amin, and Zolfagharinia
redesigned network. 2020), multi-product reverse logistics networks (e.g. Alu-
In recent years, reverse logistics networks have been mur et al. 2012; Bing et al. 2015; Sadrnia, Langarudi,
widely studied in the literature. Most studies investigate and Sani 2020), multi-period recycling (e.g. Xu et al.
the reverse logistics network design problem (RLND) 2017; Reddy, Kumar, and Ballantyne 2019; Shi et al.
that aims to build a new reverse logistics network for 2020), multi-objective decision (e.g. Pishvaee, Farahani,
emerging recovery industries (e.g. Ene and Öztürk 2015; and Dullaert 2010; Tosarkani, Amin, and Zolfagharinia
Sadrnia, Langarudi, and Sani 2020; Rentizelas et al. 2021). 2020), and the effect of uncertainty (e.g. Lieckens and
Some studies focus on the reverse logistics network Vandaele 2007; Krug, Guillaume, and Battaïa 2021). Table
redesign problem (RLNR), which aims to improve the 1 summarises the recent literature on RLND/R. This
existing network through various strategies such as the paper focuses on the multi-echelon and multi-product
insertion of new facilities or relocation of current facil- reverse logistics network redesign problem.
ities (Bing et al. 2015; Trochu, Chaabane, and Ouhim- Multi-echelon reverse logistics network depicts the
mou 2018; Gao and Cao 2020). RLNR is similar to an complicated recycling process of end-of-life products.
RLND with some existing facilities. When redesigning a Various facilities are designed in the network such as col-
reverse logistics network, these existing facilities should lection centres, disassembly centres, and remanufactur-
be considered for use. We relate our work with literature ing plants. The relevant research focuses on determining
Table 1. Comparison of the recent RLND/R research.
Authors Multi-echelon Multi-product Multi-period Multi-objective Uncertainty Model Facility location Facility capacity Approach Experiment

Jayaraman, Patterson, MILP Collection centr, Reman- Certain Concentration heuristic Numerical examples
and Rolland (2003) ufacturing
√ plant
Min, Jeung Ko, and MINLP Collection cen- Certain Genetic algorithm Numerical examples
Seong Ko (2006) tre, Remanufacturing
√ plant
Lieckens and Vandaele MINLP Reprocessing centre Multi-level Differential evolution Numerical examples
(2007) √ √
Pishvaee, Farahani, and MILP Production plant, Recovery/ Multi-level Genetic algorithm Numerical examples
Dullaert (2010) Distribution/Collection/
√ √ √ Disposal centre
Alumur et al. (2012) MILP Inspection cen- Modular A case study in Germany
tre, Remanufacturing
√ √ plant
Eskandarpour et al. MILP Collection centre, Recycling Certain Tabu search Numerical examples
(2014) centre, Disposal /Recovery
√ centre
Bing et al. (2015) MILP Reprocessing centre Certain A case study between
√ √ √ √ Netherland and China
Xu et al. (2017) MILP Sorting centre, Recycling plant Certain A case study between
√ √ Greece and China
Paydar and Olfati MILP Collection cen- Certain Genetic algorithm A case study in

INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH


(2018) tre, Remanufacturing Iran andnumerical
√ √ plant examples
Trochu, Chaabane, and MILP Sorting centre Multi-level A case study in Canada
Ouhimmou (2018) √ √
Reddy, Kumar, and MILP Inspection centre, Certain Three-phase heuristic Numerical examples
Ballantyne (2019) √ √ √ Remanufacturing centre
Sadrnia, Langarudi, MILP Collection centre, Charity Certain A case study in Iran
and Sani (2020) √ √ √ √ centre
Tosarkani, Amin, and MINLP Collection centre, Recovery Certain A two-phase fuzzy Numerical examples
Zolfagharinia (2020) centre, Remanufacturing compromise approach
√ √ plant, Supplier/Retailer
This study MILP Disassembly centre, Multi-level DEH heuristic A case study in
Remanufacturing plant China andnumerical
examples

3
4 A. CHE ET AL.

facility locations (e.g. Jayaraman, Patterson, and Rol- considered collection centres, recycling centres, and
land 2003; Tosarkani, Amin, and Zolfagharinia 2020). repair centres in their network and proposed a multi-
Most researchers consider a set of candidate facilities echelon multi-product MILP model with uncertainty
with certain capacities and propose a mix-integer pro- to decide the facility locations and transportation vol-
gramming model (MIP) to decide their locations. Heuris- umes. Although these researches have well studied the
tic approaches are widely used to solve the proposed RLND of household appliances, they still have three
model. The meta-heuristic approaches include genetic shortcomings: (1) few studies clearly describe the dis-
algorithm (Min, Jeung Ko, and Seong Ko 2006), sim- assembly and remanufacturing process; (2) most papers
ulated annealing (Pishvaee, Kianfar, and Karimi 2010), consider certain capacities rather than multi-level capac-
memetic algorithm (Pishvaee, Farahani, and Dullaert ities for facilities to decide their locations; and (3) few
2010), Tabu Search (Eskandarpour et al. 2014), fruit fly efficient approaches are developed to solve the proposed
optimisation algorithm (Fu et al. 2021), beetle anten- models.
nae Search (Khan et al. 2020), and gravitational search This study investigates the entire recycling pro-
algorithm (Wang et al. 2021). Other heuristic approaches cess of household appliances in China and redesigns
include the concentration heuristic (Jayaraman, Patter- the reverse logistics network with necessary disassem-
son, and Rolland 2003), the two-stage heuristic approach bly centres and three types of remanufacturing plants.
(Lee and Dong 2008), and the three-phase heuristic A multi-echelon multi-capacity multi-product multi-
approach (Reddy, Kumar, and Ballantyne 2019). material MILP model is proposed to determine the
Multi-echelon single-product RLND/R research locations and capacities for disassembly centres and
investigates the recycling process for one product or one remanufacturing plants. To solve the proposed model,
type of similar products such as wood, sand, office paper, we develop an efficient decomposition-and-expansion
and polyethene terephthalate bottles. The relevant studies heuristic (DEH) algorithm and use a case study and sev-
include Barros, Dekker, and Scholten (1998), Jayaraman, eral randomly generated instances to evaluate its perfor-
Patterson, and Rolland (2003), Min, Jeung Ko, and Seong mance.
Ko (2006), Zhou and Zhou (2015), Paydar and Olfati Comparing to the existing literature, this paper mainly
(2018), etc. Multi-echelon multi-product RLND/R liter- makes the following contributions. Firstly, it redesigns
ature aims to design a comprehensive reverse logistics a new reverse logistics network of household appli-
network that recycles different products such as vehi- ances that incorporates disassembly centres and various
cles, household appliances, and solid wastes. The relevant remanufacturing plants to improve the recycling ratio.
studies include Bing et al. (2015), Xu et al. (2017), Guo, As many remanufacturing plants have been established
Zhang, et al. (2020), etc. in the traditional network, it also discusses if these exist-
Many researchers have studied the RLND for house- ing facilitates should be remained in the redesigned net-
hold appliances or called Waste Electrical and Electronic work. Secondly, it considers multi-level capacity choices,
Equipment (WEEE). Shih (2001) designed a network instead of fixed capacities, for facilities and proposes a
with collection centres, storage sites, disassembly plants, MILP to decide their locations and capacities. Finally, it
and disposition facilities and proposed a multi-echelon develops an efficient DEH algorithm to solve the pro-
multi-product multi-material mix-integer linear pro- posed model.
gramming model (MILP) to decide facility locations and The remainder of this paper is organised as follows.
transportation volumes. Achillas et al. (2010) built a In Section 2, the detailed problem description and its
network with collection centres, storage stations, and mathematical formulation are demonstrated. Section 3
treatment units and proposed a multi-echelon MILP presents the DEH algorithm. Section 4 introduces the
model to decide facility locations and transportation numerical experiment of a real-life instance. Section 5
volumes. Ayvaz, Bolat, and Aydın (2015) constructed evaluates the DEH algorithm with randomly generated
a network with collection centres, sorting centres, and instances. Finally, the conclusions and future directions
recycling centres and formulated a multi-echelon, multi- are discussed in Section 6.
product and two-stage stochastic programming model
to decide facility locations and transportation volumes.
2. Problem description and formulation
Tosarkani, Amin, and Zolfagharinia (2020) considered
collection centres, recovery centres, and remanufactur- Consider a set of cities where multiple household appli-
ing plants in their network and proposed a multi- ances are recycled through the traditional method, as
echelon multi-product multi-material multi-objective shown in Figure 1(a). In the traditional network, col-
MILP model to decide facility locations and trans- lection centres collect various products and break them
portation volumes. Sadrnia, Langarudi, and Sani (2020) into valuable metal materials. Smelters reprocess the
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 5

materials and use them for remanufacturing. The net- D Set of candidate disassembly centres
work is less efficient because collection centres cannot L Set of disassembly lines
efficiently disassemble products and remanufacturing R Set of the existing remanufacturing plants
plants for plastic and glass pieces are quite inadequate. R Set of the new candidate remanufacturing plants
To improve the recycling rate, disassembly centres and K Types of remanufacturing plants
various types of remanufacturing plants should be intro- S Set of capacity level choices
duced in the network. We redesign a new network as
For simplicity, Al represents the set of end-of-life
shown in Figure 1(b).
household appliances processed in disassembly line l. Rk
In the new network, collection centres collect products
and Rk indicate the set of k type of existing remanufac-
and send them to disassembly centres. Disassembly cen-
turing plants and new remanufacturing plants, respec-
tres break products into materials and send them to dif-
tively. Mk denotes the set of materials processed in k-type
ferent remanufacturing plants. Remanufacturing plants
remanufacturing plants.
reprocess the materials and use them in remanufacturing.
Parameters
The key problem in redesigning the network is to decide
the best locations and capacities of disassembly centres on,a Output amount of end-of-life household
and remanufacturing plants. However, many remanufac- appliance a in city n
turing plants are established in the traditional network or wa Weight of household appliance a
existed in many places. Another challenge in redesigning cn Collecting cost in city n
the network is to decide if these existing remanufacturing σa Subsidy for recycling household appliance
plants should be used in the new network. a
In this paper, we consider a set of candidate disas- θ a,m Conversion rate from household appliance
sembly centres D = {1, 2, . . . , |D|} in the new network. a to material m
A = {1, 2, . . . , |A|} kinds of household appliances are col- D
vd,l Variable cost of disassembly line l in can-
lected from cities N = {1, 2, . . . , |N|} and sent to these didate disassembly centre d
disassembly centres. Each disassembly centre has several D
cl,s Capacity of disassembly line l in level s
disassembly lines L = {1, 2, . . . , |L|}. Each disassembly D
fd,l,s Fixed cost of candidate disassembly centre
line chooses a capacity level from capacity choices S = {1,
d to set the capacity of disassembly line l in
2, . . . , |S|}. M = {1, 2, . . . , |M|} types of materials are
level s
abstracted from household appliances. These materials R
vr,m Variable cost of the existing remanufactur-
are sent to K = {1, 2, . . . , |K|} types of remanufactur-
ing plant r to process material m
ing plants. A set of remanufacturing plants R = {1, 2, . . . , R
cr,m Capacity of the existing remanufacturing
|R|} have been built in the traditional network. They have
plant r to process material m
fixed capacities to process each material. A set of new
frR Fixed cost of the existing remanufacturing
remanufacturing plants R = {1, 2, . . . , |R |} are intro-
plant r
duced in the new network to prevent the shortage of these
bRr,m Unit benefit of the existing remanufactur-
existing plants. Each new remanufacturing plant chooses
ing plant r to process material m
a capacity level for each material from capacity choices 
vrR ,m Variable cost of the new remanufacturing
S = {1, 2, . . . , |S|}.
plant r to process material m
The objective of the problem is to redesign the exist- R 
cm,s Capacity of a new remanufacturing plant
ing reverse logistics network to maximise the total profit.
to process material m in level s
The key decisions include (1) locations and capacities of 
frR ,m,s Fixed cost of the new remanufacturing
disassembly centres; (2) existing remanufacturing plants
plant r to process material m in level s
that should be opened in the new network; (3) loca- 
bRr ,m Unit benefit of the new remanufacturing
tions and capacities of new remanufacturing plants; and
plant r to process material m
(4) optimal transportation quantity from one facility to D
tn,d Transportation cost from city n to candi-
another.
date disassembly centre d
To formulate the problem, the following notations are R , tR 
td,r d,r Transportation cost from disassembly cen-
defined.
tre d to the existing remanufacturing
Sets
plants r and the new remanufacturing
N Set of cities plants r
A Set of end-of-life household appliances Dmax , Rkmax Maximum numbers of disassembly centres
M Set of materials and k type of remanufacturing plants
6 A. CHE ET AL.


Decision variables  
R R
+ td,r  qd,r  ,m (1)
qD quantity of household appliance a from city n to d∈D r ∈R m∈M
n,d,a
disassembly centre d Subject to
qRd,r,m quantity of material m from disassembly centre 
d to the existing remanufacturing plant r qDn,d,a = on,a ∀n ∈ N, ∀a ∈ A (2)

qRd,r ,m quantity of material m from disassembly centre d∈D
d to the new remanufacturing plant r   
qRd,r,m + qRd,r ,m
xdD equal to 1 if the candidate disassembly centre d r∈R r ∈R
is open, and 0 otherwise 
xrR equal to 1 if the existing remanufacturing plant = qD
n,d,a wa θa,m ∀d ∈ D, ∀m ∈ M (3)
r is open, and 0 otherwise n∈N a∈A
  
xrR equal to 1 if the new remanufacturing plant r is qD D D
n,d,a ≤ cl,s yd,l,s ∀d ∈ D, ∀l ∈ L (4)
open, and 0 otherwise n∈N a∈Al s∈S
D
yd,l,s equal to 1 if the candidate disassembly centre d 
sets its capacity of disassembly line l in level s, qRd,r,m ≤ cr,m
R
xrR ∀r ∈ Rk , ∀m ∈ Mk , ∀k ∈ K
and 0 otherwise d∈D
 (5)
yrR ,m,s equal to 1 if the new remanufacturing r sets its
    
capacity to process material m in level s, and 0 qRd,r ,m ≤ R R
cm,s yr ,m,s
otherwise d∈D s∈S

The RLNR can be formulated as the following model ∀r ∈ Rk ,



∀m ∈ Mk , ∀k ∈ K (6)
(RM). qRd,r,m = 0 ∀d ∈ D, ∀r ∈ Rk ,
 ∀m ∈ M − Mk , ∀k ∈ K (7)
  
Max f = σa q D + bRr,m qRd,r,m R
n,d,a qd,r ,m = 0 ∀d ∈ D, ∀r ∈ Rk ,
n∈N d∈D a∈A d∈D r∈R m∈M
 ∀m ∈ M − Mk , ∀k ∈ K (8)
    
+ bRr ,m qRd,r ,m D
yd,l,s = xdD ∀d ∈ D, ∀l ∈ L (9)
d∈D r∈R m∈M s∈S
 
  
yrR ,m,s = xrR

∀r ∈ Rk , ∀m ∈ Mk , ∀k ∈ K
D D
− fd,l,s yd,l,s + frR xrR
s∈S
d∈D l∈L s∈S r∈R
 (10)
  R
+
 
frR ,m,s yrR ,m,s yr ,m,s =0 ∀r ∈ Rk , ∀m ∈ M − Mk ,
r ∈R m∈M s∈S
∀s ∈ S, ∀k ∈ K (11)


  xdD ≤ Dmax (12)
−⎝ D D
vd,l qn,d,a
d∈D
n∈N d∈D l∈L a∈Al   
  xrR + xrR ≤ Rkmax ∀k ∈ K (13)
R
+ vr,m qRd,r,m
r∈Rk r ∈R k
d∈D r∈R m∈M
 qD qRd,r,m ,

qRd,r ,m ≥ 0
   
n,d,a , (14)
+ vrR ,m qRd,r ,m R
xdD , xrR , x ∈ {0, 1}
r (15)
d∈D r ∈R m∈M
 D 
 yd,l,s , yrR ,m,s ∈ {0, 1} (16)
− cn wa qD
n,d,a
n∈N d∈D a∈A The objective function (1) maximises the profit of the
 new reverse logistics network including the revenue from
D
+ tn,d wa qD
n,d,a the subsidy and materials, the fixed and variable costs for
n∈N d∈D a∈A setting up facilities and processing products and materi-
 
R R
+ td,r qd,r,m als, and the collection and transportation costs for deliv-
d∈D r∈R m∈M ering products or materials from one facility to another.
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 7

Figure 2. The general framework of DEH.

Constraints (2)–(3) ensure the flow balance of disas- solution by fixing the locations for disassembly centres
sembly centres and remanufacturing plants, respectively. and gradually expanding the remanufacturing facilities.
Constraints (4)–(6) represent the capacity limits of the Thus, the DEH algorithm consists of two components: a
candidate disassembly centres, existing remanufacturing decomposition heuristic and an expansion heuristic. Its
plants, and new remanufacturing plants. Constraints general framework is shown in Figure 2.
(7)–(8) indicate that the existing and new remanufac-
turing plants cannot accept the materials beyond their
recycling scope. Constraints (9)–(10) define the relation- 3.1. Decomposition heuristic
ship between facility-location and capacity-level deci- The decomposition heuristic first determines the
sions. When a disassembly centre (new remanufactur- locations for facilities and then determines their capac-
ing plant) is open, one capacity level is selected for ity levels. It aims to quickly find an initial solution for
each disassembly line (material). Constraint (11) suggests the proposed problem. In the RM model, every candi-
that new remanufacturing plants do not choose capac- date facility is associated with several binary variables to
ity levels for the materials beyond their recycling scope. determine their optimal capacity levels. With the increas-
Constraints (12) and (13) limit the numbers of disas- ing candidate facilities, more capacity-level variables are
sembly centres and remanufacturing plants, respectively. involved in the model, which can lead to difficulty in find-
Constraints (14)–(16) restrict the regions of decision ing the optimal solution. To better solve the problem, we
variables. first decide the locations for facilities by solving a reduced
MILP model that relaxes the large number of binary vari-
ables on capacity level choices, called RM1 model. Then,
3. Decomposition-and-expansion heuristic we determine the optimal capacity levels for facilities by
In this section, we introduce an efficient DEH algorithm solving another MILP model that fixes the facility loca-
to solve the proposed problem. This problem aims to tions to the values determined in RM1 model, called RM2
determine the locations and capacities for disassembly model.
centres and remanufacturing plants. It can be decom- The RM1model is described as
posed into two single-decision problems: facility location Max f
and capacity determination problems. In the first prob- subject to:
lem, locations for facilities are decided by relaxing their (2)–(15) and
capacity-level choices. In the second problem, capacity 
levels for facilities are determined by fixing their loca- 0 ≤ yd,l,s
D
≤ 1, 0 ≤ yrR ,m,s ≤ 1 (17)
tions determined in the first problem. A high-quality
initial solution can be quickly obtained through this pro- The objective function and constraints (2)–(15) are the
cess. To further improve the initial solution, an expansion same as those in the RM model. Constraint (17) relaxes
heuristic is developed based on Jayaraman, Patterson, the binary variables of capacity level choices to real vari-
and Rolland (2003). This heuristic improves the initial ables between 0 and 1.
8 A. CHE ET AL.

The original RM model is NP-hard because the classi- remanufacturing plants. Therefore, we fix the locations
cal reverse logistics network design problem is NP-hard for disassembly centres and iteratively expand the reman-
(Pishvaee, Farahani, and Dullaert 2010; Santibanez- ufacturing plants to improve the initial solution.
Gonzalez and Diabat 2013). In this Section, the RM1 Jayaraman, Patterson, and Rolland (2003) introduced
model relaxes numerous binary variables of the original the expansion heuristic in solving the RLND. They use
RM model for deciding facility capacities, which signif- the concentration heuristic proposed by Rosing and ReV-
icantly reduces the computational complexity. Although elle (1997) to obtain an initial facility-location solution
the RM1 model is still NP-hard, it can be relatively effi- and adopt the expansion heuristic to improve the ini-
ciently solved by a standard mathematical programming tial solution. The basic logic of the expansion heuristic
solver. We solve the RM1 model with CPLEX and remem- is checking each non-open facility in the initial solu-
ber the locations for disassembly centres, existing reman- tion to identify if they can improve the initial decision.
ufacturing plants and new remanufacturing plants as x̂dD , In this paper, we consider several types of remanufac-

x̂rR , and x̂rR . Then, we determine their capacities using the turing plants in our network. Each type of remanu-
following model (RM2). facturing plants has a large number of candidate loca-
Max f tions. It is time-consuming to check each non-open
subject to: plant one by one. To improve the efficiency, we rank
(2)–(11), (14)–(16), and all non-open plants and check a group of non-open
plants together in each iteration. Therefore, there are
xdD = x̂dD ∀d ∈ D (18) two differences between the algorithms of this paper
xrR = x̂rR ∀r ∈ R (19) and Jayaraman, Patterson, and Rolland (2003). First, we
  use the decomposition heuristic, instead of the con-
xrR = x̂rR ∀r ∈ R (20) centration heuristic, to obtain an initial solution. Sec-
The objective function and constraints (2)–(11) and ond, we rank all non-open facilities and iteratively check
(14)–(16) are the same as those in the RM model. Con- a group of non-open facilities, instead of each non-
straints (18)–(20) fix the locations for disassembly cen- open facility, in expansion heuristic. The detailed pro-
tres, existing remanufacturing plants and new remanu- cedures of our expansion heuristic are described as

facturing plants to the x̂dD , x̂rR , and x̂rR obtained by the follows.
RM1 model. First, we place all open remanufacturing plants deter-
The RM2 model can be also efficiently solved by a mined by the initial solution X∗ into the checking set
standard solver because most variables are fixed to cer- CS = {OR1 , OR2 , . . . , OR|K| }, which distinguishes the
tain values. The facility-location variables are fixed to different types of remanufacturing plants by k, k∈K. Sec-
the results of RM1 model. The capacity variables and ond, we rank all non-open remanufacturing plants in
the transportation-quantity variables for non-open facil- non-decreasing order of their distance to the closest open
ities determined by RM1 model are restricted to zero disassembly centres in the initial solution, and we save
under the constraints (9)–(10) and (5)–(6). Actually, the them into the ranking set RS = {NR1 , NR2 , . . . , NR|K| }.
RM2 model only needs to determine the capacity vari- If a tie occurs, we rank the non-open remanufacturing
ables and the transportation-quantity variables for open facilities according to their transportation cost to the sec-
facilities determined by RM1 model. We solve the RM2 ond closest open disassembly centres. Third, for each
model with CPLEX and remember its solution and profit type of remanufacturing plant, we move the first Num
as X∗ and f ∗ , respectively. They are also the initial solution plants from the ranking set RS into the expanding set
and profit of the original RM model. ES = {ER1 , ER2 , . . . , ER|K| }. Finally, we solve the follow-
ing RM3 model, which fixes the locations of disassembly
centres and shuts down the remanufacturing plants that
3.2. Expansion heuristic
are not in the checking or expanding sets. If a better solu-
The expansion heuristic focuses on improving the ini- tion is obtained, we update the solution X∗ and the profit
tial solution obtained by the decomposition heuristic. value f ∗ and reset the checking set CS; otherwise, we
This solution can be further improved when the opti- repeat the third step and move the next group of non-
mal locations for disassembly centres or remanufacturing open remanufacturing plants into the expanding set. This
plants are not found by the RM1 model. We observe the process is continued until all non-open remanufactur-
RM1 model is most efficient for determining the optimal ing plants have been checked, or the maximum iteration
locations for disassembly centres, but not for reman- MaxIter is met. The RM3 model is described as
ufacturing plants because of the complicated correla- Max f
tion between existing remanufacturing plants and new subject to:
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 9

(2)–(16) and the parameters and sets, steps 2–3 illustrate the decompo-
sition heuristic, and steps 4–7 correspond to the expan-
xdD = [xdD ]∗ ∀d ∈ D (21) sion heuristic. Iter is used to count the iterations in the
expansion heuristic. X and f store the solution and its
xrR = 0 ∀r ∈
/ CS ∪ ES (22) profit value of the RM3 model in each iteration, respec-
R tively.
x r =0 ∀r ∈
/ CS ∪ ES (23)

Constraint (21) fixes the locations for disassembly 4. Case study


centres to the values [xdD ]∗ determined in the initial solu- We evaluate the redesigned network using a real case in
tion X∗ . Constraints (22)–(23) shut down the reman- Shaanxi province, China. The data is collected from pub-
ufacturing plants that are not in the checking sets or lic reports and interviews with local organisations. We
expanding sets. solve the proposed model by CPLEX and using our DEH
The RM3 model fixes locations of disassembly cen- algorithm; the algorithm is coded in C++ on a PC with
tres and restricts feasible locations of remanufacturing an Intel core i5 CPU (3.1 GHZ) and 8 GB RAM. Sensi-
plants within the checking and expanding sets. It can tivity analysis is conducted to determine how changes in
be efficiently solved using a standard solver. Compar- the parameters can affect decisions.
ing to the RM2 model, the RM3 expands the location
space of remanufacturing plants. A better solution can be
obtained to improve the incumbent solution X∗ . 4.1. Case description
During the expansion iteration, two problems may Every year, over 500,000 tons of end-of-life household
occur because the subsets NR1 , NR2 , . . . , NR|K| in the appliances are recycled in Shaanxi province, China. Most
ranking set RS have different sizes. First, small subsets products are recovered in the traditional way, which leads
finish the expansion iteration before large subsets. Sec- to serious resource waste and environmental pollution.
ond, some subsets can have non-open facilities less than To improve the recycling performance, we redesign the
parameter Num during the iteration. For the first prob- reverse logistics network based on the proposed model.
lem, we continue running the iteration by expanding only The province map is shown in Figure 3. Ten cities are
the remanufacturing plants left in the large subsets. For located in different regions. Five end-of-life household
the second problem, we move all remaining non-open appliances are considered to recycle as shown in Figure
facilities into the expanding sets and keep checking if 4. The recycled materials include three metal materi-
these facilities can improve the incumbent solution. als, three plastic materials, and one glass material. There
In addition, there are two user-defined parameters: have been 21 metal smelters, 13 plastic plants, and 8
Num and MaxIter in the expansion procedure. Num glass plants in the province, as shown in Figure 3. To
defines the expanding size of non-open remanufactur- redesign the network, we assume every city has a candi-
ing plants. MaxIter indicates the maximum iteration. date disassembly centre, smelter, plastic plant, and glass
Their value can affect the performance of the DEH plant.
algorithm. A larger Num value indicates more non-open To determine the optimal capacities for candidate dis-
remanufacturing plants are placed into the expanding assembly centres and remanufacturing plants, we design
set, which causes the RM3 model has a larger solution small, medium, and large capacity levels as shown in
space. A larger MaxIter value indicates more expanding Tables 2 and 3. The capacity for each level is obtained
sets are allowed to be checked. The expansion procedure based on the current technology and management pol-
becomes time-consuming but quite probable to find a icy. The other parameters for this case, such as subsi-
better solution when Num and MaxIter have large values. dies, benefits, and transportation costs, are available at
To balance the computational time and solution quality, https://www.mediafire.com/file/p18fsc0vwwofpq0/
we set Num based on the maximum numbers of reman- 2021_Data.zip/file.
ufacturing plants Rk max defined in the RM model, and
we set parameter MaxIter based on the total non-open
facilities and the value of Num. 4.2. Computational results
We solve the proposed model using CPLEX and the
DEH algorithm. As this section mainly evaluates the
3.3. Procedure of DEH algorithm
redesigned network, we fix the key parameters Num and
The decomposition-and-expansion heuristic can be MaxIter of the DEH algorithm to 5 and 10. In addition,
summarised in Algorithm DEH, where step 1 initialises the maximum numbers of facilities {Dmax , R1max , R2max ,
10 A. CHE ET AL.

Algorithm DEH

1. Initialise x̂dD , x̂rR , x̂rR , f ∗ , f, Iter = 0
X∗ ← Ø, X← Ø, CS← Ø, RS← Ø, ES← Ø

2. Solve RM1 model, and output the values x̂dD , x̂rR , and x̂rR
3. Solve RM2 model, and save the solution and its profit value as X∗ and f ∗
4. Add all open remanufacturing plants in solution X∗ into the checking set CS = {OR1 , OR2 , ... , OR|K| }
5. Rank all non-open remanufacturing plants in the non-decreasing order of their transportation cost to the closest open disassembly centres; save them into
the ranking set RS = {NR1 , NR2 , ... , NR|K| }
6. While Iter < MaxIter and RS is not null
6.1 Move the first Num plants for each type of remanufacturing plants in the ranking set RS to the expanding set ES = {ER1 , ER2 , ... , ER|K| }
6.2 Solve RM3 model to obtain the new solution X and its profit value f
iff > f ∗ , then
update the solution X ∗ and profit value f ∗ with the new solution X and profit f
update the checking set CS
End if
6.3 Set Iter ← Iter + 1
7. Output the final solution X ∗ and its profit value f ∗

Figure 3. City map and the existing remanufacturing facilities of Shaanxi province.

Figure 4. End-of-life household appliances and recycled materials.


INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 11

Table 2. Capacity levels for candidate disassembly centres. TimeD represent the objective value and computational
Candidate Capacity level time of the DEH algorithm, respectively; GapC calcu-
facilities Disassembly lines Small Medium Large lates the objective gap using the formula GapC = (ObjC
Disassembly TV/Computer 500,000 1,000,000 2,000,000 − ObjD )/ObjC × 100%. Table 4 shows the proposed
centres Refrigerator/Air conditioner 500,000 1,000,000 2,000,000 model has feasilbe soluitons only when {Dmax , R1max ,
Washing machine 200,000 500,000 1,000,000
R2max , R3max } ≥ {3, 3, 2, 1}, which means at least 3 dis-
Table 3. Capacity levels for candidate remanufacturing plants. assembly centres, 3 smelters, 2 plastic plants, and one
glass plant are required in this province. The objective
Candidate Capacity level (ton)
value decreases with the maximum numbers, which indi-
remanufacturing plants Materials Small Medium Large cates strong control over the facility numbers harms the
Smelters Iron 10,000 50,000 100,000 profit of the entire network. Meanwhile, all objective
Copper 10,000 50,000 100,000
Aluminium 5000 10,000 50,000 gaps are lower than 0.1% and the computational time of
Plastic plants PP plastic 5000 10,000 50,000 the DEH algorithm is shorter than 5 s, which suggests a
PS plastic 5000 10,000 50,000
ABS plastic 1000 5000 10,000 high-quality solution can be efficiently obtained by the
Glass plants Glass 5000 10,000 50,000 proposed heuristic.
For {Dmax , R1max , R2max , R3max } = {5, 5, 5, 5}, the optimal
R3max } in the model are generally given by the local gov- locations and capacities for facilities are shown in Figure
ernment. We first set {Dmax , R1max , R2max , R3max } = {7, 7, 7, 5 and Table 5. Figure 5 shows that five disassembly cen-
7} and gradually decrease them to their minimum values. tres d ∈{4, 5, 6, 7, 8} should be built to process end-of-life
The computational results are shown in Table 4, where household appliances; three candidate smelters r ∈{5, 7,
ObjC and TimeC indicate the optimal objective value and 8} and two existing smelters r∈{19, 20} should be opened
computational time of CPLEX, respectively; ObjD and to reprocess the obtained metal materials; three candidate

Table 4. Computational results for the case study.


DEH
Maximum numbers CPLEX (Num = 5, MaxIter = 10)
NO. Dmax R1max R2max R3max ObjC (RMB) TimeC (s) ObjD (RMB) TimeD (s) GapC (%)
1 7 7 7 7 1236505064 8.11 1236505064 4.10 0.00%
2 6 6 6 6 1231884683 10.24 1231673709 4.29 0.02%
3 5 5 5 5 1227013060 8.67 1227013060 4.87 0.00%
4 4 4 4 4 1221023740 9.48 1221023740 4.20 0.00%
5 3 3 3 3 1204656302 3.85 1204656302 3.10 0.00%
6 3 3 2 1 1166899533 2.40 1166899533 2.70 0.00%

Figure 5. Optimal facility locations when {Dmax , R1max , R2max , R3max } = {5, 5, 5, 5}.
12 A. CHE ET AL.

Table 5. Optimal capacity levels for candidate facilities to open.


Capacity level choices
Candidate facilities TV/Computer Refrigerator/Air conditioner Washing machine
Disassembly d = 4 Large Medium Medium
centre d = 5 Large Large Large
d = 6 Small Small Small
d = 7 Medium Large Medium
d = 8 Medium Medium Small
Smelters Iron Copper Aluminium
r = 5 Large Medium Medium
r = 7 Medium Small Small
r = 8 Medium Small Small
Plastic plants PP plastic PB plastic ABS plastic
r = 6 Small Small Small
r = 7 Medium Medium Medium
r = 8 Medium Small Small
Glass plants Glass
r = 4 Large
r = 5 Large
r = 6 Small
r = 8 Medium

plastic plants r ∈{6, 7, 8} and two existing plastic plants parameters using Taguchi’s experimental design method
r∈{2, 11} should be open to reprocess the obtained plas- (Taguchi, Jugulum, and Taguchi 2004). This method
tic materials; and four candidate glass plants r ∈{4, 5, adopts orthogonal arrays and signal-to-noise (SN) ratios
6, 8} and one existing glass plant r = 5 should be open to find the best combination of parameters. In the pro-
to reprocess the obtained glass materials. Table 5 shows posed RM model, there are five key types of parame-
these candidate facilities make different capacity choices ters: subsidy σ a , product amount on,a , transportation cost
D , t R , t R }, fixed cost {f R f R , f R }, and benefit of
{tn,d
based on their recycling demands. d,r d,r r , r ,m,s r

In addition, the result on recycling quantity indicates materials {bRr ,m , bRr,m }. We assume each type of param-
over 260, 000 tons of metals, 100, 000 tons of plastics, eter has four levels. Each level equals the value of the
and 60,000 tons of glass can be extracted from household parameter in the case study multiplied by a factor ,
appliances. Almost all useful components are disassem- where ε ∈{0.5, 1.0, 1.5, 2.0}. For example,  = 0.5 means
bled, sorted, and recycled in the new network. However, the parameter equals half of its value in the case study;
in the traditional network, according to our interview  = 1.0 indicates keeping the value constant;  = 2.0
with the local collection centres, only metal materials are implies rising the value to its double. An orthogonal array
efficiently recycled for their high benefits. The remaining L16 (45 ) that includes 16 experiments can be used in our
materials usually face poor treatments such as landfill- experiment. The SN ratio of experiments is calculated
ing, burning, and long-term storage. The recycling rate in by the formula ηi = −10log f12 , where fi is the optimal
the new network is about 91.5%, while in the traditional i
objective value of the RM model in experiment i. The
network it is only about 55.6%. They are calculated using
orthogonal array and experimental results are shown in
the formulas μnew = Wu /Wp (%) and μtra = Wm /Wp (%),
Table 6.
where μnew , Wu , Wp , μtra , and Wm denote the recycling
Table 7 summarises the computational results in
rate in the new network, the weight of useful materials
aspects of the average SN ratio of parameters in four
(metals, plastics, and glasses) in the collected products,
levels, the range between the maximum and minimum
the weight of the collected products, the recycling rate in
average SN ratio , the sum of squares S, and the signif-
the traditional network, and the weight of metals, respec-
icance rank of parameters. From Table 7, we can observe
tively. Therefore, the new network not only improves the
that the product amount plays the most role in the RLNR
recycling rate of household appliances but also reduces
decision because it determines the recycling amount of
their negative effects on the environment.
valuable materials. The benefit of materials and subsidy
play the second and third roles because they directly
affect the profit of the new network. The transportation
4.3. Sensitivity analysis of model parameters
cost and fixed cost play the fourth and fifth roles.
The RLNR decision can be largely affected by changes Five important managerial implications can be sum-
in parameters such as subsidy, fixed cost, transporta- marised from the case study. Firstly, the recycling ratio
tion cost, and benefit of materials. To investigate the of household appliances can be improved from 55.6%
effect, we conduct a sensitivity analysis for the model to 91.5% by redesigning the traditional reverse logistics
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 13

Table 6. Orthogonal array and experimental results.


 in subsidy  in product  in fixed cost  in benefit of materials  in transportation cost
  D , t R , t R }
No. {σ a } amount {on,a } {frR , frR ,m,s , frR } {bRr ,m , bRr,m } {tn,d d,r d,r ηi
1 0.5 0.5 0.5 0.5 0.5 164
2 0.5 1.0 1.0 1.0 1.0 178
3 0.5 1.5 1.5 1.5 1.5 186
4 0.5 2.0 2.0 2.0 2.0 191
5 1.0 0.5 1.0 1.5 2.0 179
6 1.0 1.0 0.5 2.0 1.5 187
7 1.0 1.5 2.0 0.5 1.0 180
8 1.0 2.0 1.5 1.0 0.5 188
9 1.5 0.5 1.5 2.0 1.0 183
10 1.5 1.0 2.0 1.5 0.5 187
11 1.5 1.5 0.5 1.0 2.0 188
12 1.5 2.0 1.0 0.5 1.5 187
13 2.0 0.5 2.0 1.0 1.5 180
14 2.0 1.0 1.5 0.5 2.0 183
15 2.0 1.5 1.0 2.0 0.5 193
16 2.0 2.0 0.5 1.5 1.0 194

Table 7. Analysis and rank of model parameters.


Average ηi in four levels

Parameters 1 2 3 4  S Rank
Subsidy {σ a } 180 183 186 188 8 137 3
Product amount {on,a } 176 184 187 190 14 416 1

Fixed cost {frR , frR ,m,s , frR } 183 184 185 184 2 6 5

Benefit of materials {bRr ,m , bRr,m } 178 183 186 189 11 236 2
Transportation cost {tn,d , td,r D R , t R } 183 184 185 185 2 14 4
d,r

network. Second, the minimum facilities for redesign- reverse logistics network. The city number N is set to 20,
ing the network can be obtained by solving the proposed 30, and 40. The existing remanufacturing plants Rk are
model. For example, at least 3 disassembly centres, given a small, medium, and large number according to
3 smelters, 2 plastic plants, and one glass plant are the city number. The maximum numbers {Dmax , R1max ,
required in the province to efficiently recycle the house- R2max , R3max } increase by 3 as the city number increases by
hold appliances of 10 cities. Thirdly, strict limitation 10. For a certain set of existing facilities and maximum
on facility numbers harms the recycling profit and numbers, we randomly generate 10 instances and calcu-
willingness. Policymakers should balance the recycling late their average computational results. All experiments
profit and recycling ratio to make the limitation deci- are run for 2 h until the optimal solution is obtained. If the
sion. Fourthly, disassembly centres and remanufactur- optimal solution is not found within 2 h, the incumbent
ing plants make different location and capacity decisions solution is output.
based on their recycling demands. Finally, parameters of In these randomly generated instances, parameters
product amount, material benefit, and government sub- wa , σ a , and θ a,m have the same values as those in
sidy play the first, second, and third roles in affecting the the case study. Other parameters are generated in the
recycling profit and decision. regions shown in Table 8. Two assumptions are consid-
ered when generating random instances. First, if capac-
5. Randomly generated instances ity level s > sêđŇ, then fd,l,s
D > fD R R
d,l,s and fr ,m,s > fr ,m,s ,
In this section, a set of large-scale instances are ran- which means higher-level capacity requires higher fixed
domly generated to further evaluate the performance of cost. Second, if city index n > n , then the fixed cost
the DEH algorithm. As the performance can be affected and variable cost in city n are generally higher than
by changes in user-defined parameters, we also make a those in city n . The generated instances are available at
sensitivity analysis of these parameters using the gener- https://www.mediafire.com/file/p18fsc0vwwofpq0/
ated instances. The best combination of parameters is 2021_Data.zip/file.
given based on the experimental results.

5.1. Data generation 5.2. Computational results


We randomly generate a set of large-scale instances with The computational results for the randomly generated
different numbers of cities and existing facilities in the instances are presented in Table 9, where two types of
14 A. CHE ET AL.

Table 8. Random regions for parameters.


Items Random regions
Candidate disassembly centres D ∈ [200000, 1600000], v D ∈ [2.0, 5.0]
fd,l,s d,l
  
Candidate smelters frR ,m,s ∈ [200000, 2000000], bRr ,m ∈ [2000, 5000], vrR ,m ∈
[500, 1000]
  
Candidate plastic plants frR ,m,s ∈ [200000, 1000000], bRr ,m ∈ [1000, 2500], vrR ,m ∈
[500, 800]
  
Candidate glass plants frR ,m,s ∈ [200000, 500000], bRr ,m ∈ [1000, 1500], vrR ,m ∈
[500, 700]
Existing smelters R ∈ [5000, 50000], f R ∈ [100000, 650000], bR ∈ [2000,
cr,m r r,m
R ∈ [500,1000]
5000], vr,m
Existing glass plants R ∈ [5000, 50000], f R ∈ [100000, 500000], bR ∈[1000,
cr,m r r,m
R ∈ [500, 800]
2500], vr,m
Existing glass plants R ∈ [5000, 50000], f R ∈ [100000, 300000], bR ∈[1000,
cr,m r r,m
R ∈ [500, 700]
1500], vr,m
Others on,a ∈ [50000, 300000], cn ∈ [1000, 2000], tn,d D , t R , t R ∈
d,r d,r
[100, 100∗N], where the region of transportation costs
increases with the city number N

Table 9. Computational results for the generated instances.


Existing facilities Maximum numbers CPLEX DEH
City |R1 | |R2 | |R3 | Dmax R1max R2max R3max ObjC TimeC (s) ObjD TimeD (s) GapC (%) GapU (%)
20 5 3 2 6 6 6 6 417211074 58.16 417208621 8.78 0.001 0.001
15 10 8 6 6 6 6 418622749 50.59 418584561 9.29 0.011 0.011
25 20 18 6 6 6 6 409909846 107.19 409909846 14.12 0.000 0.000
30 5 3 2 9 9 9 9 481256646 567.92 481150708 22.80 0.021 0.021
20 15 10 9 9 9 9 417512124 4241.32 417327921 108.91 0.043 0.043
35 30 28 9 9 9 9 487697117 3108.49 487683869 95.56 0.003 0.003
40 5 3 2 12 12 12 12 363518139 6683.31 363674990 78.32 – 0.842
25 20 15 12 12 12 12 365505227 7200.00 365491938 247.72 – 2.778
45 40 38 12 12 12 12 390001330 7200.00 389808888 378.76 – 3.666
Note: - means GapC cannot be calculated because the optimal solutions of instances are not obtained within 2 h.

gaps – GapC and GapU – are computed to evaluate the 2019). There are two user-defined parameters in the
performance of the DEH algorithm. GapC denotes the DEH algorithm: the expanding size Num and the
gap from the optimal profit obtained by CPLEX, which is maximum iteration MaxIter. We conduct a sensitiv-
calculated by the formula in the case study. GapU denotes ity analysis of these parameters using the generated
the gap from the upper bound provided by CPLEX when instances. The best configuration of parameters is
the optimal profit cannot be found within 2 h, as calcu- given based on the experimental results. Each param-
lated by formula GapU = (ObjU − ObjD )/ObjU × 100%, eter is assumed with four levels, Num ∈ {1, 5, 10, 20}
where ObjU denotes the upper bound output by CPLEX and MaxIter ∈ {10, 30, 50, 100}. Sixteen combinations
to solve RM model, and ObjD denotes the objective value of parameters are formed by these levels. There are
output by algorithm DEH. three datasets in the randomly generated instances cor-
Table 9 shows that instances for city number |N| = 20 responding to different city numbers N ∈ {20, 30, 40}.
can be solved to optimality by CPLEX within 150 s. When We evaluate the sixteen combinations of parameters
|N| increases to 30, CPLEX runs over 500 s to obtain with each dataset. A total of 1440 experiments are
the optimal solution. When |N| increases to 40, most conducted in this Section. The signal-and-noise ratios
instances cannot be solved by CPLEX within 2 h. How- of computational accuracy η1, and computational time
ever, the DEH algorithm runs all instances within 400 s η2, are calculated for each dataset and parameter
1 30
and provides high-quality solutions with GapC under combination, where η1, = −10 log 30 ,
i=1 Gapi ; η2 =
2

0.1% and GapU under 4.0%. These results demonstrate 1
−10 log 30 30 2
i=1 Timei ; Gapi is the gap between the solu-
that the DEH algorithm is efficient in solving large-scale tion of DEH algorithm and the best solution/bound of
instances. CPLEX in instance i; Timei is the computational time
of DEH algorithm in instance i. The average SN ratios of
parameters in four levels are shown in Figure 6.
5.3. Sensitivity analysis of algorithm parameters
Figure 6 shows the computational accuracy is less
The performance of an algorithm is affected by its affected by changes in user-defined parameters, which
configuration of user-defined parameters (Gao et al. demonstrates the robustness of the DEH algorithm.
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 15

Figure 6. Average SN ratios η1, (a) and η2, (b) of parameters in four levels.

However, the computational time varies with differ- 400 s and obtains high-quality solutions with gaps from
ent configurations of parameters. According to the the optimal objective value of less than 0.1% and gaps
experimental results, the best configurations of {Num, from the upper bound less than 4.0%. This demonstrates
MaxIter} should be {10, 30}, {20, 10}, and {20,10} for that the DEH algorithm can efficiently yield high-quality
three datasets with city numbers 20, 30, and 40, respec- solutions for large-scale instances. A sensitivity analy-
tively. sis of parameters in the DEH algorithm is conducted
with the generated instances. The best configuration of
algorithm parameters is given based on the experimental
6. Conclusion results.
In our future work, multiple objectives will be con-
This paper considers a reverse logistics network redesign sidered in the model to redesign the reverse logistics
problem for end-of-life household appliances. We inves- network. Meanwhile, multi-period recycling and vehi-
tigate the existing recycling process in China and cle scheduling problems will be combined in the reverse
redesign a new reverse logistics network to improve the logistics network redesign problem. Further, uncertain
recycling rate. A MILP model is proposed to determine problems in the model including the uncertain supply
the optimal locations and capacities for facilities in the of end-of-life products and the uncertain demand of
redesigned network. An efficient DEH algorithm is devel- materials in the market will be addressed.
oped to solve the model. We evaluate the redesigned
network using a real case in Shaanxi province, China. The
results indicate that the recycling rate is largely improved
Acknowledgements
by the new network in which 91.5% of household appli-
ances could be recycled, while in the traditional network, We are grateful for the valuable comments of the reviewers and
editor.
only 55.6% of the components are efficiently recycled.
A sensitivity analysis of model parameters is conducted
for the case study using the Taguchi method. The results
show that parameters of product amount, material ben- Disclosure statement
efit, and government subsidy play the first, second, and No potential conflict of interest was reported by the author(s).
third roles in affecting the recycling profit and decision.
To further evaluate the efficiency of the DEH
algorithm, a set of large-scale instances are randomly Funding
generated by increasing the city number and varying This work was supported by National Natural Science
the existing facilities. The computational results show Foundation of China [grant numbers 71871183, 71931007,
that the DEH algorithm solves all the instances within 71871142].
16 A. CHE ET AL.

Notes on contributors Alumur, S. A., S. Nickel, F. Saldanha-da-Gama, and V. Verter.


2012. “Multi-period Reverse Logistics Network Design.”
Ada Che is currently a Full Professor in European Journal of Operational Research 220 (1): 67–78.
the Department of Management Science doi:10.1016/j.ejor.2011.12.045.
and Engineering at Northwestern Poly- Ayvaz, B., B. Bolat, and N. Aydın. 2015. “Stochastic Reverse
technical University, China. He received Logistics Network Design for Waste of Electrical and Elec-
the Ph.D. degree of Mechanical Engi- tronic Equipment.” Resources, Conservation and Recycling
neering from Xi’an Jiaotong University in 104: 391–404. doi:10.1016/j.resconrec.2015.07.006.
1999. He has authored or coauthored over Barros, A. I., R. Dekker, and V. Scholten. 1998. “A Two-
50 articles in international journals, such level Network for Recycling Sand: A Case Study.” Euro-
as Omega-The International Journal of Management Science, pean Journal of Operational Research 110 (2): 199–214.
Naval Research Logistics, IEEE Transactions, European Jour- doi:10.1016/S0377-2217(98)00093-9.
nal of Operational Research, International Journal of Production Bing, X., J. Bloemhof-Ruwaard, A. Chaabane, and J. van
Research, International Journal of Production Economics, and der Vorst. 2015. “Global Reverse Supply Chain Redesign
Computers and Operations Research. He is currently serving for Household Plastic Waste Under the Emission Trad-
as an associate editor of the IEEE Transactions on Intelligent ing Scheme.” Journal of Cleaner Production 103: 28–39.
Transportation Systems and Journal of Industrial and Manage- doi:10.1016/j.jclepro.2015.02.019.
ment Optimization. His research interests include transporta- Ene, S., and N. Öztürk. 2015. “Network Modeling for Reverse
tion planning and optimisation, production scheduling, and Flows of End-of-life Vehicles.” Waste Management 38:
operations research. 284–296. doi:10.1016/j.wasman.2015.01.007.
Jieyu Lei is currently a Ph.D. student in the Eskandarpour, M., E. Masehian, R. Soltani, and A. Khosrojerdi.
Department of Management Science and 2014. “A Reverse Logistics Network for Recovery Systems
Engineering at Northwestern Polytechni- and a Robust Metaheuristic Solution Approach.” The Inter-
cal University, China. She received the national Journal of Advanced Manufacturing Technology 74
M.S. degree of Management Science and (9): 1393–1406. doi:10.1007/s00170-014-6045-7.
Engineering from Xidian University in Fleischmann, M., P. Beullens, J. M. Bloemhof-Ruwaard, and L.
2016. As a visiting researcher, she worked Van Wassenhove. 2001. “The Impact of Product Recovery on
with the OPAC group (Operations, Plan- Logistics Network Design.” Production and Operations Man-
ning, Accounting, and Control) of the Department of Industrial agement 10: 156–173. doi:10.1111/j.1937-5956.2001.tb00-
Engineering and Innovation Sciences at Eindhoven University 076.x.
of Technology (TU/e) in 2019. Her research interests include Fu, Y., M. Zhou, X. Guo, and L. Qi. 2021. “Stochastic Multi-
supply chain management, reverse logistics network design, objective Integrated Disassembly-reprocessing-reassembly
and transportation planning and optimisation. Scheduling via Fruit fly Optimization Algorithm.” Journal of
Cleaner Production 278: 123364. doi:10.1016/j.jclepro.2020.
Zhibin Jiang is currently a Full Profes- 123364.
sor in the Department of Management Gao, X., and C. Cao. 2020. “A Novel Multi-objective Scenario-
Science at Shanghai Jiao Tong Univer- based Optimization Model for Sustainable Reverse Logis-
sity, China. He is also the dean of the tics Supply Chain Network Redesign Considering Facility
Industrial and System Engineering Insti- Reconstruction.” Journal of Cleaner Production 270: 122405.
tute (Sino-US Global Logistics Institute, doi:10.1016/j.jclepro.2020.122405.
SUGLI) of Shanghai Jiao Tong Univer- Gao, S., M. Zhou, Y. Wang, J. Cheng, H. Yachi, and J. Wang.
sity. He received the Ph.D. degree of Engi- 2019. “Dendritic Neuron Model with Effective Learning
neering Management from City University of Hong Kong. Algorithms for Classification, Approximation and Predic-
He has authored or coauthored over 100 articles in interna- tion.” IEEE Transactions on Neural Networks and Learning
tional journals such as Production and Operations Manage- Systems 30 (2): 601–614. doi:10.1109/TNNLS.2018.2846646.
ment, INFORMS Journal on Computing, IEEE Transactions on Guo, X., S. Liu, M. Zhou, and G. Tian. 2018. “Dual-objective
Automation Science and Engineering, Omega-The International Program and Scatter Search for the Optimization of Dis-
Journal of Management Science, and International Journal of assembly Sequences Subject to Multiresource Constraints.”
Production Research. He is currently serving as a fellow of Insti- IEEE Transactions on Automation Science and Engineering 15
tute of Industrial and Systems Engineers and associate editor of (3): 1091–1103. doi:10.1109/TASE.2017.2731981.
International Journal of Production Research. His research inter- Guo, H., Y. Zhang, C. Zhang, Y. Zhang, and Z. Han.
ests include production and service operations management, 2020. “A Multi-commodity Location-inventory Problem in
healthcare service management, and logistics and supply chain a Closed-loop Supply Chain with Commercial Product
management. Returns.” International Journal of Production Research 58
(22): 6899–6916. doi:10.1080/00207543.2019.1686186.
Guo, X., M. Zhou, A. Abusorrah, F. Alsokhiry, and K.
References Sedraoui. 2020. “Disassembly Sequence Planning: A Survey.”
Achillas, C.,C. Vlachokostas, D. Aidonis, N Moussiopoulos, E. IEEE/CAA Journal of Automatica Sinica, doi:10.1109/JAS.
Iakovou, and G. Banias. 2010. “Optimising Reverse Logistics 2020.1003515.
Network to Support Policy-making in the Case of Electri- Guo, X., M. Zhou, S. Liu, and L. Qi. 2020. “Lexicographic Mul-
cal and Electronic Equipment.” Waste Management 30 (12): tiobjective Scatter Search for the Optimization of Sequence-
2592–2600. doi:10.1016/j.wasman.2010.06.022. dependent Selective Disassembly Subject to Multiresource
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 17

Constraints.” IEEE Transactions on Cybernetics 50 (7): Operational Research 37: 1100–1112. doi:10.1016/j.cor.2009.
3307–3317. doi:10.1109/TCYB.2019.2901834. 09.018.
Guo, X., M. Zhou, S. Liu, and L. Qi. 2021. “Multiresource- Pishvaee, M. S., K. Kianfar, and B. Karimi. 2010. “Reverse
constrained Selective Disassembly with Maximal Profit Logistics Network Design Using Simulated Annealing.” The
and Minimal Energy Consumption.” IEEE Transactions International Journal of Advanced Manufacturing Technology
on Automation Science and Engineering 18 (2): 804–816. 47 (1): 269–281. doi:10.1007/s00170-009-2194-5.
doi:10.1109/TASE.2020.2992220. Polotski, V., J.-P. Kenne, and A. Gharbi. 2017. “Set-up and Pro-
Jayaraman, V., R. A. Patterson, and E. Rolland. 2003. “The duction Planning in Hybrid Manufacturing–Remanufacturing
Design of Reverse Distribution Networks: Models and Solu- Systems with Large Returns.” International Journal of Pro-
tion Procedures.” European Journal of Operational Research duction Research 55 (13): 3766–3787. doi:10.1080/00207543.
150 (1): 128–149. doi:10.1016/S0377-2217(02)00497-6. 2017.1293863.
Khan, A. H., X. Cao, S. Li, V. N. Katsikis, and L. Liao. Reddy, K. N., A. Kumar, and E. E. F. Ballantyne. 2019. “A
2020. “BAS-ADAM: An ADAM Based Approach to Improve Three-Phase Heuristic Approach for Reverse Logistics Net-
the Performance of Beetle Antennae Search Optimizer.” work Design Incorporating Carbon Footprint.” Interna-
IEEE/CAA Journal of Automatica Sinica 7 (2): 461–471. tional Journal of Production Research 57 (19): 6090–6114.
doi:10.1109/JAS.2020.1003048. doi:10.1080/00207543.2018.1526422.
Krug, Z., R. Guillaume, and O. Battaïa. 2021. “Design of Rentizelas, A., N. Trivyza, S. Oswald, and S. Siegl. 2021.
Reverse Supply Chains Under Uncertainty: The Lexico- “Reverse Supply Network Design for Circular Economy
graphic R∗ Criterion for Exploring Opportunities.” Inter- Pathways of Wind Turbine Blades in Europe.” International
national Journal of Production Research 59 (11): 3221–3236. Journal of Production Research, 1–20. doi:10.1080/00207543.
doi:10.1080/00207543.2020.1866782. 2020.1870016.
Lechner, G., and M. Reimann. 2020. “Integrated Decision- Rosing, K. E., and C. S. ReVelle. 1997. “Heuristic Concentra-
making in Reverse Logistics: An Optimisation of Interacting tion: Two Stage Solution Construction.” European Journal
Acquisition, Grading and Disposition Processes.” Interna- of Operational Research 97: 75–86. doi:10.1016/S0377-2217-
tional Journal of Production Research 58 (19): 5786–5805. (96)00100-2.
doi:10.1080/00207543.2019.1659518. Sadrnia, A., N. R. Langarudi, and A.p. Sani. 2020. “Logistics
Lee, D.-H., and M. Dong. 2008. “A Heuristic Approach to Network Design to Reuse Second-hand Household Appli-
Logistics Network Design for End-of-lease Computer Prod- ances for Charities.” Journal of Cleaner Production 244:
ucts Recovery.” Transportation Research Part E: Logistics and 118717. doi:10.1016/j.jclepro.2019.118717.
Transportation Review 44 (3): 455–474. doi:10.1016/j.tre. Santibanez-Gonzalez, E. D. R., and A. Diabat. 2013. “Solving
2006.11.003. a Reverse Supply Chain Design Problem by Improved Ben-
Li, X., X. Cui, Y. Li, D. Xu, and F. Xu. 2021. “Optimisa- ders Decomposition Schemes.” Computers and Industrial
tion of Reverse Supply Chain with Used-product Collec- Engineering 66 (4): 889–898. doi:10.1016/j.cie.2013.09.005.
tion Effort Under Collector’s Fairness Concerns.” Inter- Shi, J., W. Chen, Z. Zhou, and G. Zhang. 2020. “A Bi-Objective
national Journal of Production Research 59 (2): 652–663. Multi-period Facility Location Problem for Household
doi:10.1080/00207543.2019.1702229. e-Waste Collection.” International Journal of Production
Lieckens, K., and N. Vandaele. 2007. “Reverse Logistics Net- Research 58 (2): 526–545. doi:10.1080/00207543.2019.1598
work Design with Stochastic Lead Times.” Computational 591.
and Operational Research 34: 395–416. doi:10.1016/j.cor. Shi, J., W. Zhang, S. Zhang, and J. Chen. 2021. “A New Bifuzzy
2005.03.006. Optimization Method for Remanufacturing Scheduling
Lu, S., W. Hou, Z. Li, J. Xia, L. Xie, and H. Su. 2021. “Design and Using Extended Discrete Particle Swarm Optimization
Optimisation of a Green Manufacturing-Recycling Network Algorithm.” Computers & Industrial Engineering 156: 107219.
Considering Heavy Metal Pollutants – an Electronic Assem- doi:10.1016/j.cie.2021.107219.
bly Case.” International Journal of Production Research, 1–20. Shih, L. 2001. “Reverse Logistics System Planning for Recycling
doi:10.1080/00207543.2021.1904160. Electrical Appliances and Computers in Taiwan.” Resources,
Min, H., H. Jeung Ko, and C. Seong Ko. 2006. “A Genetic Conservation and Recycling 32 (1): 55–72. doi:10.1016/S09-
Algorithm Approach to Developing the Multi-echelon 21-3449(00)00098-7.
Reverse Logistics Network for Product Returns.” Omega 34 Taguchi, G., R. Jugulum, and S. Taguchi. 2004. Computer-based
(1): 56–69. doi:10.1016/j.omega.2004.07.025. Robust Engineering: Essentials for DFSS. Mexico: ASQ Qual-
Park, P.-J., K. Tahara, I.-T. Jeong, and K.-M. Lee. 2006. “Com- ity Press.
parison of Four Methods for Integrating Environmental and Tang, Y., M. Zhou, and R. J. Caudill. 2001. “An Integrated
Economic Aspects in the End-of-life Stage of a Washing Approach to Disassembly Planning and Demanufacturing
Machine.” Resources, Conservation and Recycling 48: 71–85. Operation.” IEEE Transactions on Robotics and Automation
doi:10.1016/j.resconrec.2006.01.001. 17 (6): 773–784. doi:10.1109/70.975899.
Paydar, M. M., and M. Olfati. 2018. “Designing and Solving Tosarkani, B. M., S. H. Amin, and H. Zolfagharinia. 2020.
a Reverse Logistics Network for Polyethylene Terephtha- “A Scenario-Based Robust Possibilistic Model for a Multi-
late Bottles.” Journal of Cleaner Production 195: 605–617. objective Electronic Reverse Logistics Network.” Inter-
doi:10.1016/j.jclepro.2018.05.218. national Journal of Production Economics 224: 107557.
Pishvaee, M. S., R. Z. Farahani, and W. Dullaert. 2010. doi:10.1016/j.ijpe.2019.107557.
“A Memetic Algorithm for Bi-objective Integrated For- Trochu, J., A. Chaabane, and M. Ouhimmou. 2018. “Reverse
ward/Reverse Logistics Network Design.” Computational Logistics Network Redesign Under Uncertainty for Wood
18 A. CHE ET AL.

Waste in the CRD Industry.” Resources, Conservation and Emission Constraint.” Waste Management 64: 358–370.
Recycling 128: 32–47. doi:10.1016/j.resconrec.2017.09.011. doi:10.1016/j.wasman.2017.02.024.
Wang, Y., S. Gao, M. Zhou, and Y. Yu. 2021. “A Multi- Zhao, J., S. Liu, M. Zhou, X. Guo, and L. Qi. 2018. “Modified
layered Gravitational Search Algorithm for Function Opti- Cuckoo Search Algorithm to Solve Economic Power Dis-
mization and Real-world Problems.” IEEE/CAA Journal patch Optimization Problems.” IEEE/CAA Journal of Auto-
of Automatica Sinica 8 (1): 94–109. doi:10.1109/JAS.2020. matica Sinica 5 (4): 794–806. doi:10.1109/JAS.2018.7511138.
1003462. Zhou, X., and Y. Zhou. 2015. “Designing a Multi-Echelon
Xu, Z., A. Elomri, S. Pokharel, Q. Zhang, X. G. Ming, and Reverse Logistics Operation and Network: A Case Study of
W. Liu. 2017. “Global Reverse Supply Chain Design for Office Paper in Beijing.” Resources, Conservation and Recy-
Solid Waste Recycling Under Uncertainties and Carbon cling 100: 58–69. doi:10.1016/j.resconrec.2015.04.009.

You might also like