a disassembly line balancing problem with fixed number of workstations
DESCRIPTION
jurnal internationalTRANSCRIPT
![Page 1: A Disassembly Line Balancing Problem With Fixed Number of Workstations](https://reader033.vdocuments.site/reader033/viewer/2022051218/5695d4f71a28ab9b02a38076/html5/thumbnails/1.jpg)
ARTICLE IN PRESSJID: EOR [m5G;October 1, 2015;11:19]
European Journal of Operational Research 000 (2015) 1–13
Contents lists available at ScienceDirect
European Journal of Operational Research
journal homepage: www.elsevier.com/locate/ejor
Production, Manufacturing and Logistics
A disassembly line balancing problem with fixed number of workstations
Eda Göksoy Kalaycılar a, Meral Azizoglu b,∗, Sencer Yeralan c
a ASELSAN, Naval Systems Program Directorate, Ankara 06172, Turkeyb Department of Industrial Engineering, Middle East Technical University, Ankara 06800, Turkeyc Department of Industrial Engineering, Yasar University, Izmir, Turkey
a r t i c l e i n f o
Article history:
Received 30 June 2014
Accepted 3 September 2015
Available online xxx
Keywords:
Integer programming
Heuristics
Disassembly lines
Linear programming relaxation
a b s t r a c t
In this study, a Disassembly Line Balancing Problem with a fixed number of workstations is considered. The
product to be disassembled comprises various components, which are referred to as its parts. There is a speci-
fied finite supply of the product to be disassembled and specified minimum release quantities (possible zero)
for each part of the product. All units of the product are identical, however different parts can be released
from different units of the product. There is a finite number of identical workstations that perform the neces-
sary disassembly operations, referred to as tasks. We present several upper and lower bounding procedures
that assign the tasks to the workstations so as to maximize the total net revenue. The computational study
has revealed that the procedures produce satisfactory results.
© 2015 Published by Elsevier B.V.
1
a
p
r
(
R
t
l
i
v
f
m
a
o
d
w
o
T
(
t
t
I
d
m
t
d
s
s
t
t
d
T
t
s
t
q
e
u
c
f
s
r
w
E
h
0
. Introduction
The environmental regulations, customer awareness and recent
dvances in technology all together have shifted the product recovery
rocess from the act of disposing to the act of remanufacturing and
ecycling. Recycling preserves the material content of the discarded
used) products via some manufacturing and disassembly operations.
emanufacturing, on the other hand, keeps the functional content of
he used products and improves their quality up to a desired usable
evel via disassembly operations and some manufacturing.
Disassembly is the first important step of product recovery activ-
ties (McGovern and Gupta, 2011) and it is methodical extraction of
aluable parts, operations involve the separation of the reusable parts
rom the discarded products. The parts are either subject to some re-
anufacturing process or sold to suppliers. Disassembly operations
re usually performed on a disassembly line that consists of a number
f serial workstations. The first workstation takes the product to be
isassembled. The cycle terminates, i.e., the product leaves the line,
henever all its required parts are disassembled.
The Disassembly Line Balancing Problem (DLBP) assigns the set
f tasks to each workstation for each product to be disassembled.
he problem is critical in minimizing the use of valuable resources
such as time and money) invested in disassembly, and maximizing
he level of automation of the disassembly process and the quality of
he parts or materials recovered.
∗ Corresponding Author. Tel.: +90 3122102281.
E-mail addresses: [email protected] (E.G. Kalaycılar), [email protected]
(M. Azizoglu), [email protected] (S. Yeralan).
s
a
p
m
l
ttp://dx.doi.org/10.1016/j.ejor.2015.09.004
377-2217/© 2015 Published by Elsevier B.V.
Please cite this article as: E.G. Kalaycılar et al., A disassembly line balancin
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004
This study considers a DLBP with a fixed number of workstations.
t is assumed that there is a specified supply for the products to be
isassembled. For each part, a defined minimum release quantity
ust be met. The amount in excess of the minimum release quan-
ity can also be sold in the market. Hence the excess quantity is pro-
uced, provided that the part is profitable, and that there is enough
upply. The aim is to assign the tasks to the disassembly workstations
o as to maximize the total net revenue of the parts, while meeting
heir specified minimum release quantities, and without exceeding
he specified cycle time. To satisfy the minimum release quantities,
ifferent parts may be released from different units of the product.
he challenge is to use different line balances, hence use different
ask assignments to the already mounted workstations, while disas-
embling different units of the product. To the best of our knowledge,
he study is the first attempt to solve the DLBP with minimum release
uantities and fixed number of workstations.
Our study will have a direct impact on the industries that experi-
nce continuous advances in their technology. These advances nat-
rally affect the function and fashion oriented expectations of the
onsumers. One such application area that we take our motivation
rom is the electrical and electronic equipment industry. The con-
umers of the personal computers, televisions or cellular phones,
eplace their products within a few months to a few years, even
hen the products (thereby their parts) are functioning properly.
lectrical and electronic equipments consist of many different sub-
tances some of which may contain hazardous components and valu-
ble parts (Capraz, Polat, & Güngör, 2015). As the used products have
roper conditions, their valuable parts (like memories, CPUs, valuable
etals) can be used in manufacturing new models. The minimum re-
ease quantities for these parts (that we define in our models) are
g problem with fixed number of workstations, European Journal of
![Page 2: A Disassembly Line Balancing Problem With Fixed Number of Workstations](https://reader033.vdocuments.site/reader033/viewer/2022051218/5695d4f71a28ab9b02a38076/html5/thumbnails/2.jpg)
2 E.G. Kalaycılar et al. / European Journal of Operational Research 000 (2015) 1–13
ARTICLE IN PRESSJID: EOR [m5G;October 1, 2015;11:19]
e
w
d
t
a
i
t
t
m
s
t
D
o
s
t
a
c
s
D
d
T
i
p
i
s
p
m
t
b
b
f
t
D
(
t
r
B
b
r
c
c
v
a
b
s
j
c
o
a
p
(
b
g
t
b
p
K
a
m
o
(
associated to the demands from the third party like the producers of
the new models. Moreover, as we assume, the amount released ex-
cess of the minimum release quantities of the valuable parts can be
stored in the inventory for future orders. Another application area is
the automotive industry where the consumers of the luxurious cars
replace their not-so-much-used cars with new brands that make use
of the most recent technology.
The rest of the study is organized as follows. Section 2 reviews the
disassembly process and literature on the DLBP whereas Section 3
defines our problem. In Section 4 the mathematical models and
their use in finding optimal solutions are discussed. In Section 5,
we present upper bounds together with the mechanisms used to
strengthen them and present a heuristic procedure. The computa-
tional experiment is discussed in Section 6 and the study is concluded
in Section 7.
2. Disassembly process & the related literature
Güngör and Gupta (2001b) defined disassembly as a systematic
process of separating a product into its constituent parts, compo-
nents, subassemblies or other groupings. The issues in the area of
disassembly can be classified into two broad categories as design and
operational. Crowther (1999) considered a design for disassembly is-
sue and mentioned that a life cycle model that incorporates the stages
of a disassembly strategy can highlight the environmental advan-
tages of designing for disassembly, showing how it can extend service
life and thereby improve sustainability. As a practical application, he
discussed the construction industry and mentioned that experience
gained from disassemblable buildings can be used to create guide-
lines for other products. As discussed in Brennan, Gupta, and Taleb
(1994), the design aspects for disassembly have being recognized by
the industries that generate huge amount of ferrous and plastic waste
like motor cars and appliance sectors.
When the old products come to disassembly plant so that their
components can be recovered in the assembly plant or re-used, op-
erational problems arise. The operational problems that are likely to
arise are layout, resource allocation, process sequencing and disas-
sembly line balancing. As mentioned by Brennan et al. (1994), the op-
erational problems have environmental concerns that are even forced
by the governments like the recycling regulations, and limitations
on the energy consumptions. These regulations affect the operating
costs as extra costs are incurred for covering the expenses related to
the environmental matters.
In their review paper McGovern and Gupta (2011) discussed many
aspects of the operational problems with an emphasis on the disas-
sembly sequencing problem and the DLBP. The disassembly line se-
quencing problem decides on the disassembly process sequence of a
specified disassembly product and the DLBP decides on the assign-
ment of the tasks to each disassembly workstation.
Lambert (2003) provided a review of the disassembly sequencing
literature. Some noteworthy studies on disassembly sequencing are
due to Lambert (1997), Navin-Chandra (1994) and Lambert (2002).
Navin-Chandra (1994) used a modified traveling salesman problem
so as to minimize total costs while meeting obligatory reclamation of
definite parts. Lambert (1997) proposed a graph based method so as
to maximize the economic performance of the disassembly process
within given technical and environmental constraints. Both meth-
ods were applicable to the disassembly of complex consumer prod-
ucts like automotive vehicles, consumer electronics and mechani-
cal assemblies and were applied to a headlamp and ballpoint pen
disassembly products. Lambert (2002) proposed a linear program-
ming based solution procedure for the minimum cost disassembly
sequence of an electronic equipment. His method was applicable all
products having hierarchical modular structures.
The DLBP was first introduced in Güngör and Gupta (2001b).
Güngör and Gupta (2001a) mentioned several complications like
Please cite this article as: E.G. Kalaycılar et al., A disassembly line balancin
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004
arly leaving, self-skipping, skipping, disappearing and revisiting
ork pieces that could be faced in disassembly line practices. To re-
uce the effects of the complications on the disassembly process,
hey proposed a shortest path based solution algorithm. McGovern
nd Gupta (2007a) showed that the decision version of the DLBP
s NP-complete. McGovern and Gupta (2007b) considered a DLBP
o minimize the number of workstations while balancing the idle
imes between the workstations. They proposed an exhaustive search
ethod that returns optimal solutions for small sized problem in-
tances and a genetic algorithm that finds high quality solutions, for
he large sized problem instances.
Güngör and Gupta (2002) proposed a heuristic procedure for the
LBP under complete disassembly. They assumed an infinite supply
f a single product and considered the efficient utilization of the re-
ources while satisfying the minimum release quantity. They illus-
rated the proposed heuristic on an eight task personal computer dis-
ssembly example. Koc, Sabuncuoglu, and Erel (2009) considered a
omplete disassembly DLBP so as to minimize the number of work-
tations. They introduced AND/OR graphs and proposed Integer and
ynamic Programming formulations.
Altekin, Kandiller, and Özdemirel (2008) studied the DLBP un-
er partial disassembly and an infinite supply of a single product.
hey formulated their net revenue maximization model as a mixed-
nteger linear program and used its relaxations to find lower and up-
er bounds. They stated that their approach could be used in design-
ng and operating remanufacturing systems where large volumes of
imilar products should be disassembled. Altekin and Akkan (2012)
roposed a predictive–reactive approach based on a mixed-integer
odel to improve the profitability of the disassembly line. A predic-
ive balance was created and then given a failure, the line was re-
alanced. They stated that their algorithm could be used as a guide
y the disassembly workers about how to proceed in case of a task
ailure.
The recent disassembly lines research considered the uncer-
ainty of the task times and product quality. Bentaha, Battaïa, and
olgui (2014a, 2014b, 2014c, 2015) and Bentaha, Battaïa, and Dolgui
2014d) studied complete and partial disassembly models, respec-
ively. Bentaha et al. (2014a) modeled uncertainty using the notion of
esource cost and proposed a sample average approximation method.
entaha et al. (2014b) studied the joint problem of disassembly line
alancing and sequencing. Bentaha et al. 2014c proposed a lagrangian
elaxation approach to maximize the total profit. Bentaha et al. (2015)
onsidered the problem of minimizing the workstation operation
osts and additional costs for handling the hazardous parts. They de-
eloped several lower and upper bounding mechanisms. Bentaha et
l. (2014d) addressed workload balancing problem with fixed num-
er of workstations. They developed a stochastic binary program.
Ding, Feng, Tan, and Gao (2010) stated that a successful disas-
embly line often requires an integrated consideration of many ob-
ectives and proposed an ant colony algorithm to generate the effi-
ient set. They tested their algorithm with three objectives: number
f the workstations, workload balancing between the workstations
nd demand rating. They illustrated the heuristic on a 25 task cellular
hone disassembly example. Paksoy, Güngör, Özceylan, and Hancilar
2013) mentioned that in real world applications the objectives could
e fuzzy due to incomplete information and proposed fuzzy goal pro-
ramming and fuzzy multi-objective programming. They considered
hree objectives: number of the workstations, workload balancing
etween the workstations and cycle time. The approaches were ap-
lied on 10 tasks flash light and 30 tasks radio examples. Hezer and
ara (2014) introduced parallel line dissassembly balancing problem
nd proposed a network model based approach for its solution. For
ore information on the dissassembly line balancing and planning,
ne may refer to Ullerich (2014).
The most closely related published work to ours is Altekin et al.
2008)’s study. Our problem environment differs from theirs in the
g problem with fixed number of workstations, European Journal of
![Page 3: A Disassembly Line Balancing Problem With Fixed Number of Workstations](https://reader033.vdocuments.site/reader033/viewer/2022051218/5695d4f71a28ab9b02a38076/html5/thumbnails/3.jpg)
E.G. Kalaycılar et al. / European Journal of Operational Research 000 (2015) 1–13 3
ARTICLE IN PRESSJID: EOR [m5G;October 1, 2015;11:19]
s
r
o
s
t
i
r
u
f
3
p
b
c
p
i
e
k
r
r
A
d
u
i
a
u
t
A
p
A
T
b
a
t
s
F
c
a
edb
Fig. 1. AND precedence relations.
h
f
ig
Fig. 2. POR precedence relations.
b
O
4
m
u
t
X
a
l
∑
c
∑
m
fi
t
∑
t
O
X
ense that we take the number of workstations and cycle time as pa-
ameters whereas they decide on the number of workstations and
n the cycle time. We assume that there is a finite supply of disas-
embly product and that there is a lower bound on the quantity of
he part releases whereas they assume infinite supply and take min-
mum release quantities as upper bounds on the quantity of the part
eleases. Hence their solution considers a single line balance for all
nits, whereas, to satisfy the minimum release quantities, we use dif-
erent task assignments to disassemble different units.
. Problem definition
We consider a disassembly line balancing model with a finite sup-
ly of a single product. There are N tasks where task i is specified
y its cost ci and processing time ti. The tasks that release parts are
alled part releasing tasks. The minimum release quantity of each
art-releasing task i is di units. di is the associated demand for part
from the third party like the producers of the new product mod-
ls. The part-releasing task i has a revenue ri that includes its mar-
et value (for example when the part is sold to the third party) and
ecycled material value. The parts may have negative revenues that
epresent the disposal costs. Ri is the net revenue of task i, Ri = ri – ci.
t the end of the period, each unit of part i that is released excess of
i units can be held in the inventory for future orders and generates a
nit revenue ri when it is sold. Such excess releases would be favored
f they lead to increases in the total net revenue. Our problem is to
ssign the tasks to the workstations of the disassembly line for each
nit of the product so as to maximize the total net revenue. We make
he following additional assumptions.
• The workstations are already set up and there are K workstations.• All workstations are identically equipped and are capable of per-
forming any one of the tasks at the same pace.• There is a single disassembly product with a finite supply of S
units.• There is a single disassembly period of length L, that is sufficient
to disassemble all S units. That is, there is a single batch of size S
to be disassembled in L time units.• The cycle time, C, is the time between the completion of two units
consecutively leaving the disassembly line, hence it is the process-
ing workload that can be allocated to each workstation. One may
view the cycle time to be the ratio L/S .• The cycle time, C, is known with certainty and is not subject to
change as the period length L and supply S, are known with cer-
tainty and are not subject to change.• All units of the product are identical, hence contain all disassem-
bly parts.• A partial disassembly is done, i.e., a subset of -but not- all parts is
released.• The subset of parts released from one unit of the product may be
different from another unit, hence different task assignments are
possible while disassembling different units of the product.
The network comprises AND and OR type relations. Predecessor
ND (PAND) relations imply that a task cannot start before all of its
redecessor tasks are finished. The following figure illustrates the
ND precedence relations.
According to Fig. 1, tasks a, b, c and d are predecessors of task e.
ask d is the immediate predecessor of task e as there is no task in
etween.
We use AND(i) to denote the set of all AND predecessors of task i
nd PAND(i) to denote the set of all AND immediate predecessors of
ask i.
Predecessor OR (POR) relations imply that at least one task in a
pecified set should be complete before the successor task may begin.
ig. 2 illustrates POR type relations.
Please cite this article as: E.G. Kalaycılar et al., A disassembly line balancin
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004
According to the above figure, at least one task in set { f, g, h} must
e completed before task i may begin. We let POR(i) denote the set of
R predecessors of task i.
. Mathematical model and properties
The mathematical formulation given below determines the opti-
um task assignments to the workstations while disassembling each
nit that maximizes the total net revenue. The decision variables of
he model are,
iks ={
1 if task i is assigned to workstationk for s th unit of the product
0 otherwise
Constraint set (1) deals with the minimum release quantities. The
mount of each part released should be higher than its minimum re-
ease quantity.
K
k=1
S∑s=1
Xiks ≥ di i = 1, . . . , N (1)
Constraint set (2) states the cycle time limit should not be ex-
eeded at any workstation for any unit.
N
i=1
tiXiks ≤ C k = 1, . . . , K; s = 1, . . . , S (2)
Constraint sets (3) and (4) ensure that the precedence require-
ents are observed. The AND precedence relations should be satis-
ed. Thus task i cannot be assigned to a workstation that is before
he workstation of its AND predecessor task l.
K
k=1
kXlks ≤K∑
k=1
kXiks i = 1, . . . , N, l ∈ PAND(i), s = 1, . . . , S
(3)
Similarly, the OR precedence relations should be satisfied. Thus
ask i can be assigned to workstation k if and only if at least one of its
R predecessors is assigned to workstations 1 through k.
iks ≤k∑
h=1
∑j∈POR(i)
Xjhs i = 1, . . . , N; k = 1, . . . , K; s = 1, . . . , S
(4)
g problem with fixed number of workstations, European Journal of
![Page 4: A Disassembly Line Balancing Problem With Fixed Number of Workstations](https://reader033.vdocuments.site/reader033/viewer/2022051218/5695d4f71a28ab9b02a38076/html5/thumbnails/4.jpg)
4 E.G. Kalaycılar et al. / European Journal of Operational Research 000 (2015) 1–13
ARTICLE IN PRESSJID: EOR [m5G;October 1, 2015;11:19]
Y
∑
s
u
=m
w
o
M
b
t
m
t
s
l
(
s
e
∑w
e
f
e
f
ε
o
i
s
t
i
r
t
M
m
4
r
d
v
X
M
m
p
S
Constraint set (5) ensures that a task can be assigned to at most
one workstation.
K∑k=1
Xiks ≤ 1 i = 1, . . . , N; s = 1, . . . , S (5)
The assignment variables should be nonnegative as by constraint
set (6).
Xiks ≥ 0 and integer i = 1, . . . , N; k = 1, . . . , K; s = 1, . . . , S
(6)
It follows from Constraint Sets (5) and (6) that Xiks is binary.
The model, hereafter referred to as Model I, is defined by objective
function (7) below, together with constraint sets (1) through (6).
Max
N∑i=1
K∑k=1
S∑s=1
RiXiks (7)
Model I has N ∗ K ∗ S binary decision variables.
McGovern and Gupta (2007b) showed that the decision version of
the DLBP is strongly NP-complete. It follows that our DLBP problem is
NP-hard in the strong sense. We thus resort to several approaches to
solve the DLBP approximately or heuristically. We also develop theo-
retical instruments, more specifically, valid constraints, that improve
the solution effort. These topics are discussed in the remainder of the
paper. In this section, we introduce the key insight into reducing the
problem size.
In typical implementations, the model solution yields different
disassembly tasks and different part-releasing assignments for the
batch of S units. Typically, first a subset of the batch is disassem-
bled to satisfy the minimum release quantity constraints. Recall that
minimum quantities for parts are novel and distinguishing aspect of
the current work. Once the minimum release quantities are satisfied,
though, the solution understandably favors task assignments that
yield the maximum revenue. Note that, once all of the part minimum
release quantities have been met, the specific task assignments to at-
tain the maximum revenue is independent of the remaining quantity
to be disassembled. In fact, the maximum revenue yielding task as-
signments can be found quite easily by solving Model I with S = 1. The
maximum revenue yielding task assignments could be replicated as
many times as needed for the remaining parts.
The above insight immediately suggests an approach to make
problems with large S more tractable. If one had access to the number
of units needed, say Y ≤ S, to be disassembled in order to satisfy the
minimum release quantities, then Model I could be solved using Y in-
stead of S. Since Y ≤ S, and sometimes even Y << S, the computations
are simplified. The remaining task assignments are then computed by
solving Model I with S = 1 and then replicating the solution for the
remaining S–Y units.
Clearly, the above approach would also work for an upper bound
of Y. Even an approximate value of Y may be useful. If the chosen
value of Y gives an infeasible solution, then we could investigate ways
to increase Y to obtain a feasible solution. In the following sections,
we discuss how the quantity Y may be computed, exactly or approxi-
mately, and how bounds for Y may be obtained.
4.1. Computing the number of units to be disassembled to satisfy the
minimum release quantities
The formulation given below, henceforth referred to as Model II,
finds the minimum number of units to be disassembled to satisfy all
minimum release quantities.
Model II can be defined by the following additional variable and
constraint.
s ={
1 if s th unit is disassembled to satisfyminimum release quantity
0 otherwise
Please cite this article as: E.G. Kalaycılar et al., A disassembly line balancin
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004
N
i=1
K∑k=1
Xiks ≤ μ∗Ys s = 1, . . . , S (8)
Constraint Set (9) assigns Ys to 1, if there is at least once task as-
igned to any workstation while disassembling sth unit of the prod-
ct. Note that∑N
i=1
∑Kk=1 Xiks is bounded from above by μ. We use μ
N∗K. If∑N
i=1
∑Kk=1 Xiks ≥ 1 then Ys = 1 otherwise Ys is zero as we are
inimizing Y = ∑Ss=1 Ys. Y = ∑S
s=1 Ys gives the total number of units,
ith at least one task assignment. Hence it is the minimum number
f disassembly units to satisfy all minimum release quantities.
The objective function of Model II is expressed as below.
in
S∑s=1
Ys − εP
N∑i=1
K∑k=1
S∑s=1
RiXiks (9)
The above objective primarily minimizes the number of disassem-
ly units to satisfy all minimum release quantities. Among the solu-
ions that attain the minimum number of disassembly units Y, the
odel selects the one with the maximum net revenue. To maximize
he total net revenue while keeping the minimum number of disas-
embly units, some parts are released more than their minimum re-
ease requirements. The remaining S–Y units are handled by Model III
Section 4.2).
The magnitude of εp used in (9) is important in the sense that it
hould not increase Y = ∑Ss=1 Ys en route to increasing total net rev-
nue. εp should be set small enough to satisfy constraint (10).
S
s=1
Ys − εP∗Pmin ≤
S∑s=1
Ys + 1 − εP∗Pmax (10)
here Pmin (Pmax) = minimum (maximum) possible total net rev-
nue.
Hence any solution with a larger Y = ∑Ss=1 Ysvalue should not be
avored even if it leads to the maximum increase in the total net rev-
nue value. Constraint (10) implies that εP∗Pmax − εP
∗Pmin ≤ 1. This
ollows εP ≤ 1Pmax−Pmin
We use Pmax = S∗ ∑Ni=1 Ridi and Pmin = 0. Hence
P = 1
S∗ ∑Ni=1 Ridi
guarantees the maximum revenue solution among
nes with the minimum number of disassembled units.
Model II is completely defined by the objective function expressed
n (9) together with constraint sets (1)–(6) and (8).
The model satisfies all minimum release requirements by disas-
embling Y = ∑Ss=1 Ys units. For the remaining (S–Y) units, the same
ask assignments are used. To find these task assignments we mod-
fy Model I by setting S = 1, and then replicating the solution for the
emaining S–Y units. This solution, that is, Model I with S = 1 and
hen replicated as needed, is referred to as Model III. In other words,
odel III finds the maximum net revenue without considering mini-
um number of part releases.
.2. Maximum revenue solution in the absence of part minimum
elease quantities
In the following model, Model III, we drop subscript s from our
ecision variable Xiks as we are considering a single unit. The decision
ariable becomes Xik where
ik ={
1 if task i is assigned to workstation k0 otherwise
The objective function is as expressed in (11).
ax
N∑i=1
K∑k=1
RiXik (11)
We also drop the minimum release quantity constraint as they are
et by Model II. Model III is defined by the objective function ex-
ressed in (11) together with constraint sets (2) through (6) by setting
= 1.
g problem with fixed number of workstations, European Journal of
![Page 5: A Disassembly Line Balancing Problem With Fixed Number of Workstations](https://reader033.vdocuments.site/reader033/viewer/2022051218/5695d4f71a28ab9b02a38076/html5/thumbnails/5.jpg)
E.G. Kalaycılar et al. / European Journal of Operational Research 000 (2015) 1–13 5
ARTICLE IN PRESSJID: EOR [m5G;October 1, 2015;11:19]
4
c
t
s
M
a
b
w
m
4
t
i
r
I
o
u
p
s
t
u
o
f
U
W
v
i
a
T
5
t
a
t
t
5
v
m
s
v
T
t
b
o
t
T
a
P
l
a
a
s
X
v
r
fi
u
c
∑
e
c
fl
p
T
s
P
o
w∑t
t∑W
�f
r
b
E
s
c
t
s
.3. Solution procedure using Model II and Model III
Model I given above yields the optimum solution but suffers from
omputational difficulties as the problem size increases. Alterna-
ively, Model II and Model III may now be used to find an optimal
olution as given by the following procedure.
Procedure 1. Optimal Solution via Model II and Model III
Step 1. Use Model II to find the task assignments for the first Y =∑Ss=1 Ys disassembly units.
Step 2. Use Model III to find the task assignments for the last (S–Y)
disassembly units.
Step 3. The optimal total net revenue over all S units is:
Total Net revenue by Model II + (S – Y) ∗ Total Net revenue by
odel III
Model II is more complex than Model I due to the added Ys vari-
bles, hence in finding the optimal solutions, Procedure 1 should not
e favored to Model I. However, Procedure 1 is of importance, since
e use the procedure to derive upper and lower bounds on the opti-
al total net revenue value.
.4. Solution procedure using Model I and Model III
Model I and Model III together may be used to find optimal solu-
ions as well. Recall that Y = ∑Ss=1 Ys returned by Model II is the min-
mum number units to be disassembled to satisfy all minimum part
equirements. If an upper bound on Y, UB(Y), is available, then Model
can be solved with UB(Y) in place of S units. Thereafter, Model III
ptimally solves the remaining S–UB(Y) units. Procedure 2, below, is
sed to find an optimal solution using Model I and Model III. The up-
er bound UB(Y) may be guessed, or sought using a simple algorithm
uch as bisection search. We will revisit the task of finding UB(Y) in
he following sections.
Procedure 2. Optimal Solution via Model I and Model III
Step 1. Solve Model I with S = UB(Y).
Step 2. Solve Model III to find the task assignments for the remain-
ing S –UB(Y) units.
Step 3. Find the optimal total net revenue over all S units as:
Total net revenue over UB(Y) units by Model I+(S–UB(Y))
nits∗Total net revenue over a single unit by Model III
Note that there are three ways of finding an optimal solution: the
riginal Model I, Procedure 1, and Procedure 2. Procedure 2 should be
avored to original Model I, in particular when compared to S, a small
B(Y) is available.
In our experiments, we use Model I to find the optimal solutions.
e develop two upper bounds to the optimum objective function
alue that use the Linear Programming Relaxations (LPRs) of the orig-
nal model. We benefit from the ideas used in Model II and III to find
third upper bound and to develop a heuristic solution procedure.
hese are presented next.
. Solution approaches
As mentioned, our DLBP is NP-hard in the strong sense. Hence
here is little chance of finding polynomial even a pseudo-polynomial
lgorithm that solves it optimally. In the section we present our solu-
ion approaches that provide upper bounds and a heuristic solution
o our DLBP.
.1. Upper Bounds
Upper Bound 1 (UB1): An optimal solution to any relaxation pro-
ides an upper bound on the optimal objective function value of our
aximization problem.
Please cite this article as: E.G. Kalaycılar et al., A disassembly line balancin
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004
Model I is solved to optimality by relaxing the integrality con-
traints on the Xiks values. The resulting optimal objective function
alue, ZLP, provides an upper bound on the optimal total net revenue.
he only difference between the original model and its LPR is the in-
egrality requirement of the Xiksvalues. Consequently, ZLP is an upper
ound on the optimal objective function value Z∗.
Upper Bound 2 (UB2): UB2 is found by adding valid cuts to the LPR
f Model I. The valid cuts are found by investigating the properties of
he optimal solutions that may not be satisfied by the optimal LPR.
Cut1–Fixing the minimum release quantities
heorem 1. There exists an optimal solution in which di units of part i
re released from the first di disassembly units.
roof. Assume an optimal solution in which the first di units are re-
eased from di units of the product. If the tasks in any two units, say
and b, are changed, i.e., Xika is set to Xikb and Xikb is set to Xika for
ll i , then the optimal net revenue is not affected. Assume an optimal
olution in which part i is released in units a+1, a+ di . Setting Xjkr =jk(a+r) for all j, r ≤ dj and Xjk(ta+r) = Xjkr does not change the revenue
alue. Then, there is an optimal solution in which dj units of part i is
eleased in the first di units. �
Using Theorem 1, dr units for any part r can be obtained from the
rst dr disassembly units. Constraint set (12) below ensures that dr
nits of part r are released from the first dr disassembly units. To in-
rease the power of (12), we select task r such that dr=Maxi{di }.
K
k=1
dr∑s=1
Xrks = dr s = 1, . . . , dr (12)
In computational studies, we observe that the constraint set (12)
liminates many alternate optimal solutions, thereby decreasing the
omputational effort. The reduction in computational effort was re-
ected as significantly reduced CPU times.
Cut2 –Lower Bound on the workstation positions
Cut2 provides a lower bound on the workstation index that any
articular task.
heorem 2. In all optimal solutions, task i cannot be assigned to work-
tations 1, 2, . . . , Ei − 1 where Ei = � ti+Min j∈POR(i){t j}+∑j∈AND(i) t j
C �.
roof. Task i cannot start before its predecessors and at least one
f its immediate POR predecessors are completed. Hence it should
ait at least Min j∈POR(i){t j} units for its POR predecessors and
j∈AND(i) t j units for its AND predecessors to start and at least
i + Min j∈POR(i){t j} + ∑j∈AND(i) t j units to complete. When task split-
ing is allowed and other tasks are ignored, ti + Min j∈POR(i){t j} +j∈AND(i) t j units require � ti+Min j∈POR(i){t j}+∑
j∈AND(i) t j
C � workstations.
hen task splitting is not allowed and other tasks are consideredti+Min j∈POR(i){t j}+∑
j∈AND(i) t j
C � is a lower bound on the earliest station
or task i. �
In the absence of POR relations, Theorem 2 reduces to the one de-
ived in Patterson and Albracht (1975) for the classical assembly line
alancing problem.
Constraint set (13) below supports Theorem 2.
i−1∑k=1
S∑s=1
Xiks = 0 i = 1, . . . , n (13)
Constraint set (13) ensures that task i is not assigned to work-
tations 1, 2, …, Ei–1. The network in Fig. 3 is used to illustrate Ei
omputations.
Note that { t1 + t2 } is the minimum total processing load to start
ask A and Min {t3, t4}+ t5 is the minimum total processing load to
tart task B. Then Min{(t1 + t2+ tA), (Min {t3, t4} + t5 + tB)} is the
g problem with fixed number of workstations, European Journal of
![Page 6: A Disassembly Line Balancing Problem With Fixed Number of Workstations](https://reader033.vdocuments.site/reader033/viewer/2022051218/5695d4f71a28ab9b02a38076/html5/thumbnails/6.jpg)
6 E.G. Kalaycılar et al. / European Journal of Operational Research 000 (2015) 1–13
ARTICLE IN PRESSJID: EOR [m5G;October 1, 2015;11:19]
Min {SA+tA , SB+tB } = Sj
t1 + t2 = SA
Min {t3 , t4 } + t5 = SB
A
j
B
1
2
3
4
5
Fig. 3. An example network to illustrate Ei computations.
c
t
c
p
i
T
o
t
K
P
r
r
t
h
t
w
s
O
t
f
p
c
i
e
w
m
a
∑
∑
t
a
s
M
q
u
s
m
m
A
T
b
P
l
m
u
minimum total processing load to start task j.
Note that Ei =⌈
ti + Minimum total processing load tostart task i
C
⌉.
Hence Ei =⌈
ti + Min{(t1 + t2 + tA), (Min{t3, t4} + t5 + tB)}C
⌉.
The network in Fig. 4 is adapted from Bourjault (1984) to illustrate
Theorem 2.
We take K = 4, C = 63, S = 80 and the task times are as given
below:
Tasks 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
ti 13 7 1 6 7 5 15 8 17 3 14 12 17 4 11 18 15 19 4 18 16 19
E13 =⌈
t13 + Min{(t1 + t3 + t4), (t1 + t2)} + t22 + t7
C
⌉
E13 =⌈
17 + Min{(13 + 1 + 6), (13 + 7)} + 19 + 15
63
⌉=
⌈71
63
⌉=2
The theorem states that task i cannot be assigned to workstations
1, 2, …, Ei – 1. Accordingly, task 13 cannot be assigned to workstation
1 (E13– 1 = 1).
Cut3–Projecting minimum release quantities
Altekin et al. (2008) stated that in the optimal solution, the total
fraction of any task assigned is never greater than the total fraction
of each of its assigned AND predecessors and the total fractional as-
signment of all of its OR predecessors, and moreover the constraints
may not be satisfied for the optimal LPR. Their results hold for any
problem that contains PAND and POR relations and an infinite supply.
We generalize their results to our problem where the supply is finite
and impose the constraint sets (14) and (15).
K∑k=1
S∑s=1
Xiks ≤K∑
k=1
S∑s=1
Xjks i = 1, . . . , n; j ∈ PAND(i). (14)
K∑k=1
S∑s=!
Xiks ≤∑
j∈POR(i)
K∑k=1
S∑s=1
Xjks i = 1, . . . , n (15)
Cut4–Existence of a feasible solution
Our last cut uses the results of the following two theorems.
Theorem 3. If there exists a feasible schedule that processes all tasks
other than task i and its successors in r workstations then there exists an
optimal schedule in which task i is processed in workstations 1 through r.
Proof. Assume task i is assigned to workstation k such that k ≤ r and
workstations r +1 through K process the successors of task i. Taking
task i from workstation k and placing to workstation l such that l ≥r + 1 cannot increase the total net revenue as such a replacement
Please cite this article as: E.G. Kalaycılar et al., A disassembly line balancin
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004
annot allow more task assignments. This is due to the fact that all
asks in workstations r + 1 through K are successors of task i, hence
annot be processed in workstations 1 through r, since task i is to be
rocessed later. Then, there exists an optimal solution in which task i
s processed in the first r workstations. �
heorem 4. If there exists a feasible schedule that processes all tasks
ther than task i and its predecessors in the last r workstations, then
here is an optimal schedule in which task i is processed in workstations
– r + 1 through K.
roof. Assume task i is assigned to workstation k such that k ≥ K –
+ 1 and workstations 1 through K – r process the predecessors of
task j. Taking task i from workstation k and placing it to workstation
l such that l ≤ K – r cannot increase the total net revenue, as such a
eplacement cannot allow more Assignments. This is due to the fact
hat all tasks in workstations 1 through K–r are predecessors of task i,
ence cannot be processed in workstations K – r + 1 through K, since
ask i is processed earlier. Then, there exists an optimal solution in
hich task i is processed in the last r workstations. �
Note that Theorems 3 and 4 require a feasible solution. To find
uch a solution we develop the following simple rule.
Let A be a set of tasks that should be assigned to K workstations.
rder the tasks in set A according to non-decreasing task times. Take
he tasks from the order starting from the first feasible task. A task is
easible if its inclusion to the current workstation does not violate the
recedence relations and cycle time constraints. If no such task exists,
lose the current workstation and open a new one. Stop when all jobs
n set A is assigned. Let m(A) be the resulting number of workstations.
Following Theorem 3, we set A = A1 where A1 is the set of all tasks
xcept the successors of task i. Following Theorem 4, we set A = A2
here A2 is the set of all tasks except the predecessors of task j. After
(A1) and m(A2) are obtained, implementing the heuristic for sets A1
nd A2, we include one of the two cuts (16) and (17).
S
s=1
K∑k=m(A1)+1
Xiks = 0 i = 1, . . . , N (16)
S
s=1
m(A2)−1∑k=1
Xiks = 0 i = 1, . . . , N (17)
We select one of the cuts for the LPR using the following rule.
Rule. Use Cut (16) if m(A1) ≥ m(A2) – 1, else use Cut (17)
The rule selects the cut that prevents more assignments, hence
he stronger one.
We hereafter refer to the LPR using the cuts (12) through (17)
s strengthened LPR. UB2 is the total net revenue returned by the
trengthened LPR.
Upper Bound 3 (UB3): To obtain UB3, we use the idea used in
odel II. Recall that Model II aims to satisfy all minimum release
uantities (∑
i di) and minimizes Y = ∑Ss=1 Ys. Here in place of Y, we
se LB(Y), to get rid of the Ys binary variables.
First, we find LB(Y), and then increase it one by one, until a feasible
olution is reached. Theorem 5 states LB(Y).
We let dnewi = Max{di, Max j|i∈AND j{d j}}. Note thatdnewi is the
inimum release quantity of task i which is modified by its mini-
um release quantity and the minimum release quantities of all its
ND successors.
heorem 5. LB(Y) = �∑n
i=1 ti∗dnewi
K∗C � is a valid lower bound on the num-
er of units to satisfy all minimum release quantities.
roof. The total processing required to produce all minimum re-
ease quantity is∑n
i=1 ti∗dnewi Each unit can be disassembled in no
ore than K∗C time units as each workstation is available for C time
nits. �∑n
i=1 ti∗dnewi∗ � is the number of units disassembled to satisfy
K Cg problem with fixed number of workstations, European Journal of
![Page 7: A Disassembly Line Balancing Problem With Fixed Number of Workstations](https://reader033.vdocuments.site/reader033/viewer/2022051218/5695d4f71a28ab9b02a38076/html5/thumbnails/7.jpg)
E.G. Kalaycılar et al. / European Journal of Operational Research 000 (2015) 1–13 7
ARTICLE IN PRESSJID: EOR [m5G;October 1, 2015;11:19]
1
3
5
2
4
6
8
9
7
13
14
15
16
17
12
11
10
18
20
21
22
19
Fig. 4. Precedence diagram adapted from Bourjault’s ball-point pen example.
a
s
i
d
K
a
s
f
a
m
M
F
fi
w
H
r
a
t
5
l
o
i
T
t
c
i
q
a
M
s
n
s
i
p
ll minimum release quantity if task splitting between the disas-
embly units and workstations were allowed. As no task splitting
s allowed �∑n
i=1 ti∗dnewi
K∗C � is a lower bound on the number of units
isassembled. �
We illustrate Theorem 5 and UB3 on the network in Fig. 4 with
= 4, S = 80 and C = 63. The minimum release quantities, task times
nd dnewi values are given in Table 1.
Using Theorem 5 we find LB(Y) = � 13∗4+7∗4+···+19∗44∗63 � = 3. We first
et the number of units to LB(Y) = 3. However a feasible solution
or the LPR is not found. There is no feasible solution with 4 units
s well. The LPR finds a feasible solution with 5 units. Then the mini-
um number of units to satisfy all minimum release quantities, Y = 5.
odel III is used to find the line balance for the remaining 75 units.
ormally we let,
ZA = maximum net revenue for the strengthened LPR problem
with Y units
ZB = maximum net revenue for a single unit problem with
Model III.
UB3 = ZA + (S–Y)∗ZB
In our example, the strengthened LPR finds ZA as 1626. Model III
nds ZB , as 326. Then UB3= 1626 + (80 – 5) ∗ 326 = 26.076.
Note that here, UB2 = UB3. ZA is found much easier than the LPR
ith S = 80. Moreover ZB is found easily as it considers a single unit.
ence one should expect to obtain UB3 much quicker than UB2.
For this instance we find that UB1=26.796 and the optimal net
evenue = 26.076. Note that UB3, improves UB1 by 720 money units
nd catches the optimal solution in considerably small solution
ime.
Please cite this article as: E.G. Kalaycılar et al., A disassembly line balancin
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004
.2. Lower bound, heuristic solution
The lower bound uses UB3 as a stepping stone to find a feasible so-
ution. The motivation to use UB3 is its satisfactory behavior in terms
f its solution quality and time. The bound first solves UB3, then fixes
ts variables that are assigned to 1 and obtains a reduced problem.
he reduced problem has only the partially assigned or unassigned
asks of the UB3 solution. As the partial assignments are quite low
ompared to the original variables, the reduced problem is very small
n size. Since our MILP solver can handle small-sized problems very
uickly, we prefer to use it to find an optimal solution for partially
ssigned tasks.
After fixing the variables to 1, the resulting solution obtained by
ILP may violate the cycle time constraints or the precedence con-
traints. If the cycle time constraint is violated, then we increase the
umber of disassembly units one by one, until we obtain a feasible
olution. If the precedence relations are violated, the tasks whose fix-
ngs cause violation, are set to zero, and the MILP is solved again.
Below is the stepwise description of our lower bounding
rocedure.
Step 0. Let LB(Y) = �∑n
i=1 ti∗dnewi
K∗C � and let T = LB(Y)
Step 1. Solve the LPR with cuts for T units of the disassembly
product.
Fix the variables that have value ‘1’ in the LPR solution.
Step 2. Find the optimal solution by Model I for the non-fixed tasks.
If the solution is feasible then let ZC be the objective function
value and go to Step 4.
Step 3. If infeasibility is due to the cycle time constraints then let T
= T+1, go to Step 1.
If infeasibility is due to the precedence relations then set the
variable of a task to zero if one of its predecessor tasks is not
assigned to ‘1’, go to Step 2.
g problem with fixed number of workstations, European Journal of
![Page 8: A Disassembly Line Balancing Problem With Fixed Number of Workstations](https://reader033.vdocuments.site/reader033/viewer/2022051218/5695d4f71a28ab9b02a38076/html5/thumbnails/8.jpg)
8 E.G. Kalaycılar et al. / European Journal of Operational Research 000 (2015) 1–13
ARTICLE IN PRESSJID: EOR [m5G;October 1, 2015;11:19]
Table 1
The data for the Bourjault’s ball-point pen problem.
Tasks 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
di(units) 0 0 0 4 0 4 0 3 4 4 4 4 4 3 2 1 0 4 2 1 0 0
ti(time units) 13 7 1 6 7 5 15 8 17 3 14 12 17 4 11 18 15 19 4 18 16 19
dnewi(units) 4 4 4 4 4 4 4 3 4 4 4 4 4 3 2 1 0 4 2 1 0 4
ci(money units) 17 18 14 7 14 8 8 15 16 7 15 12 6 17 17 6 9 13 12 16 7 15
ri(money units) 49 10 35 37 16 12 20 37 43 32 24 41 19 35 19 18 48 21 10 40 16 11
Ri (money units) 32 -8 21 30 2 4 12 22 27 25 9 29 13 18 2 12 39 8 -2 24 9 -4
1
25
3
4
6
13
11
14
15
12
7
32
31
17
348
9
10
30
16
33
23
24
28
29
21
20
22
19 25
26
27
18
Fig. 5. Precedence diagram adapted from Lambert’s radio example.
a
Step 4. Find an optimal solution to the single unit problem using
Model III.
Let ZD be the objective function value
The heuristic gives a total net revenue of ZC + (S – T) ∗ ZD.
6. Computational experiment
In this section the data generation scheme is discussed and the
computational results are evaluated. Two networks from the previous
works are used for network generation.
1st Network (22 tasks). Bourjault (1984) illustrated a disassembly
network for a 10-part ballpoint pen. We adapt this network by
replacing the successor OR relations with successor AND rela-
tions and give the resulting network in Fig. 4.
Please cite this article as: E.G. Kalaycılar et al., A disassembly line balancin
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004
2nd Network (34 tasks). Lambert (1997) illustrated a disassembly
network for a 20-part radio. We adapt this network by replac-
ing the successor OR relations with successor AND relations
and give the resulting network in Fig. 5.
Three networks are found by extending the small networks using
dditional arcs.
3rd Network (47 tasks). 2nd network and a part of the 1st net-
work are combined with some arcs added.
4th Network (60 tasks). 3rd network and a part of the 1st network
are combined with some arcs added.
5th Network (73 tasks). 4th network and a part of the 1st network
are combined with some arcs added.
The following two sets for the number of workstations K are used.
g problem with fixed number of workstations, European Journal of
![Page 9: A Disassembly Line Balancing Problem With Fixed Number of Workstations](https://reader033.vdocuments.site/reader033/viewer/2022051218/5695d4f71a28ab9b02a38076/html5/thumbnails/9.jpg)
E.G. Kalaycılar et al. / European Journal of Operational Research 000 (2015) 1–13 9
ARTICLE IN PRESSJID: EOR [m5G;October 1, 2015;11:19]
Table 2
UB1 performances.
Variables Set C1 Set C2
Set D1 Set D2 Set D1 Set D2
Avg Max Avg Max Avg Max Avg Max
Set K1 N = 22 3,520 Percent DEV 1 2 1 2 1 2 1 2
# of frac. 722.4 1636 1120.7 1778 487.1 1356 470.6 1684
Set K2 N = 22 7,040 Percent DEV 1 3 1 3 1 3 2 3
# of frac. 2437.5 3643 2497.4 3592 1628.1 3017 1900.4 3306
Set K1 N = 34 5,440 Percent DEV 0 1 0 1 0 1 0 1
# of frac. 715.9 1950 1150.8 2251 786 2367 741 2210
Set K2 N = 34 10,880 Percent DEV 0 1 0 1 0 1 0 1
# of frac. 3803.1 4864 4238.7 5136 3004.1 4307 3102.6 4386
Set K1 N = 47 7,520 Percent DEV 0 1 0 1 0 1 0 1
# of frac. 1184 4011 2509.4 3924 584.2 2403 1152.4 3391
Set K2 N = 47 15,040 Percent DEV 0 1 0 1 0 1 0 1
# of frac. 5492.4 6865 5910.8 6966 4041.1 5888 4654.9 5946
Set K1 N = 60 19,200 Percent DEV - ∗ - - - - - - -
# of frac. 8749.8 9579 8999.5 9791 6887.3 8344 7751.1 8935
Set K2 N = 60 38,400 Percent DEV - - - - - - - -
# of frac. 13515.3 14742 13853.1 15153 11399.3 13268 12307.1 14280
Set K1 N = 73 23,360 Percent DEV - - - - - - - -
# of frac. 10876.3 11804 11249.8 11974 9137.3 10462 9716.3 10633
Set K2 N = 73 46,720 Percent DEV - - - - - - - -
# of frac. 16730.7 18440 17291.3 18654 14045.5 15689 15019.8 16147
∗ ‘-‘ illustrates that optimal solution for large networks could not be obtained.
S
S
t
c
e
b
q
q
d
g
N
2
i
u
r
r
s
f
s
t
t
b
t
r
l
1
w
w
d
v
n
o
w
W
T
h
s
3
r
d
n
u
C
a
t
r
i
v
n
K
t
w
a
D
a
i
r
Set K1. K = 2 for small networks with 20, 34, 47 tasks
K = 4 for large networks with 60, 73 tasks
Set K2. K = 4 for small networks with 20, 34, 47 tasks
K = 8 for large networks with 60, 73 tasks
The following two sets for the cycle times, C are used.
et C1.C =⌈∑
ti
K
⌉
et C2.C = 1.5∗⌈∑
ti
K
⌉The processing times are generated from a discrete uniform dis-
ribution [1, 20]. A task receives a minimum release quantity, hence
alled a part-releasing task, according to a random process. We gen-
rate a random number between 0 and 1. If the generated number is
elow 0.3, we set its minimum release quantity to zero.
Two distributions are used to generate their minimum release
uantities.
Set D1. Minimum release quantity is uniform [1, 5]
Set D2. Minimum release quantity is uniform [5, 10]
Note that Set D1 contains instances with a low minimum release
uantity, and Set D2, with a high minimum release quantity.
The number of units, S, is set to 80. The costs are generated from a
iscrete uniform distribution between 5 and 20 and the revenues are
enerated from a discrete uniform distribution between 10 and 50.
There are 40 combinations due to 5, 2, 2 and 2 alternatives for
, K, C and D respectively (3∗2∗2∗2 = 24 for small networks and∗2∗2∗2 = 16 for large networks). For each combination, 10 problem
nstances are generated. That is, a total of 400 problem instances are
sed in our experiments.
The mathematical models are solved by CPLEX 10 and the algo-
ithms are coded in Microsoft Visual C++ 2008. The experiments are
un in Intel Core2 Duo 2.00 gigahertz, 2 gigabyte RAM.
The execution of the MILP (Model I) is terminated if the optimal
olution is not returned in 3600 seconds. The optimal solutions are
ound within 3600 seconds when N = 22, 34 and 47. However, optimal
olutions were not found when N = 60 and 73 .
First the performance of our upper bounds is studied. The devia-
ion of an instance is found as %Dev = (UBi−OPT
) × 100 where OPT =
OPTPlease cite this article as: E.G. Kalaycılar et al., A disassembly line balancin
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004
he optimal net revenue and UBi = the total net revenue by upper
ound i. The average and maximum deviations and number of frac-
ional variables are reported in Table 2 and Table 3, for UB1 and UB2,
espectively.
As can be observed from the tables, the deviations are consistently
ow over all problem sets. For UB1, the average deviations are below
percent and almost all maximum deviations are below 3 percent
hen N = 22, 34 and 47. Note that the deviations do not deteriorate
ith an increase in N. However they deteriorate with an increase in K
ue to the inflation of the number of binary variables. When N and K
alues are fixed, the minimum release quantities and cycle times do
ot affect the deviations significantly.
UB2 produces slightly smaller deviations due to the power of
ur cuts. For example, the maximum deviation for UB1 is 3 percent,
hereas it is nearly zero for UB2 in sets K1, C1 and D1 for N = 22.
e also give the number of fractional variables in Tables 2, 3 and 4.
he number of the fractional variables produced by the LPR is too
igh. For example when N = 22, for sets K2, C2 and D2 (i.e., 4 work-
tations, high minimum release quantity, and high cycle time cases),
306 out of 7040 variables are found to be fractional. This value
educes to 412 when cuts are added. On the average, the cuts re-
uce the number of fractional variables from 1900.4 to 180.3. The
umber of fractional variables is generally affected by the N val-
es. For example, from Tables 2–4, it can be seen that for sets K2,
2 and D2, increasing N increases the number of fractional vari-
bles of UB2, on the average. Note that the average number of frac-
ional variables are 1900.4, 3102.6 and 4654.9 for N = 22, 34 and 47
espectively.
It can also be observed from the tables that, for fixed N, an increase
n the number of workstations increases the number of the fractional
ariables. For example for N = 22 and sets C1 and D1, the average
umber of fractional variables is 722.4 when K = 2, and 2437.5 when
= 4. This is due to the fact that more workstations lead to higher
ask splits, leading to more fractional variables. For fixed N and K,
e observe that increasing C reduces the number of fractional vari-
bles (with few exceptions). For example, when N = 47, K = 2, and
1 combination, the average number of fractional variables are 1184
nd 584.2 for low and high C, respectively. This is due to the fact that
ncreasing C gives more room for complete task assignments, hence
educing the number of fractional variables. In our experiments, we
g problem with fixed number of workstations, European Journal of
![Page 10: A Disassembly Line Balancing Problem With Fixed Number of Workstations](https://reader033.vdocuments.site/reader033/viewer/2022051218/5695d4f71a28ab9b02a38076/html5/thumbnails/10.jpg)
10 E.G. Kalaycılar et al. / European Journal of Operational Research 000 (2015) 1–13
ARTICLE IN PRESSJID: EOR [m5G;October 1, 2015;11:19]
Table 3
UB2 performances.
Variables Set C1 Set C2
Set D1 Set D2 Set D1 Set D2
Avg Max Avg Max Avg Max Avg Max
Set K1 N = 22 3,520 Percent DEV 0 0 0 0 0 0 0 0
# of frac. 65.3 160 67 160 0 0 0 0
Set K2 N = 22 7,040 Percent DEV 0 0 0 0.01 0 0 0 0
# of frac. 368.4 570 376 596 197.8 432 180.3 412
Set K1 N = 34 5,440 Percent DEV 0 0 0 0 0 0 0 0
# of frac. 19 160 22.9 158 0 0 0 0
Set K2 N = 34 10,880 Percent DEV 0 0 0 0 0 0 0 0
# of frac. 525.2 870 517.7 855 271.1 438 259.1 446
Set K1 N = 47 7,520 Percent DEV 0 0 0 0 0 0 0 0
# of frac. 49 160 50 160 0 0 0 0
Set K2 N = 47 15,040 Percent DEV 0 0 0 0 0 0 0 0
# of frac. 570.3 970 592 982 370.4 800 362.2 815
Set K1 N = 60 19,200 Percent DEV -∗ - - - - - - -
# of frac. 746 1524 765.7 1546 418.3 1068 424.8 1068
Set K2 N = 60 38,400 Percent DEV - - - - - - - -
# of frac. 1778.7 2203 1725.2 2116 1192.1 1523 1179.2 1434
Set K1 N = 73 23,360 Percent DEV - - - - - - - -
# of frac. 657.5 922 649.9 859 460.8 1027 462.4 1045
Set K2 N = 73 46,720 Percent DEV - - - - - - - -
# of frac. 1789.1 2326 1769 2244 1197.1 1931 1201.8 2008
∗ ’-‘ illustrates that optimal solution for large networks could not be obtained.
Table 4
The number of units with different line balances (out of 80).
N C1 C2
D1 D2 D1 D2
Avg Max Avg Max Avg Max Avg Max
K1 22 5.30 6.00 11.00 12.00 5.50 6.00 11.20 12.00
34 5.10 6.00 10.00 10.00 5.90 6.00 10.00 10.00
47 5.50 6.00 11.30 12.00 5.80 6.00 11.20 12.00
K2 22 5.10 4.00 11.80 13.00 5.30 6.00 10.70 11.00
34 5.00 5.00 10.90 11.00 5.10 6.00 10.00 10.00
47 5.30 6.00 11.10 12.00 5.40 6.00 11.00 12.00
i
s
t
0
t
U
C
s
d
a
f
a
c
w
i
f
s
T
t
m
K
s
F
a
T
t
h
c
t
6
i
u
h
t
do not observe a significant effect of the minimum release quantity
figures on the number of fractional variables.
Table 4 reports the number of different line balances incurred in a
period, while S units are disassembled, for small networks.
Note that the number of different line balances returned by Model
II is quite small relative to the total number of disassembly units. For
D1, there are about 5 and 6 different line balances. Increasing D from
D1 to D2, almost doubles the number of different balances, as more
units have to be used to satisfy the minimum release quantities. UB3
is based on the idea of the different line balances, hence one should
expect that it runs much faster than UB2.
The solution times are measured in central processing unit (CPU)
seconds. The average and maximum CPU times for our three upper
bounds are given in Table 5 and Table 6 for small and large networks,
respectively. Table 5 includes the CPU times of the MILP, as the small
networks could be solved to optimality.
As can be observed from the tables, the upper bounds are pro-
duced very quickly. Compared to UB1, UB2 is computed more quickly,
due to the power of the cuts, i.e., their effectiveness in reducing the
solution space.
Moreover, compared to UB2, UB3 is computed more quickly as rel-
atively fewer units, hence fewer variables, are used by the LPRs. For
example, for N = 22 and the instances in sets K2, C2 and D2, the CPU
time decreases from 0.59 to 0.25 seconds by adding the cuts, and sim-
ilarly, the CPU time decreases from 0.25 to 0.08 seconds by using up
to 10 supply units instead of 80. The CPU times of finding the up-
per bounds increase with an increase in N and K. This is due to the
Please cite this article as: E.G. Kalaycılar et al., A disassembly line balancin
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004
ncrease in the dimensions of the linear programs. For example, for
ets K1, C2 and D2, as N increases from 60 to 73, the average CPU
imes of UB1, UB2 and UB3 increase from 3.18, 0.72, 0.11 to 4.52, 0.83,
.13 respectively.
For N = 34 and sets C2 and D2, as K increases, the average CPU
imes increase from 0.43, 0.08, 0.06 to 1.18, 0.31, 0.09 seconds for UB1,
B2 and UB3 respectively.
For fixed N, K and D values, an increase in the C value decreases the
PU time. This is due to the fact that for large C, more tasks find an as-
ignment for a workstation, and hence assignment decisions are pro-
uced easier by linear programs. For example, for N = 34 and Sets K2
nd D2, increasing C from C1 to C2, decreases the average CPU times
rom 1.67, 0.46, 0.11 seconds to 1.18, 0.31, 0.09 seconds for UB1, UB2
nd UB3, respectively. When the other parameters are fixed, the in-
rease in the D value increases the CPU times slightly. This is because
hen D increases, the number of units with different assignments
ncreases. For example, for N = 34 and sets C1 and K2, increasing D
rom D1 to D2, increases the average CPU times from 1.27, 0.44, 0.08
econds to 1.67, 0.46, 0.11 seconds for UB1, UB2 and UB3 respectively.
he effect of K is more dominant for the MILP, due to the inflation of
he number of decision variables.
For fixed N, an increase in the K value increases the CPU times re-
arkably. For example for N = 22 and sets C1 and D1, an increase in
from K1 to K2 increases the CPU time from 3.83 seconds to 1078.32
econds. However the effect of N is not as significant as that of K.
or example, for K = 2, when N increases from 22 to 34 for sets C1
nd D1, the CPU time increases from 3.83 to 4.37 seconds. Note from
able 5 that for fixed N and K, an increase in the C value decreases
he CPU times remarkably. This is due to the fact that a workstation
as room to accommodate many tasks, thereby leading to easier de-
isions. However D values do not have a consistent effect on the solu-
ion time of the MILP.
Fig. 6 illustrates the CPU times of UB1 for sets K, C and D when N =0 and 73.
As can be observed from Fig. 6 the CPU times of UB1 increase with
ncreases in N and K values and decrease with an increase in C value.
To summarize, our experiments have revealed that high quality
pper bounds could be obtained even for large networks in less than
alf a minute. The upper bound deviations are not affected from
he increases in the problem sizes and CPU times to obtain them
g problem with fixed number of workstations, European Journal of
![Page 11: A Disassembly Line Balancing Problem With Fixed Number of Workstations](https://reader033.vdocuments.site/reader033/viewer/2022051218/5695d4f71a28ab9b02a38076/html5/thumbnails/11.jpg)
E.G. Kalaycılar et al. / European Journal of Operational Research 000 (2015) 1–13 11
ARTICLE IN PRESSJID: EOR [m5G;October 1, 2015;11:19]
Table 5
The CPU times (in second) of the Upper Bounds and MILP–small networks.
Set C1 Set C2
Set D1 Set D2 Set D1 Set D2
Avg Max Avg Max Avg Max Avg Max
Set K1 N = 22 UB1 0.33 0.5 0.42 0.54 0.22 0.28 0.23 0.34
UB2 0.09 0.1 0.11 0.14 0.07 0.09 0.07 0.08
UB3 0.05 0.09 0.06 0.08 0.05 0.06 0.06 0.08
MILP 3.83 10.67 10.65 73.21 1.76 2 1.77 2.15
Set K2 N = 22 UB1 0.7 1.25 0.84 1.09 0.47 0.61 0.59 0.73
UB2 0.41 0.67 0.4 0.65 0.23 0.35 0.25 0.36
UB3 0.07 0.1 0.09 0.11 0.06 0.07 0.08 0.1
MILP 1078.32 3599.85 796.07 3600.62 363.9 3599.87 365.29 3600.3
Set K1 N = 34 UB1 0.57 0.86 0.78 0.92 0.36 0.44 0.43 0.53
UB2 0.11 0.12 0.11 0.13 0.09 0.11 0.08 0.11
UB3 0.06 0.08 0.07 0.1 0.06 0.08 0.06 0.07
MILP 4.37 6.04 5.93 8.15 3.22 4.32 3.41 4.44
Set K2 N = 34 UB1 1.27 1.95 1.67 2.21 0.91 1.14 1.18 1.34
UB2 0.44 0.53 0.46 0.58 0.32 0.39 0.31 0.38
UB3 0.08 0.1 0.11 0.12 0.08 0.09 0.09 0.11
MILP 1454.94 3600.34 939.75 3417.72 373.65 3599.81 19.76 27.86
Set K1 N = 47 UB1 1.05 1.36 1.42 1.63 0.48 0.69 0.65 0.87
UB2 0.16 0.18 0.14 0.16 0.11 0.12 0.12 0.13
UB3 0.07 0.11 0.08 0.12 0.06 0.07 0.07 0.09
MILP 22.77 118.85 14.38 64.45 5.65 6.88 5.62 7.02
Set K2 N = 47 UB1 2.46 3.59 2.94 3.51 1.52 1.9 2.17 2.86
UB2 0.82 1.17 0.79 1.05 0.55 0.83 0.56 0.84
UB3 0.09 0.11 0.16 0.2 0.09 0.1 0.12 0.17
MILP 784.49 3600.0 759.57 3600.0 376.38 3600.0 385.74 3600.0
Table 6
The CPU times (in seconds) of the Upper Bounds - large networks.
Set C1 Set C2
Set D1 Set D2 Set D1 Set D2
Avg Max Avg Max Avg Max Avg Max
Set K1 N = 60 UB1 4.27 5.03 4.89 5.82 2.47 2.94 3.18 3.66
UB2 1.21 1.69 1.26 1.98 0.72 1.12 0.72 1.21
UB3 0.10 0.12 0.16 0.23 0.08 0.10 0.11 0.15
Set K2 N = 60 UB1 13.87 18.6 19.59 25.51 6.82 8.67 9.44 11.55
UB2 5.36 6.79 4.97 6.76 2.52 3.13 2.67 3.44
UB3 0.17 0.21 0.38 0.54 0.15 0.18 0.24 0.33
Set K1 N = 73 UB1 6.27 8.8 7.59 9.42 3.85 4.37 4.52 5.21
UB2 1.57 2.23 1.49 2.51 0.79 1.13 0.83 1.14
UB3 0.11 0.14 0.17 0.22 0.11 0.15 0.13 0.16
Set K2 N = 73 UB1 19.99 23.34 32.14 42.58 9.35 10.88 13.59 17.4
UB2 5.83 8.80 6.92 10.83 3.63 5.06 3.31 4.16
UB3 0.21 0.31 0.56 0.99 0.15 0.17 0.29 0.45
05
101520253035
C1 D1K1
C2 D1K1
C1 D2K1
C2 D2K1
C1 D1K2
C2 D1K2
C1 D2K2
C2 D2K2
N = 60 73
Fig. 6. The average CPU times (in seconds) of UB1.
i
M
fi
b
t
s
s
a
t
n
ncrease at a linear rate. On the contrary, the solution times of the
ILP model increase at an exponential rate and the model cannot
nd any solution in one hour, for large networks. Hence our upper
ounds could be used in place of the optimal solutions to evaluate
he performances of the lower bounds that are used to find feasible
olutions. t
Please cite this article as: E.G. Kalaycılar et al., A disassembly line balancin
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004
We finally investigate the performances of the lower bound. For
mall networks with 22, 34 and 47 tasks, the optimal solutions
re found by the MILP and the deviations are measured relative to
he optimal solutions, i.e., Percent Dev = ( OPT−LBOPT ) × 100. For large
etworks, the optimal solutions are not available and the devia-
ions are measured relative to UB2 in place of optimal solutions, i.e.,
g problem with fixed number of workstations, European Journal of
![Page 12: A Disassembly Line Balancing Problem With Fixed Number of Workstations](https://reader033.vdocuments.site/reader033/viewer/2022051218/5695d4f71a28ab9b02a38076/html5/thumbnails/12.jpg)
12 E.G. Kalaycılar et al. / European Journal of Operational Research 000 (2015) 1–13
ARTICLE IN PRESSJID: EOR [m5G;October 1, 2015;11:19]
Table 7
The Lower Bound performances.
Variables Set C1 Set C2
Set D1 Set D2 Set D1 Set D2
Avg Max Avg Max Avg Max Avg Max
Set K1 N = 22 3,520 Percent DEV 0 1 0 1 0 0 0 0
CPU 0.06 0.13 0.08 0.12 0.07 0.09 0.07 0.08
Set K2 N = 22 7,040 Percent DEV 0 1 0 1 0 1 0 0
CPU 0.1 0.15 0.12 0.16 0.08 0.1 0.1 0.14
Set K1 N = 34 5,440 Percent DEV 0 0 1 2 0 0 0 0
CPU 0.05 0.08 0.08 0.08 0.09 0.11 0.08 0.11
Set K2 N = 34 10,880 Percent DEV 1 2 1 2 1 2 1 1
CPU 0.03 0.04 0.13 0.15 0.1 0.12 0.11 0.14
Set K1 N = 47 7,520 Percent DEV 0 1 0 1 0 0 0 0
CPU 0.09 0.15 0.1 0.14 0.11 0.12 0.12 0.13
Set K2 N = 47 15,040 Percent DEV 1 1 1 1 1 1 0 1
CPU 0.12 0.14 0.19 0.23 0.12 0.14 0.15 0.21
Set K1 N = 60 19,200 Percent GAP 0 1 0.01 1 0 1 0 1
CPU 0.12 0.14 0.18 0.25 0.1 0.13 0.13 0.18
Set K2 N = 60 38,400 Percent GAP 0 1 1 2 1 1 1 1
CPU 0.2 0.24 0.41 0.59 0.18 0.23 0.28 0.38
Set K1 N = 73 23,360 Percent GAP 0 0 0 1 0 1 1 1
CPU 0.14 0.17 0.18 0.25 0.14 0.2 0.16 0.21
Set K2 N = 73 46,720 Percent GAP 1 1 1 2 1 1 1 2
CPU 0.26 0.34 0.61 1.04 0.19 0.22 0.32 0.49
0
20
40
60
80
100
22 34 47
Average
Maximum
Fig. 7. The optimal total net revenue values.
l
v
e
t
a
b
T
o
m
u
7
i
f
m
w
i
c
r
r
a
m
i
b
L
n
t
l
s
w
fi
r
r
t
fi
t
Percent Gap = ( UB2−LBUB2
) × 100. The average and maximum deviations
of the lower bounds are reported in Table 7.
Note from the table that the lower bound solutions are quite satis-
factory over all problem sets. The solutions have small deviations and
are obtained in negligible times. The performance does not deterio-
rate with an increase in N. However the deviations slightly increase
with an increase in K. For example for N = 60 and sets C1 and D2, the
average deviations of K1 and K2 are 1 percent, whereas the maximum
deviations are 1 percent and 2 percent for K1 and K2, respectively.
The CPU times to find the lower bounds increase with an increase in
N and K with only a few exceptions. This is due to the fact that the
linear and mixed-integer linear programs have more decision vari-
ables for higher N and K. For example, for sets K1, C2 and D2, as N
increases from 60 to 73, the average CPU times increase from 0.13 to
0.16 seconds. For N = 47 and sets C2 and D2, increasing K from 2 to 4,
increases the average CPU times from 0.12 to 0.15 seconds.
We could not observe any notable effect of C on the lower bound
performances and CPU times. For fixed N, K and C, increasing D, does
not change the performance, but increases the CPU times slightly.
This increase can be attributed to the increase in LB(Y). For example
for N = 60 and sets C1 and K2, the average CPU times are 0.20 seconds
and 0.41 seconds for set D1 and set D2, respectively.
To summarize, our computational experiment has revealed the
excellent performances of the bounds that do not deteriorate as the
number of tasks increases. On the other hand, the optimal total rev-
enue values increase significantly with the increases in the number of
the tasks. Fig. 7 shows the optimal total net revenue values for each
n, over all other parameter combinations, for small networks. For the
Please cite this article as: E.G. Kalaycılar et al., A disassembly line balancin
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004
arge networks, the optimal solutions, hence their total net revenue
alues are not available.
Note from the above figure that the average optimal total net rev-
nue values are around 30, 45 and 63 money units for 22, 34 and 47
asks, respectively. The respective maximum total net revenue values
re 37, 54 and 78 money units and the respective maximum lower
ound deviations are around 1 percent, 2 percent and 1 percent.
hese altogether imply that the practitioners that use the results of
ur study would lose no more than 1000 money units over the opti-
al total net revenue that are of magnitude of around 50.000 money
nits.
. Conclusions
This study considers disassembly systems that have gained signif-
cant importance in recent years. This heightened importance stems
rom the recognition of environmental issues and the advances in
anufacturing technologies. A disassembly line balancing problem,
here the disassembly line is configured with defined workstations,
s studied. The units of the product to be disassembled are identi-
al and the parts have defined minimum release quantity, cost and
evenue. It is assumed that the amount released excess of minimum
elease quantity can be sold at the end of the period. The aim is to
ssign the tasks to the workstations so that the total net revenue is
aximized.
A MILP model that could solve the problems with up to 50 tasks,
s developed. For larger sized instances, we propose upper and lower
ounds that are motivated by the satisfactory behavior of the LPR. The
PR is strengthened by same valid cuts.
The experimental results have revealed that the bounding mecha-
isms produce high quality solutions very quickly. For the maximum
rial size of 73 tasks, there is a gap of less than 5 percent between our
ower and upper bounds.
To the best of our knowledge, the study is the first attempt to
olve the DLBP with minimum release quantities and fixed number of
orkstations. The proposed model assumes that all units of a speci-
ed part bring the same revenue. The presented methods can be di-
ectly applied when the unit revenues of the first level (minimum
elease quantity) and second level (excess of minimum release quan-
ity) are different. Note that as the minimum release quantities are
xed, the contribution of the first level’s total revenue to the objec-
ive function is fixed, hence irrelevant for the model solutions.
g problem with fixed number of workstations, European Journal of
![Page 13: A Disassembly Line Balancing Problem With Fixed Number of Workstations](https://reader033.vdocuments.site/reader033/viewer/2022051218/5695d4f71a28ab9b02a38076/html5/thumbnails/13.jpg)
E.G. Kalaycılar et al. / European Journal of Operational Research 000 (2015) 1–13 13
ARTICLE IN PRESSJID: EOR [m5G;October 1, 2015;11:19]
i
s
u
a
b
R
A
A
B
B
B
B
B
B
B
C
C
D
G
G
G
H
K
L
L
L
M
M
M
N
P
P
U
We hope that the results of our study will trigger some advances
n the disassembly literature. Some extensions of the study may con-
ider the Successor OR type precedence relations, non-identical prod-
cts (different units of the disassembly product with different parts)
nd a stochastic nature of the disassembly outcome (some parts may
e found to be defective or damaged during disassembly).
eferences
ltekin, F. T., & Akkan, C. (2012). Task-failure-driven rebalancing of disassembly lines.International Journal of Production Research, 50, 4955–4976.
ltekin, F. T., Kandiller, L., & Özdemirel, N. E. (2008). Profit oriented disassembly line
balancing. International Journal of Production Research, 46, 2675–2693.entaha, M. L., Battaïa, O., & Dolgui, A. (2014a). A sample average approximation
method for disassembly line balancing problem under uncertainty. Computers andOperations Research, 51, 111–122.
entaha, M. L., Battaïa, O., & Dolgui, A. (2014b). Disassembly line balancing and se-quencing under uncertainty. Procedia CIRP, 15, 56–60.
entaha, M. L., Battaïa, O., & Dolgui, A. (2014c). Lagrangian relaxation for stochasticdisassembly line balancing problem. Procedia CIRP, 17, 239–244.
entaha, M. L., Battaïa, O., & Dolgui, A. (2014d). “Disassembly line balancing problem
with fixed number of workstations under uncertainty. In Proceedings of the 19thIFAC World Congress.
entaha, M. L., Battaïa, O., & Dolgui, A. (2015). An exact solution approach for disas-sembly line balancing problem under uncertainty of the task processing times.
International Journal of Production Research, 53, 1807–1818.ourjault, A. (1984). Contribution to a systematic approach of automatic generation of
task sequences. In PhD Dissertation, Universite de Franche-Comte, Besancon, France.
rennan, L., Gupta, S. M., & Taleb, K. N. (1994). Operations planning issues in an as-sembly/disassembly environment. International Journal of Operations and Produc-
tion Management, 14, 57–67.apraz, O., Polat, O., & Güngör, A. (2015). Planning of waste electrical and electronic
equipment (WEEE) recycling facilities: MILP modelling and case study investiga-tion. Flexible Services and Manufacturing Journal. doi:10.1007/s10696-015-9217-3.
rowther, P. (1999). Designing for disassembly to extend service life and increase sus-
tainability. In 8th International Conference on Durability of Building Materials andComponents.
Please cite this article as: E.G. Kalaycılar et al., A disassembly line balancin
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004
ing, L. P., Feng, Y. X., Tan, J. R., & Gao, Y. C. (2010). A new multi objective and colony al-gorithm for solving the disassembly line balancing problem. The International Jour-
nal of Advanced Manufacturing Technology, 48, 761–771.üngör, A., & Gupta, S. M. (2001a). Disassembly sequence plan generation using a
branch and bound algorithm. International Journal of Production Research, 39, 481–509.
üngör, A., & Gupta, S. M. (2001b). A solution approach to the disassembly line bal-ancing problem in the presence of task failures. International Journal of Production
Research, 39, 1427–1467.
üngör, A., & Gupta, S. M. (2002). Disassembly line in product recovery. InternationalJournal of Production Research, 40, 2569–2589.
ezer, S., & Kara, Y. (2014). A network-based shortest route model for parallel disassem-bly line balancing problem. International Journal of Production Research, 53, 1849–
1865.oc, A., Sabuncuoglu, I., & Erel, E. (2009). Two exact formulations for disassembly
line balancing problems with precedence diagram construction using an AND/OR
graph. IIE Transactions, 41, 866–881.ambert, A. J. D. (1997). Optimal disassembly of complex products. International Journal
of Production Research, 35, 2509–2523.ambert, A. J. D. (2002). Determining optimum disassembly sequences in electronic
equipment. Computers and Industrial Engineering, 42, 553–575.ambert, A. J. D. (2003). Disassembly sequencing: A survey. International Journal of Pro-
duction Research, 41, 3721–3759.
cGovern, S. M., & Gupta, S. (2011). The disassembly line: Balancing and modeling. NewYork: McGraw Hill, New York.
cGovern, S. M., & Gupta, S. M. (2007a). A combinatorial optimization analysis of theunary NP-complete disassembly line balancing. International Journal of Production
Research, 45, 1485–1551.cGovern, S. M., & Gupta, S. M. (2007b). A balancing method and genetic algorithm for
disassembly line balancing. European Journal of Operational Research, 179, 692–708.
avin-Chandra, D. (1994). The recovery problem in product design. Journal of Engineer-ing Design, 5, 65–86.
aksoy, T., Güngör, A., Özceylan, E., & Hancilar, A. (2013). Mixed model disassembly linebalancing problem with fuzzy goals. International Journal of Production Research, 51,
6082–6096.atterson, J. H., & Albracht, J. J. (1975). Assembly line balancing: 0–1 programming with
fibonacci search. Operations Research, 179, 166–174.
llerich, C. (2014). Advanced disassembly planning: Flexible, price-quantity dependentand multi-period planning approaches. Wiesbaden: Springer Gabler.
g problem with fixed number of workstations, European Journal of