Documente Academic
Documente Profesional
Documente Cultură
JID: EOR
[m5G;October 1, 2015;11:19]
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 xed number of workstations is considered. The
product to be disassembled comprises various components, which are referred to as its parts. There is a specied nite supply of the product to be disassembled and specied 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 nite number of identical workstations that perform the necessary 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. Introduction
The environmental regulations, customer awareness and recent
advances in technology all together have shifted the product recovery
process from the act of disposing to the act of remanufacturing and
recycling. Recycling preserves the material content of the discarded
(used) products via some manufacturing and disassembly operations.
Remanufacturing, on the other hand, keeps the functional content of
the used products and improves their quality up to a desired usable
level via disassembly operations and some manufacturing.
Disassembly is the rst important step of product recovery activities (McGovern and Gupta, 2011) and it is methodical extraction of
valuable parts, operations involve the separation of the reusable parts
from the discarded products. The parts are either subject to some remanufacturing process or sold to suppliers. Disassembly operations
are usually performed on a disassembly line that consists of a number
of serial workstations. The rst workstation takes the product to be
disassembled. The cycle terminates, i.e., the product leaves the line,
whenever all its required parts are disassembled.
The Disassembly Line Balancing Problem (DLBP) assigns the set
of tasks to each workstation for each product to be disassembled.
The problem is critical in minimizing the use of valuable resources
(such as time and money) invested in disassembly, and maximizing
the level of automation of the disassembly process and the quality of
the parts or materials recovered.
http://dx.doi.org/10.1016/j.ejor.2015.09.004
0377-2217/ 2015 Published by Elsevier B.V.
Please cite this article as: E.G. Kalayclar et al., A disassembly line balancing problem with xed number of workstations, European Journal of
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004
JID: EOR
2
ARTICLE IN PRESS
[m5G;October 1, 2015;11:19]
E.G. Kalayclar et al. / European Journal of Operational Research 000 (2015) 113
associated to the demands from the third party like the producers of
the new models. Moreover, as we assume, the amount released excess 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
denes our problem. In Section 4 the mathematical models and
their use in nding 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 computational experiment is discussed in Section 6 and the study is concluded
in Section 7.
2. Disassembly process & the related literature
Gngr and Gupta (2001b) dened disassembly as a systematic
process of separating a product into its constituent parts, components, subassemblies or other groupings. The issues in the area of
disassembly can be classied into two broad categories as design and
operational. Crowther (1999) considered a design for disassembly issue and mentioned that a life cycle model that incorporates the stages
of a disassembly strategy can highlight the environmental advantages 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 guidelines 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, operational problems arise. The operational problems that are likely to
arise are layout, resource allocation, process sequencing and disassembly line balancing. As mentioned by Brennan et al. (1994), the operational 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 disassembly sequencing problem and the DLBP. The disassembly line sequencing problem decides on the disassembly process sequence of a
specied disassembly product and the DLBP decides on the assignment 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 modied traveling salesman problem
so as to minimize total costs while meeting obligatory reclamation of
denite 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 methods were applicable to the disassembly of complex consumer products like automotive vehicles, consumer electronics and mechanical assemblies and were applied to a headlamp and ballpoint pen
disassembly products. Lambert (2002) proposed a linear programming 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 rst introduced in Gngr and Gupta (2001b).
Gngr and Gupta (2001a) mentioned several complications like
Please cite this article as: E.G. Kalayclar et al., A disassembly line balancing problem with xed number of workstations, European Journal of
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004
JID: EOR
ARTICLE IN PRESS
[m5G;October 1, 2015;11:19]
E.G. Kalayclar et al. / European Journal of Operational Research 000 (2015) 113
sense that we take the number of workstations and cycle time as parameters whereas they decide on the number of workstations and
on the cycle time. We assume that there is a nite supply of disassembly product and that there is a lower bound on the quantity of
the part releases whereas they assume innite supply and take minimum release quantities as upper bounds on the quantity of the part
releases. Hence their solution considers a single line balance for all
units, whereas, to satisfy the minimum release quantities, we use different task assignments to disassemble different units.
c
Fig. 1. AND precedence relations.
3. Problem denition
f
We consider a disassembly line balancing model with a nite supply of a single product. There are N tasks where task i is specied
by its cost ci and processing time ti . The tasks that release parts are
called part releasing tasks. The minimum release quantity of each
part-releasing task i is di units. di is the associated demand for part
i from the third party like the producers of the new product models. The part-releasing task i has a revenue ri that includes its market value (for example when the part is sold to the third party) and
recycled material value. The parts may have negative revenues that
represent the disposal costs. Ri is the net revenue of task i, Ri = ri ci .
At the end of the period, each unit of part i that is released excess of
di units can be held in the inventory for future orders and generates a
unit revenue ri when it is sold. Such excess releases would be favored
if they lead to increases in the total net revenue. Our problem is to
assign the tasks to the workstations of the disassembly line for each
unit of the product so as to maximize the total net revenue. We make
the following additional assumptions.
h
Fig. 2. POR precedence relations.
Xiks =
0
Constraint set (1) deals with the minimum release quantities. The
amount of each part released should be higher than its minimum release quantity.
S
K
Xiks di
i = 1, . . . , N
(1)
k=1 s=1
Constraint set (2) states the cycle time limit should not be exceeded at any workstation for any unit.
N
ti Xiks C
k = 1, . . . , K; s = 1, . . . , S
(2)
i=1
Constraint sets (3) and (4) ensure that the precedence requirements are observed. The AND precedence relations should be satised. Thus task i cannot be assigned to a workstation that is before
the workstation of its AND predecessor task l.
K
kXlks
k=1
K
kXiks
i = 1, . . . , N, l PAND(i), s = 1, . . . , S
k=1
(3)
Similarly, the OR precedence relations should be satised. Thus
task i can be assigned to workstation k if and only if at least one of its
OR predecessors is assigned to workstations 1 through k.
Xiks
k
X jhs
i = 1, . . . , N; k = 1, . . . , K; s = 1, . . . , S
h=1 jPOR(i)
(4)
Please cite this article as: E.G. Kalayclar et al., A disassembly line balancing problem with xed number of workstations, European Journal of
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004
ARTICLE IN PRESS
JID: EOR
4
[m5G;October 1, 2015;11:19]
E.G. Kalayclar et al. / European Journal of Operational Research 000 (2015) 113
K
N
K
i=1 k=1
Xiks 1
i = 1, . . . , N; s = 1, . . . , S
(5)
k=1
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 dened by objective
function (7) below, together with constraint sets (1) through (6).
Max
S
K
N
Ri Xiks
(7)
Ys =
0
s = 1, . . . , S
(8)
Constraint Set (9) assigns Ys to 1, if there is at least once task assigned to any workstation while disassembling sth unit of the prod K
uct. Note that N
k=1 Xiks is bounded from above by . We use
i=1
K
= N K. If N
i=1
k=1 Xiks 1 then Ys = 1 otherwise Ys is zero as we are
minimizing Y = Ss=1 Ys . Y = Ss=1 Ys gives the total number of units,
with at least one task assignment. Hence it is the minimum number
of disassembly units to satisfy all minimum release quantities.
The objective function of Model II is expressed as below.
Min
S
Ys P
s=1
Xiks Ys
S
K
N
Ri Xiks
(9)
The above objective primarily minimizes the number of disassembly units to satisfy all minimum release quantities. Among the solutions that attain the minimum number of disassembly units Y, the
model selects the one with the maximum net revenue. To maximize
the total net revenue while keeping the minimum number of disassembly units, some parts are released more than their minimum release requirements. The remaining SY units are handled by Model III
(Section 4.2).
The magnitude of p used in (9) is important in the sense that it
should not increase Y = Ss=1 Ys en route to increasing total net revenue. p should be set small enough to satisfy constraint (10).
S
Ys P Pmin
s=1
S
Ys + 1 P Pmax
(10)
s=1
P =
1
N
min
i=1 Ri di
Xik =
1
0
Max
K
N
Ri Xik
(11)
i=1 k=1
Please cite this article as: E.G. Kalayclar et al., A disassembly line balancing problem with xed number of workstations, European Journal of
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004
JID: EOR
ARTICLE IN PRESS
[m5G;October 1, 2015;11:19]
E.G. Kalayclar et al. / European Journal of Operational Research 000 (2015) 113
Model I is solved to optimality by relaxing the integrality constraints on the Xiks values. The resulting optimal objective function
value, ZLP , provides an upper bound on the optimal total net revenue.
The only difference between the original model and its LPR is the integrality requirement of the Xiks values. Consequently, ZLP is an upper
bound on the optimal objective function value Z .
Upper Bound 2 (UB2 ): UB2 is found by adding valid cuts to the LPR
of Model I. The valid cuts are found by investigating the properties of
the optimal solutions that may not be satised by the optimal LPR.
Cut1Fixing the minimum release quantities
Theorem 1. There exists an optimal solution in which di units of part i
are released from the rst di disassembly units.
Proof. Assume an optimal solution in which the rst di units are released from di units of the product. If the tasks in any two units, say
a and b, are changed, i.e., Xika is set to Xikb and Xikb is set to Xika for
all i , then the optimal net revenue is not affected. Assume an optimal
solution in which part i is released in units a+1, a+ di . Setting Xjkr =
Xjk(a+r) for all j, r dj and Xjk(ta+r) = Xjkr does not change the revenue
value. Then, there is an optimal solution in which dj units of part i is
released in the rst 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
units of part r are released from the rst dr disassembly units. To increase the power of (12), we select task r such that dr =Maxi {di }.
dr
K
Xrks = dr
s = 1, . . . , dr
(12)
k=1 s=1
ti +Min jPOR(i) {t j }+
C
jAND(i) t j
.
Proof. Task i cannot start before its predecessors and at least one
of its immediate POR predecessors are completed. Hence it should
wait at least Min jPOR(i) {t j } units for its POR predecessors and
jAND(i) t j units forits AND predecessors to start and at least
ti + Min jPOR(i) {t j } + jAND(i) t j units to complete. When task splitting is allowed and other tasks are ignored, ti + Min jPOR(i) {t j } +
ti +Min jPOR(i) {t j }+ jAND(i) t j
workstations.
jAND(i) t j units require
C
When task splitting
is not allowed and other tasks are considered
ti +Min jPOR(i) {t j }+
C
for task i.
jAND(i) t j
In the absence of POR relations, Theorem 2 reduces to the one derived in Patterson and Albracht (1975) for the classical assembly line
balancing problem.
Constraint set (13) below supports Theorem 2.
Ei 1
S
Xiks = 0 i = 1, . . . , n
(13)
k=1 s=1
Constraint set (13) ensures that task i is not assigned to workstations 1, 2, , Ei 1. The network in Fig. 3 is used to illustrate Ei
computations.
Note that { t1 + t2 } is the minimum total processing load to start
task A and Min {t3 , t4 }+ t5 is the minimum total processing load to
start task B. Then Min{(t1 + t2 + tA ), (Min {t3 , t4 } + t5 + tB )} is the
Please cite this article as: E.G. Kalayclar et al., A disassembly line balancing problem with xed number of workstations, European Journal of
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004
ARTICLE IN PRESS
JID: EOR
6
[m5G;October 1, 2015;11:19]
E.G. Kalayclar et al. / European Journal of Operational Research 000 (2015) 113
cannot allow more task assignments. This is due to the fact that all
tasks in workstations r + 1 through K are successors of task i, hence
cannot be processed in workstations 1 through r, since task i is to be
processed later. Then, there exists an optimal solution in which task i
is processed in the rst r workstations.
t1 + t2 = SA
Min {SA+tA , SB+tB } = Sj
3
B
Min {t3 , t4 } + t5 = SB
4
5
Note that Ei =
Hence Ei =
ti + Min{(t1 + t2 + tA ), (Min{t3 , t4 } + t5 + tB )}
.
C
E13 =
E13 =
17 + Min{(13 + 1 + 6 ), (13 + 7 )} + 19 + 15
71
=
=2
63
63
Xiks
k=1 s=1
S
K
k=1 s=!
S
K
X jks
i = 1, . . . , n; j PAND(i).
(14)
k=1 s=1
Xiks
S
K
X jks
i = 1, . . . , n
(15)
K
Xiks = 0
i = 1, . . . , N
(16)
Xiks = 0
i = 1, . . . , N
(17)
k=1
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
the stronger one.
We hereafter refer to the LPR using the cuts (12) through (17)
as strengthened LPR. UB2 is the total net revenue returned by the
strengthened LPR.
Upper Bound 3 (UB3 ): To obtain UB3 , we use the idea used in
Model II. Recall that Model II aims to satisfy all minimum release
quantities ( i di ) and minimizes Y = Ss=1 Ys . Here in place of Y, we
use LB(Y), to get rid of the Ys binary variables.
First, we nd LB(Y), and then increase it one by one, until a feasible
solution is reached. Theorem 5 states LB(Y).
We let dnewi = Max{di , Max j|iAND j {d j }}. Note thatdnewi is the
minimum release quantity of task i which is modied by its minimum release quantity and the minimum release quantities of all its
AND successors.
i
is a valid lower bound on the numTheorem 5. LB(Y ) = i=1KiC
ber of units to satisfy all minimum release quantities.
more than
n K C time units as each workstation is available for C time
t dnewi
is the number of units disassembled to satisfy
units. i=1KiC
n
t dnew
Please cite this article as: E.G. Kalayclar et al., A disassembly line balancing problem with xed number of workstations, European Journal of
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004
ARTICLE IN PRESS
JID: EOR
[m5G;October 1, 2015;11:19]
E.G. Kalayclar et al. / European Journal of Operational Research 000 (2015) 113
16
3
21
4
11
20
12
10
2
8
15
22
9
5
17
18
19
14
13
Fig. 4. Precedence diagram adapted from Bourjaults ball-point pen example.
all minimum release quantity if task splitting between the disassembly unitsand workstations were allowed. As no task splitting
n
t dnewi
is allowed i=1KiC
disassembled.
4
= 3. We rst
Using Theorem 5 we nd LB(Y ) = 13 4+744++19
63
set the number of units to LB(Y) = 3. However a feasible solution
for the LPR is not found. There is no feasible solution with 4 units
as well. The LPR nds a feasible solution with 5 units. Then the minimum number of units to satisfy all minimum release quantities, Y = 5.
Model III is used to nd the line balance for the remaining 75 units.
Formally 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 + (SY) ZB
t dnew
i
and let T = LB(Y)
Step 0. Let LB(Y ) = i=1KiC
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-xed 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.
Please cite this article as: E.G. Kalayclar et al., A disassembly line balancing problem with xed number of workstations, European Journal of
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004
ARTICLE IN PRESS
JID: EOR
8
[m5G;October 1, 2015;11:19]
E.G. Kalayclar et al. / European Journal of Operational Research 000 (2015) 113
Table 1
The data for the Bourjaults ball-point pen problem.
Tasks
10
11
12
13
14
15
16
17
18
19
20
21
22
di (units)
ti (time units)
dnewi (units)
ci (money units)
ri (money units)
Ri (money units)
0
13
4
17
49
32
0
7
4
18
10
-8
0
1
4
14
35
21
4
6
4
7
37
30
0
7
4
14
16
2
4
5
4
8
12
4
0
15
4
8
20
12
3
8
3
15
37
22
4
17
4
16
43
27
4
3
4
7
32
25
4
14
4
15
24
9
4
12
4
12
41
29
4
17
4
6
19
13
3
4
3
17
35
18
2
11
2
17
19
2
1
18
1
6
18
12
0
15
0
9
48
39
4
19
4
13
21
8
2
4
2
12
10
-2
1
18
1
16
40
24
0
16
0
7
16
9
0
19
4
15
11
-4
34
8
6
9
17
2
18
10
32
23
16
31
33
30
1
11
24
13
22
29
3
28
12
14
19
25
21
15
26
20
27
Please cite this article as: E.G. Kalayclar et al., A disassembly line balancing problem with xed number of workstations, European Journal of
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004
ARTICLE IN PRESS
JID: EOR
[m5G;October 1, 2015;11:19]
E.G. Kalayclar et al. / European Journal of Operational Research 000 (2015) 113
Table 2
UB1 performances.
Variables
Set C1
Set C2
Set D1
Set K1 N = 22
3,520
Set K2 N = 22
7,040
Set K1 N = 34
5,440
Set K2 N = 34
10,880
Set K1 N = 47
7,520
Set K2 N = 47
15,040
Set K1 N = 60
19,200
Set K2 N = 60
38,400
Set K1 N = 73
23,360
Set K2 N = 73
46,720
Percent DEV
# of frac.
Percent DEV
# of frac.
Percent DEV
# of frac.
Percent DEV
# of frac.
Percent DEV
# of frac.
Percent DEV
# of frac.
Percent DEV
# of frac.
Percent DEV
# of frac.
Percent DEV
# of frac.
Percent DEV
# of frac.
Set D2
Avg
Max
Avg
Max
Avg
Max
1
722.4
1
2437.5
0
715.9
0
3803.1
0
1184
0
5492.4
-
8749.8
13515.3
10876.3
16730.7
2
1636
3
3643
1
1950
1
4864
1
4011
1
6865
9579
14742
11804
18440
1
1120.7
1
2497.4
0
1150.8
0
4238.7
0
2509.4
0
5910.8
8999.5
13853.1
11249.8
17291.3
2
1778
3
3592
1
2251
1
5136
1
3924
1
6966
9791
15153
11974
18654
1
487.1
1
1628.1
0
786
0
3004.1
0
584.2
0
4041.1
6887.3
11399.3
9137.3
14045.5
2
1356
3
3017
1
2367
1
4307
1
2403
1
5888
8344
13268
10462
15689
1
470.6
2
1900.4
0
741
0
3102.6
0
1152.4
0
4654.9
7751.1
12307.1
9716.3
15019.8
2
1684
3
3306
1
2210
1
4386
1
3391
1
5946
8935
14280
10633
16147
- illustrates that optimal solution for large networks could not be obtained.
ti
Set D2
Max
Set C1. C =
Set D1
Avg
ti
The processing times are generated from a discrete uniform distribution [1, 20]. A task receives a minimum release quantity, hence
called a part-releasing task, according to a random process. We generate a random number between 0 and 1. If the generated number is
below 0.3, we set its minimum release quantity to zero.
Two distributions are used to generate their minimum release
quantities.
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
quantity, and Set D2, with a high minimum release quantity.
The number of units, S, is set to 80. The costs are generated from a
discrete uniform distribution between 5 and 20 and the revenues are
generated from a discrete uniform distribution between 10 and 50.
There are 40 combinations due to 5, 2, 2 and 2 alternatives for
N, K, C and D respectively (3 2 2 2 = 24 for small networks and
2 2 2 2 = 16 for large networks). For each combination, 10 problem
instances are generated. That is, a total of 400 problem instances are
used in our experiments.
The mathematical models are solved by CPLEX 10 and the algorithms are coded in Microsoft Visual C++ 2008. The experiments are
run in Intel Core2 Duo 2.00 gigahertz, 2 gigabyte RAM.
The execution of the MILP (Model I) is terminated if the optimal
solution is not returned in 3600 seconds. The optimal solutions are
found within 3600 seconds when N = 22, 34 and 47. However, optimal
solutions were not found when N = 60 and 73 .
First the performance of our upper bounds is studied. The deviaUBi OPT
) 100 where OPT =
tion of an instance is found as %Dev = ( OPT
the optimal net revenue and UBi = the total net revenue by upper
bound i. The average and maximum deviations and number of fractional variables are reported in Table 2 and Table 3, for UB1 and UB2 ,
respectively.
As can be observed from the tables, the deviations are consistently
low over all problem sets. For UB1 , the average deviations are below
1 percent and almost all maximum deviations are below 3 percent
when N = 22, 34 and 47. Note that the deviations do not deteriorate
with an increase in N. However they deteriorate with an increase in K
due to the ination of the number of binary variables. When N and K
values are xed, the minimum release quantities and cycle times do
not affect the deviations signicantly.
UB2 produces slightly smaller deviations due to the power of
our cuts. For example, the maximum deviation for UB1 is 3 percent,
whereas it is nearly zero for UB2 in sets K1, C1 and D1 for N = 22.
We also give the number of fractional variables in Tables 2, 3 and 4.
The number of the fractional variables produced by the LPR is too
high. For example when N = 22, for sets K2, C2 and D2 (i.e., 4 workstations, high minimum release quantity, and high cycle time cases),
3306 out of 7040 variables are found to be fractional. This value
reduces to 412 when cuts are added. On the average, the cuts reduce the number of fractional variables from 1900.4 to 180.3. The
number of fractional variables is generally affected by the N values. For example, from Tables 24, it can be seen that for sets K2,
C2 and D2, increasing N increases the number of fractional variables of UB2 , on the average. Note that the average number of fractional variables are 1900.4, 3102.6 and 4654.9 for N = 22, 34 and 47
respectively.
It can also be observed from the tables that, for xed N, an increase
in the number of workstations increases the number of the fractional
variables. For example for N = 22 and sets C1 and D1, the average
number of fractional variables is 722.4 when K = 2, and 2437.5 when
K = 4. This is due to the fact that more workstations lead to higher
task splits, leading to more fractional variables. For xed N and K,
we observe that increasing C reduces the number of fractional variables (with few exceptions). For example, when N = 47, K = 2, and
D1 combination, the average number of fractional variables are 1184
and 584.2 for low and high C, respectively. This is due to the fact that
increasing C gives more room for complete task assignments, hence
reducing the number of fractional variables. In our experiments, we
Please cite this article as: E.G. Kalayclar et al., A disassembly line balancing problem with xed number of workstations, European Journal of
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004
ARTICLE IN PRESS
JID: EOR
10
[m5G;October 1, 2015;11:19]
E.G. Kalayclar et al. / European Journal of Operational Research 000 (2015) 113
Table 3
UB2 performances.
Variables
Set C1
Set C2
Set D1
Set K1 N = 22
3,520
Set K2 N = 22
7,040
Set K1 N = 34
5,440
Set K2 N = 34
10,880
Set K1 N = 47
7,520
Set K2 N = 47
15,040
Set K1 N = 60
19,200
Set K2 N = 60
38,400
Set K1 N = 73
23,360
Set K2 N = 73
46,720
Percent DEV
# of frac.
Percent DEV
# of frac.
Percent DEV
# of frac.
Percent DEV
# of frac.
Percent DEV
# of frac.
Percent DEV
# of frac.
Percent DEV
# of frac.
Percent DEV
# of frac.
Percent DEV
# of frac.
Percent DEV
# of frac.
Set D2
Max
Avg
Max
Avg
Max
Avg
Max
0
65.3
0
368.4
0
19
0
525.2
0
49
0
570.3
-
746
1778.7
657.5
1789.1
0
160
0
570
0
160
0
870
0
160
0
970
1524
2203
922
2326
0
67
0
376
0
22.9
0
517.7
0
50
0
592
765.7
1725.2
649.9
1769
0
160
0.01
596
0
158
0
855
0
160
0
982
1546
2116
859
2244
0
0
0
197.8
0
0
0
271.1
0
0
0
370.4
418.3
1192.1
460.8
1197.1
0
0
0
432
0
0
0
438
0
0
0
800
1068
1523
1027
1931
0
0
0
180.3
0
0
0
259.1
0
0
0
362.2
424.8
1179.2
462.4
1201.8
0
0
0
412
0
0
0
446
0
0
0
815
1068
1434
1045
2008
- illustrates that optimal solution for large networks could not be obtained.
C1
C2
D1
K1
K2
22
34
47
22
34
47
Set D2
Avg
Table 4
The number of units with different line balances (out of 80).
N
Set D1
D2
D1
D2
Avg
Max
Avg
Max
Avg
Max
Avg
Max
5.30
5.10
5.50
5.10
5.00
5.30
6.00
6.00
6.00
4.00
5.00
6.00
11.00
10.00
11.30
11.80
10.90
11.10
12.00
10.00
12.00
13.00
11.00
12.00
5.50
5.90
5.80
5.30
5.10
5.40
6.00
6.00
6.00
6.00
6.00
6.00
11.20
10.00
11.20
10.70
10.00
11.00
12.00
10.00
12.00
11.00
10.00
12.00
Please cite this article as: E.G. Kalayclar et al., A disassembly line balancing problem with xed number of workstations, European Journal of
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004
ARTICLE IN PRESS
JID: EOR
[m5G;October 1, 2015;11:19]
E.G. Kalayclar et al. / European Journal of Operational Research 000 (2015) 113
11
Table 5
The CPU times (in second) of the Upper Bounds and MILPsmall networks.
Set C1
Set C2
Set D1
Set K1 N = 22
UB1
UB2
UB3
MILP
UB1
UB2
UB3
MILP
UB1
UB2
UB3
MILP
UB1
UB2
UB3
MILP
UB1
UB2
UB3
MILP
UB1
UB2
UB3
MILP
Set K2 N = 22
Set K1 N = 34
Set K2 N = 34
Set K1 N = 47
Set K2 N = 47
Set D2
Set D1
Set D2
Avg
Max
Avg
Max
Avg
Max
Avg
Max
0.33
0.09
0.05
3.83
0.7
0.41
0.07
1078.32
0.57
0.11
0.06
4.37
1.27
0.44
0.08
1454.94
1.05
0.16
0.07
22.77
2.46
0.82
0.09
784.49
0.5
0.1
0.09
10.67
1.25
0.67
0.1
3599.85
0.86
0.12
0.08
6.04
1.95
0.53
0.1
3600.34
1.36
0.18
0.11
118.85
3.59
1.17
0.11
3600.0
0.42
0.11
0.06
10.65
0.84
0.4
0.09
796.07
0.78
0.11
0.07
5.93
1.67
0.46
0.11
939.75
1.42
0.14
0.08
14.38
2.94
0.79
0.16
759.57
0.54
0.14
0.08
73.21
1.09
0.65
0.11
3600.62
0.92
0.13
0.1
8.15
2.21
0.58
0.12
3417.72
1.63
0.16
0.12
64.45
3.51
1.05
0.2
3600.0
0.22
0.07
0.05
1.76
0.47
0.23
0.06
363.9
0.36
0.09
0.06
3.22
0.91
0.32
0.08
373.65
0.48
0.11
0.06
5.65
1.52
0.55
0.09
376.38
0.28
0.09
0.06
2
0.61
0.35
0.07
3599.87
0.44
0.11
0.08
4.32
1.14
0.39
0.09
3599.81
0.69
0.12
0.07
6.88
1.9
0.83
0.1
3600.0
0.23
0.07
0.06
1.77
0.59
0.25
0.08
365.29
0.43
0.08
0.06
3.41
1.18
0.31
0.09
19.76
0.65
0.12
0.07
5.62
2.17
0.56
0.12
385.74
0.34
0.08
0.08
2.15
0.73
0.36
0.1
3600.3
0.53
0.11
0.07
4.44
1.34
0.38
0.11
27.86
0.87
0.13
0.09
7.02
2.86
0.84
0.17
3600.0
Table 6
The CPU times (in seconds) of the Upper Bounds - large networks.
Set C1
Set C2
Set D1
Set K1 N = 60
Set K2 N = 60
Set K1 N = 73
Set K2 N = 73
UB1
UB2
UB3
UB1
UB2
UB3
UB1
UB2
UB3
UB1
UB2
UB3
Set D2
Set D1
Set D2
Avg
Max
Avg
Max
Avg
Max
Avg
Max
4.27
1.21
0.10
13.87
5.36
0.17
6.27
1.57
0.11
19.99
5.83
0.21
5.03
1.69
0.12
18.6
6.79
0.21
8.8
2.23
0.14
23.34
8.80
0.31
4.89
1.26
0.16
19.59
4.97
0.38
7.59
1.49
0.17
32.14
6.92
0.56
5.82
1.98
0.23
25.51
6.76
0.54
9.42
2.51
0.22
42.58
10.83
0.99
2.47
0.72
0.08
6.82
2.52
0.15
3.85
0.79
0.11
9.35
3.63
0.15
2.94
1.12
0.10
8.67
3.13
0.18
4.37
1.13
0.15
10.88
5.06
0.17
3.18
0.72
0.11
9.44
2.67
0.24
4.52
0.83
0.13
13.59
3.31
0.29
3.66
1.21
0.15
11.55
3.44
0.33
5.21
1.14
0.16
17.4
4.16
0.45
N = 60
73
35
30
25
20
15
10
5
0
C1 D1
K1
C2 D1
K1
C1 D2
K1
C2 D2
K1
C1 D1
K2
C2 D1
K2
C1 D2
K2
C2 D2
K2
Please cite this article as: E.G. Kalayclar et al., A disassembly line balancing problem with xed number of workstations, European Journal of
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004
ARTICLE IN PRESS
JID: EOR
12
[m5G;October 1, 2015;11:19]
E.G. Kalayclar et al. / European Journal of Operational Research 000 (2015) 113
Table 7
The Lower Bound performances.
Variables
Set C1
Set C2
Set D1
Set K1 N = 22
3,520
Set K2 N = 22
7,040
Set K1 N = 34
5,440
Set K2 N = 34
10,880
Set K1 N = 47
7,520
Set K2 N = 47
15,040
Set K1 N = 60
19,200
Set K2 N = 60
38,400
Set K1 N = 73
23,360
Set K2 N = 73
46,720
Percent DEV
CPU
Percent DEV
CPU
Percent DEV
CPU
Percent DEV
CPU
Percent DEV
CPU
Percent DEV
CPU
Percent GAP
CPU
Percent GAP
CPU
Percent GAP
CPU
Percent GAP
CPU
Set D2
Avg
Max
Avg
Max
Avg
Max
0
0.06
0
0.1
0
0.05
1
0.03
0
0.09
1
0.12
0
0.12
0
0.2
0
0.14
1
0.26
1
0.13
1
0.15
0
0.08
2
0.04
1
0.15
1
0.14
1
0.14
1
0.24
0
0.17
1
0.34
0
0.08
0
0.12
1
0.08
1
0.13
0
0.1
1
0.19
0.01
0.18
1
0.41
0
0.18
1
0.61
1
0.12
1
0.16
2
0.08
2
0.15
1
0.14
1
0.23
1
0.25
2
0.59
1
0.25
2
1.04
0
0.07
0
0.08
0
0.09
1
0.1
0
0.11
1
0.12
0
0.1
1
0.18
0
0.14
1
0.19
0
0.09
1
0.1
0
0.11
2
0.12
0
0.12
1
0.14
1
0.13
1
0.23
1
0.2
1
0.22
0
0.07
0
0.1
0
0.08
1
0.11
0
0.12
0
0.15
0
0.13
1
0.28
1
0.16
1
0.32
0
0.08
0
0.14
0
0.11
1
0.14
0
0.13
1
0.21
1
0.18
1
0.38
1
0.21
2
0.49
80
Average
40
Maximum
20
0
22
34
Set D2
Max
100
60
Set D1
Avg
47
large networks, the optimal solutions, hence their total net revenue
values are not available.
Note from the above gure that the average optimal total net revenue values are around 30, 45 and 63 money units for 22, 34 and 47
tasks, respectively. The respective maximum total net revenue values
are 37, 54 and 78 money units and the respective maximum lower
bound deviations are around 1 percent, 2 percent and 1 percent.
These altogether imply that the practitioners that use the results of
our study would lose no more than 1000 money units over the optimal total net revenue that are of magnitude of around 50.000 money
units.
7. Conclusions
UB LB
2
Percent Gap = ( UB
) 100. The average and maximum deviations
2
of the lower bounds are reported in Table 7.
Note from the table that the lower bound solutions are quite satisfactory over all problem sets. The solutions have small deviations and
are obtained in negligible times. The performance does not deteriorate 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 nd 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 variables 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 xed 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 revenue values increase signicantly 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
This study considers disassembly systems that have gained significant importance in recent years. This heightened importance stems
from the recognition of environmental issues and the advances in
manufacturing technologies. A disassembly line balancing problem,
where the disassembly line is congured with dened workstations,
is studied. The units of the product to be disassembled are identical and the parts have dened minimum release quantity, cost and
revenue. It is assumed that the amount released excess of minimum
release quantity can be sold at the end of the period. The aim is to
assign the tasks to the workstations so that the total net revenue is
maximized.
A MILP model that could solve the problems with up to 50 tasks,
is developed. For larger sized instances, we propose upper and lower
bounds that are motivated by the satisfactory behavior of the LPR. The
LPR is strengthened by same valid cuts.
The experimental results have revealed that the bounding mechanisms produce high quality solutions very quickly. For the maximum
trial size of 73 tasks, there is a gap of less than 5 percent between our
lower and upper bounds.
To the best of our knowledge, the study is the rst attempt to
solve the DLBP with minimum release quantities and xed number of
workstations. The proposed model assumes that all units of a specied part bring the same revenue. The presented methods can be directly applied when the unit revenues of the rst level (minimum
release quantity) and second level (excess of minimum release quantity) are different. Note that as the minimum release quantities are
xed, the contribution of the rst levels total revenue to the objective function is xed, hence irrelevant for the model solutions.
Please cite this article as: E.G. Kalayclar et al., A disassembly line balancing problem with xed number of workstations, European Journal of
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004
JID: EOR
ARTICLE IN PRESS
E.G. Kalayclar et al. / European Journal of Operational Research 000 (2015) 113
We hope that the results of our study will trigger some advances
in the disassembly literature. Some extensions of the study may consider the Successor OR type precedence relations, non-identical products (different units of the disassembly product with different parts)
and a stochastic nature of the disassembly outcome (some parts may
be found to be defective or damaged during disassembly).
References
Altekin, F. T., & Akkan, C. (2012). Task-failure-driven rebalancing of disassembly lines.
International Journal of Production Research, 50, 49554976.
Altekin, F. T., Kandiller, L., & zdemirel, N. E. (2008). Prot oriented disassembly line
balancing. International Journal of Production Research, 46, 26752693.
Bentaha, M. L., Battaa, O., & Dolgui, A. (2014a). A sample average approximation
method for disassembly line balancing problem under uncertainty. Computers and
Operations Research, 51, 111122.
Bentaha, M. L., Battaa, O., & Dolgui, A. (2014b). Disassembly line balancing and sequencing under uncertainty. Procedia CIRP, 15, 5660.
Bentaha, M. L., Battaa, O., & Dolgui, A. (2014c). Lagrangian relaxation for stochastic
disassembly line balancing problem. Procedia CIRP, 17, 239244.
Bentaha, M. L., Battaa, O., & Dolgui, A. (2014d). Disassembly line balancing problem
with xed number of workstations under uncertainty. In Proceedings of the 19th
IFAC World Congress.
Bentaha, M. L., Battaa, O., & Dolgui, A. (2015). An exact solution approach for disassembly line balancing problem under uncertainty of the task processing times.
International Journal of Production Research, 53, 18071818.
Bourjault, A. (1984). Contribution to a systematic approach of automatic generation of
task sequences. In PhD Dissertation, Universite de Franche-Comte, Besancon, France.
Brennan, L., Gupta, S. M., & Taleb, K. N. (1994). Operations planning issues in an assembly/disassembly environment. International Journal of Operations and Production Management, 14, 5767.
Capraz, O., Polat, O., & Gngr, A. (2015). Planning of waste electrical and electronic
equipment (WEEE) recycling facilities: MILP modelling and case study investigation. Flexible Services and Manufacturing Journal. doi:10.1007/s10696-015-9217-3.
Crowther, P. (1999). Designing for disassembly to extend service life and increase sustainability. In 8th International Conference on Durability of Building Materials and
Components.
[m5G;October 1, 2015;11:19]
13
Ding, L. P., Feng, Y. X., Tan, J. R., & Gao, Y. C. (2010). A new multi objective and colony algorithm for solving the disassembly line balancing problem. The International Journal of Advanced Manufacturing Technology, 48, 761771.
Gngr, A., & Gupta, S. M. (2001a). Disassembly sequence plan generation using a
branch and bound algorithm. International Journal of Production Research, 39, 481
509.
Gngr, A., & Gupta, S. M. (2001b). A solution approach to the disassembly line balancing problem in the presence of task failures. International Journal of Production
Research, 39, 14271467.
Gngr, A., & Gupta, S. M. (2002). Disassembly line in product recovery. International
Journal of Production Research, 40, 25692589.
Hezer, S., & Kara, Y. (2014). A network-based shortest route model for parallel disassembly line balancing problem. International Journal of Production Research, 53, 1849
1865.
Koc, 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, 866881.
Lambert, A. J. D. (1997). Optimal disassembly of complex products. International Journal
of Production Research, 35, 25092523.
Lambert, A. J. D. (2002). Determining optimum disassembly sequences in electronic
equipment. Computers and Industrial Engineering, 42, 553575.
Lambert, A. J. D. (2003). Disassembly sequencing: A survey. International Journal of Production Research, 41, 37213759.
McGovern, S. M., & Gupta, S. (2011). The disassembly line: Balancing and modeling. New
York: McGraw Hill, New York.
McGovern, S. M., & Gupta, S. M. (2007a). A combinatorial optimization analysis of the
unary NP-complete disassembly line balancing. International Journal of Production
Research, 45, 14851551.
McGovern, S. M., & Gupta, S. M. (2007b). A balancing method and genetic algorithm for
disassembly line balancing. European Journal of Operational Research, 179, 692708.
Navin-Chandra, D. (1994). The recovery problem in product design. Journal of Engineering Design, 5, 6586.
Paksoy, T., Gngr, A., zceylan, E., & Hancilar, A. (2013). Mixed model disassembly line
balancing problem with fuzzy goals. International Journal of Production Research, 51,
60826096.
Patterson, J. H., & Albracht, J. J. (1975). Assembly line balancing: 01 programming with
bonacci search. Operations Research, 179, 166174.
Ullerich, C. (2014). Advanced disassembly planning: Flexible, price-quantity dependent
and multi-period planning approaches. Wiesbaden: Springer Gabler.
Please cite this article as: E.G. Kalayclar et al., A disassembly line balancing problem with xed number of workstations, European Journal of
Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.09.004