You are on page 1of 14

Available online at www.sciencedirect.

com

Computers & Industrial Engineering 56 (2009) 462475


www.elsevier.com/locate/caie

Heuristic procedures and mathematical models for cell


loading and scheduling in a shoe manufacturing company
Gursel A. Suer *, Ananthanarayanan Subramanian, Jing Huang
Industrial and Systems Engineering, Ohio University, Athens, OH 45701, USA

Available online 17 October 2008

Abstract

This paper focuses on cell loading and scheduling issues in a shoe manufacturing company. It involves a three-phase
solution methodology where rst number of cells is determined, then cells are loaded and nally detailed scheduling is per-
formed. Three dierent family denitions are used (subfamilies, families and superfamilies) during the cell loading process.
Three cell loading heuristic procedures (H1, H2, and H3), a simple cell scheduling heuristic and two-level mathematical
models are proposed and then compared with respect to makespan. An important observation is that better post-loading
schedules did not lead to better post-scheduling results. As a result, the best schedule can only be known after both cell
loading and cell scheduling functions are performed. The results showed that H1 led to the best post-scheduling makespan
values and they were on the average within 3% of the optimal solution. The execution time for H1 was less than four sec-
onds for the problems tested in the experimentation and this can be considered negligible for a manufacturing setting.
2009 Published by Elsevier Ltd.

Keywords: Cell loading; Cell scheduling; Cellular manufacturing; Shoe manufacturing; Heuristics; Mathematical models

1. Introduction

A cellular manufacturing system (CMS) comprises of individual machine groups that are considered as sep-
arate entities by themselves. The cellular manufacturing system is perhaps the most complex system to design,
because of the inherent judgments involved in deciding machine groupings and product families, which are
critical in determining the systems performance. The benets of cellular manufacturing include reduced
set-up times, lower work-in-process levels, faster throughput times, improved product quality and reduced
material handling (Wemmerlov and Hyer, 1989).
In any cellular manufacturing system, not only the part families have to be determined, but the time frames
(i.e., the start and the completion times) for manufacturing the product families and individual products have
to be determined. Any good implementation of cellular manufacturing requires a planning and control system
to take the implementation to the exemplary level. Suer, Saiz, and Gonzalez (1999a) referred to cell loading

*
Corresponding author. Tel.: +1 740 593 1542; fax: +1 740 593 0778.
E-mail address: suer@ohio.edu (G.A. Suer).

0360-8352/$ - see front matter 2009 Published by Elsevier Ltd.


doi:10.1016/j.cie.2008.10.008
G.A. Suer et al. / Computers & Industrial Engineering 56 (2009) 462475 463

and cell scheduling as two key activities of special interest in cellular manufacturing in connection with how to
control the system. Cell loading is a production planning activity that results in the assignment of products
families/products to the cells. Cell scheduling, on the other hand, involves scheduling operations in a cell once
the product has been assigned.
Cells are ideally designed in such a way that raw materials enter the cell from one end and completed prod-
ucts leave the system from the other end. However, the shoe manufacturing environment considered in this
study consists of several Connected Cell Groups (Fig. 1). Each of these cell groups consists of a Lasting Cell
(LC) to prepare the shoes for injection molding the soles, a Rotary Molding Cell (RMC) with 6 pairs of
stations (a pair of which can process one pair of shoes at any given time) and a nishing/packing cell (FC)
to remove extra material from the sole, nish the shoe and to pack the shoe. The shoes move sequentially from
the LC to the RMC and then to the FC. Each cell group is independent of the others and therefore no inter-
cell group transfer is allowed.
In this study, family formation is not considered since product families have been pre-determined. This
paper focuses on the second cell of each group, namely RMC since it is the bottleneck of the cell group
and also forms a complex loading and scheduling problem. The cell loading task deals with assignment of
product families/products to RMC in their feasible cell groups. On the other hand, cell scheduling focuses

P1

P6 P2
Desma Rotary 1 Finishing /
Lasting
Lasting
with 6 pairs of Packing
Cell
Cell12 Cell 1
positions
P5 P3

P4

P1

P6 P2
Desma Rotary 2 Finishing /
Lasting
with6 pairs of Packing
Cell 2
positions Cell 2

P5 P3

P4

More Cells

P1

P6 P2
Desma Rotary 6 Finishing /
Lasting Packing
with 6 pairs of
Cell 6 Cell 3
positions
P5 P3

P4

Fig. 1. Connected cell groups in the company.


464 G.A. Suer et al. / Computers & Industrial Engineering 56 (2009) 462475

on assigning dierent sizes of products to the rotary machine positions by including setup times, whenever
needed. Cell scheduling also involves product sequencing decisions.
The objective of this study is to develop heuristic procedures and mathematical models for cell loading and
scheduling and then compare their performance with respect to makespan measure and also explore the inter-
relations between cell loading and cell scheduling decisions.

2. Literature survey

The summary of relevant research on the topics associated with the loading and scheduling of manufactur-
ing cells is presented in this section. To the best knowledge of authors, the only work on Rotary Injection
Molding Machine Scheduling in a shoe manufacturing company was reported by Suer, Santos, and Vasquez
(1999b). Their objective was to reduce makespan but they considered only a single machine and a single prod-
uct in their work.
Saad, Gindy, and Baykasoglu (2002) proposed a multiple objective optimization technique, based on sim-
ulation and a taboo search algorithm, for solving loading and scheduling problems. Gomez (2001) proposed a
three phase hierarchical hybrid approach for simultaneous cell loading and manpower allocation and then
sequencing of jobs in cellular manufacturing systems based on an evolutionary computation methodology.
Suer (1998c) and Suer et al. (1999a) also discussed how synchronized ow can be tied to cell loading (master
scheduling) and to the MRP system in the same company to provide complete solution to the loading and
scheduling issues.
Mathematical models were also proposed for cell loading in several works (Akturk& Wilson, 1998; Ham &
Hitomi, 1981). Baykasoglu and Gindy (1999) discussed a multiple objective simulation optimization model for
loading exible cells. Suer, Saiz, Dagli, and Gonzalez (1995) introduced simple heuristics for cell loading pur-
poses. They discussed eleven rules and algorithms for cell-loading in a system of connected manufacturing
cells in a multi-cell environment. Later, Suer et al. (1999a) evaluated those cell loading rules for independent
manufacturing cells.
Some of the work focused on both cell loading and manpower allocation in labor intensive manufacturing
cells (Suer, 1996; Suer & Bera, 1998a, Suer & Bera, 1998b). Suer, Vasquez, and Cortes (2005) developed a
hybrid approach of genetic algorithms and local optimizers in cell loading to minimize the number of tardy
jobs. Suer, Arikan, and Babayigit (2008) proposed a bi-objective fuzzy mathematical modeling approach to
minimize number of tardy jobs and also total manpower in cellular manufacturing systems.
Mahmoodi and Mosier (1998) presented the results of their group scheduling research and its managerial
implications are illustrated with some group scheduling heuristics that can be applied to a small problem.
Ruben and Mahmoodi (1999) discussed scheduling ow line cells under combinations of a number of dierent
conditions and variations in factors such as arrival times, set-up times etc. Some researchers discussed cell
scheduling issues and the eectiveness of period batch control (PBC) technique in their work (Sinha & Hollier,
1984; Riezebos, 1998; Suer & Gonzalez, 1993).

3. Description of the system studied

Each customer order typically contains shoes of dierent models, each model in turn having dierent
sole types [fullshot (FS) and midsole (MS)], colors [Black (B), Honey (H) and Nicotine (N)], material
(PVC and TPR) and sizes. Each entry of the customer order becomes a job. All the jobs are assumed
to have the same due date at the end of planning period. The main structure for product families is
summarized in Fig. 2.
The rotary machine cell consists of a single rotary injection molding machine as shown in Fig. 3. The main
parts of the machine are a tank to hold the material to be injected and 6 pairs of positions for shoes. Once the
unprocessed shoes are loaded into the rotary machine, material (either PVC or TPR) is injected into the molds,
thus forming the soles. The machine can handle one material type and only one pigment color at a time. Injec-
tion machine is rotated anti-clockwise after every Injection Time seconds. There are four main stages of the
injection molding, (1) loading shoes (loading/unloading point), (2) injection molding of soles (injection point),
(3) cooling o stage (next four stations) and (4) unloading shoes (loading/unloading point).
G.A. Suer et al. / Computers & Industrial Engineering 56 (2009) 462475 465

Fig. 2. Main structure for product families.

Loading/Unloading Point

Po 2
Po 3
Po 1 - -
- Po 4
-
Injection Point

Po 12 - - Po 5
Initial position
Po 11 - - Po 6 All positions are free

- - Po 7
Po 10
- -
Po 9 Po 8

Fig. 3. Rotary injection molding machine.

The injection molding times vary based on size and sole type. Times are higher for fullshot shoes than the
midsole models and they increase as the shoe sizes increase. There are 18 dierent sizes for male shoes and 13
dierent sizes for female shoes. There is setup time whenever there is change in color or material or both.
Available Capacity per cell (MAXCAP) is dened by the scheduler. This has been introduced to allow for inef-
ciencies in the manufacturing system such as machine failures, worker absenteeism, material handling delays,
etc.
Product families are formed at three levels:

(1) Subfamilies: All the dierent sizes of a particular model for a given sole type, color, and material form a
subfamily.This information is used to determine product sequences in a family. The products in the same
subfamily are run consecutively in the injection molding machine so that theyll arrive at the nishing/
packing cell consecutively and theyll be shipped out as soon as all the products in a subfamily are com-
pleted at the nishing/packing cell. In other words, dening subfamilies does not contribute to reducing
makespan at the injection molding machine but critical for other parts of the cell group.
(2) Family: Subfamilies with the identical sole type, color and material form families regardless of the shoe
model. This denition is used to identify families to be run in a cell. By grouping subfamilies into fam-
ilies, the number of setups is minimized thus increasing time available to produce shoes.
(3) Superfamilies: All the fullshot families and midsole families form two superfamilies, respectively. This
information is used to determine number of cells needed for each of two signicantly dierent product
groups.
466 G.A. Suer et al. / Computers & Industrial Engineering 56 (2009) 462475

4. The methodology used

The methodology for loading and scheduling the jobs is a multi-phase one. The three main phases are: (1)
background phase (determining the number of cells for super-families); (2) cell loading phase and (3) cell
scheduling phase. Each phase of the methodology is a multi-step process by itself.

4.1. Determining number of cells for superfamilies

There is a need to form two superfamilies (fullshot and midsole) rst and then determine the number of cells
for each. This separation is critical due to large dierence in the processing times of these two types of shoe
soles. An example list of customer orders is given in Table 1. Table 2 summarizes the order ID numbers for the
families formed using the denitions in Section 3.
The time required to process each job is calculated multiplying its injection molding time by its demand.
The total load on a superfamily is calculated by summing up the loads of the shoes in that superfamily.
The number of cells is computed dividing the total load on a superfamily by MAXCAP value. The number
of cells can be rounded up to determine the nal cellular conguration. However, it is quite possible that
the summation of decimal parts of number of cells for both superfamilies can be less than or equal to 1. In
this case, a shared cell is created to minimize the total number of cells in use.

4.2. Cell loading heuristics

In this section, three heuristic procedures developed for cell loading are discussed in detail.

4.2.1. Calculation of subfamily processing times


The rst step in this phase is to nd the processing times for each subfamily. The processing time of sub-
family v (TSUBFv) is the sum of the processing times for every job in the subfamily.

4.2.2. Heuristic I Split subfamilies as and when required (H1)


In this strategy, subfamilies are formed rst and then they are grouped into two sets, fullshot and midsole.
Then each set is sorted by LPT. The next subfamily in the considered set is assigned to the rst available cell

Table 1
An example of customer orders.
Order ID Model ID Sole type Material Color Size Code Demand
1 K FS TPR Black 6 FTB-K 269
2 U FS PVC Black 8 FPB-U 688
3 C FS TPR Black 5.5 FTB-C 1045
4 C FS TPR Black 7 FTB-C 208
5 L MS TPR Black 5 MTB-L 881
6 T MS TPR Black 6 MTB-T 831
7 T MS TPR Black 10 MTB-T 277
8 O FS PVC Black 8 FPB-O 250
9 E FS PVC Black 8 FPB-E 636
10 W FS PVC Black 9 FPB-W 384
11 W FS PVC Black 10 FPB-W 329
12 F MS TPR Black 5 MTB-F 440
13 F MS TPR Black 9 MTB-F 321
14 N MS TPR Black 8 MTB-N 355
15 N MS TPR Black 8 MTB-N 255
16 X MS PVC Black 6 MPB-X 788
17 E FS PVC Nicotine 5 FPN-E 574
18 E FS PVC Nicotine 8 FPN-E 245
G.A. Suer et al. / Computers & Industrial Engineering 56 (2009) 462475 467

Table 2
Denition of families for dierent family levels.
Family No. Subfamilies Families Superfamilies
1 1 1, 3, 4 1, 2, 3, 4, 8, 9, 10, 11, 17, 18
2 2 2, 8, 9, 10, 11 5, 6, 7, 12, 13, 14, 15, 16
3 9 5, 6, 7, 12, 13, 14, 15
4 10, 11 16
5 3,4 17, 18
6 5
7 12, 13
8 14, 15
9 6, 7
10 8
11 16
12 17, 18

(of its type) as long as there is enough capacity. If the capacity is not sucient to cover the entire subfamily,
then it is split into two groups and the balance is assigned to the next available cell of its type. The same pro-
cedure is repeated for all subfamilies and all sets. At the end, remaining subfamilies in both sets are loaded to
the shared cell. With this heuristic, fewer subfamilies with longer runs are assigned to the rst cell; more sub-
families with shorter runs are assigned to the second cell and so on. Finally, the subfamilies with shortest runs
are kept for the shared cell. This may help to bring stability to a larger segment of the cellular system as it is
desired by supervisors on the shop oor. Finally, families are formed in each cell out of subfamilies assigned to
it. More formal description of the steps of the algorithm is given below:

(1) Determine TSUBF for all subfamilies (TSUBFv).


(2) Form sets of fullshot and midsole subfamilies, SF and SM, respectively.
(3) Sort subfamilies in SF and SM by the Largest Processing Time (LPT) rule.
(4) Select the next set of subfamilies, S, and continue with step 5. If all sets have been already considered,
then go to step 8.
(5) Select the next subfamily, v, from S, and continue with step 6. If all subfamilies have been already con-
sidered then go to step 4.
(6) Find the rst available dedicated cell k for subfamily v and continue with the next step. If no such cell is
available, go to step 4.
(7) If TSUBFv 6 (MAXCAP-LOADk), then assign subfamily v to cell k and go to step 5.Otherwise, split
subfamily into two groups Gv1 and Gv2 such that total processing time of jobs in Gv1 would not exceed
available capacity in cell k. An attempt is made to assign Gv2 to the next available cell. Go to step 5.
(LOADk indicates the current load on cell k).
(8) Assign the remaining subfamilies to the shared cell.
(9) Form families from subfamilies in each cell to minimize the setup times.

When a subfamily is split, jobs are divided into groups but no job (shoe of particular size) is allowed to be
split into two groups. Another important issue is that jobs in a subfamily are rst ordered by LPT and then
splitting of subfamily starts. We use the rst t strategy in forming group Gv1 and as a result, various com-
binations are not evaluated even though they could t to the available capacity better. This is as per
McNaughtons algorithm (McNaughton, 1959). The ordering jobs in a subfamily by LPT allows rst t
strategy to nd a feasible solution faster.

4.2.3. Heuristic II Split families when load is greater than MAXCAP (H2)
In this strategy, the subfamilies are not split regardless of their load and the capacity remaining in the cell
except for subfamilies with processing time greater than MAXCAP. In this heuristic, subfamily v with
TSUBFv > MAXCAP is split into minimum possible t groups {Gv1, Gv2, ...., Gv,t1, Gvt} such that total
468 G.A. Suer et al. / Computers & Industrial Engineering 56 (2009) 462475

processing time of jobs in Gv1, Gv2, . . ., Gv,t1 cannot exceed MAXCAP and the remaining jobs are grouped in
Gvt. All groups are treated as dierent subfamilies after this point. This procedure is repeated for each such
large subfamily. Later, subfamilies are grouped into two sets, fullshot and midsole. Then, each set is sorted
by LPT. Since subfamily splitting is not allowed and the objective is to minimize makespan, the next subfamily
in the considered set is assigned to the cell with minimum load (of its type) as long as there is enough capacity.
The reason for not allowing subfamily splitting is to maintain family integrity as expected in a GT application.
If there is not enough capacity in the cell, the subfamily is set aside to be loaded into the shared cell later. The
same procedure is repeated for all subfamilies and all sets. At the end, remaining subfamilies in both sets are
loaded to the shared cell. Finally, families are formed in each cell out of subfamilies assigned to that cell. More
formal description of the steps of the algorithm is given below:

(1) Determine TSUBF for all subfamilies (TSUBFv).


(2) Subfamily v with TSUBFv > MAXCAP is divided into t groups as described above and Gv1, Gv2, ....,
Gv,t1 are assigned to their dedicated cells and Gvt is included with the remaining subfamilies.
(3) Form sets of fullshot and midsole subfamilies, SF and SM, respectively, with the remaining subfamilies.
(4) Sort subfamilies in SF and SM by the largest processing time (LPT) rule.
(5) Select the next set of subfamilies, S, and continue with step 6. If all sets have been already considered,
then go to step 9.
(6) Select the next subfamily, v, from S, and continue with step 7. If all subfamilies have been already con-
sidered then go to step 5.
(7) Find dedicated cell k with minimum load for subfamily v and continue with the next step. If no such cell
is available, go to step 5.
(8) If TSUBFv 6 (MAXCAP-LOADk), then assign subfamily v to cell k and go to step 6. Otherwise, do not
assign it and go to step 6.
(9) Assign remaining subfamilies to the shared cell.
(10) Form families from subfamilies in each cell to minimize the setup times.

If the capacity is not sucient to cover the entire subfamily, then it is split into t groups and the group t is
not assigned to the next available cell as in H1 and instead it is ordered according to LPT like other
subfamilies.

4.2.4. Heuristic III Load jobs and then form subfamilies (H3)
This procedure diers from the other procedures because the loading is on the job level itself. In other
words, jobs are separated into fullshot and midsole groups rst. The jobs are to be loaded in the same order
as they arrive. The main motivation behind this procedure is to increase the exibility of loading cells and nd
better t schedules by having independent smaller jobs as opposed to dealing with larger subfamilies. Having
loaded jobs onto cells, rst subfamilies and then families are formed based on the jobs assigned to each cell. As
in the previous heuristics, no job splitting is allowed. The steps of the algorithm are:

(1) Determine processing times (Ti) for all jobs.


(2) Form sets of fullshot and midsole jobs, SF and SM, respectively.
(3) Select the next set of jobs, S, and continue with step 4. If all sets have been already considered, then go to
step 7.
(4) Select the next job, job i, from S, and continue with step 5. If all jobs have been already considered then
go to step 3.
(5) Find dedicated cell k with minimum load for job i and continue with step 6. If no such cell is available,
go to step 3.
(6) If Ti 6 (MAXCAP-LOADk), then assign job i to cell k and go to step 4. Otherwise, do not assign it and
go to step 4.
(7) Assign remaining jobs to the shared cell.
(8) Form subfamilies from jobs in each cell.
(9) Form families from subfamilies in each cell to minimize the setup times.
G.A. Suer et al. / Computers & Industrial Engineering 56 (2009) 462475 469

4.2.5. Estimating MAXCAP value


The best MAXCAP value is estimated by using the procedure described next. Initially, MAXCAP is set to
the theoretical minimum makespan and then the chosen heuristic procedure is applied. If it does not produce a
feasible solution, MAXCAP is incremented by one unit and tried again. The procedure continues until a fea-
sible solution is found. The minimum MAXCAP is the lowest value that produces a feasible solution.

4.3. Details of cell scheduling phase

Once the cell loading is complete, the nal step is to determine the exact start times and completion times
for every job/subfamily/family in each cell. This is where the molds and hence setup times enter into the equa-
tion. The molds used in the injection molding process and the tank holding the material have to be cleaned
each and every time the material or the coloring pigment used changes. In this study, setup times are taken
as 20 minutes regardless whether color or material or both changes. On the other hand, setup times for mold
changes are assumed negligible.
Another assumption that has been made is that the rst family processed in every cell has zero setup, i.e.,
the molds required by the family are available at time t = 0 and the tank in the injection molding machine has
been cleaned and relled in time for the rst family. So at time zero, the rst family in the list for every cell is
loaded and the processing begins.
This processing time is calculated by rst estimating the number of turns required to process every job.
Since the job will be processed simultaneously on all six positions of the rotary machine, every single complete
turn of the machine will result in the output of 6 pairs of shoes. This means that if the demand for a particular
shoe and of a particular size is not a multiple of 6, some positions will remain idle until all the shoes are com-
pleted. So, if the demand is not a multiple of 6, then the number of turns is rounded up. Since every position is
going to be processing the same job, the total processing time for job i, TPi, is dened as given in Eq. (1).
 
Di
TP i  Pi  6 1
6
Finally, a setup time for 20 min is added to the schedule between families to account for color and/or mate-
rial changes and thus the nal schedule to be implemented in every cell is established.

4.4. Mathematical modeling for cell scheduling

In this section, two mathematical models are proposed to directly determine cell schedules from the initial
load of families. The rst model assigns families to cells and the second model selects jobs for each cell when a
family is split among multiple cells.

4.4.1. Family-based cell scheduling model


In this section, a mathematical model is formulated to assign families to cells such that makespan is min-
imized. The decision variables, parameters and indices are dened as follows:
Indices
j cell index
k family index
Parameters
Sk setup time for family k
F number of families
C number of cells
M big value
FF number of families that belongs to fullshot superfamily
FC number of cells dedicated to produce fullshot superfamily
MC number of cells dedicated to produce midsole superfamily
470 G.A. Suer et al. / Computers & Industrial Engineering 56 (2009) 462475

Nk number of jobs in family k


TPik total processing time for job i in family k (computed using Eq. (1))
P
Nk
TFPk total processing time for family k, TFP k TP ik
i1
Decision variables
Xkj percentage of load of family k assigned to cell j
MS Makespan

Dependent variable
Wkj 1, if any part of family k is assigned to cell j; 0, otherwise

The formulation of the model is as follows:


Objective function :
Min Z MS 2
Subject to :
XF X
F
MS  TFP k  X kj  S k  W kj P 0 j 1; . . . ; C 3
k1 k1
M  W kj P X kj j 1; . . . ; C and k 1; . . . ; F 4
XC
X kj 1 k 1; . . . ; F 5
j1
X kj 0 k 1; . . . ; FF and j FC 1; . . . ; FC MC 6
X kj 0 k FF 1; . . . ; F and j 1; . . . ; FC 7
X kj ; W kj 2 0; 1; MS P 0; W kj integer
The objective is to minimize the makespan as given in Eq. (2). Eq. (3) guarantees that completion time for
each cell never exceeds makespan. Eq. (4) enforces setup time if family k is processed on cell j. The constraint
that each family should be completely allocated to cells is given by Eq. (5). Since we already know the number
of dedicated cells, it is assumed that Cell 1 to Cell FC are the dedicated cells for fullshot superfamily, and Cell
FC + 1 to Cell FC + MC are the dedicated cells for midsole superfamily. Similarly, it is also assumed that fam-
ily 1 to family FF belongs to fullshot superfamily, and family FF + 1 to family F belongs to midsole superfam-
ily. Eqs. (6) and (7) ensure that the dedicated cells never produce products from other superfamily.

4.4.2. Job-based cell scheduling model


In this section, the result of family assignment is used to determine job allocation in each family. The opti-
mal result of family-based cell scheduling model only provides a general solution without the detailed job allo-
cation. Since job splitting is not allowed, and jobs processing times vary signicantly, it may be dicult to
divide the capacity into the amount obtained during family-based cell scheduling. Those results are not prac-
tical in some cases. For example, assume that there are only three jobs in one family with processing times 10,
20 and 15, respectively. There is no feasible solution for job allocation, if 50% percentage of family (22.5) has
to be allocated to cell 1 and the remaining (22.5) has to be assigned to cell 3 according to family-based sched-
uling result. Thus, in job-based cell scheduling model, the objective is to minimize the dierence between
actual load after job allocation implementation and theoretical load indicated by the family-based cell sched-
uling result. The decision variables and parameters are dened as follows:
Parameters
TPi total processing time for job i
CRkj total capacity requirements for family k assigned to cell j (based on family-based cell
scheduling result)
Nk number of jobs in family k
Ck number of cells family k is assigned to
G.A. Suer et al. / Computers & Industrial Engineering 56 (2009) 462475 471

Decision variables
Xij 1 if job i is assigned to cell j; 0, otherwise
Sj slack capacity for cell j
Ej excess capacity for cell j

The formulation of model is as follows:

Objective function :
X
Ck X
Ck
Min Z Sj Ej 8
j1 j1

Subject to :
X
Nk
TP i  X ij S j  Ej CRkj j 1; . . . ; C k 9
i1
X
Ck
X ij 1 i 1; . . . ; N k 10
j1

S j ; Ej P 0; X ij 2 0; 1; X ij integer

The objective is to minimize the gap between actual load and theoretical load as given in Eq. (8). Eq. (9) de-
nes the relation between actual load and theoretical load by using slack and excess values. Eq. (10) guaran-
tees that each job can only be allocated to one cell. This model should be run for each family that has been
assigned to more than one cell.

5. Experimentation performed

The heuristic procedures developed for cell loading are compared by using the post-loading results. Next,
the performance of heuristics is evaluated based on post-scheduling results. The nal section is devoted to the
comparison of post scheduling results of heuristics with mathematical models.

5.1. Experimental conditions

Nine problems are generated and used in all experiments. The characteristics of the problems are summa-
rized in Table 3.

Table 3
Characteristics of problems considered in the experimentation.
Problem # Jobs # Fullshot # Midsole Total number # families # Superfam Total load on Total load on Total
subfam. subfam. of subfam. fullshot subfam. midsole subfam. Load
Big1 461 25 22 47 12 2 6281.276 4763.273 11044.5
Big2 403 16 9 25 10 2 8450.989 2939.567 11390.5
Big3 427 20 13 33 8 2 7452.036 3785.591 11237.6
Medium1 309 14 9 23 5 2 4834.243 3235.502 8069.7
Medium2 310 13 11 24 6 2 4960.614 3182.015 8142.6
Medium3 359 15 11 26 6 2 5786.753 3982.603 9769.3
Medium4 273 12 11 23 6 2 4874.989 3035.528 7910.5
Small1 350 11 9 20 5 2 3664.317 2590.501 6254.8
Small2 283 9 6 15 5 2 4592.516 1913.182 6505.6
472 G.A. Suer et al. / Computers & Industrial Engineering 56 (2009) 462475

5.2. Initial comparison of cell loading heuristics

In this section, an initial attempt is made to evaluate the eectiveness of heuristics in terms of minimizing
post-loading makespan values. First, the number of cells is established. Based on the number cells, the theo-
retical minimum makespan is estimated dividing the total load by the number of cells. The theoretical mini-
mum makespan indicates the lowest possible makespan value that could be obtained if the entire load could be
divided equally among all the cells. Dierent MAXCAP values are tried starting from theoretical minimum
makespan and incrementing it by one minute intervals until a feasible solution is obtained. Then the lowest
MAXCAP value that produces a feasible schedule for each heuristic is determined. Actual post-loading make-
span values are determined as given in Table 4 (lowest makespan values bolded for each problem). As one can
easily observe, H3 outperformed both H1 and H2. In most cases, H3 found makespan values close to the the-
oretical minimum makespan. It is also clear that H1 was better then H2 in all nine cases.

5.3. Comparison of heuristics after cell scheduling phase

In this section, the analysis focuses on when to compare the performance of cell loading heuristics, i.e.,
post-loading results or post-scheduling results. Based on the post-scheduling results, H1 found the lowest
makespan 8 out of 9 problems (bolded) compared to none based on post-loading results. It should be noted
that the lowest cell-loading makespan schedule did not lead to the lowest cell-scheduling makespan (see Table
5). The purpose of this section is to show that comparing heuristics based on cell-loading results may be an
immature step since the exact schedules and thus accurate makespan values can only be determined after cell
scheduling phase. This is a natural result of hierarchical scheduling. It is clear that actual makespan values
increased during scheduling as shown in Table 5. There are two sources for that; (1) setup times, (2) times
for additional cycles due to discrepancy between (number of turns * 6) and actual number of shoes produced.

Table 4
Post-loading makepan values for dierent heuristics.
Problem Number of cells (Full Theoretical minimum MAX CAP H1 MAX CAP H2 MAX CAP H3
shot + midsole + shared) makespan used used used
Big1 6(3 + 2 + 1) 1840.758 1870 1869.912 1889 1888.326 1841 1840.988
Big2 6(4 + 2 + 0) 1898.426 2137 2136.668 2174 2173.24 2114 2113.027
Big3 6(4 + 2 + 0) 1872.938 1906 1905.847 1923 1922.886 1894 1892.921
Medium1 6(3 + 2 + 1) 1344.958 1364 1363.627 1444 1443.621 1347 1346.91
Medium2 6(3 + 2 + 1) 1357.105 1376 1373.428 1420 1419.992 1358 1357.87
Medium3 6(3 + 2 + 1) 1628.226 1650 1649.709 1668 1667.603 1629 1628.9
Medium4 5(3 + 2 + 0) 1582.103 1649 1641.198 1671 1670.291 1626 1625.239
Small1 6(3 + 2 + 1) 1042.47 1059 1058.721 1089 1088.676 1043 1042.795
Small2 4(2 + 1 + 1) 1626.425 1648 1647.538 1695 1694.273 1627 1626.878

Table 5
Post-scheduling makespan results.
Problem Number of cells MAX CAP used H1 MAX CAP used H2 MAX CAP used H3
Big1 6 1870 2086.152 1889 2088.498 1841 2217.876
Big2 6 2137 2269.088 2174 2290.194 2114 2294.796
Big3 6 1906 2039.456 1923 2049.348 1894 2026.604
Medium1 6 1364 1478.558 1444 1533.484 1347 1530.19
Medium2 6 1376 1472.58 1420 1518.318 1358 1559.784
Medium3 6 1650 1756.974 1668 1805.422 1629 1846.674
Medium4 5 1649 1724.404 1671 1768.212 1626 1728.33
Small1 6 1059 1141.798 1089 1200.348 1043 1244.932
Small2 4 1648 1772.48 1695 1782.766 1627 1828.204
G.A. Suer et al. / Computers & Industrial Engineering 56 (2009) 462475 473

Table 6
Variation of heuristics from the Optimal solution.
Problem Number of cells Math mode H1 Variation (%) H2 Variation H3 Variation (%)
B1 6 1961.9 2086.152 6.33 2088.498 6.45% 2217.876 13.05
B2 6 2215.8 2269.088 2.40 2290.194 3.36% 2294.796 3.57
B3 6 1995.3 2039.456 2.21 2049.348 2.71% 2026.604 1.57
M1 6 1421.3 1478.558 4.03 1533.484 7.89% 1530.19 7.66
M2 6 1434.2 1472.58 2.68 1518.318 5.87% 1559.784 8.76
M3 6 1709 1756.974 2.81 1805.422 5.64% 1846.674 8.06
M4 5 1701.1 1724.404 1.37 1768.212 3.95% 1728.33 1.60
S1 6 1119.3 1141.798 2.01 1200.348 7.24% 1244.932 11.22
S2 4 1717.4 1772.48 3.21 1782.766 3.81% 1828.204 6.45

The latter can be explained with an example. For example, if there are 63 pairs of shoe to be injected, the total
time used during the loading phase is 63 * Pi whereas in scheduling phase 11 * Pi * 6. Obviously, H3 was very
successful in cell loading since the loading was done considering individual jobs as opposed to families and
better t could be obtained when only jobs were considered. On the other hand, H3 proved ineective after
the cell scheduling phase since families were ignored during loading and as a result more setups were required
(see Table 5).

5.4. Heuristics vs. math model post-scheduling results

In this section, the performance of heuristics is compared with the mathematical model based on post-
scheduling results. The results are summarized in Table 6 for all three heuristic procedures with percent var-
iation from the optimal solution (best results are bolded). The results indicate that the variation from optimal
solution for H1 was 1.376.33% with an average variation of 3.01%. Considering the sizes of the problems
included in the experimentation, the results can be considered as satisfactory.
The computer running time for each problem is reported in Table 7. In H3, since feasible solutions are
always found when MAXCAP is close to lower bound without repeatedly increasing MAXCAP, the CPU
times are very short. For all heuristic procedures, the main factor that aects computation time is the gap
between lower bound of MAXCAP and feasible MAXCAP and not the problem size. The bigger the gap,
the higher the number of iterations will have to be done, thus increasing computational time. We can nd that
the longest running time by using any heuristic procedure is 10 s.
Recall that, job-based cell scheduling model should be developed for each product family as long as this
family has been assigned to more than one cell. Number of mathematical models and total CPU time required
for each problem is also shown in Table 7. As shown in this table, mathematical models take longer time than
heuristic procedures. CPU times increase signicantly when the problem size increases. Another experiment
showed that it took 7026 s when the size of problem Big 1 was doubled.

Table 7
Comparison of CPU times.
Problem H1 H2 H3 Mathematical model
CPU time (s) CPU time (s) CPU time (s) Number of models Total CPU time
Big1 4 6 <1 1+5 358
Big2 2 9 <1 1+3 69
Big3 1 4 <1 1+3 40
Medium1 <1 10 <1 1+3 38
Medium2 2 8 <1 1+5 28
Medium3 1 6 <1 1+3 30
Medium4 4 6 <1 1+3 14
Small1 2 6 <1 1+5 15
Small2 1 7 <1 1+3 12
474 G.A. Suer et al. / Computers & Industrial Engineering 56 (2009) 462475

6. Conclusions and further remarks

The work presented in this paper is a three-phase hierarchical approach for solving cell loading and sched-
uling in a shoe manufacturing company. This study covers only injection molding cells; however, some deci-
sions have been made considering lasting and packing cells as well.
The family classication is done at three levels; superfamilies, families, and subfamilies. Dierent family
denitions are used at dierent stages in the planning and scheduling process. This problem lands itself to such
a detailed family classication and the proposed methodology takes the full advantage of this property of the
problem on hand. Authors are not aware of any other work that considers these multiple level family deni-
tions. Superfamily classication is used to determine the number of cells of each type. This is important since
processing requirements of fullshot and midsole shoes at lasting, packaging and injection molding cells are
signicantly dierent. Family classication is utilized once the products are assigned to cells. Families are
formed to minimize the number of setups in the cells. Subfamily classication is used before cell loading starts
when heuristic procedures 1 and 2 are implemented. Such a detailed multi-level family classication simplies
computations and captures required level of similarity among products at dierent stages of the planning
process.
An observation that can be made in this study is that MAXCAP value signicantly aects the performance
of heuristic procedures. Therefore, heuristic procedures should not be compared at arbitrarily selected MAX-
CAP values. The best approach is to identify the lowest MAXCAP value where each heuristic procedure will
be producing a feasible solution. It is best to determine the lowest theoretical makespan and then increase
MAXCAP value on a trial-error basis until a feasible solution is found.
A signicant conclusion obtained in the work is that Heuristic 1 is the best procedure since it found the
lowest post-scheduling makespan values among three procedures even though it did not result in the lowest
post-loading makespan values. Another conclusion that can be drawn from the results is that the comparison
of heuristics based on cell loading results may not lead to accurate conclusions. Therefore, it is more accurate
to make the comparisons based on post-scheduling results. We believe that this is a valuable conclusion and
researchers should keep this possibility in mind especially when they isolate and tackle a single level in a multi-
level hierarchical problem. Obviously, the implications of this phenomenon are beyond planning, scheduling
and manufacturing.
One has to be also cautious in setting MAXCAP values since there is a possibility that even though post-
loading makespan does not violate MAXCAP the corresponding post-scheduling makespan may. The sched-
ulers will have to develop their own insight as to what MAXCAP values are likely to generate feasible post-
scheduling results. Overall, one has to establish MAXCAP carefully not to generate infeasible post-scheduling
solutions. Possible delays due to machine downtimes, material shortages, material transfer times, setup times,
etc. have to be factored in determining the maximum possible value.
The allocation of product families to dierent cells has to be done regularly due to variation in product mix
and the desire to use cells eciently. Unfortunately, the last cell for a superfamily may not be utilized fully in
some periods. In this case, a shared cell may be allowed for both fullshot and midsole products. The products
in the shared cell are grouped as fullshot and midsole superfamily rst and then processed. However, the
sequence of running superfamilies in the shared cell is not important.
When compared with mathematical models, heuristic procedures produced satisfactory results and rela-
tively fast. The mathematical model can still be used for small problems eectively but the same cannot be
stated for large problems.
Future work includes developing schedules for lasting and nishing cells in synch with RMC cells, possi-
bility of running dierent sizes of shoes simultaneously on the rotary machine, adding mold restrictions
and cost analysis of minimum makespan solution vs. maximum idle time solution for the shared cell (to
explore part-time work opportunity).

References

Akturk, M. S., & Wilson, G. R. (1998). A hierarchical model for the cell loading problem of cellular manufacturing systems. International
Journal of Production Research, 36(7), 20052023.
G.A. Suer et al. / Computers & Industrial Engineering 56 (2009) 462475 475

Baykasoglu, A., & Gindy, N. Z. (1999). Loading exible cells: Tabu search based simulation optimization approach. In Proceedings of the
15th international conference on production research (Vol. 2, pp. 14411444).
Gomez, A. M. (2001). A hierarchical hybrid approach to cell loading, manpower allocation and job sequencing in cellular manufacturing.
Unpublished M.S Thesis, University of Puerto Rico, Mayaguez, Puerto Rico.
Ham, I., & Hitomi, K. (1981). Group technology applications for machine loading under multi-resource constraints. In: Proceedings of IX
North American manufacturing conference and 1981 SME transactions (pp. 515518). Society of Manufacturing Engineers.
Mahmoodi, F., & Mosier, C. T. (1998). Scheduling rules for cellular manufacturing. In J. M. Kay & N. C. Suresh (Eds.), Group technology
& cellular manufacturing: A state-of-the-art synthesis of research and practice (pp. 321338). Boston, USA: Kluwer Academic
Publishers.
McNaughton, R. (1959). Scheduling with deadlines and loss functions. Management Science, 6, 112.
Riezebos, J. (1998). Production planning systems for cellular manufacturing. In J. M. Kay & N. C. Suresh (Eds.), Group technology and
cellular manufacturing: a state-of-the-art synthesis of research and practice (pp. 289308). Boston, USA: Kluwer Academic Publishers.
Ruben, R., & Mahmoodi, F. (1999). Scheduling cellular manufacturing systems. In Shahrukh A. Irani (Ed.), Handbook of cellular
manufacturing systems (pp. 141154). John Wiley & Sons.
Saad, S. M., Gindy, N. Z., & Baykasoglu, A. (2002). A new integrated system for loading and scheduling in cellular manufacturing.
International Journal of Computer Integrated Manufacturing, 15(1), 3749.
Sinha, R. K., & Hollier, R. H. (1984). A review of production control problems in cellular manufacturing. International Journal of
Production Research, 22(5), 773789.
Suer, G. A., & Gonzalez, W. (1993). Synchronization in manufacturing cells: A case of study. International Journal of Management and
Systems, 9(3), 313337.
Suer, G. A., Saiz, M., Dagli, C. & Gonzalez, W. (1995). Evaluation of manufacturing cell loading rules for connected cells. In: A. K.
Kamrani, H. R. Parsaei, D. H. Liles (Eds.) Planning, design, and analysis of cellular manufacturing systems (pp. 97127). Berlin: Elsevier
Science B.V.
Suer, G. A. (1996). Optimal operator assignment and cell loading in labor intensive manufacturing cells. Computers and Industrial
Engineering, 31(1-2), 155158.
Suer, G. A., & Bera, I. S. (1998a). Multi-period cell loading and cell size determination. Computers and Industrial Engineering, 35(1-2),
8588.
Suer, G. A., & Bera, I. S. (1998b). Optimal operator assignment and cell loading in labor intensive manufacturing cells when lot-splitting is
allowed. Computers and Industrial Engineering, 35(3-4), 431434.
Suer, G. A. (1998c). Operation and control of cellular systems at Avon Lomalinda, Puerto Rico. In J. M. Kay & N. C. Suresh (Eds.),
Group technology & cellular manufacturing: a state-of-the-art synthesis of research and practice (pp. 339361). Boston, USA: Kluwer
Academic Publishers.
Suer, G. A., Saiz, M., & Gonzalez, W. (1999a). Evaluation of manufacturing cell loading rules of independent cells. International Journal
Production Research, 37(15), 34453468.
Suer, G. A., Santos, J., & Vasquez, R. (1999b). Scheduling rotary injection molding machines. In Proceedings of the second Asia-Pacic
conference on industrial engineering and management systems (pp. 319322).
Suer, G. A., Vasquez, R., & Cortes, M. (2005). A hybrid approach of evolutionary programming and local optimizers in cell loading.
Computers and Industrial Engineering, 48, 625641.
Suer, G. A., Arikan, F., & Babayigit, C. (2008). Bi-objective cell loading problem with non-zero setup times with fuzzy aspiration levels in
labour-intensive manufacturing cells. International Journal Production Research, 46(2), 371404.
Wemmerlov, U., & Hyer, N. L. (1989). Cellular manufacturing in the US industry: A survey of users. International Journal of Production
Research, 27, 15111530.

You might also like