..
I. . .
. . . . ‘ Hindi-UV”!
. r . . ‘ ‘ . . . «9'... “until!
-' 4 . ‘ . v I V V
v9.1.2,: . v
‘ai .>I.-». . ‘ :nvu.
gum. 1A [rdelafihviir P .3. . . 0.! in:
. .. . a {3.115 v94 fihtthH. . . .
man. Run». the: £8. .
I. ‘

 

 

 

. ail , .
I: 4 ‘
’9: h‘fii '

.77. 5%..

9

I
.3!
K.

k L .V
fifb L FINE...” mu...

3; 0'53, n' . biil
0!.

.I shop...

I . .1... .
Ii... 7:. Iv. fir»;
.1. . . r

.I...‘

(a .

. .153 Ill...‘
11!,‘cVuu . 2.5 .I. ..
7)! tits; ..
9...!51, v cv-
vv 7 t l thOt

3. L.
a :TSflE... maniflwwu
. 2.!
.. . , . .u..:a...§n 2%! K
. I‘ v. uh . .2‘...i;f..l 5:17
.2... , .1. t \;

 

 

 

 

1.2.9.31
:.:..- v.|

V 1.} .
;.|. . .fv-.l.‘f c1.
.bb: .II. «1|

IO. «3?: I:
L; If. ivv.

. .1. s; .
,l .lf «4);
{7.46...I...

A ‘ .tIl’
LII. t...‘\
0 -‘V.’ \l but 3‘ , n .9
‘ :1;
:45...- c a».
I.
12.x. ....‘a:o.:!(.. v.1.
, a! 9:7,? vibrlyf: r: ’~
.....lo.0.\v17.{

Mi ‘

.Iv .‘VAoO.
alttil.. ‘

arm?"-
,v...«
1“"
.V'
:5..-

.11".-.
(v .\.. twin
.. $1.5

. '1'
I.‘
A
liu‘
“.1
"I!
2'
l
..

Q:
.‘

é,

-1}: .3... ‘
7’57..er {I'VE ‘ . . A . 1. up”...
_ . 7 . . . T v ‘
‘ s- I u if: t}, h

 

STATE UNIV RS SITY

illllLllllllllllllll IN! N

31293 00904 5802

 

 

 

 

 

 

 

 

 

 

 

 

 

                                             

This is to certify that the

dissertation entitled

Batch Production Using Dynamic Cellular
Manufacturing Systems

presented by

Vijay R. Kannan

has been accepted towards fulfillment
of the requirements for

Ph . D . degree in Operat ions Management

 

 

MAM/Z

 

Major professor

Date 7/2d/€3
/ T

MSUIS an Affirmative Action/Equal Opportunity Institution 0-12771

 

 

LIBRARY
Michigan State
Unhorslty

 

 

 

PLACE IN RETURN BOX to remove We checkout from your record.
TO AVOID FINES return on or before date duo.

DATE DUE. DATE DUE DATE DUE
SEP 2 9 1992'

 

 

 

 

 

 

 

NOV 0 5 fig?

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

MSU Is An Affirmative Action/Equal Opportunity Institution
empmS-pd

 

BATCH PRODUCTION USING

DYNAMIC CELLULAR MANUFACTURING SYSTEMS

By

Vijay R. Kannan

A DISSERTATION

Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of

DOCTOR OF PHILOSOPHY

Department of Management

1993

ABSTRACT

BATCH PRODUCTION USING
DYNAMIC CELLULAR MANUFACTURING SYSTEMS

By

Vijay R. Kannan

Demand patterns in the batch manufacturing environment are increasingly characterized
by greater variety, frequent design changes, and lower volumes. These trends place a
premium on short lead times and small batch sizes. Production methods that are
commonly used in this environment, are limited in their ability to provide both the
, flexibility and efficiency needed to meet these nwds. Job shops provide the flexibility to
respond to changes in demand, but their use of frequent setups is not conducive to the
repetitive production of small batches. Cellular manufacturing exploits similarities in

production needs, but is inflexible due to its rigid physical layout.

Dynamic Cellular Manufacturing (DCM) systems allow cellular manufacturing to be
operationalized without the layout constraints imposed by traditional cellular systems.
Manufacturing cells are formed on a real time basis, based on prevailing production
needs. These cells can evolve, expand, contract, or dissolve, depending on the needs of
specific part families and machine availability. This allows the principles of family based
production to be implemented with the flexibility required to meet current demand

patterns. This is accomplished without physically changing an existing job shop.

The use of DCM is compared to that of traditional job shop and cellular production
methods under a range of shop conditions. In addition, DCM is examined under a
broader range of conditions in order to identify conditions that appear conducive to its
use. The results show that the combination of flexibility and setup efficiency embodied
in DCM, enables it to meet the nwds of small/ medium batch production more effectively
than the other production methods. DCM outperforms these traditional production
methods over a wider range of operating conditions than anticipated. It also appears to

be more robust to certain kinds of variability than currently used production methods.

Copyright by
VIJAY RAGHAVAN KANNAN

1993

DEDICATED TO MY PARENTS

' ACKNOWLEDGEMENTS

I wish to thank several people without whose help and support I would not have been

able to complete this dissertation.

Professor Soumen Ghosh, Chairman of the Dissertation Committee, has played a major
role, both as friend and advisor. Throughout the dissertation process, he has been
readily available as a source of knowledge, wisdom, and encouragement. His advice,
insights, and editorial comments have added significantly to the quality of this

dissertation.

Professors Robert Handfield, Steven Melnyk, and Gary Ragatz, the members of the
dissertation committee, have provided comprehensive feedback on a regular basis. Their

timely suggestions and advice have been invaluable.

A number of fellow doctoral students in the Management Department have helped at
various stages of this work, despite their own struggles through the Ph.D process. In
particular, Michael D’Itri, Steven Lyman, David Mendez, and Keah Choon Tan have

made important contributions, as well as made being a Ph.D student survivable.

vi

Last but certainly not least, I wish to thank my soon-to-be wife, Karin de Jonge. Without
her continual support, encouragement, and telling me that I can do it, while she too

works on a Ph.D, I would not have reached this point.

vii

TABLE OF CONTENTS

LIST OF TABLES
LIST OF FIGURES

CHAPTER 1 : INTRODUCTION AND PROBLEM STATEMENT
1.1 Introduction
1.2 Problem Statement
1.3 Organization of Dissertation

CHAPTER 2 : BACKGROUND AND LITERATURE REVIEW
- 2.1 Introduction
2.2 Research on Cellular Manufacturing (Group Technology)
2.2.1 Comparative Studies of Cellular Manufacturing Performance
2.2.1.1 Cellular Layouts vs. Process Layouts
2.2.1.2 Hybrid Cellular Layouts vs. Process Layouts
2.2.2 Other Operational Issues in Cellular Manufacturing
2.2.3 Extensions to Cellular Manufacturing Systems
2.2.3.1 Group Scheduling
2.2.3.1.1 Group Scheduling in Job Shop Cells
2.2.3.1.2 Group Scheduling in Flow Shop Cells
2.2.3.1.3 Sequence-Dependent Scheduling in Job Shops
2.2.3.2 Lot Splitting in Cellular Manufacturing
2.2.3.3 Alternate Routings in Cellular Manufacturing
2.2.4 Cell Formation Techniques in Cellular Manufacturing
2.2.4.1 Non-Analytic Methods
2.2.4.1.1 Descriptive Methods
2.2.4.1.2 Manual Methods
2.2.4.2 Analytic Methods
2.2.4.2.1 Similarity Coefficients/Cluster Analysis
2.2.4.2.2 Block Diagonal Methods
2.2.4.2.3 Other Analytic Methods
2.2.4.3 Other Studies on Cell Formation
2.2.5 Summary of Research on Cellular Manufacturing
2.3 Flexibility Issues in Manufacturing
2.3.1 Types of Manufacturing Flexibility

xi

ext».—

{Ssoooooq

14
15
15
17
22
25
27
28
30
31
31
31
32
32
34
35
36
38
40
41

2.3.2 The Impact of Resource Flexibility
2.3.3 Flexibility, Manufacturing Performance, and Competitiveness
2.3.4 Summary of Research on Manufacturing Flexibility

CHAPTER 3 : DYNAMIC CELLULAR MANUFACTURING (DCM)
3.1 Introduction
3.2 Manufacturing Environment for DCM
3.3 DCM vs. Cellular and Process Layouts
3.3.1 Shop Layout
3.3.2 Routing Flexibility
3.3.3 Setup Issues
3.3.4 Part/Volume Mix (Demand Structure)
3.4 Limitations of DCM
3.5 DCM vs. Flexible Manufacturing Systems (FMS’s)
3.6 Summary of DCM
3.7 Research Statement

CHAPTER 4 : RESEARCH DESIGN
4.1 Introduction
4.2 Research Methodology
4.2.1 Statistical Issues in Simulation
4.2.1.1 Initialization Bias
4.2.1.2 Autocorrelation
4.2.1.3 Normality
4.2.1.4 Sample Size
4.2.1.5 Variance Reduction
4.2.2 Pilot Runs
4.3 Simulation Environment
4.4 Performance Measures
4.5 Experimental Stage I
4.5.1 Experimental Factors
4.5.1.1 Shop Configuration
4.5.1.2 Setup Time
4.5.1.3 Part Mix Variability
4.5.2 Simulation Environment
4.5.3 Experimental Design & Research Hypotheses
4.6 Experimental Stage 11
4.6.1 Experimental Factors
4.6.1.1 Utilization
4.6.1.2 Job Dispatching
4.6.1.3 Volume Mix Variability
4.6.1.4 Part Mix Variability
4.6.2 Simulation Environment
4.6.3 Experimental Design

43

46

47
48
50
50
51
52
53
53
55
56
57

61
62
63
63

65
65
67
67
67
68
73
74
75
76
82
82
82
83

85
85

4.7 Summary of Research Design

CHAPTER 5: EXPERIMENTAL RESULTS
5.1 Introduction
5.2 Appropriateness of Analysis of Variance
5.3 Analysis of Stage I Data
5.3.1 Introduction
5.3.2 Residual Analysis
5.3.3 Analysis of Effects
5.3.3.1 Mean Flow Time
5.3.3.2 Mean Tardiness
5.3.3.3 Mean Work in Process
5.3.3.4 Standard Deviation of Work in Process
5. 3. 4 Summary of ANOVA Results
5. 3. Analysis of A Priori Hypotheses
..5 1 Hypothesis 1
Hypothesis 2
Hypothesis 3
Hypothesis 4
Hypothesis 5
Hypothesis 6
Hypothesis 7
Hypothesis 8
Hypothesis 9
Summary of Research Hypotheses
Analysis of Secondary Performance Measures
Discussion of Stage I Results
9 Summary of Stage I Results
5.4 Analysis of Stage 11 Data
5.4.1 Introduction
5.4.2 Residual Analysis
5.4.3 Analysis of Effects
5.4.3.1 Mean Flow Time
5.4.3.2 Mean Tardiness
5.4.3.3 Mean Work in Process
5.4.3.4 Standard Deviation of Work in Process
5.4.4 Analysis of Secondary Performance Measures
5.4.5 Discussion of Stage II Results
5.5 Summary of Stage II Results
5.6 Summary of Experimental Results
5.7 Conclusions & Future Directions

MEJIM
#WN

EJIUD
\lO‘

WWUWWWNUW
M ill
on M

M
\O

5
5.
5.
5.
5.
5.
5.
5.
5.
5.
6
7
8

5..3
5..3
5..3
5..3

BIBLIOGRAPHY

85

87

87

89

89

89

91

91

96

97

103
104
104
108
108
111
114
114
117
117
120
120
123
126
139
143
144
144
145
146
146
149
159
167
167
178
184
184
185

190

Table 1.
Table 2.
Table 3.
Table 4.
Table 5 .
Table 6.
Table 7.

Table 8.
Table 9.

Table 10.
Table 11.

Table 12.
Table 13.
Table 14.
Table 15.
Table 16.
Table 17.
Table 18.
Table 19.

Table 20.
Table 21.
Table 22.
Table 23.

LIST OF TABLES

Analysis of Variance for Mean Flow Time
Treatment Means for Mean Flow Time

Analysis of Variance for Log Mean Tardiness
Treatment Means for Mean Tardiness

Analysis of Variance for Mean Work in Process
Treatment Means for Mean Work in Process
Analysis of Variance for Log Standard Deviation of
Work in Process

Treatment Means for Standard Deviation of Work in Process
Treatment Means for Mean Utilization

Treatment Means for Mean Proportion Tardy
Treatment Means for Mean Setup Time Proportion
(Mean Setup Time)

Treatment Means for Mean Queue Time Proportion
Analysis of Variance for Log Mean Flow Time
Treatment Means for Mean Flow Time

Analysis of Variance of Log Mean Tardiness
Treatment Means for Mean Tardiness

Analysis of Variance for Log Mean Work in Process
Treatment Means for Mean Work in Process
Analysis of Variance for Log Standard Deviation of
Work in Process

Treatment Means for Standard Deviation of Work in Process
Treatment Means for Mean Setup Time Proportion
Treatment Means for Mean Queue Time Proportion
Treatment Means for Mean Proportion Tardy

xi

91
92
96
97
100
100

103
104
129
136

137
138
147
148
152
153
160
161

168
169
173
175
177

Figure 1.
Figure 2.
Figure 3.
Figure 4.
Figure 5.
Figure 6.
Figure 7.
Figure 8.
Figure 9.

Figure 10.
Figure 11.
Figure 12.

Figure 13.
Figure 14.
Figure 15.
Figure 16.
Figure 17.
Figure 18.
Figure 19.
Figure 20.
Figure 21.
Figure 22.
Figure 23.
Figure 24.
Figure 25.
Figure 26.
Figure 27.
Figure 28.
Figure 29.
Figure 30.
Figure 31.
Figure 32.
Figure 33.
Figure 34.

LIST OF FIGURES

Comparison of DCM, Process and Cellular Layouts
Tradeoffs Between Flexibility and Setup Efficiency
Part Family Affiliations

Simulation Environment and Performance Measures
Configuration of Cellular Layout

Configuration of Process Layout

Routings in Process Layout

Stage I Experimental Design

Stage I Treatment Means

Stage II Experimental Design

Stage I Treatments

Tukey Multiple Comparisons of Shop Configuration by
Setup Time x Part Mix

Mean Flow Time by Setup Time x Part Mix

Mean Tardiness by Setup Time x Part Mix

Mean Work in Process by Setup Time x Part Mix
Standard Deviation of Work in Process by Setup Time x Part Mix
Treatment Numbers by Hypothesis

Tukey Multiple Comparisons for Hypothesis 1
Shop Performance for Hypothesis 1

Tukey Multiple Comparisons for Hypothesis 2
Shop Performance for Hypothesis 2

Tukey Multiple Comparisons for Hypothesis 3
Shop Performance for Hypothesis 3

Tukey Multiple Comparisons for Hypothesis 4
Shop Performance for Hypothesis 4

Tukey Multiple Comparisons for Hypothesis 5
Shop Performance for Hypothesis 5

Tukey Multiple Comparisons for Hypothesis 6
Shop Performance for Hypothesis 6

Tukey Multiple Comparisons for Hypothesis 7
Shop Performance for Hypothesis 7

Tukey Multiple Comparisons for Hypothesis 8
Shop Performance for Hypothesis 8

Tukey Multiple Comparisons for Hypothesis 9

xii

54
57
65
68
69
70
71
77
78
86
89

93

94

98

101
105
107
108
109
111
112
114
115
117
118
120
121
123
124
126
127
129
130
132

Figure 35.
Figure 36.
Figure 37.
Figure 38.
Figure 39.
Figure 40.
Figure 41.
Figure 42.
Figure 43.
Figure 44.
Figure 45.
Figure 46.

Figure 47.

Shop Performance for Hypothesis 9

Summary of Hypothesis Conclusions

Stage H Treatments

Tukey Multiple Comparisons for Log Mean Flow Time
Mean Flow Time by Part Mix x Volume Mix

Tukey Multiple Comparisons for Log Mean Tardiness
Mean Tardiness by Dispatching Rule x Part Mix

Mean Tardiness by Dispatching Rule x Volume Mix

Tukey Multiple Comparisons for Log Mean Work in Process
Mean Work in Process by Dispatching Rule x Part Mix
Mean Work in Process by Part Mix x Volume Mix

Tukey Multiple Comparisons for Log Standard Deviation of
Work in Process

Standard Deviation of Work in Process by Dispatching Rule
x Part Mix x Volume Mix

xiii

133
135
145
149
150
154
155
157
162
163
165

170

171

CHAPTERI

INTRODUCTION AND PROBLENI STATEMENT

1.1 INTRODUCTION

Between 60 and 80% of manufacturing takes place in a batch production environment
(Chevalier, 1986). However, recent changes in the batch manufacturing environment
have brought into question the use of the process layouts or job shops that are often used
for this type of production. According to Hyer (1982), 60-80% of manufacturing in the
US takes place in shops that utilize a process layout. Such shops, organized as groups
of functionally similar machines, provide routing flexibility by allowing any available
machine of the required type to be used to process a part. In this environment, the ratio
of setup times to processing times is typically large, making the use of frequent setups
inefficient. In addition, jobs often encounter long delays waiting for machines to become
idle and then be setup. Jobs spend as much as 95% of their flow times in queues and in
transit between machines (Tersine, 1985). To reduce these inefficiencies, batch ‘sizes are

generally large, increasing work in process and finished goods inventories.

Increasingly however, demand is characterized by greater variety, lower volumes,
frequent design changes and short cycle manufacturing. Demand is also more uncertain
than in the past. These factors make reductions in lead times and batch sizes essential to
improve responsiveness and competitiveness. This has been demonstrated by the

performance of manufacturing systems based on the just in time philosophy. Quality and

2

reliability are also major considerations given the increased level of competition. Greater
diversity in materials, and higher material and energy costs add to the need for a more

efficient production system.

Given the limitations of job shops, considerable attention has been paid recently to
Cellular Manufacturing (CM). CM is one facet of Group Technology (GT). This is a
philosophy of production based on the principle of exploiting similarities in part design
and manufacture. CM specifically deals with the manufacture of families of parts, parts
that have been grouped based on processing similarity, within manufacturing cells, sets
of machines which have been dedicated to the manufacture of specific families. The
suggested gains from this are:

Improved control and monitoring of jobs, since they move within a limited
' physical area, and follow a clearly defined route ‘

Shorter lead times and work in process inventory, since the benefits of a flow line
can be attained

Faster quality feedback, since with the use of a flow line, the source of a problem
can be detected more quickly

Lower setup times and tooling requirements due to the greater homogeneity of
parts produced within a cell

Smaller lot sizes due to the ability of jobs to share setups
Learning benefits due to repetitive processing and worker specialization

Increased operator satisfaction due to the enlargement of job assignments to cover
family processing.

3

Despite the potential however, the evidence increasingly suggests that CM is not the
solution to the problems of small batch manufacturing. The anticipated gains from its use
are more than offset by its limitations. The cost of physically reorganizing the shop floor
and adding new equipment which often results, is high. In addition, there is a cost
associated with lost production during the reorganization. Since reorganization is based
on existing demand patterns, it may not be possible to absorb subsequent changes in
demand without additional reorganization and/or investment. The dedication of
equipment to cells reduces routing flexibility. Since cells are dedicated to individual
families, demand patterns that create uneven cell workloads result in bottlenecks in some
cells, and idleness in others. This leads to long queues and consequently increases the
mean and variance of flow time. The evidence suggests that the range of conditions when
a cellular layout performs at a comparable level to a process layout, is extremely limited.
In summary, although production control and focus might be improved by effectively
creating plants within plants consistent with Skinner’s concept of focus (1974), shop

performance is poorer.

1.2 PROBLEM STATEMENT
Based on the existing evidence, a need exists for greater manufacturing flexibility.
Swamidass (1988) defined flexibility as

"the capacity of a manufacturing system to adapt successfully to changing

environmental conditions and process requirements” and ”the ability of the
production system to cope with the instability induced by the environment. "

4

While traditional process layouts offer a high degree of flexibility, they are otherwise
inefficient for part family production. CM improves efficiency, but at the cost of reduced
flexibility. As an alternative to these extremes, this research operationalizes a
manufacturing system for small/ medium batch family production that offers a higher
degree of flexibility than traditional CM without significantly compromising its
efficiency. The system pr0posed is a hybrid between CM and job shop production that
takes advantage of the benefits of each, yet attempts to minimize or eliminate their
individual limitations. This is accomplished by manufacturing in an environment
characterized. by dynamically formed manufacturing cells, or Dynamic Cellular
Manufacturing (DCM). It is envisioned that DCM will enable contemporary production
needs to be met more effectively, by better managing the apparent tradeoff between
flexibility and setup efficiency. Furthermore, higher flexibility in DCM is obtained

without necessitating the high capital investment as in a FMS.

DCM focusses on current manufacturing trends, for example shorter product life cycles,
frequent product revisions, and new product introduction. It does this by using the
principles of CM, but without the physical shop reorganization. Scheduling mechanisms
are used that recognize part family affiliations, and based on these, temporarily dedicate
machines. The result is the formation of logical production cells based on need, that exist
only as long as the need prevails. Since no permanent machine dedication is involved,
the underlying flexibility of a process layout is maintained while simultaneously

establishing the flow pattern dominance and family orientation of CM. This facilitates

5

lower flow times and work in process, reduced bottlenecks, and more balanced
utilization. Since there is no physical shop reorganization, there is no financial cost
associated with the introduction of DCM, nor a need for either a total or partial shutdown
of the production system. In addition to providing an alternative to traditional cellular and
job shop manufacturing, DCM provides a vehicle for testing the appropriateness of CM

in a given production environment.

This study compares the performance of DCM to that of a traditional job shop, and a
shop organized using the principles of CM. The objective is to identify whether DCM
ean meet the needs of small/ medium batch production more effectively, and if so, under
what conditions. The questions to be addressed are :

a. Do setup conditions exist where DCM’s use of the part family concept and
efficient use of setups, is more beneficial than the flexibility of a traditional
job shop. If so, what setup conditions are conducive to DCM.

b. Can the recognition of part families by DCM, make it more effective in
dealing with different part mix compositions. If so, for what part mix
characteristics is DCM preferable.

c. Can the greater flexibility of DCM allow it to overcome the setup efficiencies
of permanent machine dedication in traditional CM, and if so, under what

setup conditions.

d. Does the greater flexibility of DCM make it more responsive to changes in
part mix than traditional CM, and if so, under what part mix conditions.

e. Does the information used to form dynamic cells affect their performance.
f. Does shop load have a significant impact on the effectiveness of DCM.
g. Is DCM sensitive to changes in job size.

h. Does job dispatching affect the performance of DCM.

6

A simulation model is used to compare the behavior of the three systems under various

sh0p conditions, and to provide answers to these questions.

1.3 ORGANIZATION OF DISSERTATION

Chapter two discusses the literature that addresses the use and design of CM systems,
evidence of their performance, and mechanisms used to improve their effectiveness. This
will illustrate the shortcomings of CM for small/medium batch production, and address
the need to incorporate greater flexibility into existing cellular production methods.
Chapter three describes in more detail, the concept of DCM, how it differs from existing
production methods, and the specific issues and questions to be addressed by this
research. Chapter four describes in detail the experiments carried out, the methodology
used, and the techniques used to answer the research questions. The results of the
experiments and analysis of their implications are discussed in Chapter five, followed by

a discussion of the conclusions of the study and directions for future research.

CHAPTER2

BACKGROUND AND LITERATURE REVIEW

2.1 INTRODUCTION

Considerable attention has been paid recently to improving the performance of
small/ medium batch production systems. One attempt to deal with this problem has
focussed on the principles of GT. One aspect of GT that has received particular attention
is CM. A number of articles have discussed the advantages and disadvantages of CM
(e.g., Greene & Sadowski, 1984, Suresh & Meredith, 1985). However, though case

evidence suggests gains from the use of CM, a larger body of literature refutes this.

This chapter discusses the literature that has addressed the merits of CM and compared
its use to that of a process layout. Mechanisms that have been proposed to overcome
CM ’5 limitations are identified and their impact examined. The results will demonstrate
the nwd for a new approach to small/medium batch manufacturing. The literature on the
formation of manufacturing cells is also discussed. This is an important issue in CM

since cell formation is a key component in CM system design.

The chapter concludes by addressing the issue of manufacturing flexibility. As the
evidence will show, it is CM ’3 lack of flexibility in contrast to that of a process layout,
’ that is the driving force behind the nwd for an alternative approach to small/medium

batch production.

8
2.2 RESEARCH ON CELLULAR MANUFACTURING (GROUP TECHNOLOGY)

2.2.1 ° ' ll 1 ' rf
Evidence exists, primarily from surveys of CM users, that its use can significantly
improve shop performance compared to that of a process layout (Hyer, 1982). However,

several questions exist regarding the validity of these observations.

Leonard & Rathmill (1977) concluded that comparisons between cellular and process
layouts are often made between efficiently organized and operated cellular layouts, and
process layouts that were not optimized in the same way. This makes direct comparison
of their performance inconclusive. Their research also suggested that contrary to
expectations, CM yields lower utilization, more complex production control, and reduced
job satisfaction. Craven ( 1977) drew attention to the fact that gains from the use of CM
depend on a myriad of design considerations, constraints, and trade-offs. Flynn & Jacobs
(1986) noted that during the time it takes to implement CM systems, other variables such
as product mix, can change. This again makes comparison with the original process

layout inappropriate.

Wemmerlov & Hyer (1987) stated that performance improvements arising from the
conversion from a process to a cellular layout, do not reflect the cost and effort involved.
In addition, CM typically co—exists with machines organized on a functional basis. Any
gains from the use of the manufacturing cells may come at the expense of the functional

component of the shop. They also suggested that one of the reasons for poor CM

9

performance is the use of poor routing information in cell design. Morris & Tersine
(1989) indicated that when CM is introduced, it is frequently accompanied by the
addition of new equipment. This new equipment might have improved the performance

of the existing process layout.

As well as the evidence from industry, several simulation studies have also concluded
that CM performance is not superior to that of process layouts, and is in fact inferior on

several dimensions.

2.2.1.1 W

Cummings (1980) study is the only one to consistently support the use of a cellular
layout. It compared the performance of cellular and process layouts under different levels
of utilization, and with no and low labor absenteeism. Results showed that absenteeism
did not yield poorer performance when using a cellular layout except at very high
utilization. However, when using the process layout, performance decreased as
absenteeism increased at all levels of utilization. With or without absenteeism, the
cellular layout performed considerably better, particularly as utilization increased.
Despite these findings, the lack of recognition of family and setup characteristics, makes

their value limited.

Flynn (1984), Flynn & Jacobs (1986, 1987) conducted several comparisons of process

and cellular layouts. They compared four layouts; a process layout, a process layout with

10
machines dedicated to specific parts, and two cellular layouts that differed in how

machines were organized. They showed that although the shops using dedicated machines
yielded improved performance with respect to setup, utilization and material handling,
they performed poorly for most other measures, in particular flow time and queue related
measures. Of the layouts using dedicated machines, the process layout performed the best
for queue related measures. They concluded that it was the dedication of machines rather

than the layout itself that led to differences in performance.

Flynn (1984) also considered two alternate routing strategies as a means of improving
the performance of the cellular layouts. According to the first strategy, jobs had at most
one alternate machine for a given operation, the machine that was physically closest to
the primary machine. Jobs were rerouted to the alternate machine if there were more
than twenty parts in the queue of the primary machine. Although this strategy generally
led to improved performance, performance was still inferior to that of a process layout.
The second strategy allowed work to be rerouted to any similar machine when the total
work content at the primary machine exceeded a certain level, defined in terms of
numbers of days of work. The results again showed that shop performance improved, but
only when cells were designed based on material flows between machines and not cells.
Larger critical queue lengths yielded better results, by increasing the accumulation of
jobs in a queue and increasing the potential to share setups. Though this strategy
outperformed the first, the improvement was again not enough to make the performance

of the cellular layout comparable to that of a process layout.

11
Morris (1988), Morris & Tersine (1990) also found that the use of CM led to a

degradation in shop performance. They sought to identify operating conditions that are
conducive to CM. Results showed that a cellular layout performed at a comparable or
superior level to a process layout only when setup and material handling times were very
high, demand patterns stable, and job flow unidirectional. The difference in flow times
under other conditions was large enough that even with a ten fold increase in material
handling time, the process layout, which does not lend itself to material handling

efficiencies, had lower flow times.

Suresh (1992) demonstrated using queuing theory that partitioning a job shop into cells
necessarily leads to a decrease in flow time and work in process performance, and lower
utilization. Only if setup times within the cells are significantly reduced can the cells
generate improvements in performance. These conclusions were tested using a simulation
study of a process layout and two cellular layouts that differed in size and number of
families processed within each cell. Based on the batch size that yielded the most
efficient process layout operation, the results showed that under high setup times, the
cellular layouts were unstable at this or smaller batch sizes. Bottlenecks at even a single
machine were enough to render the entire cell unstable. Reductions in setup times within
the cells were necessary to induce stability. Further reductions enabled the cellular
layouts to outperform the process layout. Using larger batch sizes, the cellular layouts
were initially stable but yielded considerably poorer performance than the process layout

operating with the optimal batch size. The performance of the cellular layouts again

l2

improved as setup times within the cells were reduced. When cells were larger, flow
time performance was always better due to the reduction of bottlenecks. However,

utilization was unchanged due to the increase in setup frequency.

The ability to move work between cells was also shown to yield improvements in
performance. Inter—cell movement in the shop with small cells induced stability under
conditions that were previously unstable. Performance improved but was still poorer than
that when using the process layout. When batch sizes were small, smaller reductions in
setup time were needed to obtain stability in the shop and to generate improved
performance compared to that of the process layout. However, for large reductions in
setup time, performance began to deteriorate as move times more than offset the gains

from lower setup times.

2.2.1.2 W
Hyer (1982) found that in shops using CM, none was organized entirely as cells. Shops

. contained a combination of cells and machines organized by function. Most users
produced at least 45 96 of parts outside the cell. Burgess (1988) suggested that
determining the extent to which cells should be used in such a hybrid layout was of

greater importance than determining whether they should they be used at all.

Christy & Nandkeolyar (1986) investigated the percentage of jobs that must be completed

within the cellular component of the shop in order for it to outperform a pure process

13
layout. They showed that for percentages between 17.5 and 22.5 % , the hybrid yielded

lower flow times. Consistently however, the hybrid layout yielded higher mean tardiness.
Utilization was generally higher within the cellular component of the shop. For the
optimal proportion of jobs passing through the cell, the performance of the hybrid
improved when setup, operation and material handling times were reduced by 60-65 % ,

though greater reductions were needed to yield improvements in tardiness performance.

Using a labor constrained environment, Burgess (1988) showed that contrary to Christy
& Nandkeolyar’s results, the proportion of jobs passing through the cell in the hybrid
layout had to exceed 40% in order for the hybrid layout to outperform the traditional
process layout. This was true even when setup times in the process layout were reduced
by as much as 25%. Reducing the proportion of jobs passing through the cell had a
significant impact. For example, the process layout outperformed the hybrid layout if
only 30% of jobs passed through the cell even when setup times in the cell were reduced
by 90%. Burgess concluded that in such hybrid layouts, the relative allocation of
resources between the two components of the shop has a significant impact on shop

performance.

Queuing theory was used by Suresh (1991) to explain why shop performance must
decrease when a job shop work center is partitioned into the kind of hybrids described
above. Flow time and work in process were shown to increase due to the increase in

queues that result from machine dedieation. Smaller batch sizes were also shown to be

14

infeasible within the cellular component of the shop. When setup times within the cellular
component were reduced, not only were smaller batch sizes feasible, but the performance
of this component of the shop was better than that of the original un-partitioned work
center. However, performance of the shop as a whole was inferior due to the negative
impact on the functional component of the shop. Conditions for effective partitioning
were defined and confirmed using a simulation model. The results reiterated the
potential for machine dedication at low batch sizes if setup times are reduced, but also
highlighted the need to resolve the difficulties this creates within the functional

component of the shop.

2.2.2 r ' n in 11 l f ‘
Crookall & Lee (1977) and Lee (1985) showed that CM systems that utilized large cells,

few families, and small batch sizes, generally yielded better performance than those that
did not. The presence of multiple servers in large cells more than offset any increases
in setup frequency. Fewer families decreased the need for setup changes. Smaller batch
sizes allowed jobs to move through the shop faster, though the resulting increase in
setups caused increases in utilization. A similar study was carried out by Gupta &
Tompkins (1982) who used a simulation model to examine the tradeoffs between cell
size, material movements, and number of inter-cell moves, and between batch size and
setup times. Though intra-cell moves increase with larger cell size, these are generally
preferable to inter-cell moves. They showed that as expected, larger cells yielded fewer

inter-cell moves. Karmarkar et al. (1985b) showed analytically and using a simulation

15

model that reductions in batch size reduced flow time and work in process. However,
batch sizes that were too small yielded more frequent setups, increased queues, and
poorer performance. Additional studies such as Sinha & Hollier (1984), and Wemmerlov
& Hyer (1987), have addressed additional issues that can affect the performance of CM

systems, but which have yet to be examined.

2.2.3 ' n 11 l f ri

Given the limitations of CM particularly with respect to flexibility, a number of
approaches have been suggested to improve its performance. These fall into three main

areas; group scheduling, repetitive lots, and alternate routing.

2.2.3.1 W
According to Mosier & Taube (1985), group scheduling is the least addressed topic

related to CM. It refers to scheduling rules that exploit family processing similarities,
primarily in setups, between jobs in a queue. Similar sequencedependent scheduling

rules have been used in job shops where part families were not explicitly considered.

One stream of group scheduling research is analytic in nature. A number of these studies
consist of optimal family and job sequencing rules in a single machine facility. Hitomi
& Ham (197 8) used mathematical programming to maximize production rate. Foo &
Wager (1983) developed a dynamic programming formulation to minimize setup time for

a single part family with sequence-dependent job setup times. Ozden et a1. (1985) used

l6

dynamic programming to minimize total setup cost where job and family setup times

were sequence—dependent.

Analytic models have also been applied specifically in a cellular environment. A branch
and bound solution to family and job sequencing was used by Hitorrri & Ham (1977) to
minimize makespan. A similar approach was used by Ham et al. (1979) to minimize
makespan with the minimum number of tardy jobs. Sundaram (1983) developed two
static heuristics to minimize makespan. The first heuristic selects the family to be
processed next on a machine based on earliest family completion date at that machine.
Within a family, a Gantt chart is used to schedule jobs to minimize makespan. According
to the second heuristic, the family is selected based on shortest family processing time.
Only on completion of family processing at the machine is the family loaded at the next
machine. The second heuristic was shown to perform the better of the two, yielding an

optimal solution for the data set used, though optimality is not guaranteed.

Most of the group scheduling research specific to cellular environments consists of
heuristics applied in simulation studies of single job shop and flow shop cells. They are
generally scheduling rules that use different criteria to select a family for processing, then
process all jobs in the queue from this family prior to resetting the machine. This takes

advantage of the similar setup needs of jobs from the same family, thereby reducing

setup frequency.

17
2111.1 MW
Vaitlrianathan & McRoberts (1982) defined five heuristics for family selection. These

consider lowest slack/processing time ratio, lowest family slack, highest setup time,
lowest setup time, and highest similarity of jobs (using similarity coefficients). Within
a family, jobs are dispatched using the shortest processing time (SPT) rule. Compared
to scheduling based on SPT alone, the heuristics yielded lower flow times and number

of setups per job. However, due date performance was very poor.

Mosier (1983) and Mosier et a1. (1984) compared three mechanisms for family selection.
These select families based on highest family work content (WORK), highest average job
priority (AVE, five dispatching rules were used to prioritize jobs), and the economic
benefit of changing the current setup or continuing to use the existing setup (ECON).
This rule makes it possible to change the setup even though jobs remain that require the
existing setup. They showed that overall, WORK yielded the best performance followed
by ECON. They also showed that family rules performed well with respect to mean flow
time and mean lateness, but not for mean tardiness and percent tardy. Dispatching using
either the SPT or minimum slack rule generally yielded the best results. Kelly at. al
(1986) compared WORK and ECON to two cost based family selection rules. They

showed that cost based heuristics performed poorly for flow time and tardiness measures.

Flynn (1987) applied the repetitive lots (RL) procedure (Jacobs & Bragg, 1988) in a

multi-cell shop. This procedure, designed to minimize setups in job shop scheduling, is

l8
equivalent to FCFS family selection. Within a family, jobs are also dispatched using the

FCFS criterion. Flynn also considered a truncated form of RL that limits the number of
batches processed with the same setup, in order to prevent long waits for jobs from other
families. The study compared the use of a cellular layout using both forms of RL, with
a process layout, a process layout with machines dedicated to particular parts, and a
cellular layout using FCFS dispatching alone. The results showed that in the cellular
shop, both forms of the RL rule, though indistinguishable in their performance,
outperformed FCFS dispatching for all performance measures. The relative performance
of the three layouts when RL was used was similar to that of Flynn & Jacobs previous
works (1986, 1987). The shops with dedicated machines performed better with respect
to setup time and utilization, but poorer than the pure process layout on queue related
measures and flow time. However, the difference in performance between the cellular
and pure process layouts was significantly smaller. No differences in performance existed
between the two layouts with dedicated machines. This further suggests that machine

dedication rather than layout, has a more significant effect on performance.

Mahmoodi et al. (1990) considered three family selection rules. FCFAM selects the
family containing the first job in the queue, DDFAM the family containing the job with
the earliest due date, and MSFAM the family that minimizes future sequence-dependent
family setups. They showed that MSFAM and DDFAM performed well for most
performance measures. MSFAM performed poorly only for mean tardiness. Their

relative performance depended on other conditions such as load and setup time/run time.

19
They also showed that dispatching jobs within a family using either the SPT rule or a

processing time/ slack hybrid (SI‘), yielded the best performance, similar to the findings
of Mosier et al. (1984). Mahmoodi et a1. (1988) also showed that the use of DDFAM
and FCFAM in conjunction with SPT and FCFS job dispatching, always yielded superior

flow time and tardiness performance than scheduling using the dispatching rules alone.

Mahmoodi & Dooley (1991) compared DDFAM and MSFAM, which they categorized
as exhaustive rules, to two non-exhaustive rules that do not require all jobs using the
current setup to be processed prior to a setup change. SLFAM processes jobs until the
total slack of another family in the queue becomes negative. The setup is then changed
to that of the more urgent family. If more than one family in the queue has negative
slack, the family selected is that with the most jobs in the queue. DKFAM processes jobs
until the time remaining until due date of the most urgent job, is more than C units
greater than that of the most urgent job in another family in the queue. The setup is then
changed to facilitate the new family. They showed that MSFAM always yielded the best
flow time performance, and DKFAM the worst. DKFAM always performed as well as,
if not better than the other rules for mean tardiness. MSFAM and SLFAM performed
poorly. Proportion tardy was lowest for MSFAM, and highest for DKFAM. They again
showed that dispatching using either the SPT or 81‘ rules yielded the best overall
performance. They concluded that although exhaustive rules as expected, generally

perform better, there are benefits associated with non—exhaustive rules.

20
Ruben et a1. (1993) examined factors that affect the performance of family scheduling

heuristics. They showed that a rule that selects families based on minimum setup time
and dispatches jobs based on SPT (MSSPT) performed better than existing family and
non-family based scheduling rules. Significant gains in (flow time performance were
obtained by using group scheduling rules, though these were smaller for mean tardiness.
The extent of gains depended on shop conditions, increasing with high utilization,
setup/processing time ratio, and stable demand patterns. Though the DDSI‘ rule, which
selects families based on most imminent job and dispatches using the 81" rule yielded the
best due date performance, scheduling using 81" alone performed well, though this again
depended on other conditions. MSSPT again performed best for proportion tardy, and
FCFS based family scheduling for lateness. The authors concluded that by minimizing
setups, group scheduling rules are in general more robust to shop load and
setup/processing time ratio. However, they are also responsible for increases in tardiness

when part families are large, by discriminating against parts from smaller families.

Wemmerlov (1992) conducted a comprehensive analysis of farmly and non-family based
scheduling on a single machine. The study considered two family based rules, one
equivalent to FCFAM, and one that selects families {and jobs based on SPT. These were
compared to scheduling using FCFS and SPT alone. The results showed that as the
number of families was decreased, the resulting reduction in number of setups yielded
improved flow time performance for all rules. However, at low setup times, the use of

the SPT based family rule resulted in an increase in flow time. As the impact of setup

21
times became less significant, the difference between this rule and the SPT dispatching

rule increased. Unlike past studies, this demonstrated that when setup times are low,
dispatching rules can outperform their family based counterparts. When demand was
biased towards specific families, flow times were as expected lower, particularly when
setup times were high and there were few families. Again, this was attributed to the
increased ability to reduce setup frequency. Decreases in setup time were again shown

to yield reductions in flow time, particularly at high utilization levels.

The research demonstrated the benefits of reducing variance. When processing time and
arrival rate variance were reduced, the mean and variance of flow times were lower, as
was the difference in performance between job and family based scheduling. In addition,
greater benefit was obtained from reducing arrival rate variance, though failure to reduce
processing time variance did on occasion lead to a degradation in performance of the two
- SPT based rules. The benefits of family scheduling rules also depended on the stability
of the environment. When the environment was unstable, FCFAM performed better than
FCFS. However, the SPT based family rule performed better than SPT only when
processing time variance was reduced. Under unstable conditions, family scheduling rules
were able to generate significant capacity increases and simultaneously improve flow time
performance. In addition to instability, conditions of high utilization, setup times, and

few families were shown to be most conducive to family based scheduling.

22

22.3.1.2 WW

Hitomi et al. (1977) compared two group scheduling heuristics to traditional dispatching
rules in a flow shop, job shop, and flow shop where flow patterns differed by family
affiliation. The two heuristics select families based on minimum family setup time, and
on the travelling salesman problem, where minimization of the sequence-dependent
family setup times is the objective. Both rules performed well with respect to flow time
measures, though not necessarily outperforming rules that did not recognize setups.
However, their relative performance improved when utilization was high and the
setup/processing time ratio large. For large ratios, they yielded the best performance.
The interaction of utilization and setup/processing time ratio had a similar effect on flow
time measures as the setup/processing time ratio. Heuristic methods were also used by
Manivannan et al. (1987), and Abin & Mohamed (1987). Manivannan et al. developed
a rule to minimize mean flow time given the optimal makespan. Abin & Mohamed

developed a rule to minimize total setup time.

Wemmerlov & Vakharia (1992) compared a number of dynamic and static job scheduling
rules to their family based counterparts. They confirmed that for each job based rule, the
corresponding family based rule yielded superior performance. Of the family based
procedures, minimum slack and FCFS based family selection generally yielded the best
flow time and tardiness performance when used with FCFS dispatching. The performance
of the FCFS based family selection rule is contrary to its performance in a job shop cell

(Mahmoodi et al., 1990). As a group, family based rules outperformed job based rules

23

though there was little discrimination between individual rules. In less that 40% of cases
did the best job based rule perform significantly poorer than the worst family based rule.
The gains from using family oriented rules were greater when utilization was high. This
concurs with the results of Hitomi et al. (1977). Contrary to the evidence on job shop
cells (e.g., Wemmerlov, 1992, Ruben et al., 1993), the number of part families did not
affect the relative performance of job and family rules. This was significant only when
both utilization and setup times were high. Similarly, the ratio of setup time to processing

time was generally insignificant.

Mahmoodi et al. (1992) compared four family based scheduling rules to FCFS and SPT
dispatching. They examined their performance under different shop load,
setup/processing time ratio, due date tightness and inter-arrival time distribution
conditions. Their results showed that for mean flow time, the MSSPT rule (Ruben et al. ,
1993) always yielded the best performance. Under certain conditions, this was matched
by a rule that selects the next family based on job slack and dispatches jobs using 81".
This rule was shown to perform well in a job shop cell (Mahmoodi et al., 1990). As
expected, this rule consistently performed best for due date measures. These two rules

were also shown to be the most robust to changes in shop environment.

The ECON rule that had been shown to perform well in a job shop cell (Mosier, 1983,
Mosier et al. 1984), performed poorly. It also proved to be the least robust of the family

rules used. Contrary to Wemmerlov & Vakharia (1992), FCFAM also performed poorly.

24

The research indicated that for each performance measure, differences in performance
between family heuristics were small when utilization was low, but increased at higher
utilization. Consistent with past research on flow shop cells (Wemmerlov & Vakharia,
1992), and job shop cells (Mahmoodi et al., 1988), group scheduling rules yielded
superior performance than their corresponding dispatching rules, particularly when

utilization and variance were high.

Russell & Philipoom (1992) examined the effect of due date setting procedures on the
performance of family scheduling rules. They showed that a procedure that considers
how many setup changes occur before a job is processed, consistently yielded the best
performance. They also demonstrated the importance of selecting scheduling mechanisms
in conjunction with the due date setting procedure particularly when setup times were
high. Overall, they showed that for flow time, the best scheduling rule was one that
selects families based on lowest processing time per job, and dispatches jobs using the
SPT rule. For this heuristic, no due date setting mechanism dominated. Mosier (1983),
- Mosier et al. (1984) showed that in a job shop cell, family selection based on average
family priority tended to perform poorly. For other heuristics, relative performance did
depend to a greater degree on which due date setting procedure was used, but again,

none dominated.

For due date performance, the relative performance of the rules depended on setup times.

When setup times were high, three heuristics generally performed best; FCFS family

25
selection with slack based job dispatching (FCFS-SLK), due date dispatching where jobs

not belonging to the current family have a constant added to their due date to penalize
additional setup changes (EDD-T), and a rule that requires a setup change after a fixed
amount of time has elapsed in addition to when no jobs with the current setup remain
(Sawicki, 1973). For this rule (SAW-T), the authors showed that the best family and job
selection rule depended on the performance measure of interest and the due date setting
mechanism used. Overall however, FCFS family selection and slack based dispatching
performed best. When setup times were low, performance was generally best for the

EDD-T rule.

2.2.3.1.3 W

In addition to the group scheduling literature, limited research exists on sequence-
dependent scheduling in job shops. However, Wemmerlov (1992) makes the distinction
that unlike group scheduling rules that attempt to avoid setups, sequence-dependent
scheduling rules typically consider only that changeover times are dependent on the

existing setup, and do not explicitly try to avoid setups.

Gavett (1965) considered the use of a scheduling rule that processes the job with the
lowest setup time relative to the job just completed, as well as two variants of this rule.
For a finite number of jobs, he showed that these rules performed significantly better
than random rules, but were frequently not optimal. The gains from the use of these rules

as well as their relative performance, depended on parameters such as distribution and

26

variance of setup times, and batch size. Haynes et al. (1973) examined how these
parameters caused the heuristics to perform differently to an optimal sequence. They
showed that the rules yielded results closer to optimality when either a gamma or normal
distribution was used for setup times, and when batch sizes were small. The use of a

uniform distribution led to poor results.

Hollier (1968) compared a dispatching rule that selects the next job based on its using
the current setup, to dispatching rules that do not recognize setups. This rule was shown
to perform well for a number of measures, sometimes outperforming the other rules.
Wilbrecht et al. (1969) evaluated three sequence-dependent scheduling heuristics. These
select the job with the lowest setup time relative to that of the job just completed, the job
with the lowest process time (setup time plus run time), and the job with the highest
process time. They showed that for a number of performance measures, these rules
performed as well as or better than rules that do not consider sequence dependencies.

The first two rules exhibited particularly good performance.

White & Wilson (1977) developed a regression model that allows setup times to be
predicted, based on the assumption that actual setup times are not always known. Given
these predictions, a heuristic was used to sequence jobs in order to minimize the number
of more time consuming setups, and total setup time. This heuristic was shown to

generate good results even though actual setup times could not always be used.

27
The Repetitive Lots (RL) procedure (Jacobs & Bragg, 1988) proposes splitting jobs into

transfer batches smaller than their original release quantity. This promotes more efficient
material flow and scheduling by allowing transfer batches to move independently. At
each machine, transfer batches using the existing setup are processed first on a FCFS
basis. When no such jobs remain, the setup is changed to that required by the. first
remaining job in the queue. This is similar in principle to the FCFAM family scheduling
rule. The use of RL was shown to yield significant improvements in flow time
performance over a range of release and transfer batch sizes. In addition, for smaller

release batch sizes, it induced stability in shops that were previously unstable.

2.2.3.2 W
Lot splitting has been to used in CM to improve the efficiency of material handling and

setup use similar to the principles of repetitive lots. Morris & Tersine (1989) considered
splitting jobs into transfer batches of size one in conjunction with the use of cell loading
(Mosier, 1983). A cell using cell loading processes a single job at any given time, unlike
the more common machine loading, where a number of jobs compete for machines.
Mosier showed that cell loading yielded low utilization and poor performance. Morris
& Tersine however applied cell loading in conjunction with transfer batches. Their results
indieated that at low utilization, cell loading yielded performance superior to that of a
process layout and a cell using machine loading. However, as utilization increased, the
shop using cell loading was more sensitive to increased congestion. The less efficient use

of machines resulted in performance that was inferior to either of the two other layouts.

28

Sassani (1990) showed that reducing transfer batch size led to reductions in setup time
and proportion tardy. However, the performance of individual cells was sensitive to the

processing characteristics of jobs processed within them.

2.2.3.3 W
Alternate routing has been proposed as a means of reducing problems in CM of

bottlenecks and imbalances in cell utilization, by rerouting jobs from overloaded
machines to less busy machines in other cells. However, this does result in an increase
in inter-family setups and the complexity of material handling. Widespread use of
alternate routing makes the operation of CM similar to that of a process layout, since the
material handling and setup benefits of CM are lost. Typically 20% of parts encounter
some inter-cell movement in practice (Wemmerlov & Hyer, 1987), though this may not

be attributable solely to alternate routing.

Ang & Willey ( 1984) considered several alternate routing heuristics. In addition they
considered routing work to idle machines from those that were not necessarily congested,
in order to balance loads. Their results indicated that a number of these heuristics led to
improved performance. In particular, a rule that transfers jobs from their primary
machine if average workload at the machine is greater than a critical value, and sends
them to the alternate machine with the lowest average workload that can process the job
immediately, showed the greatest improvement. Mean flow time, standard deviation of

lateness, and mean tardiness, all improved. However, performance gains decreased as

29

the amount of rerouting increased. They also considered the impact of returning a
transferred job to its primary machine after it had been processed at an alternate
machine. Though this also led to performance improvements, these were not as large.
Alternate routing led to performance gains regardless of which dispatching rule was used,
shop configuration (i.e. , number of cells, cell size), changes in product mix and demand
patterns. The results showed that simple heuristics can yield significant performance
improvements if used sparingly, and that the gains are more the result of balanced
workload than rerouting itself. However, the time involved in rerouting was not

explicitly considered nor was any comparison made to a process layout.

Garza (1990) and Garza & Smunt (1991) showed that limited alternate routing enabled
CM to outperform a process layout. They showed that CM performance could better that
of a process layout when batch sizes were small, setup times high, and run time variance
low. They also showed that small ratios of minor (intra-family) to major (inter-family)
setup time led to improved CM performance by increasing the impact of fewer major
setups. In addition, they showed that when the impact of material handling in the process
layout was large, the use of alternate routing was consistently beneficial, regardless of
the extent of its use. Alternate routing was also examined by Flynn (1984) and Suresh
(1992) in studies that compared CM using alternate routing to the use of process layouts

(Section 2.2.2.1).

30

Though the literature on alternate routing in CM is limited, considerable evidence of its
benefit in a job shop exists, e.g., Wayson (1965), Russo (1965), Goodman (1972), Tilak
(1978), and Khatour & Moodie (1979). Bobrowski & Mabert (1988) investigated the
effect of adding routing flexibility at the process planning stage. They showed that
increased routing flexibility led to performance benefits, but that these followed the law
of diminishing returns. With additional flexibility, a tradeoff exists due to the increased

tooling and fixtures required.

2.2.4 Cell Eumau'uu fI'fihnigugs in Cellular Manufacturing

In addition to the literature on the performance of CM, a signifieant body of research has
examined the cell formation process. Cell formation involves the grouping of parts into
families based on production similarity, and allocating machines to individual families
to form cells. The current research is not concerned with traditional cell formation since
permanent cells are not formed. However, the separation of parts into families is

important since the existence of part families is the basis for forming dynamic cells.

A number of taxonomies exist for classifying approaches to cell formation (e.g. ,
Wemmerlov & Hyer, 1986, Vakharia, 1986). These represent comprehensive surveys of
the cell formation literature. The approach taken here is to briefly summarize some of

the more significant contributions using a framework similar to that of Vakharia.

31
2.2.4.1 MW

2.2.4.1.1 W
Wemmerlov & Hyer (1986) cited implementations of CM where part families were

formed based on part name or function, i.e. , a valve manufacturer might treat a valve
stem as a part family. Categorization by visual inspection of shape or size is another
means cited. Part coding, an important component of GT, ean also be used to identify

similar parts, by capturing shape, size and machining characteristics.

214.12 mm

Burbidge’s (1975) Production Flow Analysis (PFA) uses the part/ machine matrix that
defines production requirements, and by manual rearrangement, obtains clusters of
mutually exclusive part/ machine groupings along the diagonal. Groupings that do not
yield precise partitions are used as the basis for the final cell configuration. Similar
approaches have been suggested by El Essawy & Torrance (1972), de Beer et al. (1976),
Malik & Dale (1977), and de Beer & de Witte (1978). El Essawy & Torrance’s
Component Flow Analysis sorts parts twice, based on the order in which they use
machines and the minimum number of machines required. This sorting is the basis for
forming machine groups, taking into consideration machine, part and shop constraints.
Cells are formed around groups requiring the most machines. Detailed analysis of within
cell flow patterns is earried out to ensure a feasible design. de Beer et al. and de Beer
& de Witte defined Production Flow Synthesis in which families are formed in a similar

manner to PFA, but operations defined also in terms of the number of machines that can

32
be used to process them. Machine clusters are formed depending on how many machines

can be used for a particular part. Malik & Dale suggested forming product groups based
on processing requirements, and allocating the required number of machines to each

group. Machines not allocated form a remainder cell.

Tilsley & Lewis’s (1977) Flexible Production Cells also forms cells based on processing
requirements, but also takes into account demand variability. It is based on computer
analysis of routings to identify machines that occur together frequently. Burbidge (1977)
also proposed Nuclear Synthesis, where machines used by only a few parts are identified
and represent the nuclei of cells. Once these nuclei are identified, processing
requirements of parts using them are analyzed, and parts with similar processing
requirements added to the group. Corresponding machines are then allocated to the

groups to form cells.

2.2.4.2 AnaliniLMetths

2.2.4.2.1 SEW
Similarity coefficients, first proposed by Jaccard (Sokal & Sneath, 1973) numerically

define the similarity between pairs of items. McAuley’s Single Linkage Cluster Analysis
(1972) defines the similarity between two machines as the ratio of number of parts using
both machines to the number of parts using at least one. Cluster analysis is used to
determine the optimal grouping of machines based on these similarities. Machines are

added to a cluster if their similarity with existing machines in the cluster exceeds a

33

threshold value. As fewer machines remain, a clustering algorithm is used to
systematically reduce the threshold value until all machines are alloeated. Similar to this
approach is the Bond Energy Algorithm, (McCormick et al. , 1972). Cells are identified
by reordering the binary part/ machine matrix based on bond energy, the product of
adjacent element values, and maximizing total bond energy. Similarity coefficients were

also used by de Witte (1980) in an extension of an earlier work.

Rajagopalan & Batra (1975) used graph theory to form cells. Arcs of the graph represent
the strength of similarity between machines. These form cliques or groups of machines
so that strong relationships exist within groups, and weak relationships between groups.
Chandrasekharan & Rajagopalan (1986a, 1986b) formulated the problem as a bipartite
graph and used a non-hierarchical clustering algorithm to obtain diagonal groupings

within the part/ machine matrix.

Carrie ( 1973) and Vakharia & Wemmerlov (1990) applied similarity coefficients to parts
rather than machines. Carrie’s method groups parts together if their similarity is above
a threshold value, and a specified minimum number included in a family. This prevents
the formation of unduly small cells. A clustering algorithm is used to systematically
reduce the threshold value until all parts are allocated to a family. Vakharia &
Wemmerlov explicitly considered intra-cell material flows and machine load in cell
formation. Their algorithm distinguishes between parts based on the need for operation

backtracks (non-sequential operations on the same machine) and number of operations.

34

In addition, it distinguishes between parts using the same machines in the same sequence,
and those using the same machines in a different order. Operation sequence was also

considered by Choobineh (1988).

Kusiak (1987) proposed the p—median problem in which cells are formed using an integer
programming formulation, whose objective is to maximize total similarity. The maximum
number of part families to be formed can be expressed as a constraint in this formulation.
Kusiak also presented a formulation that allows alternate process plans for a part. Kusiak
& Cho (1992) developed two formulations that consider alternate process plans both

when there are and are not bottleneck parts or machines.

More comprehensive surveys of the literature regarding similarity coefficients and cluster

analysis can be found in Chu (1988), and Shafer & Rogers (1993).

22.422 W
Several algorithms form cells by reordering the binary part/machine matrix to yield

mutually exclusive clusters along the diagonal of the matrix similar to Production Flow
Analysis. King’s ( 1979) Rank Order Clustering (ROC) re-orders the matrix by attributing
binary values to the rows and columns of the matrix. These are converted to decimal
equivalents and the rows and columns re—ordered based on decreasing decimal values.
The process is repeated until no changes in the matrix occur. King (1980) modified this

so that the rows and columns can be ordered directly from their binary values. King &

35
Nakomchai (1982) proposed ROC 2 which is similar to ROC but computationally less

demanding. This algorithm places all rows with a ‘l’ in the last column at the top of the
matrix, then does the same with the columns, moving columns with a ‘1’ in the last row
to the left of the matrix. This is repeated until there are no further changes in the matrix.
Chan & Milner’s (1982) Direct Clustering Algorithm is similar to ROC 2 but re-orders

rows and columns based on decreasing number of entries in the row or column.

Boe & Cheng (1991) and Askin et al. (1991) addressed the limitation of procedures based
on ROC that they do not guarantee a diagonal matrix structure after reordering. They

proposed new algorithms that do produce such a structure.

22.4.2.3 MAW

Combinatorial grouping and mathematical programming have been used by Purcheck
(1974, 1975a, 1975b), and Oliva-Lopez & Purcheck (1979). Machines needed to process
a part and any parts whose routing is a subset of its are identified. Based on constraints
such as cell workload, groupings of sets of similar parts are found and merged.
Corresponding machines are allocated to the merged sets to form cells. Mathematical
programming was also used by Shtub (1989) who modelled the cell formation problem

as a generalized assignment problem.

In addition to the traditional approaches to cell formation, newer approaches have been

developed recently to overcome some of the limitations of existing methods (Chu, 1993).

36
Amongst these are neural networks (Kaparthi & Suresh, 1992, Chu, 1993), fuzzy

clustering (Xu & Wang, 1989, Chu & Hayya, 1991), syntactic pattern recognition (Wu

et al., 1989), expert systems (Kusiak, 1988), and simulated annealing (Boctor, 1990).

2.2.4.3 MW

Although several cell formation procedures exist, their impact on shop performance is
largely unclear. Most procedures do not consider their impact on shop performance.
Though some have been evaluated on the performance they yield, evidence of their
relative performance is limited (e.g., Morris, 1988, Shafer, 1988, Shafer & Meredith,

1990, Chu & Tsai, 1990).

Several studies have incorporated information on additional shop characteristics as well
as those of the parts and machines themselves. Vannelli & Kumar (1986) developed a
heuristic that minimizes the number of bottleneck cells. Ballakur & Steudel (1987)
considered the impact on cell formation of factors such as cell utilization and workload.
, Seifoddini (1987) and Balasubramaniam & Panneerselvam (1993) incorporated
information on production volumes in cell formation. Nagi et al. (1990) formulated cells
while incorporating multiple routings and capacity constraints. Rajamani et al. (1990)
also considered the availability of alternate process plans. Sule (1991) developed a
heuristic that considers capacity requirements as well as equipment costs and the costs
associated with inter-cell material movements. The impact of operating costs, lot size and

production planning, was also considered by Chakravarty & Shtub (1984). Rajamani et

37
al. ( 1992) formulated a mixed integer programming model that evaluates the trade-off

between investment in additional machines and setup costs where setups are sequence-

dependent.

A number of articles have focussed on the issue of exceptional elements, parts that do
not fit into identified cells. Surveys of CM users (e. g Pullen, 1976, Wemmerlov & Hyer,
1989) suggest that despite the intent of CM to obtain independent cells, most
implementations have large numbers of parts requiring processing in multiple cells.
Waghodekar & Sahu (1984) developed a heuristic to minimize the number of exceptional
elements. Kumar & Vannelli (1987) developed a method for identifying parts that can
be subcontracted so that those remaining belong to well defined cells. Wei & Gaither
(1990) formulated an integer programming model that minimizes the opportunity cost
associated with manufacturing exceptional parts. Kern & Wei (1991) and Shafer et al.
(1992) developed models that consider the costs of eliminating exceptional elements (i.e. ,
by inter-cell movement, machine duplication, or sub-contracting) once a cell

configuration has been identified.

A number of formulations focus specifically on the issue of inter-cell movement of work
and machine duplication. Harhalakis et al. (1990) and Wu & Salvendy (1993) minimized
the number of inter-cell moves by combining cells using heuristic and network
approaches respectively. Vohra et al. (1990) also formulated the cell formation problem

as a network to minimize interactions between cells. Logendran (1990) considered the

38

effects of both inter and intra—cell movement as well as workload imbalances within cells.
This approach was extended (Logendran, 1991) to incorporate the impact of operation
sequence and cell layout. Song & Hitomi (1992) developed a quadratic assignment
problem to minimize inter-cell movement. Okogbaa et al. (1992) allow inter-cell
movement so that the variance of busy times of identical machines is similar. Dahel &
Smith (1993) formulated integer programming models to minimize inter-cell movement

and to minimize inter-cell movement while simultaneously maximizing routing flexibility.

The issue of duplicate machines was considered by Seifoddini & Wolfe (1986),
Seifoddini (1989), and Logendran (1992). Seifoddini & Wolfe developed a similarity
coefficients approach to cell formation that duplicates bottleneck machines. Seifoddini’s
model evaluates the machine duplication decision based on the tradeoff between increased
equipment cost and reduced material handling cost. Logendran formulated an integer
programming model that explicitly considers budgetary constraints in permitting machine

duplication.

2.2.5 WWW:

Past research on CM allows a number of conclusions to be made about its effectiveness.
It is evident that the process of machine dedieation either in a cellular or process layout,
leads to signifieantly reduced shop flexibility and severe utilization problems. The result
is performance that is inferior to that yielded by a pure process layout. Only when non-

processing components of flow time (i.e., setup, material handling) are large, does the

39

potential exist for CM to outperform a process layout. Even the more common
process/cell hybrid layout performs comparably to a process layout only under limited

circumstances.

Of the three procedures outlined to improve CM performance, none is designed to
overcome its inherent limitations. The objective of group scheduling is to take advantage
of batch similarities with respect to setups. As currently implemented, it is not concerned
with machine configuration or routing issues, and thus fails to address the issue of
flexibility. Lot splitting, though improving the efficiency of material flows, suffers from
the same limitations. Alternate routing, though to some degree alleviating problems of
unbalanced utilization and reduced flexibility, does not overcome the problem of machine
dedication that Flynn & Jacobs (1986) suggested is the primary cause of poor
performance. Each approach is also short term and narrowly focussed in how it tries to
improve performance. None adopts a long term perspective, taking into account the
downstream consequences of their actions, nor do any address problems of changing

product mix and volume.

Although these mechanisms enhance CM performance, the magnitude of the machine
dedication problem appears too large to be overcome within a cellular layout. Though
many approaches to cell formation have been proposed, they typieally do not consider
resulting shop performance. Those that do are faced with the problem of trying to satisfy

often conflicting goals. The result is that the impact of cell formation on shop

40

performance is not clear. Even if it were and cells formed accordingly, the fact that
machines are dedicated implies that flexibility is lost. As long as this situation remains,
so will limits on shop performance. As Lewis (1973) stated, the ability to use a
production system to its best advantage is predetermined by how it is conceived and
designed. Given the available experimental and ease evidence, it is apparent that the
processing of part families must be viewed from an alternative perspective that does not

impose the restrictions placed by traditional CM.

2.3 FLEXIBILITY ISSUES IN MANUFACTURING

Based on the evidence, it is the loss of flexibility that limits the ability of CM systems
to generate improvements in performance. Not only does this loss of flexibility
compromise the production of existing products, but it makes it unresponsive to a
changing environment. As Buffa (1984) stated, in the present manufacturing climate,
there is a premium on flexibility. Harrigan (1985) suggested that organizations need to
be flexible because of technologically driven shorter life cycles and global competition,
and can be most responsive if facilities are designed with flexibility in mind. It is evident
therefore that the flexibility of the manufacturing process is the key to the ability to

respond to a uncertain environment.

Flexibility has been suggested to be a component of manufacturing strategy (Buffa, 1984,
Wheelwright, 1984). It thus represents one of the distinctive competencies that can be

used to obtain competitive advantage. Numerous definitions of flexibility in

5w
eff

ldt

of

Wei
YOU
cha
be c

new

41
manufacturing exist (e. g., Mandelbaum, 1978, Hall, 1983, Swamidass & Newell, 1987,

Swamidass, 1988). At the core of these is the ability of a production system to respond
effectively to a changing environment. Mandelbaum (1978), and Buzacott (1982)
additionally characterized flexibility as the ability to respond to change, and the ability
to continue to perform despite the change. Slack (1987, 1990) made the distinction
between range flexibility and response flexibility. Range flexibility refers to the breadth
of change that can be accommodated. Response flexibility is the ease with which change

can be made.

23.1 Wu:

The concept of flexibility in manufacturing has been used across the entire spectrum of
the production process, from product design to processing to delivery. Swamidass (1988)
identified twenty terms associated with flexibility in the operations management
literature. A number of typologies exist that identify the aspects of manufacturing that
flexibility must address. These include Mandelbaum (1978), Buzacott (1982), Zelenovic
(1982), Gerwin (1983, 1987), Slack (1983), Browne .et al. (1984), Swamidass (1988).
These are summarized by Alder ( 1985). Common to a number of these frameworks as
well as the perceptions of managers (Slack, 1987, 1990) are part, part mix, volume and
routing flexibility. In order to be competitive, an organization needs to be responsive to
changes in demand in terms of the types of parts it produces, their mix and volumes. To
be consistent with these needs, the production system must possess the flexibility to meet

new process plans, and to alter routings to accommodate changes in production

42

schedules, capacity requirements and disturbances to the system such as machine

breakdowns.

As Gerwin (1987) suggested, manufacturing flexibility can be considered at a number of
levels. Amongst these are the flexibility of individual machines, the manufacturing
process, or the manufacturing system as a whole. Slack’s studies of managers (1987,
1990) indicated that managers are cognizant of the value of flexibility, but have a limited
view of it. They also prefer to deal with only a limited amount of flexibility. Managers
tend to view flexibility from a resource perspective, typically focussing on the flexibility
of a single resource, rather than that of the production system as a whole. Flexibility is
seen typically as a means towards an end rather than an end in itself. It is sought
primarily to meet the specific needs identified earlier, the ability to produce new parts,
modify part mix, change the level of output, and in addition, the ability to change
delivery dates. The evidence suggests that managers are more concerned with response
flexibility than range flexibility, particularly the time needed to bring about change. The
. limited evidence on the relationship of flexibility and performance from both empirical
and simulation studies, confirms the importance and increasing recognition of flexibility
as a competitive tool, an important shift away from the traditional focus on cost and

productivity.

43

2.3.2 W
Consistent with the perceptions of managers (Slack, 1987, 1990), evidence from the

literature suggests that resource flexibility, specifically machine and labor flexibility, is
central to the discussion of manufacturing flexibility (Slack, 1990). Malhotra & Ritzman
(1990) tested the hypothesis that increased resource flexibility in a multistage
manufacturing environment, allows an organization to improve its performance when
confronted with a changing environment. Machine flexibility was modelled by defining
a shop with a fixed number of machines and changing the number of departments they
were allocated to. Fewer departments implies departments must process a greater number
'of items with a larger number of more general purpose machines, making them more
flexible. Labor flexibility was modelled by changing the number of machines a worker
could operate. Two environments were examined, a benign environment characterized
by small lot sizes and a large capacity cushion, and a hostile environment characterized

by large lot sizes and a low capacity cushion.

Even with change limited to lot sizes and capacity, their results showed the value of
flexibility. In the benign environment, greater machine flexibility led to modest
improvements in customer service as measured by past due demand, but in the hostile
environment, there were substantial gains. The effect on inventory was less significant
in each environment. The benefits of increased labor flexibility were also lower in each
case. When both forms of flexibility were introduced, the performance gains were only

marginally greater than when only one existed. Machine flexibility is thus an important

44

mechanism in responding to a changing environment, though there is a related cost in

terms of the purchase of more flexible but less efficient general purpose machines.

These results are similar to those obtained by Bott & Ritzman (1983). They showed that
in an MRP environment, the allocation of equipment to a few, large, general purpose
departments rather than several, small, specialized departments led to significantly
improved performance. Customer service, measured by past due demand, was
significantly lower. Inventory and the occurrence of bottlenecks were also reduced. The

impact of greater flexibility was of particular significance when demand was unstable.

2.3.3 WW
Swamidass & Newell (1987) surveyed a number of managers as part of a study of the

relationships between environmental uncertainty, manufacturing strategy and business
performance. The organizations concerned all used small batch manufacturing processes.
One of the issues investigated was the effect of flexibility. Using a path analytic model,
they found there to be a strong positive correlation between flexibility and performance.
The benefits associated with greater flexibility were also positively correlated with
environmental uncertainty. The authors, commenting on the reported gains of Japanese
producers who designed repetitive production lines with flexibility in mind (Schonberger,
1982), concluded that flexibility is important regardless of the manufacturing process

being used.

45
Roth & Miller (1990) as part of a broader study into the relationships between

manufacturing and managerial strengths and business performance, surveyed
manufacturing executives about the strength of their competitive capabilities relative to
their competitors. Using factor analysis, the authors identified five independent
dimensions of manufacturing strength. One of these was flexibility, specifically, new
product, volume and design change flexibility. They categorized companies as superstars,
middlemen, and weaklings, based on their competitive strengths, and compared these
groups with respect to the importance they placed on flexibility. The results showed that
both superstars and middlemen placed greater importance on flexibility than weaklings,
though there was no difference between superstars and middlemen. In addition, they
categorized the companies as winners and losers based on economic performance, and
again compared the two groups to determine whether differences in attitudes to flexibility
existed. As expected, the winners were shown to place greater emphasis on flexibility

than losers.

In a comparative study of Japanese, European and American manufacturing
organizations, De Meyer et al. (1989) identified differences in competitive priorities and
courses of action of organizations in each environment. European and American
producers still consider quality, reliability, and to a lesser degree cost as their
competitive priorities. However, the Japanese, having already addressed these issues,
consider flexibility to be the top priority. Their ability to shift focus is made possible by

the fact that they have attained what they consider to be appropriate levels of quality,

46

cost and reliability, and now have a significant cushion relative to their competitors on
these dimensions. This allows them to concentrate on what the authors call the next
competitive battle. An important finding of the research is that actions taken by Japanese
‘ producers are consistent with their stated concerns and competitive priorities. This is
increasingly true of American producers, but less so of Europeans. This consistency has

been suggested to be a critical determinant of manufacturing success (Hill, 1989).

2.3.4 fR n f rin ' '1'

The evidence on the effect of manufacturing flexibility on performance, though sparse,
clearly demonstrates its value. Equally important is the finding that only a limited amount
of flexibility is required to improve performance. Further increases in flexibility may
have limited value. This is entirely consistent with results from studies that have
implicitly, if not explicitly, considered the effect of increased flexibility (Bobrowski &
Mabert, 1988, Ang & Willey, 1984). Given the costs associated with greater flexibility,

this is significant.

CHAPTER3

DYNAMIC CELLULAR MANUFACTURING (DCM)

3.1 INTRODUCTION

Given the limitations of CM, the value of resource flexibility, and the need for greater
responsiveness to changing market demands, a need exists for a more flexible production
system for small/ medium batch production that allows the advantages of part family
production to be attained. Such a system should not permanently dedicate machines but
maintain flexibility in machine allocation. A system of this nature can be characterized
by a layout in which cells are not viewed as a physical grouping of machines as they are
in traditional CM. Instead, cells are temporary entities that are formed and destroyed on
a continual basis by allocating machines to families based on current need and
availability. The parts that constitute a family are those that have similar processing

requirements, thus allowing the number of setups to be kept to a minimum.

The concept of a cell that is not a physical ordering of machines was initially suggested
by McLean et al. (1982) and Simpson et al. (1982). They defined a ‘virtual cell’ to be
a set of machines, which, though physically separated, exist together as a logical entity
for scheduling purposes. In real time, virtual cells are created to meet current processing
needs, then dissolved on completion. The virtual cell is a routing mechanism where the
required machines are claimed before processing begins, and where machines are

dedicated to a given processing requirement only as long as needed. Irani et al. (1993)

47

48
used the term virtual cell to refer to cells created by the sharing of machines in a shop

physically organized as a cell/process hybrid layout similar to that described earlier.
They proposed a shop layout in which cells with overlapping machine requirements are
located physically adjacent to each other. Likewise, machines used by several cells are
organized functionally and located physically close to the cells which require them. This
physical organization facilitates machine sharing without the need for machine duplication
or increasing the complexity of material handling. Individual parts can be processed
outside their primary cell creating the illusion of a cell since machines outside the

primary cell are temporarily dedicated to the corresponding family.

3.2 MANUFACTURING ENVIRONMENT FOR DCM

Dynamic cell formation involves examining the set of jobs awaiting processing at each
process department, and identifying their part family affiliations. When machines in the
department become available, they are temporarily allocated to families requiring them
using family based scheduling rules. It is this temporary allocation of machines to
. families that creates the illusion of a cell. Machines allocated to a family define a path
through the shop. While this path continues to exist, parts from the family are routed
along it to the specific machines they require. Unlike traditional job shops where the
allocation of machines to jobs is essentially random in nature, in DCM, machines are

to a greater degree pre-assigned as they are in traditional manufacturing cells.

49
Cells formed in this way are dynamic and virtual. They are dynamic since they are

formed on an ongoing basis based on current processing needs and machine availability.
They are virtual since they cease to exist after the need for them passes. Over time, the
machines making up a cell change based on machine availability. This yields a more
efficient utilization of machines than in traditional CM. In addition, the size of a cell can
change over time. A cell begins to evolve once a single machine is allocated to a part
family. As parts from the family progress through the shop, machines from other process
departments may be allocated to them, increasing the size of the cell. Eventually,
machines from all departments visited by family members, may be held simultaneously.
This represents the greatest length of the cell. Beyond this, the cell can expand only if
multiple machines from the same process department are allocated to it. This can occur
f the additional machines are not required by other families. The capacity of the cell can
thus adjust to better meet the processing needs of the family without compromising the
processing needs of other families. Conversely, machines no longer required by the
family may be released, causing a contraction in the size of the cell. Cells may not
always evolve to their maximum length if machines are released at a faster rate than they
are added. Cells also need not consist of a continuous path if they do not contain
machines in the interior of the routing. In this case, the cell exists as disjoint cell

segments.

The primary benefit of forming cells in this manner is that family processing needs are

met without the sacrifice of flexibility. Machines are constantly assigned or reassigned

50
to cells based on family need. This dynamic allocation overcomes the problem of

unbalanced load in traditional CM. This in turn makes the configuration more responsive
to changes in volume, family composition, and family size. In addition to offering an
alternative mode for part family production, DCM makes it possible for a manufacturing
concern considering conversion to a cellular layout, to investigate whether it ean benefit
from such a change. Since DCM applies the family concepts of CM, it can be used as
a mechanism to study the potential gains from using a cellular layout before any physical

change or investment takes place.

3.3 DCM vs. CELLULAR AND PROCESS LAYOUTS

3.3.1 W

A major advantage of DCM is that it does not require the long term or permanent
physical shop reorganization required by traditional CM. Traditional CM is founded
upon physical reorganization of machines and their dedication to part families. This
way, a line flow or similar simplified routings can be obtained within each cell. This in
principle should yield improved control, lower work in process, and more efficient
material handling. In addition, CM typically strives for cells to be independent with
machines allocated to only a single cell. Consequently, additional equipment purchases
are often needed to make this possible, adding to the cost of reorganization. Implicit is
the fact that reorganization takes time, which will likely render the shop less than 100%
operational. Morris (1988) suggested that the need to physically reorganize a shop may

discourage product innovation in favor of process convenience.

51
Physically, DCM can use the existing process layout. The only physical difference is in

the dedication of machines. Since DCM cells are not physical groupings of machines,
there is no need to physically reorganize the shop floor. This is particularly significant
with shorter product cycles, changes in part mix and the need for short lead times. Given
time and cost considerations, modifications to a traditional cellular layout to
accommodate such change may not be possible nor advisable. Since there are no physical
cells, duplication of machinery to achieve cell independence is not an issue. The need for
no additional investment is important not only in terms of dollars saved, but also given
the current emphasis on short term financial decision making. As Voss (1986) suggested,
investment decisions may not consider non-quantifiable factors such as increased
flexibility and improved competitiveness. This alone may preclude investment in CM

projects.

Since DCM cells are not fixed entities nor their machines located adjacent to each other,
there is a loss of some of the benefits of CM. In particular, the material handling benefits
of CM are lost, and production control is more complex. However, the benefits of more

efficient machine utilization can be expected to more than offset these losses.

3.3.2 Raufinsflcaihilim
With traditionally formed cells, all machines required by a part family are dedicated
permanently to that family. The result is that at times, some machines in a cell may be

idle, while functionally similar machines elsewhere may have long queues in front of

52
them. The result is that jobs in congested cells may be delayed (unless alternate routing

strategies are employed). The aim of DCM is to exploit the routing flexibility of a
process layout. The process layout makes it possible for any machine of a given type to
be used to process a job. However, in DCM, machines are dedicated to a part family for
as long as it needs them. Once a machine is no longer needed, it can be assigned to a
different cell. Since any available machines of the required type can be allocated to a
cell, routing flexibility is increased. This eliminates the nwd for alternative routing

strategies.

33.3 W

In a process layout, each job requires a major setup at each machine in its routing (unless
sequence-dependent scheduling is used). These setups cannot take place until the machine
is assigned to the job, thus the job must wait while the machine is being setup. In
traditional CM, since machines are dedicated to part families, once machines are initially
setup for a family, no major setups are required. Only minor setups are required to
recognize differences between jobs in the same family. With DCM, setup requirements
lie between these two extremes. Since a dynamic cell is dedicated to a family, a major
setup is required at each machine only when it is allocated to a family. After that, only

minor setups are required, to recognize differences between parts within a family.

53

3.3.4 W
CM is inflexible to changes in part mix and volume. Since the shop is physical organized

based on a particular part mix and workload, it cannot respond effectively or rapidly if
new parts are introduced that do not fit into existing families, and thus an existing cell.
Likewise, if the workload of a cell changes, the cell cannot adapt. Such changes may
require the additional purchase of machinery or relocation of existing machinery, which
as explained earlier, may not be possible. Alternatively, parts may have to be produced

using a combination of cells. This compromises the ability to reap the benefits of CM.

With DCM, this problem is moot. Since cells are not rigid, there is no problem of
matching new parts with existing cells. New families can be created and cells formed to
meet their processing needs without compromising the shop layout. Available equipment
can be assigned and reassigned to cells as the needs of families change. Since machines
within a process department are homogeneous and located physically adjacent to each
other, routing jobs to a secondary machine does not result in the loss of control that

might occur with physically separated cells.

3.4 LIMITATIONS OF DCM

From an operational standpoint, DCM does have certain limitations compared to existing
production methods. Scheduling in the DCM environment is more complex than in
traditional CM. In a traditional cellular environment, the scheduling problem is limited

to jobs within a given cell. In DCM, the scheduling problem encompasses the entire shop

54

 

 

 

 

 

 

 

 

 

 

 

 

 

Change

 

 

 

Shop Configuration l
Process layout Cellular Layout DCM I
Requires Shop No Yes No
Reorganization
Requires New No Possibly No
Equipment 1‘
Machine None Permanent Temporary
Dedication
Shop Floor Highly Complex Least Complex Moderately
Control Complex
Scheduling High Low Medium
- Complexity
Routing High Low Medium
Flexibility
Material High Low High
Handling
Type of Setups Major & Minor Minor Major & Minor
Frequency of One/Machine/J ob None One/ Machine! Cell
Maj or Setups
Responsive to Yes No Yes

 

 

Figure 1 : Comparison of DCM, Process and Cellular layouts

55

and all current jobs. However, since jobs can only be routed to machines that are idle
or already setup for the corresponding family, scheduling effort is lower than in a
process layout, where all machines of the same type must be considered. As previously
mentioned, the material handling benefits of traditional cellular systems will also be lost,

due to the machines in a cell being physically distant.

From a behavioral standpoint, whether DCM has particular merit depends on its
implementation. One of the suggested gains of CM is that since parts are processed
within a cell, operators, if allocated to a cell rather than a machine, have a wider variety
'of tasks. The result is greater job satisfaction, and improved quality. In DCM (and a job
shop), the same benefits can be obtained if cross training exists, and operators allocated
to jobs rather than machines. However, the scope of these benefits might be limited by
the machines in a jobs routing not being physically adjacent as they are in traditional

cells.

3.5 DCM vs. FLEXIBLE MANUFACTURING SYSTEMS (FMS’s)

Flexible manufacturing systems (FMS’s) attempt to obtain the same benefits as DCM,
namely greater flexibility and higher utilization. However, though FMS’s may be able
to achieve these benefits more efficiently, they impose additional constraints.
Specifically, FMS’s are characterized by complex planning and scheduling environments.
They require expensive machining centers, sophisticated tooling systems, and advanced

material handling systems to provide the degree of automation sought. Overall control

56

of the system is governed by complex and expensive computer hardware and software.
The costs associated with investment in capital and training are significant. It also takes
time to install and test the system. Investment and time are factors that make even
traditional CM difficult to justify unless a successful implementation can be guaranteed.
Given evidence from existing FMS implementations, this is far from certain. From an
operational standpoint, F MS’s are inappropriate in an environment with long setup times
which is characteristic of the environment being considered here. The intent of FMS’s
is to take advantage of production flexibility in environments with short setup times and

where tooling changes can be automated.

Although DCM may not be able to provide the same level of flexibility as FMS’s, it can
attain a significant degree of flexibility without physical reorganization of the shop floor
or new asset acquisition. Furthermore, it can do so without the introduction of the
complexity or cost associated with FMS’s. Given this tradeoff, the problems associated
with investment decision making described earlier, and the need for rapid introduction

of flexibility, DCM offers an attractive alternative.

3.6 SUMMARY OF DCM

In summary, DCM offers the benefits of CM while using a process layout, by means of
scheduling as opposed to machine layout. The investment, physical reorganization and
permanent machine dedication associated with traditional cellular systems are eliminated.

CM ’s recognition of family processing needs and linear routings are retained and

57

 

 

 

me: O 0
Process FMS
Layout
flexibility O
DCM
Cellular
Layout
Low 0
Low
Setup Eficiency

 

 

 

Figure 2. Tradeoffs Between Flexibility and Setup Efficiency

combined with the flexibility of a process layout. This allows the shop to respond more
quickly and efficiently to changes in production needs. DCM represents a trade-off
between the benefits of traditional CM and a process (or job shop) layout. It is also a

tradeoff between flexibility and setup efficiency.

3.7 RFSEARCH STATEMENT

This research examines the impact of DCM on small/ medium batch production in a
closed shop environment. In a closed shop with repeat orders for a standard set of parts,
considerable scope exists for the application of CM. It is possible and beneficial to
identify similarities in part processing requirements and to exploit these in the production

process, particularly since these parts and families will exist over a period of time. In an

58
open shop with different parts being produced without repetition, part families are less

clearly defined. With the composition of demand constantly changing, the make-up of
families also changes, making the application of traditional CM inefficient. Though there
is less potential to exploit production similarities in an open shop, possibilities may still
exist. One of the advantages of DCM is that since cells are not fixed entities, greater
flexibility exists in defining families and thus cells. Unlike traditional CM, DCM may
thus have applicability in an open shop. Even if a family consists of a single part, loading
the corresponding cell is equivalent to cell loading. Though cell loading in a traditional
cell was shown to yield poor performance (Mosier, 1983), the increased flexibility and

utilization of DCM can be expected to make cell loading more attractive.

The research addresses a number of questions regarding the potential of DCM. These
address two major issues. The first is the tradeoff that DCM represents between the
flexibility of a process layout, and the family processing and setup efficiency of
traditional CM. Five questions relating to this issue are investigated using specific
hypotheses:

a. Do setup conditions exist where DCM’s use of the part family concept and
efficient use of setups, is more beneficial than the flexibility of a traditional
job shop. If so, what setup conditions are conducive to DCM.

b. Can the recognition of part families by DCM, make it more effective in
dealing with different part mix compositions. If so, for what part mix
characteristics is DCM preferable.

c. Can the greater flexibility of DCM allow it to overcome the setup benefits of

permanent machine dedication in traditional cellular manufacturing, and if so,
under what setup conditions.

59

(1. Does the greater flexibility of DCM make it more responsive to changes in
part mix than traditional cellular manufacturing, and if so, under what part
mix conditions.

e. Does the information used to form dynamic cells affect their performance.

The second issue is the robustness to change of a production system that physically has
a process orientation, but is operated as if it had a product orientation. Three additional

questions are examined:

f. Does shop load have a significant impact on the performance of DCM.
g. Is DCM sensitive to changes in job size.

h. Does job dispatching affect the performance of DCM.

These questions are addressed by first comparing the performance of DCM to that of a
traditional process and cellular layout, and then examining the behavior of DCM in
greater detail. The next chapter explains the research design used to carry out these

studies.

CHAPTER 4

RESEARCH DESIGN

4.1 INTRODUCTION

The research is conducted in two stages. DCM is first compared to production using
traditional process and cellular layout methods. The objective is to identify whether DCM
yields better performance with respect to throughput and due date measures, and under
what conditions its use might be appropriate. Stage two investigates DCM in more detail
by examining other conditions that influence its performance and suggest potential for
its use. This permits a better understanding of DCM and a greater awareness of when

it might be used or when other alternatives are more appropriate.

4.2 RESEARCH METHODOLOGY

The research is conducted using computer simulation models. Simulation is a commonly
used tool in research of this kind. It enables research to be conducted under controlled
conditions defined by the researcher. This eliminates the risk of other factors affecting
‘ the validity of conclusions. Cook & Campbell ( 1979) refer to this as internal validity.
Simulation also facilitates replication of an experiment. This allows sufficient data to be
collected for statistical conclusions to be made with an appropriate degree of certainty,

or statistical conclusion validity (Cook & Campbell, 1979).

61
4.2.1 WWW

Five issues need to be addressed in simulation research to ensure its statistical validity.
These are initialization bias, independence and normality of observations, sample size,
and variance reduction. In non-terminating simulations such as those used in this
research, the system being modelled begins in a state of no activity. However, the system
is evaluated once it has reached steady state, or its long run level of activity.
Observations collected prior to the system reaching steady state have a biasing effect
since the system behaves differently initially compared to when it reaches steady state.
To eliminate this initialization bias, the time at which steady state has been reached must

be identified and observations prior to this point discarded.

Each time the simulation is run from start, observations collected during the initialization
period must be discarded. This results in a large number of discarded observations. To
reduce this, one long run can be carried out and batch sampling used. This leads to the
problem of autocorrelation. To be valid measures, observations must be independent. The
progress of a job is affected by that of jobs in the system at the same time, since they
affect shop load, queue sizes, etc. However, jobs separated by a large enough time lag
are not affected in this way. If the batch size is large enough, the mean response of
adjacent batches can be shown to be independent (Kleijnen, 1987). This batch size must

be determined.

62

In order to meet the assumptions of the statistical tests to be used in data analysis, the
distribution of batch means must be approximately normal. The central limit theorem
states that for large sample sizes (n 2 30), the distribution of sample means is
approximately normal even for non-normally distributed populations (Law & Kelton,
1982). However, the quality of the approximation depends on the population distribution.
larger batch sizes improve the quality of the approximation. An appropriate batch size

must therefore be identified.

For statistical tests to be carried out with a high degree of power, a large enough sample
size must be obtained. An appropriate number of batches must therefore be run for each
treatment. Finally, the validity of statistical conclusions is compromised by the
introduction of variance other than that due to the experimental treatments themselves.

Additional sources of variation must therefore be minimized or eliminated.

4.2.1.1 InitializaticaBias
The method used here is that of Schruben et al. (1983). If there is no significant

difference between the mean of N observations, and the mean of the first k (k < N), a
steady state response has been obtained. They defined a test statistic for this difference
based on the t distribution. Since initialization bias is likely to cause an under-estimate
of the steady state response, a one-sided hypothesis for mean difference is tested. If

steady state is not reached within the first k observations, k is increased and the test

repeated.

 

lll'r

63
4.2.1.2 Amman

The procedure used here is the Von Neumann statistic (q) whose use is suggested by
Kleijnen (1987). If batch means are independent and normally distributed, the expected
value of q is known and its variance can be computed as a function of n, the number of
batch means used to compute q. The statistic q is distributed normally. Kleijnen et al.
(1982) suggest it be at least 100 since for small n the test has low power. A value of n
= 100 is thus used. If the null hypothesis of independence is not accepted, the batch size
is increased and the test repeated. In this research, an initial batch size of one hundred
is arbitrarily selected and the batch size increased by one hundred each time the null

hypothesis is not accepted.

4.2.1.3 M

An assumption of analysis of variance (AN OVA) which is used to analyze the data is that
observations are normally distributed. However, Neter et al. (1990) state that ANOVA
is robust to small departures from normality. In order to establish whether batch sample
means are approximately normal, the Probability Plot Correlation Coefficient Test is used
(Filliben, 1975). This computes the correlation between the ordered batch means and the
order statistic medians from a standard normal distribution. If the distribution of means
is normal, the correlation coefficient should be close to one. The significance of the
correlation is evaluated by comparison with percent points of the normal probability plot
correlation coefficient. If the hypothesis of normality is not accepted, the batch size is

increased and the test repeated. In this study, the initial batch size is that which satisfies

64

the assumption of independence. Increments in batch size of one hundred are used if

normality is not obtained.

4.2.1-4 W

Assuming a normal distribution of sample means, the sample size required to obtain a
confidence interval for the mean response can be computed as a function of the
population variance, and the half width of the required interval. Pilot runs are conducted
to estimate the mean and variance of flow time for each treatment. These are used to
establish the sample size required to obtain non—overlapping confidence intervals for all
treatment means. Schmeiser (1982) suggests using between ten and twenty batch means

to estimate the confidence interval. Twenty batch means are therefore used.

4.2. 1.5 W

In order to eliminate variance other than that due to the treatments, common random
numbers are used (Kleijnen, 1987). For each treatment, the same random number stream
is used for the corresponding input process. This ensures that the random numbers are
not a source of variance. Glasserman & Yao (1992) demonstrated that the use of common
random numbers guarantees variance reduction and is optimal for a wider class of
simulation models than previously assumed. One random number stream is not
synchronized. This ensures that samples are independent (Mihram, 1974) which is an

assumption of the procedures to be used to analyze data.

65
4.2.2 Mans

In order to conduct the above tests, pilot runs are carried out for each treatment of the
two stages of the research. These identify the initialization period and batch sizes to meet
assumptions of autocorrelation and normality. For each stage of the research, the
initialization period used during actual experiments is the longest identified from the
corresponding pilot runs. Likewise, the batch size used is the smallest required to meet

the assumptions for all corresponding treatments.

4.3 SIMULATION ENVIRONMENT

To facilitate comparison, the simulation environment used here is similar to that used by
Morris (1988). This also allows the simulation models to be validated. However, for
experimental purposes, certain parameters are changed to create a more suitable research

environment. This section describes shop features common to both stages of the research.

A total of forty part types, partitioned into five families, are considered (Figure 3). Each

 

Family Part Numbers

33, 34, 35, 36, 37, 38, 39, 40

19, 20, 21, 22, 23, 24, 25, 26

27, 28, 29, 30, 31, 32

9, 10, ll, 12, l3, 14, 15, 16, 17, 18
=1, 2, 3, 4, 5, 6, 7, 8

 

 

 

 

 

 

 

1
2
3
4
5
. ==

 

Figure 3. Part Family Affiliations

 

66

family contains between six and ten parts. Parts have between four and six operations.
Jobs arrive according to a poisson process, with inter-arrival times exponentially
distributed. This is a commonly used arrival process in job shops (Law and Kelton,
1982). Jobs are for a single part type. Operation processing times consist of a constant
and a stochastic component. These are 33.33 and Normal (1, 0.25) minutes per batch of
size 100. Due dates are set using the Total Work Content (TWK) rule (Conway et al.,
1967). This defines due dates as the arrival time of the job plus a multiple, k, of the job
processing time. Baker (1984) has shown this to be an effective procedure with respect
to tardiness performance over a range of conditions. Similar to Morris, k = 3 is used
here. Weeks and Fryer (1977) showed that for a range of conditions, k values between
2.5 and 2.75 were optimal but that for small departures from optimality, performance

did not change significantly.

A total of thirty machines are used. According to Baker (1974), no conclusive evidence
exists to suggest that the number of machines in a shop affects its performance. The shop
. floor covers an area of 10,000 (100 x 100) square feet, each machine allocated an area
of 225 (15 x 15) square feet. Layouts are defined using the CRAFT algorithm (Buffa et
al., 1964). Forklift trucks are available for material handling purposes. These move at
five miles per hour and are an unconstrained resource. Loading and unloading times are

uniformly distributed in the interval 1 to 5 minutes.

67
4.4 PERFORMANCE MEASURES

Any comparative study of production systems must consider the effect they have on
throughput performance and the ability to meet due dates. These determine the ability of
the system to complete orders in a timely fashion. To accomplish this, mean flow time
and mean tardiness are measured. In addition, the mean and standard deviation of work
in process (WIP) are measured. WIP is defined in terms of number of minutes of work.
WIP provides a surrogate measure for shop congestion. Changes in WIP can also be
expected to correlate positively with flow time variance, which in turn affects tardiness
variance. These are the primary performance measures due to their combined effect of

' appropriately gauging the overall performance of the system.

To more fully understand shop behavior, average utilization, proportion of time jobs
spend during setups and in queues, and the proportion of tardy jobs are also measured.
The intent of DCM is to reduce the impact of setup times relative to a traditional process
layout, and to overcome problems caused by unbalanced utilization and long queues in

a cellular layout. The secondary measures are used to identify if these objectives are met.

4.5 EXPERIMENTAL STAGE I

4-5-1 ExncdmentaLEactcrs

Stage one compares production using DCM to that using traditional process and cellular
layout methods. The intent is to identify whether DCM performance differs from that

obtained when using these layouts, and to determine when DCM might be preferred.

68

 

 

 

 

 

Number of Machines 30

Number of Parts 40

Number of Operations/Part 4-6

Job Arrivals Exponential
Due Date TWK, k = 3

 

Operation Processing Times 33.33 + Normal (1,0.25) minutes / batch size 100
Loading/Unloading Times Uniform (1,5)
Material Handling Forklift Truck, 5 mph

 

 

 

.
Performance Measures Mean Flow Time, Mean Tardiness, Mean WIP,
Standard Deviation of WIP

Mean Utilization, Proportion Tardy,
Setup Time Proportion, Queue Time Proportion

 

 

 

 

Figure 4 : Simulation Environment and Performance Measures

Three factors are examined: shop configuration, setup times, and part mix variability.

4.5-1.1 mm

Seven shop configurations are examined, a traditional process layout, traditional cellular
layout, and five configurations based on DCM. As described earlier, each shop consists
of thirty machines. The traditional cellular layout consists of five cells, each containing
between four and eight machines (Figure 5). Cell sizes are consistent with evidence of
actual CM implementations (Wemmerlov & Hyer, 1989). Within cells, no machine
duplication exists. Parts are fully processed within a single cell. Material handling times
are not considered since cellular layouts are designed to make material handling

inconsequential.

 

f

69

 

 

 

 

 

 

 

Cell/Family Machines Part No. Routing
1 18, 25, 13, 3, 23, 10, 16 34 18, 25, 13, 3, 23, 10
40 18, 25, 3, 23, 16
38 18, 25, 3, 23
39 25, 13, 23, 10
33 25, 3, 10
36 13, 3, 23, 10, 16
37 13, 23, 10, 16
35 13, 10, 16
2 26, 2, 15, 7, 17, 4, 20, 12 24 26, 2, 15, 7, 17, 4

20 26, 2, 15, 7, 17, 4
19 26, 7, 20, 12
23 26, 20, 12
26 2, 15, 7, 17, 4
22 2, 15, 7, 17, 4
21 2, 17, 4, 20, 12
25 17, 4, 20, 12

22, 8, 28, 24, 9, 21 32 22, 8, 28, 24, 9
30 22, 8, 28, 24, 9
27 22, 28, 24, 9
31 22, 28, 24
28 8, 28, 24, 9, 21
29 8, 9, 21

29, 14, 6, 19,27 17 29, 14, 6, 19
15 29, 6, 19, 27
13 29, 19, 27
9 29, 19, 27
18 14, 27
16 14, 27
12 14, 27
10 6, 19, 27
14 6, 19
11 6, 19

ll, 1, 30, 5 7 ll, 1, 30, 5
6 11, 1, 5
4 11, 30, 5
2 11, 1, 30
3 11, 1
8 l, 30, 5
5 30, 5
1 30, 5

 

 

 

 

 

 

 

Figure 5 : Configuration of Cellular layout

 

 

70

The process layout consists of eight process departments (Figure 6). Each contains three

 

Process Department Machine Numbers l

8, 18, 19, 26
2, 25, 27, 28
11, 13, 15, 24
1, 3, 7, 9
17, 21, 23, 30
4, 5, 10, 29
14, 16, 20

6, 12, 22

 

 

OOQChUt-kwNi—t

 

Figure 6 : Configuration of Process Layout

or four machines. Routings using the process layout are defined in Figure 7. In addition,

both shops contain a shipping and receiving department.

The configurations based on DCM have the same physical layout as the process layout.
It is the temporary dedication of machines to families that distinguishes DCM from the
process layout. Machines could be allocated to families using the group scheduling rules
described earlier. However, these are typically local in nature. Most of these rules
consider processing characteristics of families only at the machine of interest, and not
elsewhere in the shop. In addition, they focus solely on exploiting sequence dependencies
in scheduling decisions, giving the appearance of a job shop using sequence-dependent

scheduling.

 

 

 

 

 

Process Departments Visited

 

'i 7

 

1 5,6 .
2 3,4,5 1
3 3,4 ‘
4 3,5,6 .
5 5,6 ,
6 3,4,6 1
7 3,4,5,6 ~
8 4,5,6 1
9 1,2,6 .
10 1,2,8 ,
11 1,8

12 2,7

13 1,2,6

14 1,8

15 1,2,6,8

16 2,7

17 1,6,7,8

18 2,7

19 1,4,7,8

20 1,2,3,4,5,6

21 2,5,6,7,8

22 2,3,4,5,6

23 1,7,8

24 1,2,3,4,5,6

25 5,6,7,8

26 2,3,4,5,6

27 2,3,4,8

28 1,2,3,4,5

29 1,4,5

30 1,2,3,4,8

31 2,3,8

32 1,2,3,4,8

33 2,4,6

34 1,2,3,4,5,6

35 3,6,7

36 3,4,5,6,7

37 3,5,6,7

38 1,2,4,5

39 2,3,5,6

4o 1,2,4,5 l

 

Figure 7 : Routings in Process Layout

72

In order to address this limitation, three selection rules that consider family processing
needs at machines other than the machine to be assigned, are considered, in addition to
two traditional family selection heuristics. These new rules embrace the intent of DCM
to consciously create complete, continuous cells. This way, dynamically formed cells
more closely resemble traditional cells in which all machines required by a family are
available for its use, and form a clearly defined routing. Each rule is first applied to
families without access to a machine in the process department in question. If no such
families exist, all remaining families are considered. This promotes the development of
multiple cells and the simultaneous processing of all families. In addition, it minimizes
the risk of some cells not having access to a machine of a given type, while others have

multiple machines of the same type.

The family selection rules based on past research are :

DCM 1 : The family with the lowest average job slack. This is similar to the
DDFAM rule of Mahmoodi et al. (1988) that selects the family
containing the job with the earliest due date, but explicitly considers
remaining processing time and the urgency of the family as a whole
(Mosier, 1984).

DCM 2 : The family containing the most jobs in the queue. This is similar to
the WORK rule of Mosier et al. ( 1984), that selects the family with
the greatest work content. This facilitates families with the greatest
ability to minimize major setups.

The rules that incorporate information on family processing elsewhere in the shop are:

DCM 3 : A family is selected which also has parts currently being processed at
its immediate predecessor departments. If more than one such family
exists, the family with the most jobs in the current queue is selected.
This rule facilitates the incremental building of cells, thereby reducing
potential setups and queuing delays.

73

DCM 4 : The family requiring the fewest machines to complete a cell is
selected. Similar to DCM 3, this facilitates the formation of complete
cells, and reduces potential major setups and queuing delays.

DCM 5 : When no jobs remain from the family currently using the machine, the
immediate predecessor departments of this family are examined to
determine whether jobs from the family are currently being processed
there. If they are, the machine is not reassigned to a new family, but
remains idle so that these jobs can use it without incurring an
additional major setup. If there are no such jobs, the machine is
assigned to the family with the most jobs in the current queue. This
rule goes further in maintaining the structure of a cell once it has
begun to evolve.

The family selection rules differentiate the five DCM shop structures among themselves,

and also from the process and traditional cellular layouts.

4.5.1.2 Mme
Setup time can be expected to affect the relative performance of DCM and traditional

process and cellular layouts. As demonstrated in past comparisons of CM and process
layouts (e.g., Morris, 1988) and in other work on setup times (e.g., Karmarkar et al.,
1985a), setup times have an important effect on shop performance. Setups require
machines to be busy but do not themselves add value to manufactured products. Any
delays due to setups therefore reduce the capacity of the production system. As described
earlier, it is the frequency of major setups when using a process layout, that makes its
use inefficient. Likewise, it is their avoidance when using a cellular layout, that makes
CM more efficient. In the context of the present study, it is the ability of DCM to
minimize the frequency of major setups without significantly compromising flexibility,

that gives it a potential advantage.

74

Both major and minor setups are considered. Major setups between families are typically
more time consuming. Minor setups between parts in the same family are generally of
shorter duration and require less extensive tooling change. Both major and minor setups
are assumed to be sequence independent. This is a common assumption in research of
this kind. Two factor levels are considered. At the low setting, major setup time is one
third of the processing time, and at the high setting, two thirds (Mahmoodi et al. , 1992).
These yield setup times of 11.33 and 22.66 minutes. Minor setup time is one quarter of
the major setup time (Flynn, 1984). This is consistent with evidence from users of
cellular systems (Wemmerlov & Hyer, 1989). The ratio of minor to major setups is not

an experimental factor in this study. There is no setup time between jobs that are for the

sameparttype.

4.5.1.3 W

The primary property of a process layout that allows it to perform well is its flexibility.
This also allows it to respond effectively to changes in the mix of parts to be produced,
since a machine’s use has not been predetermined. CM is unable to respond effectively
to such change since cells are designed to meet expectations of a given part mix. If the
mix changes, shop performance deteriorates since jobs are required to be processed in
specific cells which may not be designed to handle more than a certain load. Since it
makes less rigid assignments of machines to families, DCM offsets this loss of flexibility

while retaining the family recognition property of CM. Though a closed shop is being

75

examined, the mechanism generating actual orders may create a variable part mix

environment, for example MRP.

TWO levels of this factor are considered. Under balanced part mix conditions, each
family has the same demand probability (0.20). Within a family, parts have the same
demand probability. This mix is consistent with the design of the cellular layout in that
the workload of each cell is proportional to the number of machines it has. Under
unbalanced part mix conditions, three families have a combined demand probability of -
0.70, equally distributed between the families. The remaining two have a combined
demand probability of 0.3, again equally distributed between the families. Within each
family, individual parts have the same demand probabilities. Though no basis for this
specific partitioning of demand between families exists, a similar approach was used by

Wemmerlov (1992).

4.5 .2 Sr'urulutign Envirunment

In this stage of the research, mean inter-arrival times are set in order to obtain a load of
approximately 80% when using the traditional process layout. This is a load that has been
used in past research, and that is found commonly in practice (Baker, 1974). Jobs are
dispatched using the minimum job slack rule. This has been shown in past research to
yield good flow time and tardiness performance in both process (Conway et al. , 1967)
and cellular shops (Mosier et al. , 1984) if due dates have been established in an

appropriate manner. It is also representative of rules used in research and in practice.

76
The rule is slightly modified for use in the process layout by first giving priority to jobs

that are identical to those just completed. If there are no such jobs, job selection is then
based on minimum job slack. Cellular approaches to manufacturing have a built in
mechanism to recognize inter-family sequence dependencies. This modification
compensates for the minimum slack rule failing to recognize sequence dependencies, and
makes the implementation of the rule in the process layout more representative of actual
use. The focus of this stage of the research is on shop configurations rather than

scheduling issues.

4-5-3 WW

Stage one of the research is carried out using a full factorial design with twenty eight (7
x 2 x 2) treatments. These are defined in Figure 8. The objective of the research is to
identify conditions when DCM shows potential as an alternative to production using
traditional job shop and cellular methods. The intent is not to predict the behavior of
DCM in different environments. To accomplish this objective, stage one of the research
investigates nine a priori hypotheses. These are formulated as non-orthogonal linear
contrasts and are evaluated using ANOVA ‘and paired comparisons. This is an
appropriate approach to use since only the presence of effects and not their magnitudes

is of interest.

ANOVA is first conducted to identify the presence of significant main and interaction

effects. If there are significant main effects and no significant higher order interactions,

 

 

 

 

 

 

 

 

 

 

F: =
r Setup Time Low High
Part Mix Balanced Unbalanced Balanced Unbalanced
DCM 1 1 8 15 22 l
DCM 2 2 9 16 23 fl
DCM 3 3 10 17 24
Shop
Config. DCM 4 4 11 18 25
DCM 5 5 12 19 26
Process Layout 6 13 20 27
Cellular Layout 7 14 21 28

 

 

 

 

 

 

 

 

 

Legend : 1 - 28 = Treatment Numbers
Low Setup Time = 11.33 minutes, High Setup Time = 22.66 nrinutes
Balanced Part Mix = Part families have equal demand probabilities
Unbalanced Part Mix = Three part families have demand probabilities of
.233, two have demand probabilities of .15

Figure 8. Stage I Experimental Design

Kirk (1982) suggests the use of the Bonferroni procedure (Dunn, 1961) to test the
significance of the contrasts. This test guarantees that if the error rate when testing each
of C contrasts is a/C, the error rate for all C contrasts cannot exceed a. Neter et a1.
(1990) suggest that when only a small subset of all main effect contrasts is of interest,

this test is more powerful than other tests such as the Scheffe or Tukey tests.

If significant higher order interactions exist, contrasts are no longer meaningful, since
factor effects differ at different levels of other factors (Kirk, 1982). Under these
conditions, the hypotheses are examined using paired comparisons of all treatment means

using the Tukey method (Neter et al., 1990). Neter et al. suggest that this is a more

78
powerful test to use when the number of comparisons is large. The inability to utilize the

contrasts does however mean that any conclusions regarding factor effects must be

viewed taking into account the effect of interactions.

The following are the nine a priori hypotheses to be investigated. Treatment means are

 

 

 

 

 

 

 

 

 

defined in Figure 9.

Setup Time Low High

Part Mix Balanced Unbalanced Balanced Unbalanced
DCM 1 11 12 13 14
DCM 2 21 22 23 24
DCM 3 31 32 33 34 1

Shop

Config. DCM 4 41 42 43 44
DCM 5 51 52 53 54 fl
Process Layout 61 62 63 64 ll
Cellular Layout 71 72 73 74

 

 

 

 

 

 

Figure 9. Stage I Treatment Means

(#13)

1. When setup time is low, the process layout outperforms DCM.
. s ' £42 a £1
avg-g; 10 -§ 2 :0
11.1090

Since the time associated with each major setup is low, the effect of greater
setup frequency when using the process layout is relatively small. Under these

79

conditions, the greater flexibility of the process layout should enable it to
compensate for the increase in setup frequency that it incurs. Acceptance of
the null hypothesis suggests that DCM overcomes this flexibility premium
even when setup times are not expected to be critical.

2. When setup time is high, DCM outperforms the process layout.

5 ‘ £42 ‘ fir
3.3050

The primary benefit of DCM over the process layout is that it reduces the
frequency of major setups. When setup time is high, the benefit of fewer
major setups is greater. Acceptance of the null hypothesis suggests that the
minimization of setups cannot compensate for the reduced flexibility of DCM.

3. When part mix is balanced, the process layout outperforms DCM.

s s
1 F "’ 1| )
H ,¢ _ z; 11 g 13 _ (”61*963) 50
°' 3 10 2

3.3090

When part mix is balanced, there are fewer parts from the same family and
thus less scope to share setups. Under these conditions, the inability of the
process layout to do this should not compromise its performance. Its greater
flexibility should continue to give it an advantage. Acceptance of the null
hypothesis indicates that even under conditions less suited to family
-recognition, DCM performs better.

4. When part mix is unbalanced, DCM performs better than the process layout.

5 s
( “13+ “14)
”rifle" § 10g _ (Parks) 20

H.:¢.<0

80

When part mix is biased towards certain families, there is greater scope to
share setups. By reducing the number of major setups, DCM is in a better
position to take advantage of these conditions. Acceptance of the null
hypothesis indicates that this setup reduction is not sufficient to overcome the
lower flexibility of DCM.

5. When setup time is low, DCM outperforms the cellular layout.

9 3 2.41 a 1‘11.
H.:¢s.§ .1 1° -; 2 20

H.30,<O

When setup time is low, the greater number of major setups incurred by
DCM has a relatively small effect. Under these conditions, DCM’s greater
flexibility should give it an advantage over the cellular layout. Acceptance of
the null hypothesis suggests that this increased flexibility is insufficient to
compensate for the elimination of major setups in the cellular layout.

6. When setup time is high, the cellular layout outperforms DCM.
s ‘ £42 ‘ £22
H,:¢‘=§; ro -; 2 :0

3.3000

When setup time is high, DCM’s need to use major setups has a greater
adverse effect. By eliminating the need for major setups, the cellular layout
is less affected by high setup time. Acceptance of the null hypothesis suggests
that DCM’s greater flexibility more than offsets the effect of high setup time.

7. When part mix is balanced, DCM outperforms the cellular layout.

81

s s
1 F ‘1 ll )
H _ g s: g u __ (911*973’ 20
“'3’ 10 2

H.:O,<O

8. When part mix is unbalanced, DCM outperforms the cellular layout.
5 S
Ho:¢.- (g “13:0; "10) - (""34“") 20

H.:¢,<O

Cellular layouts are unable to respond to change in part mix and perform well
only if cell workload is consistent with cell capacity. Even when part mix is
balanced, DCM has the flexibility to adjust to short term imbalances in
workload distribution. Acceptance of the null hypotheses for hypotheses 7 and
8 suggests that this flexibility is not able to compensate for the increase in
setups incurred by DCM.

9. Dynamically formed cells that recognize work flow patterns are more
effective than those that do not.

‘
5 0 3 ‘
H.:¢’- (gépu) - (Fig IA”) 20

H.r¢,<0

The intent of DCM is to provide the benefits of part family production within
a process layout. It also aims to make production more responsive to
prevailing work patterns. Cell formation that explicitly considers the flow of
work should therefore be more effective.

82
4.6 EXPERIMENTAL STAGE H

4-6-1 Bancmrentalfactcrs

Unlike stage one whose purpose is to compare the performance of DCM to other
small/ medium batch production methods, the objective of stage two is a more detailed
sensitivity analysis of DCM alone. Stage two examines the effect of factors, which, based
on evidence from existing studies of small/medium batch production, may affect its
performance. This extends the understanding of the appropriateness of DCM in different
production environments. It also allows the behavior of DCM to be contrasted more fully
with what is known about production using traditional job shop and cellular methods.
The factors included are utilization, job dispatching, volume mix variability, and part mix

variability.

4.6.1.1 1111112311211

Past research on job shops (e. g., Baker, 1984) and CM (e.g., Hitomi et al., 1977) have
shown shop performance to depend on utilization. As utilization increases, queues build
up at machines. This increases the delays encountered by jobs, thus increasing flow times
and leading to reduced on-time job completion. Job shops face the additional problem
that increases in the arrival rate of jobs also increases the frequency of setups. This adds
further to the problem of delays. CM faces a problem of low overall utilization due to
the uneven distribution of work between machines. It is reasonable to expect that
utilization will also affect DCM. However, given DCM’s particular characteristics, it

may respond differently.

83
Three levels of utilization will be considered, 70% , 80% and 90%. As described earlier,

80% utilization is common in practice and in prior job shop research. The remaining two
levels have also been used in past job shop research (e.g., Baker, 1984) and allow the

shop to be operated under conditions of lesser and greater congestion.

4.6.1.2 Whine
Past research on job shop scheduling (e. g., Conway et al., 1967) and scheduling in CM

(e. g. , Mosier et al. , 1984) has demonstrated the impact of dispatching rules on shop
performance. The order in which jobs are processed at a machine affects the extent to
which queues build. It also determines the extent to which individual jobs are made to
wait. To examine the impact of dispatching on DCM, three rules common in practice and
in past research are examined. These prioritize jobs based on a range of characteristics

that have been shown to have an objective rationale.

FCFS : Jobs are dispatched based on earliest arrival time at the machine
or process department. This is used frequently in practice (Conway
et al., 1967) based on its intuitive fairness.

SPT : Jobs are dispatched based on minimum operation processing time.
SPT is an example of a processing time based rule. It has been
shown in the past to yield good mean flow time performance
(Conway et al., 1967). SPT reduces the build up of queues by
processing jobs that can be completed quickly.

MINSLK: Jobs are dispatched based on minimum job slack. This is an
example of a due date based rule. It has been shown in the past to
yield good performance, particularly for due date measures
(Conway et al. , 1967). MINSLK explicitly tries to process jobs
whose on time completion is compromised.

4.6.1.3 MW

Changes in size of incoming jobs directly affects the stability of dynamically formed
cells. Smaller jobs implies that there are more jobs in the system simultaneously. For the
same utilization level, this suggests that a greater proportion of time is spent by machines
while they incur setups. This reduces the extent to which individual cells are utilized.
Likewise, greater variance in job size increases the variance of cell life, and thus the
extent to which the benefits of the cellular structure can be exploited. While one of the
benefits of DCM is the flexibility it introduces to CM, a tradeoff exists with setup
frequency. If the potential for cells to change is too great, this may offset the benefits of

increased flexibility.

These effects are examined by defining three levels of this factor. The first level
corresponds to the scenario in stage one where jobs have a constant batch size of one
hundred (Morris, 1988). The effect of variance is captured by defining batch sizes to be
normally distributed, with a mean of 100, and a coefficient of variation of 0.1 (Bott &
_ Ritzman, 1983, Krajewski et al., 1987). A batch size of 100 is characterized as a large
batch size. The impact of small job size is represented by jobs with a constant batch size

of fifty, and a corresponding increase in arrival rate.

4.6.1-4 W

Part mix variability is carried over from stage one due to the potential interaction it has

with volume mix variability. Both affect cell workload and in turn the stability of cells,

85

frequency of setups and productive capacity. The factor is defined the same way as in
stage one. Part mix is either balanced in which case all families have the same demand,

or it is unbalanced and demand skewed in favor of three families.

4.6.2 Simulaticnfinximmem

Only one DCM implementation, DCM 4, is included in stage two. This is one of the
better performing implementations based on the results of stage one, and one that
embraces the intent of DCM to consciously create complete cells. Setup times are not
included as an experimental factor. Setup times are fixed at the low level from stage one,
or 11.33 minutes per major setup. Utilization levels are based on the balanced part mix,

minimum slack dispatching scenario.

4-63 W

Stage two is carried out using a full factorial design with fifty four (3 x 3 x 3 x 2)
treatments. These are defined in Figure 10. As described earlier, stage two of the
research is exploratory in nature. No specific hypotheses are tested. AN OVA is used to
identify the presence of significant main and interaction effects. Tukey multiple

comparisons are used to identify the source of specific differences.

4.7 SUMMARY OF RESEARCH DESIGN
The study uses computer simulation models to show whether DCM is a viable alternative

to production compared to traditional process and cellular layout methods. This is

 

    

volume Mix ee l _ so .
Utilization(%) 7o|80j9ol7olso|9ol7olso|9o
PartMix

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Balanced 1713192531374349
Unbalanced 2 14 20 26 32 38 44 50 §
Balanced 3 9 15 21 27 33 39 45 51
Unbalanced 4 lo 16 22 28 34 4o 46 52
Balanced 5 11 17 23 29 35 41 47 53
Unbalanced 6 12 18 24 3o 36 42 48 54

 

Legend : 1 - 54 : Treatment Numbers
Dispatching Rule : FCFS - First Come First Served, SPT - Shortest
Processing Time, Minslk - Minimum Job Slack
Volume Mix : 100 - Job Size = 100, N(100,10) - Job Size = N(100,10), 50
- Job Size = 50
Balanced Part Mix : Part families have equal demand probabilities
Unbalanced Part Mix : Three part families have demand probabilities of .233,
two have demand probabilities of .15)

Figure 10. Stage 11 Experimental Design

accomplished by comparing the different shop configurations under a range of shop
conditions, then examining additional factors expected to influence DCM performance.
The research questions posed are examined using ANOVA, linear contrasts, and multiple

comparisons. The results of these analyses are discussed in the next chapter.

CHAPTER 5
EXPERINIENTAL RESULTS

5.1 INTRODUCTION

The data collected from the simulation runs was analyzed in several stages. For each of
the primary performance measures, analyses of variance were conducted to identify the
presence of significant main and interaction effects. In each analysis, data were blocked
by replication number. This allows the independence of samples to be verified since
common random numbers were used (Mihram, 1974). Residual analysis was used to
verify the assumptions of normality and homogeneous residual variances underlying the
use and validity of ANOVA. The sources of specific differences associated with the
significant main and interaction effects were evaluated using Tukey multiple comparisons.
Since signifieant interactions were found in the stage one data for all primary
performance measures, the nine a priori hypotheses were analyzed using Tukey multiple
comparisons of treatment means. All statistical analysis was eanied out using SAS (SAS
Institute) and SYSTAT (SYSTAT Inc.) statistical software. Statistical tests were carried

out at the a = .05 level.

5.2 APPROPRIATENFSS OF ANALYSIS OF VARIANCE
The appropriateness of ANOVA models was evaluated by examining whether
assumptions of normally distributed residuals and homogeneous residual variances were

met. Neter et al. (1990) state that minor violations of these assumptions does not

87

88
necessarily compromise the validity of ANOVA. The impact of non-normally distributed

residuals is to marginally increase the actual significance level and marginally decrease
the power of the test. This is defined to be the probability of correctly failing to accept
a false hypothesis. This effect is not significant for large sample sizes. The effect of non-
homogeneous residual variances is the same as long as sample sizes are equal. Neter et
al. (1990) suggest that data transformations be used to reduce or eliminate more
substantial violations of these assumptions. Specifically they suggest the use of log,

square root and reciprocal transformations, depending on the nature of the violation.

The implieation for this research is that since sample sizes are large and balanced,
inferencesbasedonANOVAcanbeassumedtobevalid eveninthepresenceofminor

violations of assumptions, except where observed p values are close to 0.05 .

The assumption of normally distributed residuals was tested using the Probability
Correlation CoefficientTest (PCCT) described in Chapter 4. Homogeneity of variances
was tested using the Hartley test (Neter et al., 1990). This considers the ratio between
the maximum and minimum treatment residual variances and accepts the hypothesis of
homogeneity if the ratio is not significantly different from one. Neter et a1. (1990) state
that small significance levels are justified when using this test. A significance level of

a = .01 was therefore used.

89
5.3 ANALYSIS OF STAGE I DATA

5.3.1 Introduction

The twenty-eight treatments in stage one are shown again for convenience.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Setup Time LOW High :

l Part Mix Balanced Unbalanced Balanced Unbalanced
DCM 1 1 8 15 22 j

. DCM 2 2 9 l6 . 23

l DCM 3 3 10 17 24

, Shop

. C onfig. DCM 4 4 11 18 25
DCM 5 5 12 19 26
Process Layout 6 13 20 27
Cellular layout 1 7 14 21 28

Legend : l - 28 = Treatment Numbers
Low Setup Time = 680 minutes, High Setup Time = 1360 minutes
Balanced Part Mix = Part families have equal demand probabilities
Unbalanced Part Mix = Three part families have demand probabilities of
.233, two have demand probabilities of .15)

Figure 11. Stage I Treatments

53.2 Wish

For mean flow time and work in process, there was a good fit between the data and the
AN OVA models. For mean flow time, though only ten of the twenty-eight treatments had
normally distributed residuals, all but one (Treatment 7) yielded PCCI‘ values within 4%
of that required to accept the hypothesis of normality. The remaining treatment was

within 7% . Nineteen of the treatments had residual variances that were homogeneous.

90
Of the remaining nine (Treatments 3,5,7,12,14,17,20,21,28), three were treatments

whose variances were outliers (Treatments 7 , 14,28). None of the nine were amongst the

better performing treatments.

For mean work in process, eleven treatments had normally distributed residuals. All but
three had PCCI‘ values within 2 % of that required to accept the hypothesis of normality
(Treatments 7, 14, 28). These were within 6% of the critieal value. Nineteen of the
treatments had homogeneous residual variances. The heterogeneous variances again came

from poorer performing treatments (Treatments 3, 5, 6, 7, 12, 14, 20, 23, 24).

For mean tardiness none of the treatments yielded normally distributed residuals and only
nine had homogeneous residual variances. In order to overcome this, the three
transformations suggested earlier (log, square root, and reciproeal) were used. The log
transformation significantly improved the fit of the data with the assumptions of
ANOVA. All but four treatments (Treatments 1, 2, 8, 9) yielded PCCI‘ values within
6% of that required to accept the hypothesis of normality. Nineteen treatments had
homogeneous residual variances. Again, treatments with non-homogeneous residual
variances were either outliers or other poor performing treatments (Treatments 3, 7, 10,

12, 14, 21, 24, 27, 28).

For the standard deviation of work in process, though all but three treatments had PCCI‘

values within 7% of the critical value and eighteen had homogeneous residual variances,

91
a log transformation significantly improved the fit of the model. All treatment residuals

yielded PCCT values within 4% of the critical value, and twenty-five of twenty-eight had
homogeneous residual variances. Again, treatments with heterogeneous residual variances

were poorer performing treatments (Treatments 3, 5, 14).

5.3.3 Analxsiutfiffects
5.3-3.1 Wm:
ANOVA results for mean flow time are reported in Table 1. Since all higher order

Table 1. Analysis of Variance for Mean Flow Time

SS MS
Random Numbers 731960 7394
Shop Configuration 11375365 1895894
80978 ‘ 80978
293933 293933

 

 

 

 

 

74373 19

1239553

 

Shop *Mix

5330985

888497

 

Setup *Mix

55 1862

551862

 

527601

Shop * Setup * Mix 3165604
6515853 2438

 

 

 

 

 

 

 

R2 = 0.82

interactions are signifieant, Tukey multiple comparisons were carried out for each shop
configuration for the four combinations of setup time and part mix. The rationale for this

is the fact that over a short time horizon, setup time and part mix are factors that

92

management can exercise some control over through the planning system. Only over a
longer time horizon can management exercise control over the shop configuration.

Treatment means are reported in Table 2.

Table 2. Treatment Means for Mean Flow Time

 

 

 

 

 

 

 

 

 

 

 

Setup Time Low High
Part Mix Balanced Unbalanced Balanced Unbalanced
DCM 1 206.67 193.77 237.58 223.67
DCM 2 205.76 193.42 236.07 222.66
1 DCM 3 214.77 200.08 249.93 232.39
mg, DCM 4 208.51 196.30 240.50 227.04
DCM 5 228.71 210.86 249.00 232.98
Process Layout 225.47 208.44 268.15 247.98
Cellular Layout 322.11 749.10 252.01 293.42

 

 

 

When setup time is low and part mix balanced, the best performance is yielded by DCM
14 (Figures 12, 13a). No significant differences exist between these configurations. The
performance of the process layout is similar to that of DCM 5, even though DCM 5 is
the most far-sighted of the DCM implementations. The performance of the cellular layout
is poorer than that of the other configurations. When setup time is low and part mix
unbalanced, performance is indistinguishable between all implementations of DCM and
the process layout (Figures 12, 13b). Performance of the cellular layout is extremely
poor. When setup time is high and part mix balanced, DCM 1, 2 and 4 yield the lowest

flow times (Figures 12, 13c). DCM 3 and 5 and the cellular layout yield similar

Setup Time

93

 

 

 

 

 

 

 

 

 

 

 

 

Part Mix Balanced Unbalanced Balanced Unbalanced
DCM 2 -= DCM 2 - DCM 2 DCM 2
DCM 1 DCM 1 DCM 1 DCM 1 ]
Mean DCM 4 DCM 4 DCM 4 DCM 4
FlowTime DCM3 DCM3 DCM5 l DCM3}
Process ] Process DCM 3 DCM 5
DCM 5 DCM 5 Cellular Process
Cellular Cellular Process Cellular
DCM 2 ' DCM 2 'j DCM 2 ] DCM 2 ]
DCM 1 J DCM 1 DCM l DCM 1
Log Mean DCM4 DCM4 ] DCM4 DCM4
Tardiness DCM 3 DCM 3 ] DCM 5 ] DCM 5
Process DCM 5 DCM 3 DCM 3
DCM 5 Process 1 Process Process
Cellular Cellular Cellular Cellular
DCM 2 DCM 2 DCM 2 DCM 2
DCM1]r DCMl DCMI] DCMI]
MeanWork DCM4 J DCM4 DCM4 - DCM4 1
in Process DCM 3 DCM 3 DCM 3 DCM 3 1'
Process ] Process Cellular ] DCM 5 |
DCM 5 DCM 5 - DCM 5 Process
Cellular Cellular Process Cellular
DCM 2 DCM 2 DCM 2 DCM 2
Log DCM l l DCM l I DCM 1 I DCM 1 1
Standard DCM4 DCM4 DCM4 1 DCM4 *
Deviation DCM3? DCM3 =1 DCM5 l DCM5]
ofWIP DCM5: DCM5 ] DCM3 ' DCM3 -
l l

 

 

 

 

 

 

Figure 12. Tlukey Multiple Comparisons of Shop Configuration
by Setup Time x Part Mix

 

W3

Cortiguation

2

Shop

low Setup Time, Balanced Part Mix

a

 

X

Unbalanced Part

Time,

low Setup

b

Mean Flow Time by Setup Time x Part Mix

13.

Figure

 

Balanced PartMix

Time,

h Setup

18

H

C.

 

SlopCorf

X

, Unbalanced Part '

13

11. High Setup Time

(cont'd)

igure

F

96

performance. The process layout yields the poorest performance. When setup time is
high and part mix unbalanced, DCM 1, 2 and 4 again perform best (Figures 12, 13d).

All DCM implementations outperform the process and cellular layouts.

5.3.3.2 MeanIanliness

ANOVA results for log mean tardiness are reported in Table 3. Treatment means for the

Table 3. Analysis of Variance for Log Mean Tardiness

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

: SOURCE or: SS MS F p

' Random Numbers 99 980.28 9.90 6.72 0.0001 I

. Shop Configuration 6 902.58 150.43 102.06 0.0001 1
Setup Time 1 582.07 582.07 394.89 0.0001

PartMix 1 3.83 3.83 2.60 0.1069
Shop'Setup 6 412.24 68.71 46.61 0.0001 5
Shop*Mix 6 22.60 3.77 2.56 0.0180

‘ Setup * Mix 1 0.33 0.33 0.22 0.6378
Shop * Setup * Mix 6 12.25 2.04 1.39 0.2165 i

‘ Error 2673 3939.97 1.47 *-

R2 = 0.42

untransformed data are reported in Table 4. Pairwise comparison of shops by setup
time/part mix conditions show that for low setup time, balanced part mix conditions,
DCM implementations 1, 2, and 4 yield the best tardiness performance, followed by
DCM 3 (Figures 12, 14a). Similar to the result for mean flow time, the performance of

DCM 5 and the process layout is indistinguishable, and the cellular layout yields the

97

Table 4. Treatment Means for Mean Tardiness

; Setup Time

 

Unbalanced Unbalanced ,
DCM l . 0.137
DCM 2 . 0.142
DCM 3 . 0.686
DCM 4 . 0.426
DCM 5 . 0.758 -
Process Layout . 0.635
Cellular layout

 

 

 

 

 

 

 

 

 

 

 

 

 

highest tardiness. For low setup time, unbalanced part mix conditions, DCM l and 2
outperform other DCM implementations (Figures 12, 14b). The process layout again
performs poorly as does DCM 5, but not as poorly as the cellular layout. For both high
setup time scenarios, DCM 1 and 2 again perform best, followed by DCM 4 (Figures
12, 14c, 14d). The process and cellular layouts perform poorer than all DCM

implementations.

5.3.3.3 WW

ANOVA results for mean work in process are reported in Table 5. Treatment means are
reported in Table 6. Multiple comparison results are, as expected, largely similar to those
for mean flow time. The exceptions are that the performances of DCM 2 and 3 are not
indistinguishable when setup time is low and part mix balanced, and DCM 4 and 5

perform differently when setup time is high and part mix unbalanced (Figures 12, 15a-d).

 

 

IL

Corfiglntion

2

Shop

Time, Balanced Part Mix

a low Setup

 

ShopConr

Mix

Unbalanced Part

Time,

Low Setup

14

b

X

Time x Part

by Setup

mess

. Mean Tard

igure

F

x r

m»
as...

w

‘ ’4

33¢ .
{MSJNsufi \
x c

x as

“we... r

A“

 

Corfiguation

Slop

gh Setup Time, Balanced Part Mix

Hi

c.

\\»_.) . (x
”1'31

.>,« ;.\

 

Slano

Unbalanced Part Mix

High Setup Time,

(1

igure 14. (cont'd)

F

100

Table 5. Analysis of Variance for Mean Work in Process

  

 

   

 

    

 

   

 

 

 

    

 

  

 

 

 

 

 
 

 

 

 

 

 

 

 

 

DF SS MS F
‘ Random Numbers 99 223048038 2253010 8.44
Shop Configuration 6 1158547167 193091195 723.67
Setup Time 1 195185982 195185892 731.52
_ Part Mix 1 22896 22896 0.09
Shop .- Setup 6 807971467 134661911 504.69
Shop * Mix 6 386391299 64398550 241.35
Setup * Mix 1 36492126 36492126 136.77
. Shop * Setup * Mix 6 249934824 41655804 156.12
R2 = 0.81

Table 6. Treatment Means for Mean Work in Process

Low

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

‘ Setup Time A
Part Mix Balanced Unbalanced Balanced Unbalanced
DCM 1 2916.70 2630.50 2816.46 2563.98
Shop
, Config. DCM 2 2884.44 2611.62 2785.34 2541.40
DCM 3 3083.43 2752.23 3012.25 2685.35
DCM 4 2926.47 2653.40 2841.66 2593.50
DCM 5 3362.13 2970.69 3062.79 2765.63
Process layout 3331.16 2941.83 3349.51 2960.70
i ellul 4487.57 7989.86 __ 3.96

mesa

 

5

H.

ms
81an n

5.. aux

. e 5....
Eruwfimswas .

2

Low Setup Time, Balanced Part Mix

a

 

Mix

Low Setup Time, Unbalanced Part

b.
15

Time x Part Mix

Process by Setup

1n

. Mean Work '

Figure

 

mes

Ins:
Slnp Corf

W1

W2

11

me, Balanced Part Mix

Ti

gh Setup

c.Hi

 

Slanom_

igh Setup Time, Unbalanced Part Mix
15

. H

d

d)

(cont

Figure

 

103
5.3.3.4 WW
ANOVA results for log standard deviation of work in process are reported in Table 7.

Table 7. Analysis of Variance for Log Standard Deviation of Work in Process

U
as

j SOURCE
? Random Numbers

 

 

Shop Configuration

 

 

 

 

 

 

OSHOSON—v—OSS

 

 

 

 

 

 

 

R2 = 0.83

Treatment means for the untransformed data are reported in Table 8. For both low setup
time scenarios, the best performance is Obtained when DCM 1-4 are used followed by
DCM 5 (Figures 12, l6a,b). The poorest performance is obtained when the cellular
layout is used. When setup time is high and part mix balanced, DCM 1, 2 and 4 perform
best followed by DCM 3 and 5 (Figures 12, 16c). The process layout performs poorest.
The result is the same when setup time is high and part mix unbalanced except that the
cellular layout performs poorer. than DCM 5 but similar to the process layout (Figures
12, 16d).

104
Table 8. Treatment Means for Standard Deviation of Work in Process

 

Unbalanced
684.23 "
DCM 2 . 675.09
DCM 3 . 747.38
DCM 4 . 693.47
DCM 5 . _ 743.17

Process layout . 840.25
854.20

 

 

 

 

 

 

 

 

 

 

 

 

 

The ANOVA results indicate that the relative performance of the shop configurations

depends on specific setup time and part mix conditions. They also show that under each
set of conditions examined, DCM generally performs better than the process and cellular
layouts. Only under one set of conditions is there no distinct advantage to be Obtained
by using DCM. The relative performance of different DCM implementations remains
largely unchanged as shop conditions change, DCM l, 2 and 4 generally performing

best.

5.3.5 MW

As described earlier, the presence of higher order interactions makes the interpretation
of linear contrasts inappropriate. Instead, the a priori hypotheses were evaluated using

Tukey multiple comparisons Of treatment means. This was done by identifying those

 

me,

Ti

a. Low Setup

 

Mix
inProcessbySetupxPart

Unbalanced Part

low Setup Time,

b

Standard Dev

Mix

ation of Work

16

Figure

Sarrkrdlkv'ntimcfw

§§§§§§§§

a

..
8

 

 

(1. High Setup Time, Unbalanced Part Mix

Figure 16. (cont'd)

107

Hypothesis

Process layout outperforms DCM
when setup time is low.

treatments included in each hypothesis (Figure 17), and comparing treatment means for

Treatments
1 - 6, 8 - 13

 

DCM outperforms process layout
when setup time is high.

Process layout outperforms DCM
when part mix is balanced.

DCM outperforms process layout
when part mix is unbalanced.

DCM outperforms cellular layout
when setup time is low.

Cellular layout outperforms DCM
when setup time is high.

DCM outperforms cellular layout
when part mix is balanced.

DCM outperforms cellular layout
when part mix is unbalanced.

15-20,22-27
1-6, 15-20
8-13,22-27
l-5,7-12, 14
15-19, 21 -26, 28
1-5,7,15-19,21

8 - 12, 14, 22 - 26, 28

 

DCM that recognizes material flows
outperforms DCM that does not.

 

 

l-5,8-12,15-19,22-26

Figure 17. Treatment Numbers by Hypothesis

all appropriate treatments (e.g., Hypothesis 1, Treatments 1-6, 8-13). The Significance
or otherwise of multiple comparisons can provide evidence to make certain conclusions
regarding the hypotheses. If not, they can still yield information regarding underlying

trends contained within the data.

108
5.3.5.1 11111191111:qu

Hypothesis 1 states that when setup time is low, the process layout outperforms DCM.

The data does not support this (Figures 18, 19a-d). DCM always performs at least as

   

 

 

 

Mean Flow Time

  

 

Mean Work in

 

        

 

  

      
   
   

    

    

 

 

 

log Std. Dev. of
i Tardiness Process WIP '
1 DCM2/Unbal DCM2/Unbal - DCM2/Unbal DCM2/Unbal
j DCMl/Unbal ] DCMl/Unbal DCMl/Unbal DCMl/Unbal
1 DCM4/Unbal ] DCM2/Bal DCM4/Unbal DCM4/Unbal
DCM3/Unbal J DCMl/Bal ] DCM3/Unbal ] DCM3/Unbal ]
i DCM2/Bal DCM4/Ba] DCM2/Ba1 DCM2/Bal
DCMl/Bal DCM4/Unbal ] DCMl/Bal DCMl/Bal
. Process/Unbal DCM3/Unbal DCM4/Bal DCM4/Bal
j DCM4/Bal DCM3/Ba! Process/Unbal DCM5/Unbal
' DCM5/Unbal J DCM5/Unbal DCM5/Unbal ] DCM3/Ba]
j DCM3/Bal Process/Unbal ] DCM3/Bal Process/Unbal]
' Process/Bal ] Process/Bal ] Process/Ba] ] DCM5/Ba] ]
’ DCM5/Ba] DCM5/Dal DCM5/Hal Process/Bal

 
           

 

Figure 18. Tukey Multiple Comparisons for Hypothesis 1

well as the process layout with the exception of the flow time performances of DCM 3
and 5, and the mean work in process performance of DCM 5. DCM 1, 2 and 4 always
outperform the process layout for mean tardiness. DCM 2 also outperforms the process

layout for the log of the standard deviation of work in process.

5.3.5.2 Hmthesiu
Hypothesis 2 states that when setup time is high, DCM outperforms the process layout.

The results support this for DCM 1, 2 and 4, and for the tardiness performance of DCM

5 (Figures 20, 2la-d). DCM 3 and 5 never perform poorer than the process layout.

dense...

tantra emmflmmmmeMfi

 

mmmmmmabmmmmm

ion

SlcpCortiguat

AWMMX IsUhlarudMMx

Time

a Mean Flow

 

WDMMMMMMMMH‘MM

Shop Corfiguation

aflmdMMthMMx

mess

. Mean Tard

b

1s1

Shop Performance for Hypothes

gure 19.

F1

 

 

n

SlcpConr

LWMMX huflnarll’me

Process

111

Mean Work '

C

a. om .

madman“
Eating—Elana

a. n

   

madman-lumba-

ShopCorf

LWMW RWMMX

Process

Standard Deviation of Work in

d

d)

igure 19. (cont'

F

111

 

 

Mean Flow Time Log Mean Mean Work in Log Std. Dev. of
Tardiness Process WIP
DCM2/Unbal DCM2/Unbal DCM2/Unbal j DCM2/Unbal
DCMl/Unbal] DCMl/Unbal DCMl/Unbal DCMl/Unbal h
DCM4/Unbal] DCM2/Bal ] DCM4/Unbal :l DCM4/Unbal 1
DCM3/Unbal DCMl/Bal DCM3/Unbal = DCM2/Bal j
DCM5/Unbal DCM4/Unbal DCM5/Unbal DCM5/Unbal
DCM2/Bal DCM4/Hal DCM2/Bal DCMl/Bal
DCMl/Bal ] DCM5/Unbal DCMl/Bal DCM3/Unbal '
DCM4/Ba1 DCM5/Ba] ] DCM4/Bal DCM4/Bal '1
DCM3/Unbal ] Process/Unbal DCM3/Ba1
J DCM3/Bal DCM3/Bal ] DCM5/Bal ‘]
Process/Unbal DCM5/Ba] Process/Unbal
Process/Ba] Process/Bal Process/Bal

 

 

 

 

 

 

 

 

 

 

Figure 20. Tukey Multiple Comparisons for Hypothesis 2

 

  

5.3.5.3 amnesia
Hypothesis 3 states that when part mix is balanced, the process layout outperforms DCM.

The results do not support this (Figures 22, 23a-d). For mean flow time, the process
layout never outperforms any DCM implementation under the same setup time condition,
being outperformed in all but one case (DCM 5 under low setup time conditions). The
tardiness performance of DCM l and 2 is always better than that of the process layout.
DCM 4 always performs at least as well as the process layout and DCM 5 never poorer
than the process layout. DCM 1, 2 and 4 always yield better work in process
performance than the process layout. DCM 3 also yields lower mean work in process.

DCM never yields poorer work in process performance than the process layout.

.2355

3%“an

 

ion

Shop Corfiglnt

amMMx hWMMx

Time

Mean Flow

a.

 

mmmbmmmmmmmm

Shop Corfrguation

ammeflitflx hwmm

iness

Mean Tard

b.

rs2

Shop Performance for Hypothes

Figure 21.

sees...“ a...

 

mmsmmmmmmmmmm

Ship Corflglntion

aflmdhnMx hW-eedMMx

in Process

Mean Work

C.

 
 

  

Rent

mmmmm

m m
$3ng

  

lb

mammal.

mamabmmmm

Slop Corfigu'ation

ANIMMMK RWMMx

Process

rn

Deviation of Work

Standard

(1.

(0

(cont

igure 21.

F

  

  

Mean Work in

  

      

 

  

Q Process/High

 

 

 

Process/High

 

 

DCM5/Low

1 Mean Flow Time Log Std. Dev. of 7'

? Tardiness Process WIP "
DCM2/Low DCM2/Low DCM2/High 1 DCM2/High1

T DCMlILow] DCMl/Low ] DCMllHigh DCMllHigh r
DCM4/Low ] DCM4/low DCM4/High DCM4/ High

: DCM3/Low DCM2/High DCM2/Low DCM2/low 1
Process/Low] DCMllHigh ] DCMl/Low ] DCMl/Low

. DCM5/Low DCM3/low ] DCM4/Low ‘ DCM4/Low ‘

‘ DCM2/High DCM4/High DCM3/High ] DCM5/High
DCMllHigh ] Process/Low ] DCM5/High ] DCM3/High "

. DCM4/High DCM5/Low J DCM3/Low . DCM3/Low -'

. DCM5/High ] DCM5/High ] Process/low Process/High]
DCM3/High DCM3/High Process/High] DCM5/low ]

Process/Low

 

 

          
           
 

 

   

 

   

 

       
 

  

 

Figure 22. Tukey Multiple Comparisons for Hypothesis 3

535.4 mm

Hypothesis 4 states that when part mix is unbalanced, DCM performs better than the

process layout. The results support this for DCM l and 2 except for their mean flow

time performance, and for the work in process performance of DCM 3 and 4 (Figures

24, 25a-d). DCM 1-4 always perform better than the process layout under the same setup

time conditions and DCM 5 at least well as the process layout.

5.3.5-5 W

Hypothesis 5 states that when setup time is low, DCM outperforms the cellular layout.

The results support this (Figures 26, 27a-d). For all measures, the cellular layout always

yields the poorest performance.

tween

gE-Wskth:

A

“natal.

2

he.

a.

a.

.2.»

362.9- .

 

mmmmmmmmmmmm
SlsspCorflglntlon

1151133191111. 11 “dauphin

Time

Mean Flow

a.

 

mmmmmmmmmmmm

Ship Contention

alumna. smattering

Mean Tardiness

b
Shop Performance for Hypoth

3

esis

Figure 23.

s

<.

Mean Work in Process

0

Standard Deviation of Work in Process

(1

rd)

(cont

igure 23.

F

   

.1. .0160...

mmmmmmmmmmmo
gassed;

a. lawns. 11 “mm

 

lefihp'flaa ls Hfim‘fln

1

Mean Work in

Log Std. Dev. of 5

 

 

 

 

 

 

 

 

 

1 Mean Flow Time Log Mean

i Tardiness Process WIP }
DCM2/Low DCM2/Low DCM2/High ‘ DCM2/High i '
DCMl/Low ] DCMl/Low DCMllHigh DCMllHigh

, DCM4/Low DCM4/Low DCM4/High 1 DCM4/High r

‘ DCM3/Low 3 DCM2/High DCM2/Low DCM2/Low

, Process/Low] DCMllHigh ] DCMl/low DCMl/Low i
DCM5/Low DCM3/Low DCM4/Low “ DCM4/low ‘

‘ DCM2/High DCM5/Low DCM3/High - DCM3/Low
DCMllHigh ll Process/Low ] DCM3/Low DCM5/High ‘

1 DCM4/High - DCM4/High DCM5/High DCM3/High .
DCM3/High ] DCM5/High ] Process/Low DCM5/Low

? DCM5/High DCM3/High Process/High] Process/Low ]
Process/High Process/High DCM5/low Process/High

Figure 24. Tukey Multiple Comparisons for Hypothesis 4

5.3.5.6 Hmtheiisfi
Hypothesis 6 states that when setup time is high, the cellular layout outperforms DCM.

The results do not support this (Figures 28, 29a-d). DCM 1, 2 and 4 always outperform
the cellular layout. DCM 3 and 5 always perform at least as well as the cellular layout.

The cellular layout always yields the poorest tardiness performance.

5.3.5.7 Hmthcsisl
Hypothesis 7 states that when part mix is balanced, DCM outperforms the cellular

layout. The results support this for DCM 1 and 2 except for their mean work in process
performance, for DCM 4 for mean flow time and tardiness, and for DCM 3 and 5 for
mean tardiness (Figures 30, 3la-d). DCM never performs poorer than the cellular layout

with the exception of the mean work in process performance of DCM 5.

. w

,5.

.. .. seeds 4

a

$3.5.

aztamnwwmwfimmmwmmfiafifi

 

n

ShOpCO

alumna h Mandi”

 

mmmmmmmmmmmu

Slnp Corfiglntion

shalom h Hfimm

Mean Tardiness

e

b

4

Shop Performance for Hypothesis

igure 25

F

 

[Us

mmsmmmmmmmm

ShopConr

alumna: hl-l‘rhp‘flm

Mean Work in Process

0

 

alumna. hHfl-m'fln

in Process

Standard Deviation Of Work

d.

. (cont'd)

25

igure

F

120

 

 

 

 

 

 

 

 

i Mean Flow Time Log Mean Mean Work in . Log Std. Dev. of 1
' Tardiness Process WIP i
, DCM2/Unbal 1 DCM2/Unbal DCM2/Unbal - DCM2/Unbal 7

‘ DCMl/Unbal DCMl/Unbal DCMl/Unbal DCMl/Unbal
DCM4/Unbal DCM2/Bal ] DCM4/Unbal DCM4/Unbal -.
DCM3/Unba1 DCMl/Bal ] DCM3/Unbal '1 DCM3/Unbal "

; DCM2/Bal DCM4/Ba] DCM2/Bal DCM2/Ba1 1

, DCMl/Bal DCM4/Unbal ] DCMl/Bal DCMl/Bal

‘ DCM4/Ba] DCM3/Unbal DCM4/Ba] DCM4/Bal
DCM5/Unbal DCM3/Bal ] DCM5/Unbal ‘ DCM5/Unbal "

j DCM3/Ba] DCM5/Unbal DCM3/Bal .1 DCM3/Bal
DCM5/Bal - DCM5/Bal DCM5/Bal DCM5/B31
Cellular/Bal Cellular/Bal Cellular/Bel Cellular/Bal

; Cellular/Unbal Cellular/Unbal Cellular/Unbal Cellular/Unbal

Figure 26. Tukey Multiple Comparisons for Hypothesis 5

535.8 M
Hypothesis 8 states that when part mix is unbalanced, DCM outperforms the cellular

layout. The results support this with the exception of the work in process performance
of DCM 5 (Figures 32, 33a-d). With this one exception, the cellular layout always yields

the poorest performance for all measures.

5.3.5.9 Hmthesisj

Hypothesis 9 states that DCM implementations that recognize material flows in cell
formation (DCM 3-5) yield better performance than those that do not (DCM l & 2). The
results do not support this (Figures 34, 35a—d). Given the range of shop conditions

examined, it is understandable that no single implementation consistently yields the best

 

 

(' ”Av.“
cicada

«43'

 

mmmmmmmmmmm cut
Slanorfigtntron

tHluIlMMx RWMMX

a Mean Flow Time

Man'l‘arrilm

 

mammmmmmmmmmab

SlvpConfisrnhon

amMMx hwmm

b. Mean Tardiness

Figure 27. Shop Performance for Hypothesis 5

.4:

.8.

a...

fiamscwwwfirewwtw

. is. .35.? .
news... BREE

 

mmbmmmmmmmmmm

Shop Corflglntio

aHmdMMx hlflimdP-th

in Process

Mean Work

c.

sham.“

. . “we?!

 

(lb

mm.mmmmmmmmm

Corfiglmion

Shop

tantalum hlflflamedhth

Process

Standard Deviation of Work in

d

. (cont'd)

27

Figure

 

   

123

 

 

 

  
 
  
 
 
 
 
   
 
 

Cellular/Unbal

 

 

_1—=

 

; Mean Flow Time Log Mean Mean Work in Log Std. Dev. of

‘, Tardiness Process WIP

‘ DCM2/Unbal DCM2/Unba1 DCM2/Unbal DCM2/Unbal

, DCMl/Unbal ] DCMl/Unbal DCMl/Unbal] DCMl/Unbal ]
DCM4/Unbal DCM2/Bal DCM4/Unbal DCM4/Unbal ]

, DCM3/Unbal ] DCMl/Bal DCM3/Unbal J DCM2/Bal
DCM5/Unbal DCM4/Unba1 DCM5/Unbal] DCM5/Unbal

; DCM2/Bal DCM4/Bar DCM2/Bal DCMl/Bal

‘ DCMl/Bal ] DCM5/Unbal DCMl/Bal DCM3/Unbal j
DCM4/Bal DCM5/Ba] ] DCM4/Bal DCM4/Ba]

w DCM5/Bal DCM3/Unbal ] DCM3/Bal DCM3/Bal ]
DCM3/Bal ] DCM3/Bal Cellular/Bal ] DCM5/Bal

l Cellular/Bal Cellular/Bal DCM5/Bal Cellular/Bal

Cellular/Unbal Cellular/Unbal] Cellular/Unbal

 

 

 

 

Figure 28. Tukey Multiple Comparisons for Hypothesis 6

performance. However, DCM 1, 2 and 4 typieally yield the best performance under a

given set Of conditions. Of these, only DCM 4 recognizes material flows.

53.6 Symmcfmrchjmtheses
The information yielded by the multiple comparisons indicates that DCM performs well

under a wider range of conditions than anticipated. In comparison to the process layout,

DCM generally performs better regardless of setup time conditions. DCM performs

better, as expected, when setup time is high. It also performs well when setup time is

low, a scenario in which setup time was not expected to greatly compromise the

performance of the process layout. With respect to part mix, the results suggest that

DCM performance is in general better, but not conclusively so. Comparing DCM to the

cellular layout, the results suggest that DCM performs better not only when setup time

2
~§$r§§33§§£> w‘"
‘ $3 "

3.5“
$3.888:
W’MNN
“audits-:19”;
1?.0

3.9333.
!I.'W
W

a

4.55:1;

r1394»
:6 ville ”K
”'

awaits-3’
(me? "'

”(A
>5.
$5

{kin-z

Nhanliow'l'lm

MW

.1.» M'sfi’mhiaaaaewn
new ,

mime
N wen-wr-
area.

ma!- _,
Zr ’

54%:
93x64:

41'»

415')

3K

M

 

a. “landfill“: [IN-nod” M!

 

LEMMA“ kWh“:

b. Mean Tardiness

Figure 29. Shop Performance for Hypothesis 6

 

Minibmmmmmumman Ina-cu
SlanOlfrgrntlon

LNMMMx hUhIMudP-er

0. Mean Work in Process

WWdWIP
assesses

 

mmbmmmmmmmmmm

Slop Corfrgrntion

AHMMMx kWMMx

d. Standard Deviation of Work in Process
Figure 29. (cont'd)

           

 

 

  
  
  
  
   
   
 
  
 
 
 

 

 

 

 

 

   

 

 

 

 

Mean Flow Time Log Mean Mean Work in log Std. Dev. Of

Tardiness Process WIP
DCM2/Low DCM2/low DCM2/High ‘1 DCM2/High -
DCMl/low ] DCMl/low ] DCMl/High DCMllHigh *
DCM4/Low ] DCM4/low DCM4/High DCM4/High
DCM3/Low DCM2/High DCM2/Low DCM2/Low 1
DCM5/Low DCMllHigh ] DCMl/Low DCMl/Low
DCM2/High ] DCM3/low ] DCM4/low i DCM4/low a
DCMllHigh ] DCM4/High ] DCM3/High DCM5/High
DCM4/High ] DCM5/low J Cellular/High DCM3/High -'
DCM5/High DCM5/High J DCM5/High DCM3/Low ‘
DCM3/High ] DCM3/High DCM3/low ‘ Cellular/High ]
Cellular/High Cellular/High DCM5/Low DCM5/Low
Cellular/ Low Cellular/ Low Cellular/ Low

 

 

Figure 30. Tukey Multiple Comparisons for Hypothesis 7

is low, as expected, but also when setup time is high. Under these conditions, the cellular
layout was expected to have an advantage. As hypothesized, DCM performed better for

both part mix conditions.

5.3.7 W

In addition to the primary performance measures discussed SO far, data was also collected
for secondary performance measures. DCM is able to obtain the benefits described above
while simultaneously increasing effective capacity (Table 9). Mean utilization for DCM
ranges from 0.4 to 5% lower than that for the process layout, depending on setup and
part mix conditions. As expected, the cellular layout consistently yields low utilization,
varying from 61% to 70%. The utilization of the different DCM implementations is

essentially similar with the exception of DCM 5 which yields utilization that is about 2%

a a...
9,9 “atria
533' "

NEE;
an; -

9.0:.

I
(a

R

 

mmmm mmmm

ms. m malt new
ShopCorflguatlon
811me htH‘tatw'firm

a. Mean Flow Time

 

alarm“ hflfiwm

b. Mean Tardiness

Figure 31. Shop Performance for Hypothesis 7

 

mmmmmmmmmm

mum

ion

Shop Corflguat

alcwstttprure hustttpnme

in Process

Mean Work

C

 

mmbmmmmmmmaamaa

ion

Slnp Corflgrnt

alumni: hHflrhpTln

on of Work in Process

31

Standard Deviati

d

(cont'd)

igure

F

129

 

 

 

Mean Flow Time Log Mean Mean Work in Log Std. Dev. of
Tardiness Process WIP
DCM2/Low 1 DCM2/Low ] DCM2/High '1 DCM2/High -
DCMl/Low DCMl/Low DCMllHigh DCMllHigh -
DCM4/low DCM4/low DCM4/High DCM4/High
DCM3/Low DCM2/High J DCM2/Low '1 DCM2/Low
DCM5/Low DCMllHigh DCMl/Low DCMl/Low
DCM2/High DCM3/Low ] DCM4/low DCM4/Low -*
DCMllHigh DCM5/Low. J DCM3/High DCM3/Low
DCM4/High . DCM4/High DCM3/Low DCM5/High -
DCM3/High DCM5/High ] DCM5/High - DCM3/High ] -
DCM5/High DCM3/High DCM5/low ‘ DCM5/Low
Cellular/High Cellular/High Cellular/High Cellular/Hith
Cellular/Low Cellular/ low Cellular/Low Cellular/ low

 

 

 

 

 

 

 

 

 

 

 

 

Figure 32. Tukey Multiple Comparisons for Hypothesis 8

Table 9. Treatment Means for Mean Utilization

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Setup Time Low High
I Part Mix Balanced Unbalanced Balanced Unbalanced

DCM 1 0.790 0.763 0.784 0.757
DCM 2 0.789 0.762 0.782 0.755
DCM 3 0.795 0.768 0.791 0.762

222%, DCM 4 0.791 0.765 0.786 0.759
DCM 5 0.770 0.747 0.754 0.733
process layout 0.799 0.772 0.804 0.774 I
Cellular Layout 0.698 0.677 0.628 0.607

 

 

mmmmmmmbmmm

alumna: h Hamm-

$5

.' ”it” ‘.
his-1;.
orig ‘2 .m- 43.13%»

3'.

2%

alarflqflh hfl‘anpfln

b. Mean Tardiness

Figure 33. Shop Performance for Hypothesis 8

ea .

C117

 

s

a“

rememfi

\a
m
\

teammates s.
.\. a 5\

‘43.

w

mmsmmmmmmmmm

ion

Shop Conngtnt

Mean Work in Process

C

(11

mmbmmmmmmmmm

ion

Slanorfigu'at

Process

In

. Standard Deviation of Work

(1

Figure 33. (cont'd)

was
wt

..
. 33¢. .2.
see are.“

c. «5.. x.
v muss

c. its

alumni: hH‘rW'firm

 
 

alumna: thlhrpTin

132

 

 

 

 

 

 

 

a £858»: C8 288860 nae—=2 zoo—Pr .vm Bani

 

 

 

 

 

 

 

 

 

.5 .5 m3

 

are? :82

 

=82»:

 

533328 53558 555328 55558

53558 533328 5525328 Hm 529558

525328 523558 _ 523558 525328

525558 * 525528 5525\58 525228

. 553558 5533828 53558 525528

535428 . 53558 H . 555348 : 5529558

533228 e M 533228 5523558 T 555558

. a 53558 . 53558 - 553558 53558

. 5225428 . 5225428 53558 5525528
5525328 2 555228 . 525228 5525228 _

555228 . 522558 ... 525558 5529558

. 5525558 . 5523558 5533328 n 533528

55525558 e e 5533328 5525228 553558

525558 . 5529558 . 5525328 533328

5535428 e5e5:93:28 e e 5533328 533228
5533228 e 5533228 4 533328 H 53558 x
55252.28 5533328 533228 H 553558 .

F 553558 5:255:28 53558 5539528

2 5555228 5 5555228 5533828 _H 5:233:28
5525828 5555558 e5533228 553558 ,

.53 Co 5.88; 5 ”8525
085. 32m :82

    

man. mu m m We m
m m m um: m
fianoxfigtnnon
I. mmm Handex h Immune. WMMMX
c.HdIMme Wham d H‘ISwTum. UlilnedMMx

a. Mean Flow Time

 

Alumnus. NMMMX hmmflm WMMX
cflfimm MMMMLdH‘ImflmWMMx

b. Mean Tardiness

Figure 35 . Shop Performance for Hypothesis 9

 

LLGISIIQTIIB. sewn-rm h mmm mmm
c.mmprmwmm «mm-rm. Wham

c. Mean Work in Process

5

Samhdlkv'nflmd‘WIP
§§§§§§§§

 

I. MMMHMMMX h lawn-n, UlimdMMx
emu-pun. WNW d MWMWMMMX

d. Standard Deviation of Work in Process

Figure 35. (cont'd)

528855 3850:»: .8 8555 6m 052m

 

 

 

 

 

588 8: on 86: 388:: 83:88: 5:: 209 d _
.8855: a: 8:: :8: 8:3
.53 .5 8:88: 8:82 :8 :85 :8 n Sun :8 588 E83. 593 85=8 :88898 Sun .w
8:55: :8 85: 3o: :85 :8 v 20a :8 ::o8<
.8858: :85 :8 m 8 m 209 :8 ::o8< .885: m: 8:: :8: 8:3
8:3 8:: :88 8585 8:85:88: =~ :8 N d _ 28 :8 ::8o< 5o»:— 85=8 :88898 Sun .h 1
:8: a: 2:: :53 8:3
1 4:88 8: :0 20a «8889:: 59:: 85:00 6

 

135

r .583:

.32 a: 2:: :58 8:3
59¢: 85:8 «8889:: 20m—

 

d; .3 585:8: 68:53.: :8 8:: :8 v a. m SUD :8 «:83:

68:28:: 8 8:: :8: 8:3

 

 

 

 

.05: 3o: :85 E88 858:: 88888: =: :8 N on _ Sun :8 ::8o< 59:: 808: ”8889:: 20m .v
.555 m: 5: 3 8:3

4:83. 52 on Sun :88895 59?. :88: .m
88:58: :85 :8 n Son :8 583. :9: m: 2:: :53 8:3

.858:— 88888: =: :8 v 8 N J Sun :8 583:. 5o»:— ..88:: 8:88:5o 20n— .N

 

 

 

4:88 5: on

.32 a: 0:5 :38 8:3
20D «888:5: 50»:— 888:

 

 

:88880

 

 

38:89»:

 

136

lower. This is due to the enforced machine idleness that it permits. DCM 5 also yields
the poorest overall utilization of the five DCM implementations. Considering only DCM
1-4, the maximum difference in utilization between DCM and the process layout is about
2.2%. The ability of DCM to increase effective capacity is as anticipated higher when
setup time is high. Not only are utilization levels lower when setup time is high (with the
exception of the process layout upon which the 80 % utilization level was established),

but they are also lower when part mix is unbalanced.

With few exceptions, DCM yields lower proportions of jobs tardy than either the process

or cellular layouts (Table 10). Whereas the process layout has proportions tardy between

Table 10. Treatment Means for Mean Proportion Tardy

‘ Setup Time

 

Unbalanced Unbalanced =

0.003 . 0.007 :
DCM 2 . . 0.007
DCM 3 . . 0.022
DCM 4 . . 0.014
DCM 5 . . 0.022
Process layout . . 0.046
Cellular Layout . . 0.253

 

 

 

 

 

 

 

 

 

 

 

 

 

1.3 and 6.5% and the cellular layout between 35.4 and 60%, DCM has at most 3.1%

tardy, with the proportion generally much lower. When setup time is high, DCM

137

implementations always yield lower proportions tardy than the process layout. The

relative benefit of DCM is, as expected, higher when setup time is high and when part

mix is unbalanced.

As anticipated, the proportion of time spent incurring setups is lowest when using the

cellular layout (Table 11). Conversely, jobs in the cellular layout also spend the greatest

Table 11. Treatment Means for Mean Setup Time Proportion (Mean Setup Time)

 

 

 

 

I Setup Time Low High
, Part Mix Balanced Unbalanced Balanced Unbalanced
DCM 1 0.113 0.115 0.203 0.206
(23.35) (22.28) (48.23) (46.08)
DCM 2 0.113 0.114 0.202 0.206
(23.25) (22.05) (47.67) (45.87)
DCM 3 0.116 0.118 0.205 0.211
(24.91) (23.61) (51.24) (49.03)
Config. (23.77) (22.77) (48.82) (49.99)
DCM 5 0.090 0.095 0.170 0.179
(20.58) (20.03) (42.33) (41 .70)
Process Layout 0.113 0.117 0.196 0.204
(25.48) (24.39) (52.56) (50.59)
Cellular layout 0.030 0.020 0.060 0.053 f
(966 191-99 (69> «669 1

   

 

 

   
  
 

 

    
 

 

    
 

 

     
 

 

    
 

 

 

 

 

 

 

   

 

    
 

proportion of time in queue even under conditions known to be conducive to CM (high

setup time, balanced part mix). There is little difference in the proportion of time spent

in setups between DCM and the process layout when setup time is low, with the

138

exception as expected of DCM 5, where jobs spend less time incurring setups. Curiously
however, when setup time is high, DCM, again with the exception of DCM 5, yields
marginally higher proportions. DCM 5 consistently has setup time proportions 2-3%
lower than the next best configuration other than the cellular layout. This suggests the
potential of DCM 5 as setup time increases further. Part mix has little effect on the

proportion of time spent in setups.

DCM yields large improvements in proportion of time spent in queues (Table 12). When

Table 12. Treatment Means for Mean Queue Time Proportion

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Setup Time Low High

Part Mix Balanced Unbalanced Balanced Unbalanced
DCM 1 0.233 0.215 0.232 0.214
DCM 2 0.233 0.216 0.229 0.213 E
DCM 3 0.242 0.221 0.244 0.220 I

Emu. DCM 4 0.237 0.221 0.237 0.220
DCM 5 0.291 0.264 0.272 0.248
Process Layout 0.265 0.239 0.282 0.252
Cellular Layout

 

 

 

 

 

part mix is balanced, the proportion for DCM is generally of the order of 23-24% with

the exception of DCM 5. DCM 5 yields higher proportions due to the potential delay in
re-allocating machines. For the process layout, this figure is between 26 and 28% . When

part mix is unbalanced, the proportion is around 21-22% for DCM and 24-25% for the

139

process layout. When setup time is high, even DCM 5 yields marginal improvements

over the process layout.

5.3.8 121W
The analysis of effects and a priori hypotheses demonstrate the benefits of DCM. When

the shop configurations are compared under different operating conditions, DCM always

performs as well as, if not better than, the traditional process and cellular layouts.

As expected, the cellular shop, being rigid and inflexible, consistently performs poorly.
Similar to existing findings (e.g., Morris, 1988), its relative performance is good only
when setup time is high and part mix is balanced. Though the cellular layout outperforms
the process layout with respect to mean flow time and work in process under these
conditions, it cannot outperform any of the DCM implementations, and is consistently
outperformed by most of them. The impact of reduced setup frequency in the cellular
layout is small compared to the considerable loss of routing flexibility, providing
, additional evidence of the effects of permanent machine dedication. Even under
supposedly conducive conditions, the cellular layout yields the worst due date
performance, likely the result of large flow time variance. When part mix is unbalanced,
performance is particularly poor, the result of uneven utilization, frequent bottlenecks,

and ever increasing queues.

140

The failure to accept hypothesis 6 shows that even under conditions that have been shown
to be conducive to CM, the addition of flexibility to CM systems has a significant impact
on their performance. Although the reduction of setup frequency can have a beneficial
effect by reducing queue sizes, if this is done while permanently dedicating equipment,
the benefits are significantly lower than when flexibility is present. If the location of a
bottleneck were to remain constant and the cell configuration designed to accommodate
this, the cellular layout can be expected to perform better. Typically however,
bottlenecks are non-stationary. By letting cells evolve by shrinkage and growth to adjust
to this, a cellular configuration can overcome this problem. Alternatively, the planning
. system must consider conditions within individual cells when making decisions regarding

job release to those cells.

The performance of the process layout compared to DCM is more interesting since it
does not have the same problem of inflexibility. However, as the results demonstrate, its
lack of recognition of part families is a significant factor. Even when this might have
been expected to be a relatively minor problem, i.e., when setup time is low, DCM,
despite being relatively less flexible, performs better. Indeed, the relative performance
of the process layout is only marginally better under low setup time conditions than under
high setup time conditions. Comparing the process layout to DCM under different part
mix conditions, there is again little difference in relative performance. These observations
suggest that under normal and common operating environments, DCM is a better choice

than a process layout.

141

The results demonstrate that the lower flexibility of DCM compared to the process layout
does not compromise material flows through the shop. When machine dedication is
permanent as in the cellular layout, forcing jobs to utilize specific machines often leads

to problems of long queues. However, these problems are not encountered in DCM.

Clearly all DCM configurations do not perform the same, though in general, differences
in performance between them are small. This is particularly true for flow time related
performance. DCM 1-4 consistently perform better than DCM 5. This shows that
increasing the degree of permanence of cells, even dynamically formed cells, has a
detrimental effect on performance. This provides additional evidence to support the
assertion of Flynn & Jacobs (1986) that machine dedication is a limiting factor in the
performance of traditional cellular systems. It also suggests that forcing machines to
remain idle may not be beneficial. However, it is also clear that the relative performance
of DCM 5 improves as setup time increases, as might be expected. This suggests that
under extreme setup time conditions, increasing machine dedication may be beneficial,
WWW. Given the simplicity of the cell
formation heuristic used by DCM 5 and the fairly small decrease in performance when
it is used when setup time is high (flow time is 5% higher than the best DCM
implementation when part mix is balanced), a similar but more efficient heuristic that
utilizes greater cell permanence, may make such cells more viable. One way to
accomplish this is to consider the length of time a machine is allowed to remain idle. If

this is longer than the time it takes to carry out a major setup, immediate reallocation

142

of the machine may be more appropriate. As it is currently implemented, DCM 5 does

not consider this trade-off.

Of the remaining DCM implementations, the observation that DCM 3 yields relatively
poor performance under high setup time conditions is interesting. DCM 3 explicitly
attempts to promote the flow of jobs by extending cells forward to their successor
departments, allowing the cell to grow. However, when setup time is high, this
implementation performs poorly relative to the two more myopic implementations, DCM
l and 2. On the other hand, DCM 4, which also considers shop-wide family processing
requirements, performs relatively well. This may be due to DCM 3 being too myopic
itself by not considering the extent to which the newly allocated machine can be used by
a family. The potential exists for the machine to be allocated to a family with only a
single job in the current queue and only the job currently being processed in the
predecessor department. Under this extreme case, only two jobs take advantage of the
major setup incurred. On the other hand, the current queue may contain families without
jobs in process in their respective predecessor departments, but more jobs or more urgent
jobs in the current queue. DCM 4 is again likely to be more effective under these
conditions since it considers processing requirements of the family throughout the shop,
not just at the current and predecessor departments. This suggests that DCM 3 may yield
better performance if its implementation is modified to make it recognize material flows
more globally. This could be done by considering the total number of jobs in both the

current and predecessor departments that can use the new setup.

143
Comparing DCM 3 and 5 as a group to DCM l, 2, and 4, the results show that

differences in mean flow time when setup time is high, are, though statistically
significant, small in magnitude. DCM 3 and 5 have flow times that are less than 4%
higher than that required for them not to be statistically significantly different from the
other implementations. The difference in flow times is due to increases in both setup time
and time spent in queues. However, the increase in time spent in queues is relatively
higher. This lends further support to the contention that DCM implementations that
recognize information about material flows may indeed perform better than those that do
not, if they are designed effectively. Recognizing material flows provides a mechanism
to route work more efficiently. By making available all machines required by a family,
the potential for delays while jobs await major setups is reduced. This enables jobs to

pass through the shop with the fewest obstacles.

An observation concerning all DCM implementations is that they generally perform better
when part mix is unbalanced. This suggests that under these conditions, the greater
ability of parts from high demand families to share setups more than offsets the increase
in setups caused when corresponding machines are reassigned to families with low

demand.

5.3.9 W

The results of stage one demonstrate that DCM is a more effective means of production

than that using a traditional process or cellular layout under certain conditions. For the

144

conditions examined, DCM consistently outperforms these alternatives when conditions
are not suited to them. Under conditions conducive to these alternatives, DCM performs
at least as well as them, and often better than them. The increased setup efficiency of
DCM allows the process layout to be operated more efficiently than at present, despite
the loss of some degree of routing flexibility. The increased flexibility of DCM compared
to traditional CM, enables family based production to be carried out so that it is more
responsive to change, without being compromised by the decrease in setup efficiency.
Given the tradeoff between flexibility and setup efficiency that exists in a small/medium
batch production environment, DCM offers an alternative between the extremes of high
flexibility/low efficiency (process layout), and high efficiency/low flexibility (cellular
layout). The results suggest that some sacrifice along one dimension is justified and in
fact beneficial, if it is substituted with an increase in the other. Furthermore, it appears
that it is more beneficial to sacrifice setup efficiency than flexibility. DCM allows the

tradeoff to be made without changing the physical nature of the shop.

5.4 ANALYSIS OF STAGE II DATA

5.4.1 W

The treatments included in stage two are re-stated in Figure 37. Of the five DCM shop
configurations used in stage one, DCM 4 was selected for use in stage two. DCM 4
consistently performed well in stage one and is also one of the implementations that
utilizes shop information on a more global scale, actively seeking to complete the

machine requirements of part families.

 

 

145

 

 

 

 

 

 

 

  
 

 

 

    

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Legend : 1 - 54 : Treatment Numbers

Dispatching Rule :

Processing Time, Minslk - Minimum Job Slack
Volume Mix : 100 - Job Size = 100, N(100,10) - Job Size = N(100,10), 50

- Job Size = 50

Balanced Part Mix : Part families have equal demand probabilities

Volume Mix 100 I N(100,10) I 50 I
Utilization(%) 7oIsoI90I7o 80I90I70I80I90I

Disp. Rule Part Mix

FCFS Balanced l 7 13 19 25 31 37 43 49
Unbalanced 2 14 20 26 32 38 44 50

srrr Balanced 3 9 15 21 27 33 39 45 51
Unbalanced 4 10 16 22 28 34 40 46 52ll
Balanced 5 11 17 23 29 35 41 47 53
Unbalanced 6 12 13 24 3o 36 42 4s 1.3—1..

 

FCFS - First Come First Served, SPT - Shortest

Unbalanced Part Mix : Three part families have demand probabilities of .233,
two have demand probabilities of .15)

Figure 37. Stage II Treatments

5.4.2 mm;

For mean flow time, though there was a reasonable fit with the assumption of normality,

this was improved considerably by the use of a log transformation. For the raw data, all

but eight treatments yielded PCCI‘ values within 5 % of that required to accept the

hypothesis of normality. However, three were more than 20% less than the critical value.

When a log transformation was used, all but three of the treatments yielded PCCT values

within 5 % of that required to accept the hypothesis of normality, and all were within 7%.

None of the three transformations used increased the homogeneity of residual variances.

146

However, examination of the residuals showed that heterogeneity increased as utilization

increased and performance deteriorated.

The analysis for mean tardiness again showed that using a log transformation on the data
yielded the best fit with the assumption of normality. All but ten treatments yielded
PCCT values within 5% of that required to accept the hypothesis of normality. The fit
with the raw data was poor. Homogeneity of variances also increased when this
transformation was used. When a log transformation was used for the two measures of
work in process, the fit with the assumptions was again better than that with the raw
data. For mean work in process, all treatments yielded PCCT values within 2.5% of the
critical value, and for the standard deviation, all but one was within 4% . However, as
with the other measures, the variance of residuals increased as utilization increased, but

this also led to a deterioration in shop performance.

5.4.3 Analxsiuififfmts
5.4.3.1 Mmflmlime

ANOVA results for the log of mean flow time are reported in Table 13. Treatment
means for mean flow time are reported in Table 14. In order to examine the impact of
the significant interactions, Tukey multiple comparisons were carried out at each level

of utilization for each combination of part mix and volume mix (Figures 38, 39a-c).

Table 13. Analysis of Variance for Log Mean Flow Time

SOURCE

147

SS

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Random Numbers 99 4.976 0.050 7.56 0.0001 3
Dispatching Rule (D) 2 0.047 0.024 3.54 0.0292
Part Mix (16) 1 5.381 5.381 809.33 0.0001 :
[Volume Mix (V) 2 61.292 30.646 4609.69 0.0001 ;
Utilization (U) 2 101.471 50.735 7631.47 0.0001 .
D 16 P 2 0.019 0.009 1.41 0.2446 :
D 9- v 4 0.030 0.007 1.12 0.3474
F) =6 U 4 0.063 0.016 2.38 0.0495
16 1- v 2 0.599 0.300 45.05 0.0001
16 11 U 2 15.198 7.599 1142.99 0.0001
v 6 U 4 0.023 0.006 0.87 0.4780
D .. 16 11 v 4 0.007 0.002 0.28 0.8922
D 11 P 6 U 4 0.039 0.010 1.48 0.2065
D .. v 11 U 8 0.036 0.004 0.67 0.7192
P 61 v 4 U 4 0.924 0.231 34.74 0.0001 I
D 11 16 .. v 16 U 8 0.013 0.002 0.24 0.9823
[32ml- 5247 34.883 0.007 I

 

R2 = 0.84

 

148

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

E: 32] “3.16.ch 3M: 2% 28 6.6: 685.25

6.8: 5H: 3: 6.2m 6.2: :8: 9:: an: new: 6855: 5::

3% 6.2: 32 3% 3.8 3.5: :6: 58 me: 6858::

2.: 32 2: an: 58 $5 :8 6.2: as: 6858 :8

8mm ”.8: 32 3% 2.68 we: cam“ 38 6.8: 68585

«.8: 3.3 2: new we: 2:: we: 3.: new: 6858 m8:
525: 23:95

 

8_8_2._8_8_E_8_8_2.

e883:

 

05F 38: :82 :8 882 50558:. .3 035,—.

818:2

 

52 2:28,

 

 

149

 

 

 
  
   
   

 

 

 

Log Mean Flow Time by Part Mix x Volume Mix I
Utilization = 70% Utilization = 80% Utilization = 90% I
Unbal/SO Unbal/SO ] Bal/SO
Bal/SO Bal/SO Bal/N(100,10)
Unbal/lOO ] Unbal/lOO J Hal/100 ]
Unbal/N(100,10) Unbal/N(100,10) UnballSO
BalliOO ] Bal/lOO ] Unbal/lOO ]
Bal/N(100,10) Bal/N(100,10) Unbal/N(100,10)

 

 

Figure 38. Tukey Multiple Comparisons for Log Mean Flow Time

The results show that at low utilization levels (70%), flow time is as expected, lowest
when jobs are of batch size 50, with flow times lowest if part mix is unbalanced. For
jobs of batch size 100, performance is the same when part mix is unbalanced, regardless
of whether job size is constant or variable. Performance deteriorates when part mix is
balanced, though again it is not affected by variability in job size. As utilization increases
to 80%, these results repeat themselves with the exception that when job size is 50, there
is no difference if part mix is balanced or unbalanced. At high utilization levels (90%),
the results change dramatically. Flow time is lowest when job size is 50 and part mix
balanced. However, there is now no difference between jobs of size 50 when part mix
is unbalanced, and jobs of size 100 when part mix is balanced. The poorest performance

is obtained when jobs are of size 100 and part mix unbalanced.

5.4.3.2 Mines:
ANOVA results for log mean tardiness are reported in Table 15. Treatment means for

mean tardiness are reported in Table 16. The significant effects were examined by

2t
. $3.“
www.mtomit.
mama .
t...
.
t

.r

 

mmmuu-uummm

”Sac-Inv-

c.

”’53-”

bin‘humlm. 50-

= 70%

lization

a. Uti

lidmsxtwsr .
xx x

k
v

 

wumm-mmmm

bin-mull} ”-flSI'”

C-bh'm V

80%

tilization =

b.U

Mix

Time by Part Mix it Volume

. Mean Flow

39

Flgure

. 2%..
tanwwwowntwmt . .2.
Km?“ . . “R...

;,\'n\

m
i

«n»

-.
z

t
.‘
é

 

m—mmmw-mmmm

C-len-m V-bfiu-Nlmlq SO-HSBISO

c. Utilization = 90%

Figure 39. (cont'd)

152

Table 15. Analysis of Variance for Log Mean Tardiness

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

  

 

  

 

  

 

    

 

 

SOURCE DF 35 MS 1: p
Random Numbers 99 490.68 4.96 5.38 0.0001
Dispatching Rule (D) 2 924.11 462.05 501.99 0.0001
Part Mix (p) 1 53.84 53.84 58.50 0.0001
Volume Mix (V) 2 99.30 49.65 53.94 0.0001
Utilization (U) 2 4031.04 2015.52 2189.73 0.0001
1) * p 2 9.70 4.85 5.27 0.0052
D 4 v 4 114.57 28.64 31.12 0.0001
D * U 4 1423.32 355.83 386.59 0.0001
P * v 2 0.11 0.05 0.06 0.9435

[R * U 2 5.72 2.86 3.11 0.0447

I v * U 4 492.89 123.22 133.87 0.0001

F) * p * v 4 1.16 0.29 0.32 0.8674
D * P .. U 4 7.42 1.86 2.02 0.0895
1) .. v .. U 8 171.29 21.41 23.26 0.0001 .
p * v * U 4 3.60 0.90 0.98 0.4187 ,
D * p * v . U 8 13.36 1.67 1.81 0.0695
Error 5247 4829.56 0.92

 

 

 

 

 

 

R2 = 0.62

 

153

 

 

 

 

 

._. _.___ A_¥v _—. ,#_k --——7 , #744

 

 

 

 

 

 

 

 

 

 

 

 

--,,wm.mw._ mm .8... S :N z; x was So: an; 55 Rosana: 110
as ”no 85 3.2 Rd :3 2.: one 55 .6853. ans:
3.3 «an Rd :53 a; mac 3.8m 26 was 8653.5
2.8 8d o; 3% m3 mg 9.3 92. Rd 8258 ha
Rama 2; :d 3.3 an; 8... 3.3 s... 85 8653::
an: 85 85 «Sm 86 85 8.2 a; 85 8653. moon
525.— 23— 95
8fi8_cn_8_ow_oa_8%ow flop 823%:
8_

818:2

”8:63p. :82 .5“ «:82 8258:. .2 033—.

 

.32 082.5

 

154

carrying out Tukey multiple comparisons by utilization for each combination of

dispatching rule and part mix, and dispatching rule and volume mix (Figures 40, 41a-c,

42a-c).

Log Mean Tardiness by Dispatching Rule x Volume Mix

 

‘ Utilization = 70%

Utilization = 80%

Utilization = 90%

 

, MINSLK/100
MINSLK/N(100,10)
; FCFS/100

, FCFS/N(100,10)

f MINSLK/50

, FCFS/50

, SPT/N(100,10)

. SPT/100

J
l

J

 

MINSLK/N( 100, 10)
MINSLK/100
MINSLK/50

FCFS/ 100
FCFS/N(100,10)
FCFS/50

SPT/50

SPT/N(100, 10)
SPT/ 100

l
J

 

MINSLK/50
FCFS/50
MINSLK/100
MINSLK/NOW, 10)
SPT/50

FCFS/ 100
FCFS/N(lOO, 10)
SPT /N(100, 10)
SPT/ 100

 

 

Log Mean Tardiness by Dispatching Rule x Part Mix

Utilization = 70%

 

Utilization = 80%

Utilization = 90%

 

' MINSLK/Bal
MINSLK/Unbal

7 FCFS/Ba]

FCFS/Unbal

: SPT/Ba]

, SPT/Unbal

 

MIN SLK/Bal
FCFS/Bal
MINSLK/Unbal
FCFS/Unbal
SPT/Ba]
SPT/Unbal

 

 

MINSLK/Bal
FCFS/Ba]
SPT/Ba]
MINSLK/Unbal
SPT/Unbal

FCFS/Unbal

Figure 40. Tukey Multiple Comparisons for Log Mean Tardiness

At low utilization levels, tardiness is as expected lowest whenever the minimum slack

dispatching rule is used. The FCFS rule outperforms the SPT rule. For both of these

rules, tardiness is lower when part mix is balanced. Slack based dispatching generally

yields lower tardiness when job size is large, regardless of whether it is constant or

variable. At 80% utilization, tardiness is lowest when slack based dispatching is used and

Sad

M1113, ax-

m
[ltd-mammal:

= 70%

' tion

. U

a

5..

c .5: ...
“Mama.
4

a. ram

‘2“?

x
x

“:4

$1
a?

“a“

8d

Minn-Mi.-

m
llflmedMMx

on = 80%
Dispatchin

Utilizafi

b.

Mean Tardin

Mix

essby g RulexPart

41

““33?
x
‘As "'

. gmufifi “Rain?
. .3.&...»«% a
2“ tm“

WMMLLHII
-WMMS.LIH

m-mmmwm-

m-EICh-MMS’I-

 

Figure

u“

a.

x...

to .. t. at?
a “gamma”
.

I...
a.“

:2
.2...
. too

. w.)
.mmoa .5
t Rattlazi
5‘

s

as
a...
x

. . moat
twaifiwnvas

44.
s“

 

“M mm mm S’l‘

m-mmmMm-mnmnm SK-Mn'nunflack
H-WMML W-lltdmedlmm

c. Utilization = 90%

Figure 41. (cont'd)

5“,...) égwa aways?
> Evy;

 

Q

S’I'C

 

m

“1)

s $4

. V

”38'”

«mm. ”-

’Aw.“ .
\vnx'w: ,

“IV-HS}!

Hafiz:-

C-

= 70%

Utilization

a

 

6
S
4
3
2
l
0

 
 

 

a... a, ., 8.1a.
.. aim?» .. “watvaawa
a... \<. \s .2 s..\

.. Rama.

3m 2w... #5 are ..
..
.

 

a.

n
a.
.w.

m
m

 

.32..“
ix

GR

~.-.
'0

“MN‘:
tun:

M

ii: .
mean»?
Rana».
z.

3%.

Its.-

 

£333....
mm

IJW
m mammalian-mama... 1K-

a
.“93.

3:3“

c-aan-mv-uau-mmlqso

80%
Dispatchin

Utilization

b

Mean Tard

1X

g Rule x Volume M

mess by

1gure 42.

F

       

arr-so
Sad

INC

              

“V

“can“?
Wnauwxwa 44.4
zawa. ..MM. Mm“:
4454.4.

     

. MW.

4 ”44 44.. 444.4%

..4
4 44
$31144

um m!)
m-mmmws’r-mmm ax-

C-l‘bflu-m V-bSu-Nlmlq, SO-th-S)

c. Utilization = 90%

Figure 42. (cont'd)

159
part mix balanced. There is however no difference between slack based dispatching when

part mix is unbalanced and FCFS dispatching when part mix is balanced. The relative
performance of the remaining three scenarios is the same as that when utilization is 70%.
Slack based dispatching always yields the best performance regardless of job size, though
again performance is better when jobs are of size 100. The FCFS dispatching rule
performs better than SPT for all job sizes, but unlike slack based dispatching, there is
no difference due to job size. When SPT dispatching is used, small jobs yield relatively
better due date performance than large jobs. At high utilization levels, there is no
difference between minimum slack and FCFS dispatching when part mix is balanced.
Tardiness is always lower when part mix is balanced. Only at 90% utilization do small
jobs yield relatively lower tardiness, and this only if minimum slack or FCFS dispatching
are used. Overall, slack based dispatching as expected yields better performance, and

SPT dispatching performs poorly.

5.4.3.3 WES.

ANOVA results for log mean work in process are reported in Table 17. Treatment means
for mean work in process are reported in Table 18. Tukey multiple comparisons were
carried out for each utilization level, for each combination of dispatching rule and part
mix, and part mix and volume mix (Figures 43, 44a-c, 45a-c). When utilization is 70 or
80% , there is no difference in work in process based on particular dispatching rule and
part mix combinations. However, at 90% utilization, a balanced part mix always yields

lower work in process. For each part mix, there is no difference due to dispatching rule.

160

Table 17. Analysis of Variance for Log Mean Work in Process

SOURCE

U1
"‘1:

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Random Numbers 99 9.818 0.099 28.40 0.0001
Dispatching Rule (D) 2 0.002 0.001 0.34 0.7113
Part Mix (P) 1 2.414 2.414 691.19 0.0001
Volume Mix (V) 2 101.203 50.601 14488.80 0.0001
Utilization (U) 2 164.557 82.278 23558.92 0.0001
D ... p 2 0.028 0.014 4.06 0.0173
D * v 4 0.010 0.002 0.71 0.5832
D * U 4 0.032 0.008 2.31 0.0552
P * v 2 0.876 0.438 125.36 0.0001

; P * U 2 7.939 3.970 1136.60 0.0001
v * U 4 0.067 0.017 4.83 0.0007

I D .. P r v 4 0.004 0.001 0.26 0.9023
D 4 p * U 4 0.058 0.014 4.13 0.0024
D a v 4 U 8 0.011 0.001 0.40 0.9189
R v v * U 4 0.969 0.242 69.37 0.0001 I
D r p ... v . U 8 0.007 0.001 0.24 0.9831 I
Error 5247 18.325 0.003 J

 

 

 

 

 

 

2 = 0.94

161

E

 

E

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

8:. 32 2: £8 9% on; 38 38 EN Beige:
:2 a: 3: 45 Nam 52 :fi 38 menu 365.8 6:452
83. a: a: N3... 8% EN «3. 8% ”SN 8653:: .
Sum no: on: 3.: 88 wemm N4: 88 8Q 8653 Em
:9. a: ma: 9% 28 EN 5e 8% 88 Boise:
EN :3 on: NE 28 :2 new 88 as 86543 wave
52 Ba use 95
8_8_2_8 Mew—oe—oo 8 2 85%:
on 818:2 8_ a: oE=_o>

 

 

 

888m 5 3.83 :82 ecu 382 Sega—h .m— 23H

162

 

Log Mean Work in Process by Part Mix x Volume Mix

 

 

   
   
  

 

Utilization = 70% Utilization = 80% Utilization = 90%
Unbal/SO Bal/SO J Bal/SO

Bal/SO Unbal/SO Unbal/SO
Unbal/lOO ] Unbal/100 ] Dal/100 ]
Unbal/N(100,10) Unbal/N(100,10) Bal/N(100,10)
Bal/IOO ] Bal/IOO Unbal/lOO
Bal/N(100,10) Bal/N(100,10) Unbal/N(100,10)

 

 
 
 

   

 

 

   

 
    

 

   
   

 

 

 

 

Utilization = 70% Utilization = 80% Utilization = 90%
SPT/Unbal ‘l FCFS/Unbal '1 SPT/Bal
FCFS/Unbal SPT/Unbal FCFS/Ba]
MINSLK/Unbal MINSLK/Unbal MINSLK/Ba]
FCFS/Bal SPT/Ba! MINSLK/Unbal
SPT/Ba] FCFS/Ba] FCFS/Unbal
MINSLK/Ba] " MINSLK/Ba] - SPT/Unbal

 

 

Figure 43. Tukey Multiple Comparisons for Log Mean Work in Process

Work in process is as expected, always lowest when job size is small. At 70%
utilization, an unbalanced part mix yields lower work in process than a balanced part mix
when jobs are of size 50. For jobs of size 100, work in process is lower whenever part

mix is unbalanced, though for a particular mix, variability in job size is not significant.

These observations repeat themselves at higher utilization levels with few exceptions. At
80% utilization, there is no difference between a balanced and an unbalanced part mix
when job size is 50, but for jobs of size 100 under balanced part mix conditions, constant

job size yields lower work in process. At 90% utilization, jobs of size 50 yield the lowest

 

mm
m-mmmwwr-mmnm sx-Mn'mnm
H-WMM’Llflll-LHHMMMX

a. Utilization = 70%

§

3111 . » . _ “.3“
' . .4 xii;

342:: 4:4

\K-ziflkafi‘;

4
44

44

444m:
494:4

0

 

W
m-hmmmm-mmmu-Mfimm
m-mmmw-mmmm

b. Utilization = 80%
Figure 44. Mean Work in Process by Dispatching Rule x Part Mix

afifiwfiyfi

 

WW

WEB/L
m-mwmmm-mmnm ax-

H-WMM‘W—WMMMX

and

c. Utilization = 90%

Figure 44. (cont'd)

u.

4
4.
v4
«N
«.4

...... . ...
. {>434 4. arias
. . am.

mm. m. mafifigu

4 .4524. “w.

a 4445
4 4M.
4 4%“memu .3“ 4

...A. 4.4 4.4.4 .0442 ..

44 4'2
4 . .44
3:?

4.4
3‘.

 

V

 

1)

olfldmdMMx

. . 44
.44444.x LR
<44 44...

444

a)
sirloin smith-30

WMMLUIII
hh-mv-bh

H
c.

70%

a. Utilization =

\v
5.4
tau
M4a44
4 4444
..4 .

“4...

. .
m4“ inn“ 4.4.“.4.» .
...”.Rafimaiama 4.44.4
.fi {$454 .44 4,444,449“? . . .44
. .....a a 4.

..

vs xx \ 4.
... 41344
..

aacwovv

.2 “wavium v.4..4.“..4am.m»d4 .
44% H 5.444. ...mawxa.» «.4. «WAR .
$44.... 45.. .44. M 345444. 4444 4 . . . .
4“: «$4. 444 4“.) 4 . u u . .
Momwrsa.%mem ”@4444. . ..mM

5.4

4.44444h
4m .44...

. 4

-
«

\ 4 \ V Vv

amass.

\
4
4
4

..Wx

as:

4

‘ 9$4
§~M

. 4333:
3544
..

.
§§

4
\‘n.
‘4

‘3-

1%?

4a!

mmmmmmm

4 .
. 434.. ~523me

 

~mMMx4 [Ii-WMMX
”Sn-m V-bh-Nflnlfi. ”-lbll'”

H
c.

80%

tilization =

b.U

Mix

Mix x Volume

Process bme

1n

Mean Work

lgure 45.

F

4..
44

:5
. ...4. “33%....

R

44.43.49»

“44
4

 

V

BIL-V

1)

Mr!)
m-mmmuu-mmedmm

C-ltfiu-lm V-HSu-Nlmlo, SO-I'bfin-SO

90%

c. Utilizafion

Figure 45 . (cont'd)

167
work in process, particularly when part mix is balanced. For jobs of size 100, a balanced

part mix yields lower work in process than an unbalanced part mix.

5.4.3.4 WW

ANOVA results for the log of the standard deviation of work in process are reported in
Table 19. Treatment means for the standard deviation of work in process are reported
in Table 20. Tukey multiple comparisons were carried out for each utilization level for
all combinations of dispatching rule, part mix and volume mix (Figures 46, 47a-c). The
comparisons show that the standard deviation of work in process is always lower when
job size is 50. For this job size, no differences exist based on part mix or dispatching
rule. When utilization is 70 or 80%, there is no difference in work in process based on
particular dispatching rule and part mix combinations. However, at 90% utilization, such

differences do exist. Performance is generally lower when part mix is unbalanced.

5.4.4 WWW

‘ The proportion of time spent in setups decreases as utilization increases, falling from
12.6% to between 6 and 8% for jobs of size 100, and from 21.6% to between 10 and
14% for jobs of size 50 (Table 21). Most of this decrease occurs when utilization goes
from 80 to 90%. The proportion is also higher for jobs of size 50, typically of the order
of 9% higher than for jobs of size 100. This falls to only 6% as utilization increases. For
jobs of size 50, setup time proportion at high utilization levels is also around 2-3%

higher when part mix is balanced. The SPT dispatching rule yields poorer performance

168

Table 19. Analysis of Variance for Log Standard Deviation of Work in Process

 

 

 

 

 

  

 

 

 

 

 

 

 

  

 

  

 

 

  

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

SOURCE DF ss MS F p I
Random Numbers 99 23.383 0.236 69.13 0.0001
Dispatching Rule (D) 2 0.062 0.031 9.14 0.0001
Part Mix (P) 1 0.764 0.764 223.51 0.0001
Volume Mix (V) 2 121.311 60.655 17752.16 0.0001
Utilization (U) 2 75.250 37.625 11011.79 0.0001
D * p 2 0.001 0.001 0.20 0.8208
D * v 4 0.014 0.004 1.04 0.3828
D * U 4 0.035 0.009 2.60 0.0345
F * v 2 0.350 0.175 51.26 0.0001
9 * U 2 0.157 0.079 23.03 0.0001
v * U 4 0.454 0.114 33.25 0.0001
D * P r v 4 0.074 0.019 5.42 0.0002
D * p * U 4 0.001 0.000 0.10 0.9840
D * v * U 8 0.035 0.004 1.28 0.2467 II
p * v * U 4 0.315 0.079 23.06 0.0001 II
D * p * v * U 8 0.150 0.019 5.48 0.0001 i
Error 5247 17.928 0.003 1

 

R2 = 0.93

 

{lulflflflfl i

169

mg 5m 3m hN— _ 85 an mug m2. 05 805395

 

vwm 5m mum 09: Sum eve am: 0mm 3.0 Bonn—«m 53:2

 

bnn mwm 3m 2Q fit. 50 cmm— mi. 20 coon—WEED

 

 

 

En mam Nun 9.9 Sm 3.0 3: VS N3 @3575 Em
a3 mwm N3 2 : Eh as nmu— wk. 20 Bog—3.5
whm own man 33 mg 3.0 83 mum :6 coo—83m mmum

 

 

 

 

 

 

 

 

 

en: fin 23— 6.5

 

8 8 E_8_8_E_8 8 E Sega:

 

 

 

 

 

 

on _ 618:2 _ 8_ , 6:: oeso>

I L ‘ ll

 

285 5 6.63 e6 Season Raga é 28: .8532... .8 2.3

 

170

 

 

 

 

 

«.885 5 x83

.6 .5qu 2355 ms 8.“ macaw—3860 032:2 hoe—E. .3 23E

. 8:39.523
818324.828

fl 831555?—

8.432%
818:2:«EEm
833.524sz
8:3..Emmum
SSSSELm

§ .8:z=2eEEm
818:2:2524523
r 818:2:3e2mmom
.. omEeDDszE
anfieeaamncm
3.28%;sz
332%
82820..
823525

I'—

 

If

 

81835.84??? -

 

j

 

 

  
 

8382435
818:2imiqmz—2
SSoEmmum
8:3:Em
adsosziemmuu
s_.8:z=om\Em
oozaoeagqmzfi
818324352523
8. 332435430...
8_.8:z\_S=EEm
8S2=2Em
83352an
5.8545sz
8:25:sz
33328
omzoeeazem
3.82%
8435428

61835324525 ,

 

 

8:39qu5
s_.8:z\§\mn_um
Giacczfifitm
8_\§\Em
cozaemmum .

§ .8:z\_3e:§4_mz_2
§.8:z\_sea\mmum
8:3..25ng ,
8_.8:z\aee2.Em
cozaeeEEm .
835.2%?—
Rim—24%;.
535% .

om\_dm\mn_0m _.
8:352..ng .
amineEEm
8435mm“:

 

 

«.8 .1. Sega:

 

 
 

 

*8 u conga:

 

   

*2. fl 59355

 

 

52 oeao> u «.2 Em x 63 “56.83 3 a? .6 son .3 ”3

 

IK-BAb-C

SX-BMrV

w M
um um
mm W.
mm mm
................H mm ... mm ...
fimfiem. “W m a m
vu m % W m
M m m
mm“ .m m M
“mm m M 4
mm m U m m
w . M
mm. . a w. o
. m
u
m

uni-Immune

.“ “4 a
44 $§a

[ALSO mm»

4 .444

24 a $4.. .44
44 .
3 34m

. .
.“ .uv4 “n ..< 4,».

91-.
mmmwm-mmmu-umm

ET-UNBAIAO mm”
HIOIBMMM-mmm ax-

. n. \ tr“ \.\
4444.444 4.. 4

. .4 4... .mm.
unmumvam .44 a
.5 . $3.3“. 4.3 «353.34..

  

m5 arr-1mm “WAD” SPT-UNBAbc “WV 81me

mms SK-UPBMAO 91'

. . . a)444

. .... I. :

mmmmmmmmmr
Haggai—um

WMMLUH
-WMMX,UIH

KB-
m.
m-
H

 

=80%

Utilization

b.

Rule x Part Mix

tching

Di

in Process by
x Volume M

on of Work

47. Standard Deviati

igure

F

ix

4
4 G4
4

..44w««4

we“.

54;...

3.4645449: 4

$4 454
..444v4u.44.>»4.<

 

err-11mm,” 9133M,” mmmsmwamv m-UNBALV MAL-C mm WT-BALC m-BMIV

”'BAIIV

SIC—BM

mmso Eli-BAD” “MAL” alt-{MALV mm “mm mmv

m-n-mumws’r-mnmnu. ax-

Sad

all-manna.w-mmmmc-bSu-mv-uh-mmm so-uu-so

c. Utilization = 90%

Figure 47. (cont'd)

173

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

818:2

 

 

£2 oEEo>

cow—88m oEF aaom :82 8m «:82 .5838; . _ N 035.

 

 

83 aflmd 83 83 x3 gdflflflafio 8053::
R3 83 :3 N83 :3 53 ~83 :3 33 38.58 32.2
23 83 :3 2.3 23 #3 R3 23 :3 805.25
2.3 33 23 383 23 83 383 23 83 805.8 Em
33 S3 :3 83 83 a3 83 83 $3 805.25
133 83 £3 :3 23 £3 :3 23 83 8°53 Eu“
525m 23— 95 .
8_8_2_8_8_E_8_8_E 838%: ,
on 8_

\\\\\

174

when part mix is unbalanced, increasing the proportion by 1% when job size is 100, and

2% when it is 50.

Results for queue time proportion mirror those for setup time proportion (Table 22). The
proportion increases as utilization increases, going from 17-19% to as high as 53%. Jobs
of size 50 are in queues proportionately longer when utilization is low. At 70%
utilization, they consistently spend about 2% longer in queues than jobs of size 100.
However, at medium and high utilization levels, the proportion is higher only when part
mix is unbalanced. At 80% utilization this is only about 1% higher, but at 90%
«utilization it is 4-5% higher. Queue time proportion is consistently higher for an
unbalanced part mix particularly when utilization is high, except when SPT dispatching
is used. For jobs of size 100, this proportion is 4-5% higher. For jobs of size 50, an
unbalanced part mix yields an 11% increase when FCFS or minimum slack dispatching
is used, and 4% when SPT dispatching is used. This increase is only 2% when utilization

is 80%.

The choice of dispatching rule yields differences, again primarily at high utilization, with
SPT dispatching yielding lower proportions. When job size is a constant 100, SPT
dispatching yields 6% lower queue time proportions when part mix is balanced, and 10%
lower when part mix is unbalanced, compared to the FCFS rule. Compared to slack
based dispatching, these differences are 5 and 8%. Compared to jobs of size N(100,10),

the differences are 7 and 11% and 6 and 9% respectively. For jobs of size 50 they are

175

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

a $3 83 33 $3 $3 $3 $3 83 8053::
793 83 83 33 $3 83 £3 :3 83 .353 “as:
=33 S3 83 83 S3 83 33 S3 83 38535
83 R3 83 83 $3 83 $3 33 83 8°53 Em
£3 £3 33 83 $3 $3 83 83 $3 8053::
$3 83 83 :3 83 :3 $3 $3 $3 8053 Eon
_ 52 5.. 23. .95 n
— 8 _ E _ 8 _ 8 _ E 823%:
M 818:2 52 2.3.5

 

EVER—Em 08:. 03:0 :82 ..8 2.82 2.258; .mm 033.

 

 

176
4 and 10%, and 3 and 9% respectively. At 80% utilization, SPT dispatching typically

yields only a 1-1.5% improvement over other rules.

As expected, proportion tardy increases significantly as utilization increases (Table 23).
For jobs of size 100, it goes from near zero at 70% utilization, to 1-4% at 80%
utilization, to 14—30% at 90% utilization. For jobs of size 50, it rises from near zero to
2-6% to 17-40% at utilization levels of 70, 80 and 90% respectively. Again, smaller jobs
consistently fare poorly, particularly when part mix is unbalanced. At 90% utilization,
when part mix is unbalanced, proportion tardy is typically 9% higher for jobs of size 50.
Proportion tardy is slightly higher when job size is N(100,10) compared to a constant
100. Again, small increases in pr0portion tardy are caused by an unbalanced part mix.
These rise as utilization increases. At 80% utilization, an unbalanced part mix yields
increases in proportion of about 1% when job size is 100, and 2% when job size is 50.
However at 90% utilization, these rise to 10 and 20% respectively, except when SPT
dispatching is used. In this case the increases are of the order of 2 and 7% respectively.
At 70 and 80% utilization levels, the FCFS and minimum slack dispatching rules yield
similar proportions tardy. SPT dispatching yields 1-2% poorer performance at 80%
utilization, but at 90% utilization, it is more effective. For jobs of size 100, it yields 5-
6% lower proportions tardy when part mix is balanced, and 13-14% lower when
unbalanced. For jobs of size N(100, 10), these are about 2% higher in each case, and for

jobs of size 50, about 2% lower.

177

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

_ 83 aflad 913313433--- a3! ammo! .3131-.-&o.to Mimcwl J,
:33 e3 83 83 23 83 83 23 83 8053 332
:33 83 23 R3 33 83 33 33 83 8053::
—:.3 3.3 23 £3 33 83 $3 33 33 8053 #5
=83 33 83 R3 £3 83 83 33 83 805.25 _
A 83 83 83 23 £3 83 83 23 83 8053 96.:
£2 ta.— 23— damn— i
8_8_2_8_8_2_8_8—S 833%: 1
on §.8:z 82 a: 2:25

 

 

3.8% 5an95 532 .28 882 Eon—«~29. .mm 035.

178
5.4.5 W

The results demonstrate a number of characteristics of DCM, some of which are similar
to those of traditional process and cellular shops, and some which appear to be specific
to DCM. Utilization, as expected, has a considerable effect on performance. Whereas
an increase in utilization from 70 to 80% does not significantly affect flow times, an
increase to 90% has a dramatic effect, particularly when part mix is unbalanced. Mean
flow time increases by about 10-15% when utilization is increased to 80%. When
utilization is 90%, flow times (compared to those at 70% utilization), are typically 70
or 44% higher when part mix is balanced, for jobs of size 100 and 50 respectively.

When part mix is unbalanced, these are 190 or 250% respectively.

Similar increases in work in process are found. Mean work in process increases by 25-
30% when utilization increases to 80%. At 90% utilization, work in process is 127 or
216% higher (compared to work in process at 70% utilization) when part mix is
balanced, for jobs of size 100 and 50 respectively. When part mix is unbalanced, these
are 216 and 282% respectively. The corresponding figures for the standard deviation of
work in process are a 2030% increase at 80% utilization regardless of part mix, and

increases of 120 or 80% at 90% utilization.

This in turn explains the large increases in tardiness that occur at higher utilization
levels. At 80% utilization, the percentage increases are large due to the negligible

tardiness at 70% utilization, though actual tardiness is low, of the order of less than

179

seven minutes. However, at 90% utilization, tardiness is of the order of 10-60 minutes
for a balanced part mix, and over 200 for an unbalanced part mix. This behavior is to
be expected since it is well known that shop performance deteriorates under high

utilization levels (Baker, 1984).

The result of increased utilization and the cause of the. decreased performance is as
expected a dramatic increase in time spent in queues. At 80% utilization, the impact of
queues is relatively small since there is sufficient capacity to absorb the extra workload.
At 90% utilization, this is no longer true. The benefits of setup efficiencies are eroded
by the volume of work. However, this does not detract from the increase in effective
capacity yielded by reducing major setups. As was seen in stage one, even at 80%
utilization, DCM outperforms a traditional process layout. One can surmise that at higher
utilization levels, the greater setup efficiency of DCM will yield relatively larger
improvements over the traditional process layout. Indeed, DCM performance at 90%
utilization is not considerably poorer than comparable process layout performance at 80%
. utilization (see stage one results). Even at 90% utilization, flow time performance with
a balanced part mix does not increase in an explosive manner. Given that existing CM
systems are more conducive to an environment characterized by a balanced part mix, the
results show that the incorporation of flexibility increases the potential of manufacturing
based on the cellular concept. The traditional cellular layout, as stage one demonstrates,

performs poorly even when it is operated at utilization levels of 60-70%.

180

An important point to recognize is the impact of minor setups on DCM performance.
DCM attempts to schedule jobs within a cell based on a single stage dispatching rule and
treats minor setups as relatively insignificant. As utilization increases, though the
frequency of major setups may decrease due to an increase in number of jobs in a queue
from the same family, the frequency of minor setups necessarily increases. By reducing
this frequency, for example by using sequence-dependent scheduling within a cell, or by
reducing minor setup times, potential exists to reduce the negative impact of higher
utilization further. A possible drawback with sequence-dependent scheduling however,
is that it discriminates against jobs requiring a change in minor setup. A large number
of identical jobs in the queue could thus lead to an increase in flow time variance and
thus tardiness, by making jobs that require a setup endure long waits. This problem will
be exacerbated if the number of machines in each process department is small in relation

to the number of part families.

The response of DCM to part mix and volume mix variability is an important
characteristic of it. At high utilization levels, unbalanced part mix leads to a degradation
in flow time related performance, not unlike traditional CM. Though the setup time
proportion does not change significantly except when job size is 50, queue time
proportion does. Two explanations are plausible. First, the permanence of cells
corresponding to high demand families may be resulting in low demand families
competing for few remaining machines. In addition, when high demand families gain

access to multiples of the same machine, these machines must be relinquished when

181

required by cells lacking that machine type. The result is that more frequent setups occur
on these two groups of machines, i.e., those used by low demand families and ‘shared’
machines. This may be resulting in longer waits for jobs competing for these machines.
No adverse impact on average setup time proportion is seen since cells corresponding to
high demand families incur few major setups. An alternative explanation is that since
multiple machines must be relinquished by more permanent cells as demand elsewhere
dictates, these cells, though able to increase their capacity when machines are available,
generally have fewer machines than necessary to efficiently reduce the queues in front
of them. More likely, the cause of poor performance under unbalanced part mix
conditions is a combination of these two explanations. DCM therefore exhibits behavior
similar to a traditional cellular layout with respect to part mix, but only at much higher
levels of utilization. At lower utilization levels, DCM is unlike traditional CM,
performing better when part mix is unbalanced. It appears that not only is there a greater
ability of high demand families to retain multiple machines of the same type, but
competition among low demand families for machines not already allocated, is lower.
This reduces the impact of setups and queues. These observations further substantiate the
claims of Flynn & Jacobs (1986) regarding the effects of permanent machine dedication.
Reducing machine dedication and allowing machines to supplement cells or be re-

assigned between them, makes CM more responsive to changes in production needs.

The impact of small jobs is also evident. Although as expected flow times are lower at

low and medium utilization levels when jobs are small, these flow times are greater in

182

proportion to job size than for large jobs. At low and medium utilization levels, flow
times of jobs of size 50 are 60% or more of those of jobs of size 100, indicating the
absence of returns to scale. Consistent with the findings for part mix, this becomes worse
at higher utilization levels when part mix is unbalanced. Flow times grow to as much as
70% of that for jobs of size 100. The only instance when flow times of jobs of size 50

are in proportion to their size is when utilization is high and part mix balanced.

Less pronounced effects exist for work in process. At 70 and 80% utilization levels,
mean work in process is as expected, half that for jobs of size 100. At 90% utilization
however, this is only 44% when part mix is balanced, but 62% when unbalanced. The
standard deviation of work in process is in all cases close to the one half expected. The
poor performance when jobs are small can be attributed to the increased impact of
setups. Setup time is larger in proportion to processing time when batch size is small.
This can be seen by the higher setup time proportions when jobs are of size 50. This also
translates into longer times spent in queues by jobs awaiting setups, increasing the queue
time proportion further. In order for small jobs to be processed efficiently, further setup
time reduction is required. It should however be pointed out that since for each utilization
level the arrival rate was established based on processing and setup times, this has an
important effect on small jobs. Due to more frequent setups, the productive capacity of
the shop is lower when batch size is small. The arrival rate for jobs of size 50 is
consistently about 58% of that for jobs of size 100, not one half. This implies that if the

impact of setups on small jobs is held constant, the arrival rate will decrease, and work

183

in process will increase. What remains to be seen is whether the increase in arrival rate
will further degrade performance, or whether the reverse will happen. Increased arrival

rate increases the potential number of jobs that can share a setup, and may thus reduce

setup frequency.

The absence of differences in performance due to introducing variability in job size when
job size is 100 does suggest DCM to be robust with respect to uncertainty in volume
mix. One would expect a reduction in variance to yield better results, yet the results
consistently show this not to be so. Given this and the observations regarding part mix
variability, DCM appears to be an effective method of production even in an environment

characterized by the kind of uncertainty modelled here, except when utilization is high.

As expected, DCM exhibits similar behavior to process and cellular layouts with respect
to job dispatching. Minimum slack dispatching yields better due date performance when
this is measured by mean tardiness. In addition, it performs well at high utilization levels
when part mix is unbalanced. However, SPT performs relatively better with respect to
proportion tardy. This is no surprise based on evidence from past job shop research. For
other performance measures, dispatching rule has little effect on shop performance. This
can be attributed to the fact that dispatching rules are often used in job shops in situations
where more effective planning would have been more appropriate. Since DCM explicitly

considers family production characteristics prior to scheduling jobs, it compensates for

184

poor planning to a greater extent than traditional job shops do. This reduces the impact

of dispatching rules.

5.5 SUMMARY OF STAGE 11 RESULTS

The results suggest that DCM performs well under a wide range of shop conditions and
is fairly robust against usual forms of variability found to adversely affect other shop
configurations. Performance is good at both low and medium utilization levels, and
shows substantial deterioration only when utilization is at a high level. Even then DCM
appears to provide improved performance compared to traditional methods of small batch
production. DCM is robust to part and volume mix variability except when the shop is
subject to heavy loads. However, DCM does not perform well when batch size is small.
DCM is in general not overly sensitive to different dispatching rules, but when it is, it
exhibits behavior similar to that found in traditional process and cellular layout

production.

5.6 SUMMARY OF EXPERIMENTAL RESULTS

The results indicate that DCM is an effective configuration for small/medium batch size
production under the conditions investigated. Not only does it outperform traditional
process and cellular layout production under conditions when it was expected to, but also
when conditions were thought to be more conducive to these alternative configurations.
Further, while these other configurations have in the past been shown to be sensitive to

variability in the shop environment, DCM is not affected in the same way.

185

Though one of the objectives of DCM is to make small batch production more efficient,
DCM does not perform well when batch sizes are decreased. DCM resolves some of the
difficulties inherent in existing production methods, particularly with respect to
scheduling issues, but it does not address others such as setup time. DCM focusses on
operating the shop more effectively, not changing the physical characteristics of the shop.
Only by simultaneously addressing both can additional improvements in performance be
obtained. Despite this, DCM appears to be more effective than existing small/medium

batch production methods for the same batch size, setup time environment.

5.7 CONCLUSIONS AND FUTURE DIRECTIONS

A preponderance of evidence suggests that manufacturing systems that physically embody
the principles of CM perform poorly in an environment characterized by small batch
sizes, changing demand patterns, and an emphasis on short lead times. This is true
regardless of whether the system is composed either fully or partially of manufacturing
cells, or if it is designed to compensate for the limitations of cellular production methods.
- This research shows that the principles of CM can be utilized effectively if two of its
main properties, layout and similarity in part design/processing needs, are separated. The
study shows that if the production system embraces CM’s philosophy with regard to part
design/similarity in isolation from CM’s layout requirements, it has the potential to
perform well under a variety of conditions that can be found in contemporary batch

production environments.

186

The research carried out in this study has examined relatively simple implementations of
DCM under a limited set of conditions. However, this does not undermine the value of
DCM. To the contrary, by keeping the design of DCM simple and not introducing
potentially confounding factors, it emphasizes how important the separation of the layout
and design/processing aspects of CM is. In addition, it demonstrates that reaping the
benefits of DCM is a realistic objective for manufacturing organizations, since DCM
does not add significant constraints, either operational or financial, to existing job shop

production environments.

The results suggest a number of areas where additional investigation may add to the
benefits and understanding of DCM. As discussed earlier, one of the reasons why DCM
implementations that recognize material flows do not perform as well as anticipated is
that their design is possibly overly simplistic and myopic. Similar but more far-sighted
cell formation methods may more clearly demonstrate the benefits of recognizing material
flows. In addition to using heuristics to form dynamic cells, the formation of dynamic
cells using optimization methods may further improve the performance of DCM. This
might be accomplished by using optimization models similar to those used for traditional

CM cell formation.

Given that DCM cells are not fixed entities, part families do not have to be pre-
determined as they must be in traditional CM. This suggests that ‘families’ or groups of

parts that will be processed together, can be created in real time based on parts currently

 

187

awaiting processing. This also eliminates the problem of parts that do not naturally fit
into existing families. Instead of identifying families based on setup requirements, they
might be distinguished based on the machines individual parts use, and/or the sequence
in which they are used. This is similar in principle to how similarity coefficients have
been used in traditional cell formation models. This also allows the potential for DCM

to be evaluated in an environment characterized by less repetitive production.

The study indicates that DCM becomes less effective as batches become smaller. This
is due to the relatively greater impact of setup time. One can surmise that as setup time
is reduced, this will be less of a problem. Closer examination of the impact of setup time
may make it possible to determine whether smaller batches can be produced as effectively
as larger batches, or whether lower setup time still yields relatively better performance
with larger batches. Another way to evaluate the effect of setup time and batch size is
to consider the use of lot splitting. One would expect that some reduction in batch size
may improve throughput by making processing more continuous. Since DCM recognizes
family processing requirements, problems of increased setup frequency found in other
studies on lot splitting should be reduced. This can be extended to incorporate transfer

batch scheduling based on the repetitive lots logic.

Additional factors that define the physical characteristics of the shop can also be expected
to affect DCM performance. The impact of these factors needs to be investigated. These

factors include shop size, family size, and the number of families. Any negative effects

188
of dedicating a particular machine to a family are likely to be reduced if the number of

machines of that type is increased. More machines of the same type should result in less
competition for remaining machines of this type. Similarly, fewer families should reduce
the competition for available machines. The size of a family may be important since the
larger a family is, the larger is the number of its parts that may require the use of a
machine at any instant. This may increase the extent to which the family retains use of

machines, thus increasing the life of the cell.

The research also provides insight into how traditional CM implementations might
perform more effectively. DCM’s separation of the processing and layout properties of
CM shows that CM is effective if the shop is not viewed as individual cells that do not
interact. Part of the reason for this layout effect may be due to the planning system
releasing jobs to the shop, not to individual cells. The cells however are independent and
have their own processing, capacity and workload constraints that change over time.
Releasing work to the shop without regard to the status of individual cells may place
demands on cells that are inconsistent with their current capabilities. In contrast, DCM
and job shops consist of process departments that are linked by the routings of individual
jobs, and by prevailing processing requirements. Since there are no permanent layout
restrictions, the shop as a whole assumes the demands placed by prevailing work
patterns. Whereas a planning system that releases jobs to the shop as a whole is
consistent with this environment, this may not be so for a shop that consists of

independent elements. Consistent with the concepts of focus and plants within plants

 

189

(Skinner, 1974), the infrastructure of a manufacturing facility needs to be consistent with
the demands placed by its individual components. In traditional CM, this appears not to
be the case. One way to resolve this issue is to release work to cells in a traditional
cellular layout based on their individual workload or utilization, not on the behavior of

the shop as a whole.

BIBLIOGRAPHY

BIBLIOGRAPHY

    

Abin, M.F., Mohamed, N. S., 1987, Simulation Support Methodology for Group
Scheduling in FMS, a _ - . - . . .
Research

Alder, P., 1985, Managing Flexibility. A Selective Review of the Challenges of
Managing the New Production Technologies Potential for Flexibility, 8mm
01°21 .-_ -on a -1010 o 01- .. n 1' ,| v nu rfo c iv °

.
a 3.0.1'

AAAA

Ang, C. L. ,Willey, P. C. T., 1984, A Comparative Study of the Performance of Pure and
Hybrid Group Technology Manufacturing Systems Using Computer Simulation,

MW 22 (2)

Askin, R. G, Cresswell, S. H. ,Goldberg, J. B. Vakharia, A. J., 1991, A Hamiltonian
Path Approach to Reordering the Part Machine Matrix for Cellular Manufacturing,

MW 29 (6)
Baker, K.R., 1974, Introduction to Sequencing and Scheduling, WM

Baker, K.R., 1984, Sequencing Rules and Due Date Assignments in a Job Shap,
Managememficience. 30 (9)

Balasubramaniam, K. N., Panneerselvam, R, 1993, Covering Technique Based
Algorithm for Machine Grouping to Form Manufacturing Cells, WW

- W 31 (6)

Ballakur, A., Steudel, H. J. 1987, A Within Cell Utilization Based Heuristic for
Designing Cellular Manufacturing Systems MW
25 (5)

Bobrowski, P.M., Mabert, V.A., 1988, Alternative Routing Strategies in Batch
Manufacturing : An Evaluation, Wm, 19, (4)

Boctor, F. F. ,,1991 A Linear Formulation of the Machine-Part Cell Formation Problems,
MW 29 (2)

190

191

Boe, W. J. Cheng, C. H, 1991, A Close Neighbor Algorithm for Designing Cellular
Manufacturing Systems MW 29 (10)

Bott, K. N., Ritzman, L. P, 1983, Irregular Workloads withMRPSystems: Some Causes
and Consequences MW 3 (4)

Browne, J., Dubois, D., Rathmill, K., Sethi, S., Stecke, K., 1984, Classification of
Flexible Manufacturing Systems, W, 2 (2)

Buffa, E. S, 1984, Meeting the Competitive Challenge: Manufacturing Strategy for U. 8.
Companies W

Buffa, E.S., Armour, G.C., Vollman, TE, 1964, Allocating Facilities with CRAFT,
Wu. 42(2)

Burbidge, J ., 1975, The Introduction of Group Technology, Heinemm

Burbidge, J ., 1977, A Manual Method of Production Flow Analysis, Won
Engineer, 56 (10)

Burgess, A. G, 1988, Evaluation of Group Technology/Cellular Manufacturing: A
Computer Simulation Wm

Buzacott, J. A, 1,982 The Fundamental Principles of Flexibility in Manufacturing
Systems,ao.ev- ' t -1‘ n 1401101111 or ‘.0 .-. 1"!

Sim

Carrie, A. 19,73 Numerical Taxonomy Applied to Group Technology and Plant layout,
MW 11 (4)

Chakravarty, A. K., Shtub, A, 1984, An Integrated layout for Group Technology with
In-Process Costs MW 22 (3)

Chan, H, Milner, D, 1982, Direct Clustering Algorithm for Group Formation in
Cellular Manufacturing, MW 1 (1)

Chandrasekharan, M. P, Rajagopalan, R., 1986a, An Ideal Seed Non-Hierarchical

Clustering Algorithm for Cellular Manufacturing, WWII
Search 24(2)

Chandrasekharan, M. P, Rajagopalan, R., 1986b, ZODIAC- An Algorithm for
Concurrent Formation of Part Families and Machine Cells, W

W 25 (6)

 

192

Chevalier, P.W., 1986, Group Technology as a CAD/CAM Integrator in Batch
Manufacturing, ; _ f ' _. . g - £.- -

 

Choobineh, F., 1988, A Framework for the Design of Cellular Manufacturing Systems,
Wm, 26 (7)

Christy, D. P., Nandkeolyar, U, 1986, A Simulation Investigation of the Design of
Group Technology Cells WW

Chu, C. H., 1988, Cluster Analysis in Manufacturing Cell Formation, QMEGA,
WW 17(3)

Chu, C. H., 1993, Manufacturing Cell Formation by Competitive Learning, International
MW 31 (4)

Chu, C. H, Tsai, M., 1990, A Comparison of Three Array Based Clustering Techniques
for Manufacturing Cell Formation MW 28 (8)

Chu, C. H., Hayya, J. C., 1991, A Fuzzy Clustering Approach to Manufacturing Cell
Formation WW 29(7)

Conway, R.W., Maxwell, W.L., Miller, L.W., 1967, Theory of Scheduling, AngLisQn
My

Cook, T.D., Campbell, D.T., 1979, Quasi-Experimentation: Design & Analysis Issues
for Field Settings, finnganLfilin

Craven, F. W., 1977, Some Constraints, Fallacies and Solutions in GT Applications,
'10..‘v.‘v.t_1 c. .1‘ .4- 1‘11.-. a 11 0.1 D.‘ '1 1- "..‘-:=.- 1 1111.

Crookall, J.R., Lee, LC, 1977, Computer Aided Performance Analysis and Design of
Cellular Manufacturing Systems, CW. 23(11)

Cummings, G. F, 1980, Simulation Model to Compare Group Technology and
Functional layout, '1 ' . . _ ,

    

U. .51. l l ‘

Dahel, N. E., Smith, S. B, 1993, Designing Flexibility into Cellular Manufacturing
Systems, WW 31(4)

de Beer, C., van Gerwen, R., de Witte, J., 1976, Analysis of Engineering Production
Systems as a Base for Product Oriented Rwonstruction, W, 25

de Beer, C., de Witte, J., 1978, Production Flow Synthesis, W, 27

 

193

De Meyer, A., Nakane, J, Miller, J. G, Ferdows, ,1989, Flexibility: The Next

Competitive Battle, The Manufacturing Futures Survey, WM,
10 (2)

de Witte, J., 1980, The Use of Similarity Coefficients in Production Flow Analysis,
WW. 18(4)

Dunn, O..,J 1961, Multiple Comparisons Among Means, WW
Wu. 56

El- -Essawy, I., Torrance, J, 1972, Component Flow Analysis. An Effective Approach
to Production Systems Design, Wm, 51

Filliben, J .J ., 1975, The Probability Plot Correlation Coefficient Test for Normality,
Mummies. 17(1)

Flynn, B. B., 1984, Group Technology Versus Process layout: A Comparison Using
Computerized lob Shoo Simulation HEW

Flynn, B. B, 1987, Repetitive Lots: The Use of a Sequence Dependent Setup Time
Scheduling Procedure 1n Group Technology & Traditional Shops, Winn:

Wfluafl)

Flynn, B. B. J,acobs, F. R., 1986, A Simulation Comparison of Group Technology With

Traditional Job Shop Manufacturing. W. 24
(5)

Flynn, B.B., Jacobs, F.R., 1987, An Experimental Comparison of Cellular (Group
Technology) layout With Process layout, W, 18 (4)

Foo, F. C, Wager, J. G, 1983, Set-up Times in Cyclic and Acyclic Group Technology
Scheduling System. W. 21 (l)

Garza, O., 1990, The Effect of Intercell Flow on the Performance of Cellular
Manufacturing Systems, WWW

Garza, O., Smunt, T. L., 1991, Countering the Negative Impact of Intercell Flow in
Cellular Manufacturing. WW. 10(1)

Gavett, J.W., 1965, Three Heuristic Rules for Sequencing Jobs to a Single Production
Facility. MW. 11 (8)

Gerwin, D, 1983, A Framework for Analyzing the Flexibility of Manufacturing
Processes, 10.001 1'. 1'51111 111.101 1' 40 ...01111 -...‘-.

194

Gerwin, D., 1987, An Agenda for Research on the Flexibility of Manufacturing

 

Glasserman, P., Yao, D.D., 1992, Some Guidelines and Guarantees for Common

Random Numbers. MauagcmcuLScicucc. 38 (6)

Goodman, SH, 1972, The Effects of Machine and Worker Assignments on the
Performance of a Dual Constrained Job Shop Possessing Flexibility in Both Resources,

2122' 'E liS Ili'

Greene, T. J., Sadowski, R. P., 1984, A Review of Cellular Manufacturing Assumptions,
Advantages & Design TechniqueS.lournal_cLQrzcratrcus_Mauegcmcu1.4 (2)

Gupta, R.M., Tompkins, J .A., 1982, An Examination of the Dynamic Behavior of Part
Families in Group Technology. JutcmaticuilcumalcflHrcducticEResearch. 20(1)

Ham, 1., Hitomi, K., Nakamura, N, Yoshida, T., 1979, Optimal Group Scheduling and
Machining Speed Decision Under Due Date Constraints, W
mmLMechauicaLEugiueers. 101 (5)

Hall, R.W., 1983, Zero Inventories, I293! Innes-Irwin

Harhalakis, G., Nagi, R. ,,Proth J. M., 1990, An Efficient Heuristic in Manufacturing

Cell Formation for Group Technology Applications, WW
Research. 28 (1)

Harrigan, K.R., 1985, Strategic Flexibility, Lexingtgnjmks

Haynes, R.D. , Komar, C.A. , Byrd, J. , 1973, The Effectiveness of Three Heuristic Rules
for Job Sequencing in a Single Production Facility, W, 19 (5)

Hill, T., 1989, Manufacturing Strategy, Text and Cases, 1min

Hitomi, K, Ham, 1., 1977, Group Scheduling Technique for Multiproduct, Multistage

Manufacturing Systems,
99 (3)

 

Hitomi, K., Ham, 1., 1978, Machine Loading and Product Mix Analysis for Group
Technology, .. ' , . .. - , - 100(3)

 

Hitomi, K, Nakamura, H., Yoshida, T., Okuda, K., 1977, An Experimental

Investigation of Group Production Scheduling, WWW
W11

195

Hollier, R.H. , 1968, A Simulation Study of Sequencing in Batch Production, W
ResechQuerterly. 19 (4)

Hyer, N. L, 1982, Group Technology: Diffusion, Development and Application in the
U S and International Context. HERDisscuaticEJudiauaJluixersinr

Irani, S. A, Cavalier, T. M., Cohen, P. H., 1993, Virtual Manufacturing Cells. Exploiting
layout Design and Intercell Flows for the Machine Sharing Problem, Internatinnal

lcuruachHrcducticuRcscarch. 31 (4)

Jacobs, F. R. B,ragg, D., 1988, Repetitive Lots: Job Flow Considerations 1n Production
Sequencing & Batch Sizing. W. 19 (2)

Kaparthi, S, Suresh, N. C., 1992, Machine-Component Cell Formation in Group

Technology A Neural Network Approach. lulcmetmmumalctHrmycricuResearch.
30 (6)

Karmarkar, U.S., Kekre, S., Kekre, S., 1985a, Lotsizing in Multi-Item Multi-Machine
Job Shout». IlEIrausacticus. 17(3)

Karmarkar, U.S., Kekre, S., Kekre, S., Freeman, S., 1985b, Lot Sizing and Lead Time
Performance in a Manufacturing Cell, Intenfaces, 15 (2)

Kelly, F. H., Mosier, C. T, Taube, L. R, 1986, Impact of Cost Structures on Group
Technology Implementation. W

 

Kern, G. M. ,Wei, J. C, 1991, The Cost of Eliminating Exceptional Elements 1n Group
Technology Cell Formation. W. 29 (8)

Khatour, S, Moodie, C., 1979, A Machine Loading Procedure Which Considers Part
Complexity and Machine Capability. W. 17
(1)

King, J., 1979, Machine Component Group Formation in Group Technology, 2131281115.;
5 l I . l E E E l . B I

King, J.,1980, Machine Component Group Formation 1n Production Flow Analysis: An

Approach Using a Rank Order Clustering Algorithm, WW
Research. 18 (2)

King, J., Nakomchai, ,,1982 Machine Component Group Formation for Group
Technology: Review andV Extension. WW. h20 (2)

196

Kirk, R.E. , 1972, Experimental Design: Procedures for the Behavioral Sciences,
Eadsmcrth

Kleijnen, J.P.C., 1987, Statistical Tools for Simulation Practitioners, W
Inc.

Kleijnen, J. P. C, van der Ven, R., Sanders,B ,l,982 Testing the Independence of
Simulation Subruns: A Note on the Power of the Von Neumann test, W

cchcreticuachsearch. 9(1)

Krajewski, L.J., King, B.B., Ritzman, L.P., Wong, D.S., 1987, Kanban, MRP, and
Shaping the Manufacturing Environment, Wm, 33 (1)

Kumar, K. R., Vannelli, A, 1987, Strategic Subcontracting for Efficient Disaggregated
Manufacturing. EtcruaticriaucurualflrcducficERcsearch. 25 (12)

Kusiak, A, 1987, The Generalized Group Technology Concept, WW
Hrcductrcujcsmrch. 25 (4)

Kusiak, A, 1988, EXGT-S: A Knowledge Based System for Group Technology,
W. 26 (5)

Kusiak, A., Cho, M. 1,992 Similarity Coefficient Algorithms for the Group Technology
Problem. MW. 30 (1 1)

Law, A.M., Kelton, W.D., 1982, Simulation Modeling and Analysis, W

Lee, L. C, 1985, A Study of System Characteristics in a Manufacturing Cell,
IntcmaticuaLlcumachHMucticrLResearcE 23(6)

‘ Leonard, R, Rathmill, K., 1977, The Group Technology Myths, W,
January

Lewis, F. A., 1973, Some Factors Affecting the Design of Production Systems 1n Batch
Manufacture, ' - - . .. . ° .

 

Logendran, R., 1990, Workload Based Model for Minimizing Total Intercell and

Intracell Moves 1n Cellular Manufacturing. hitcmaticuallcumalcthcducticukcsearch.
28 (5)

Logendran, R, 1991, Impact of Sequence of Operations and layout of Cells 1n Cellular
Manufacturing. mtcmaticuallcurualctHMucu’cuRcmh. 29 (2)

197

Logendran, R., 1992, A Model for Duplicating Bottleneck Machines 1n the Presence of
Budgetary Limitations in Cellular Manufacturing, WWW
Research. 30(3)

Mahmoodi, F. Dooley, K. J, Starr, P. J. 1988, A Comparison of Dynamic Scheduling
Heuristics for Cellular Manufacturing. WWW

Mahmoodi, F., Dooley, K. J., Starr, P. J., 1990, An Investigation of Dynamic Group
Scheduling HeuristiCS. W. 28(9)

Mahmoodi, F., Dooley, K.J., 1991, A Comparison of Exhaustive and Non-Exhaustive

Group Scheduling Heuristics in a Manufacturing Cell, MW
Research. 29 (9)

Mahmoodi, F., Tierney, L.J., Mosier, OT, 1992, Dynamic Group Scheduling
Heuristics in a Flow Through Cell Environment, W, 23 ( 1)

Malhotra, M.K., Ritzman, L.P., 1990, Resource Flexibility Issues in Multistage
Manufacturing. W. 21 (4)

Malik, M.Y. , Dale, B.G., 1977 , Cell Formation for a Group Technology Manufacturing
System. Machinery. 131

Mandelbaum, M. 1,978 Flexibility 1n Decision Making: An Exploration and Unification,
NEW

Manivannan, 8., Watson, E.F., Ham, 1., 1987, A Heuristic Rule to Minimize the Mean
Flow Time as a Secondary Measure to Makespan in a Group Technology Cell,

as..e-‘v-t_1' 0 .17 1°1h 1.‘ 1-01. 11fr--1-- 01710.1. 101 l‘r-_- 1

McAuley, J ., 1972, Machine Grouping for Efficient Production, W
Engineer. 51 (2)

McCormick, W., Schweitzer, P., White, T., 1972, Problem Decomposition and Data
Reorganization by a Clustering Technique, WM, 20 (5)

McLean, C.R., Bloom, H.M., Hopp, TH, 1982, The Virtual Manufacturing Cell,

31.-agar. ’ o .1‘ 1 1-1 1 ‘F' 0171‘ v-‘or 10111-01 0110 110.711 1

Mihram, G. A, 1974, Blocking 1n Simular Experimental Designs, W
W. 3

198

Morris, J. S., 1988, Forming Cellular layouts in Group Technology. A Simulation
Evaluation of Alternative Approaches. HhflissertaticEJluiyersithLleahcma

Morris, J .S., Tersine, R.J., 1989, A Comparison of Cell Loading Practices in Group
Technology. tournachMauutamriucflccratianMauaccmem. 2

Morris, J .S., Tersine, R.J., 1990, A Simulation Analysis of Factors Influencing the
Attractiveness of Group Technology Cellular layouts, ManagemenLScience, 36 (12)

Mosier, C.T. , 1983, The Job Shop Using Group Technology: A Simulation Investigation,
2111]? . Hiv' [111C 1

Mosier, C. T., Elvers, D. A. ,DKelly, 1,984 Analysis of Group Technology Scheduling
HeufisuchhternahLaLlcllrualeLHrcducucERcmrch. 22(5)

Mosier, C. T. ,T,aube L., 1985, The Facets of Group Technology and Their Impacts on
Implementation: A State of the Art Survey, WW
Managcrheucficjcuce. 13

Nagi, R. Harhalakis, G., Proth, J. M, 1990, Multiple Routings and Capacity

Considerations in Group Technology Applications, WW
Research. 28 ( 12)

Neter, J., Wasserman, W., Kutner, M.H., 1990, Applied Linear Statistical Models,
1min

Okogbaa, O. G, Chen, M. T., Changchit, C., Shell, R. L, 1992, Manufacturing System
Cell Formation and Evaluation Using a New Inter-Cell Flow Reduction Heuristic,

lutcmaticualloumachHrsmcticEResearch. 30 (5)

Oliva-Lopez, E., Purcheck, G. F. K, 1979, Load Balancing for Group Technology

Planning and Control. EmmmmLMachiuelchDcuguaniResearch. 19
(4)

Ozden, M., Egbelu, P. J., Iyer, A. V., 1985, Job Scheduling in a Group Technology
Environment for a Single Facility. WW. 9 (1)

Pullen, RD, 1976, A Survey of Cellular Manufacturing Cells, IhLELanntjnn
Engineer. 55(9)

Purcheck, G.F.K., 1974, Combinatorial Grouping - A lattice Theoretic Method for the
Design of Manufacturing Systems, MW, 4 (3)

199

Purcheck, G.F.K., 1975a, A Mathematical Classification as a Basis for the Design of
Group Technology Ce118. W. 54 (1)

Purcheck, G.F.K., 1975b, Linear Programming Method for the Combinatorial Grouping
of an Incomplete Power Set, W, 5 (4)

Rajagopalan, R., Batra, J., 1975, Design of Cellular Production Systems. A Graph
Theoretic Approach. WW. 13(6)

Rajamani, D., Singh, N., Aneja, Y. P., 1990, Integrated Design of Cellular
Manufacturing Systems 1n the Presence of Alternate Process Plans, Internatinnaflnnmal
cLHrqihch'cERcsearch. 28(8)

Rajamani, D, Singh, N., Aneja, Y. P., 1992, A Model for Cell Formation in

Manufacturing Systems with Sequence Dependence, WW
Research. 30(6)

Roth, A.V., Miller, LG, 1990, Manufacturing Strategy, Manufacturing Strength,

Managerial Success, and Economic Outcomes, in ‘Mnnnfacmningmy’, (Editors:
J.E. Ettlie, M.C. Burstein, A. Fiegenbaum)

Ruben, R. A. ,Mosier, C. T., Mahmoodi, F, 1993, A Comprehensive Analysis of Group

Scheduling Heuristics in a Job Shop Cell, MW,
31 (6)

Russell, G. R, Philipoom, P. R, 1992, Sequencing Rules and Due Date Setting
Procedures in a Flow-line Cell Shop with Family Setups, W

Management. 10 (2)

Russo, F .J .A., 1965, A Heuristic Approach to Alternate Routing in a Job Shop, M59.
II . 1 1 I I . f I I I
Sassani, F., 1990, A Simulation Study on Performance Improvement of Group

Technology CeIIS.Iu1crnaticuaLlcurnach_Hrcducticu_Research. 28(2)

Sawicki, J .D., 1973, The Problems of Tardiness and Saturation in a Multi-Class Queue
with Sequence Dependent Setups, MW 5 (4)

Schmeiser, B. W., 1982, Batch Size Effects in the Analysis of Simulation Output,
QpcraticueRcsmrch. 30(3)

Schonberger, R.J., 1982, Japanese Manufacturing Techniques: Nine Lessons in
Simplicity. IheEreeHress

200

Schruben, L., Singh, H, Tierney,L ,1983, Optimal Tests for Initialization Bias in
Simulation Output. W. 31 (6)

Seifoddini, H., 1987, Incorporation of Production Volume 1n Machine Cells Formation

in Group Technology Applications Hrccflhgscflheflthlmcruaucuaulcufcrcueecu
HuxlucticERcsearch

Seifoddini, H., 1989, Duplication Process in Machine Cells Formation in Group
Technology. llEIrausacticns. 21 (4)

Seifoddini, H., Wolfe, P.M., 1986, Application of the Similarity Coefficient Method in
Group Technology. mm 18(3)

Shafer, S.M., 1988, A Comparison of Manufacturing Cell Formation Techniques, 1111.2
WW

Shafer, S. M., Meredith, J. R., 1990, A Comparison of Selected Manufacturing Cell
Formation TechniqueS. WW. 28(4)

Shafer, S. M, Kern, G. M., Wei, J. C, 1992, A Mathematical Programming Approach
for Dealing With Exceptional Elements 1n Cellular Manufacturing, Internntjnnalflnnnnl
cLHrcducuculesearch. 30(5)

Shafer, S. M. Rogers, D. F., 1993, Similarity and Distance Measure for Cellular
Manufacturing Part1 A Survey. InternamallcurnachHrcduchcuRcsarch. 31 (5)

Shtub, A, 1989, Modelling Group Technology as a Generalized Assignment Problem,
lutcruahchaLlcuruaLcLHrcducticuRescarch. 27(5)

Simpson, J. A, Hocken, R. J., Albus, J. S, 1982, The Automated Manufacturing
Research Facility of the National Bureau of Standards, LQnmaJJLManufacmnng

Systems. 1(1)

Sinha, R. K, Hollier, R. H., 1984, A Review of Production Control Problems 1n Cellular
Manufacture. IntcruaucuallcurualeLHrocucucuRcscarch. 22 (5)

Skinner, W.H., 1974, The Focused Factory, W, May-June

Slack, N, 1983, Flexibility as a Manufacturing Objective, W
W. 3(3)

Slack, N., 1987, The Flexibility of Manufacturing Systems, W
W. 7(4)

201

Slack, N ., 1990, Flexibility as Managers See It, in ‘mnnmmmm
Management’, John Wiley & Sons

Sokal, R.R., Sneath, P.H.A., 1973, Principles of Numerical Taxonomy, W

Song, S., Hitomi, K., 1992, GT Cell Formation for Minimizing the Intercell Parts Flow,
IutcruaticuaLlcumachHrcducticERemrch. 30 ( 12)

Sule, D. R., 1991, Machine Capacity Planning in Group Technology, International
WW. 29(9)

Sundaram, R. M., 1983, Some Scheduling Algorithms for Group Technology

Manufacturing System. in ‘CcmcutcLAcchcaucuuLHrcduchLauLEugrmuhg’.
(Editor E. A. Warman)

Suresh, N.C., 1991, Partitioning Work Centers for Group Technology: Insights From
an Analytical Model. DccisicEScichces. 22 (4)

Suresh, N.C., 1992, Partitioning Work Centers for Group Technology: Analytical
Extension and Shop Level Simulation Investigation, W, 23 (2)

Suresh, N. C., Meredith, J. R, 1985, Achieving Factory Automation Through Group
Technology Principles. lcuruaicLQucraticusMauagemeht. 5(2)

Swamidass, P.M. , 1988, Manufacturing Flexibility, WW
MmrcgrapElicJ

Swamidass, P.M., Newell, W.T., 1987, Manufacturing Strategy, Environmental
Uncertainty and Performance: A Path Analytic Model, Wm, 33 (4)

Tersine, R.J., 1985, Production/Operations Management: Concepts, Structures and
AnalysiS. Nerthficllahc

Tilak, A.O., 1978, Job Shop Scheduling with Routing and Resource Flexibility, £11.12
11° . I I I . I II . .

Tilsley, R., Lewis, F., 1977, Flexible Cell Production Systems - A Realistic Approach,
CIRHAnnalcu. 26 (l)

Vaithianathan, R., McRoberts, K. L., 1982, On Scheduling 1n aGT Environment, Journal
cLMauufachrriuLSmcms. 1

Vakharia, A. J, 1986, Methods of Cell Formation 1n Group Technology: A Framework
for Evaluation. leurnachQpcraticucMauagcmeut. 6(3)

 

202

Vakharia, A.J., Wemmerlov, U., 1990, Designing a Cellular Manufacturing System: A
Materials Flow Approach Based on Operation Sequences, W 22 (1)

Vannelli, A. ,Kumar, K. R. 1986, A Method for Finding Minimal Bottleneck Cells for
Grouping Part-Machine Families ImemauchaLlcumachHMucticnjcsearch. 24(2)

Vohra, T. ,Chen, D. S, Chang, J. C., Chen, H. C., 1990, A Network Approach to Cell

Formation in Cellular Manufacturing. luteruaticualicumalefHuxlhcticuResearch. 28
(l 1)

Voss, C.A. , 1986, Managing New Manufacturing Technologies, Wagering);
I i . 1 1 I 11 I

Waghodekar, P. H. Sahu, S., 1984, Machine Component Cell Formation in Group
Technology MACE. WW. 22(6)

Wayson, RD, 1965, Effects of Alternate Machines on Two Priority Dispatching
Disciplines in the General Job Shop, Master’s Thesis. Qnmell University

Weeks, J.K., Fryer, J.S., 1977, A Methodology for Assigning Minimum Cost Due
DateS. Wm. 23 (8)

Wei, J.C., Gaither, N. , 1990, An Optimal Model for Cell Formation Decisions, Minn
Sciences. 21 (2)

Wemmerlov, U., 1992, Fundamental Insights Into Part Family Scheduling: The Single
Machine Case. W. 23 (3)

Wemmerlov, U. H,yer, N. L, 1986 Procedures for the Part/Machine Group Identification
Problem in Cellular Manufacturing, MW, 6 (2)

Wemmerlov, U., Hyer, N. L. 1987, Research Issues in Cellular Manufacturing,
lhtcruauchaLlcurhalchucucERcsearch. 25(3)

Wemmerlov, U, Hyer, N. L., 1989, Cellular Manufacturing in the U. S. Industry: A
Survey of UserS. Internaticual_lc_uruach_Hrcducticu_Research. 27(9)

Wemmerlov, U., Vakharia, A.J., 1992, Job and Family Scheduling of a Flow Line
Manufacturing Cell: A Simulation Study, Wm 23 (4)

Wheelwright, SC, 1984, Manufacturing Strategy: Defining the Missing Link, m;
Mauagcmcuclcurnal. 5

 

203

White, C.H., Wilson, RC, 1977, Sequence Dependent Setup Times and Job
Sequencing. MW. 15 (2)

Wilbrecht, J.K., Prescott, W.B., 1969, The Influence of Setup Time on Job Shop
Performance, ManagemehLSciehec. 16(4)

Wu, H. L. ,Venugopal, R. Baras,h, M. M, 1989, Design of a Cellular Manufacturing

System: A Syntactic Pattern Recognition Approach, W,
5

Wu, N., Salvendy, G., 1993, A Modified Approach for the Design of Cellular
Manufacturing SystemS. IhtcmahealfiuruachflcducthResearch. 31 (6)

Xu, H. ,Wang, H. P, 1989, Part Family Formation for GT Applications Based on Fuzzy
Mathematics. MW. 27 (9)

Zelenovic, D. M. 1982, Flexibility: A Condition for Effective Production Systems,
lurematicuaucurhaLQLHMucuceRcsearch. 20(3)

"iiiiliiiiiiiiiiii