Sunteți pe pagina 1din 6

Manufacturing Cell Formation - Dr.

Chao-Hsien Chu

Page 1 of 6

A comparison of three array-based clustering techniques for manufacturing cell formation


Chao-Hsien Chu and Mayshing Tsai
(Condensed from: Int. J. of Prod. Res., 1990, Vol. 28, No. 8, 1417-1433)

Contents

Abstract
Introduction
Review of related literature
Computational algorithms
Research methodology
Computational results
Conclusions
References

Back to M418 Page

Abstract
This paper examines three array-based clustering algorithms - rank order clustering (ROC), direct clustering analysis (DCA), and bond energy analysis (BEA) - for
manufacturing cell formation. According to our test, bond energy analysis outperforms the other two methods, regardless of which measure or data set is used. If
exceptional elements exist in the data set, the BEA algorithm also produces better results than the other two methods without any additional processing. The BEA can
compete with other more complicated methods that have appeared in the literature.
Back to Top of the Page

Introduction
Cellular manufacturing has been recognized as one of the most recent technological innovations in job-shop or batch-type production to gain economic advantages similar
to those of mass production. The design of a cellular manufacturing system must take many structural and operational issues into consideration (Wemmerlov and Hyer
1987). One of the first and most important problems faced in practice is to select and group parts or machines with similar features into families; this process is called cell
formation. During the past decade, there has been a tremendous interest in this type of problem from practitioners as well as from academics. Numerous heuristic or
analytical methods have been developed and applied, but very little effort has been put into selecting the best or most appropriate method. Since different clustering
approaches may produce conflicting results when different performance criteria are measured, there arises a major concern with the issue of which measure and clustering
technique a company should use.
This paper proposes a procedure for evaluating three alternative array-based clustering techniques within the manufacturing domain. A considerable number of data sets
from literature are tested, from which some managerial guidelines for the selection of these methods are developed. Specifically, the study is intended to provide answers to
the following questions: (1) How should clustering algorithms for cell formation be evaluated? (2) What kind of measures can be used for the comparison? (3) Which arraybased clustering technique performs better? (4) Which array-based method has the best ability to deal with exceptional elements?
Back to Top of the Page

Review of related literature


Cell formation involves the process of analyzing parts or machine populations, grouping those parts with similar design features or manufacturing routings into families,
and forming the corresponding machines into cells. There has been extensive work done in this area, and numerous clustering algorithms have been applied or developed in
practice. Table 1 summarizes some of the clustering algorithms which appeared in the literature. A comprehensive review of the problem can be found in (Chu 1989, King
and Nakornchai 1982, Waghodekar and Sahu 1984, Wemmerlov and Hyer 1986).
One of the simplest approaches, called the ocular or eye-balling method, is to examine the data and perform the classification using the human eye. Though this method is
perhaps the easiest to perform, its success is highly dependent on human experience, knowledge, and preference. Because of human limitations, the number of parts that can
be handled is limited as well. In the case of more than a hundred parts, for example, it would be impossible to tell which part goes with which family via this method.
A classification and coding system improves on these deficiencies. It scrutinizes the design features of parts from product codes. Those parts with similar codes are, then
formed into the same family. Much time and effort goes into coding parts and creating an elaborate data base; nevertheless, this technique produces a weak connection
between component features and machine groupings.
Production flow analyses (Burbridge 1971), on the other hand, group parts into families according to routing sheets or process plans. By virtue of the routing information,
this method is quick and sufficiently accurate for a company to rearrange the shop floor into independent manufacturing cells. Both manual and computerized methods have
been well developed and are widely used in the industry.
Recently, the use of agglomerative clustering methods, such as the single linkage and complete linkage methods, etc., in cell formation has been examined (Carrie 1973,
McAuley 1972, Mosier 1989, Tarsuslugil and Bloor 1979). This approach involves the process of calculating the similarity coefficients between pairs of machines/parts.
Those machine/parts with closer similarity coefficients are arranged into the same group following a specific algorithm. Several problems, such as the effect of similarity
coefficients on performance, and how to decide the appropriate number of part families etc., yet remain to be solved (Anderberg 1973, Chu 1989).
Another simple approach, called array-based clustering, has also been the subject of extensive study recently (Chan and Milner 1982, Gongaware and Ham 1981, Jacobs
1985, Khator and Irani 1987, King 1980a,b, King and Nakornchai 1982). One common feature of this approach is that all the methods sequentially rearrange the columns
and rows of the machine/part matrix according to an index, until some cross diagonal or irregular blocks are generated. This approach is easier to understand, but the results
vary from algorithm to algorithm.
============================================================================
Method
Criteria
References
---------------------------------------------------------------------------(1) Array-based approach:
(a) Rank order clustering
Binary value
[13, 14]
(b) Direct clustering analysis [5, 29]

http://net1.ist.psu.edu/chu/course/cm.htm

6/29/2014

Manufacturing Cell Formation - Dr. Chao-Hsien Chu

(c)
(d)
(e)
(f)
(g)

Bond energy analysis


ROC2
MODROC
Jacob's
Occupancy value

Bond energy
Binary word
Binary value
Operation time
Occupancy value

Page 2 of 6

[10, 20]
[15]
[7]
[11]
[12]

(2) Agglomerative clustering:


(a) Single linkage
Similarity coefficient [4, 19, 21, 27]
(b) Average linkage
Similarity coefficient [27]
(c) Complete linkage
Similarity coefficient [21, 27]
(d) Centroid
Similarity coefficient [21, 27]
(e) Median
Similarity coefficient [27]
(f) McQuitty's
Similarity coefficient [27]
(g) Ward's
Similarity coefficient [21, 27]
(h) Lance and Williams'
Similarity coefficient [27]
(3) Non-agglomerative clustering:
(a) C&R
[6]
(b) A divisive procedure
[26]
(4) Graphic theoretical methods:
(a) Purcheck
[22]
(b) Mathematical classification [23]
(c) R&B
Similarity coefficient [24]
(5) Mathematical programming:
(a) Multiobjective programming Similarity coefficient [10]
(b) 0-1 Integer programming
[17]
(6) Heuristic and other methods:
(a) WUBC
[3]
(b) Cluster identification
[18]
(c) Cost analysis
[18]
(d) MACE
Similarity coefficient [28]
============================================================================

Table 1. Summary of clustering algorithms for cell formation.


There has also been some interest in using more complicated mathematical approaches (Gongaware and Ham 1981, Kusiak 1987, Purcheck 1974, 1975, Rajagopalan and
Batra 1975) and other heuristics (Ballakur and Steudel 1987, Kusiak and Chow 1987), to increase the efficiency and effectiveness of the clustering. For example, a settheoretic approach was developed by Purcheck (1974). However, as the set is amplified, the diagram grows exponentially; consequently, its usefulness is purely illustrative
Back to Top of the Page

Computational algorithms
While the literature yields about seven array-based clustering algorithms, only three of them - rank order clustering (ROC), direct clustering analysis (DCA), and bond
energy analysis (BEA) - were selected for the comparison. ROC2 and MODROC were mainly extended or modified from ROC and the Jacob's algorithm is far less
efficient according to our preliminary test.
Rank order clustering (ROC)
ROC is a well-known clustering technique that attempts to create a block diagonal form by repeatedly reallocating the columns and rows of a machine/part matrix
according to binary values (King 1980 a, b). Although ROC is easy to apply, it has several disadvantages. First, the quality of the results is strongly dependent on the initial
disposition of the machine/part matrix. Second, the binary value (a power of 2) that is used for the reallocation restricts the size of the problem that the technique can
handle. This results in a storage problem. A revised version, called ROC2, has been developed by King and Nakornchai (1982) to overcome this limitation and to increase
computational efficiency.
Direct clustering analysis (DCA)
DCA rearranges the rows with the left-most positive cells to the top and the columns with the top-most positive cells to the left of the matrix (Chan and Milner 1982). After
several iterations, all the positive cells will form diagonal blocks from the top left corner to the bottom right corner. Our study follows the algorithm with Wemmerlov's
correction (1984). This method allows more flexibility in the size of the problem. Furthermore, the sensitivity of the ROC algorithm to the initial matrix is eradicated
because DCA initiates the procedure by counting the number of positive cells instead of depending on intuition.
Bond energy analysis (BEA)
The bond energy analysis algorithm (McCormick et al. 1972) attempts to identify and exhibit the interrelations within each cell and the associations among the clustered
groups by means of total bond energy. A bond is claimed to exist between each pair of the neighboring rows and columns if they have positive cells in the machine/part
matrix. BEA begins with an arbitrarily selected column (or row). It then places that column with the greatest contribution to the total bond energy beside the assigned
column (row). It repeats the same procedure for all the columns and rows. The method is applicable to problems of any size because the BEA has nothing to do with
calculating the binary values. However, since the first step of the algorithm is determined by intuition, many possible solutions can be generated; that is, the solution
depends on the initial row (or column) selected for starting the process. In this study, we designed the program that allows a user to choose initial row/column and select a
better clustering result.
Back to Top of the Page

Research methodology
In this section, a general procedure for evaluating alternative clustering techniques is presented. These algorithms are coded in the BASIC programming language using
Microsoft Quick Basic compiler and run on an IBM PS/2 model 50 machine.
Evaluation procedure
One major problem with clustering arises when a large number of components need to be processed in a machine (such a machine is called a bottleneck machine) or when
exceptional elements exist in the machine/part matrix. Under these conditions, the parts or machines may not be divisible into mutually exclusive groups; in this case,
additional processing or human interfaces are needed.
The evaluation follows these steps. First of all, each problem set is clustered by each algorithm. If the results indicate that bottleneck machines or exceptional elements
exist, additional procedures such as those described in this section are then applied. Such procedures, which introduce a human interface with the algorithm, offer much
greater flexibility in dealing with exceptional elements and bottleneck machines.

http://net1.ist.psu.edu/chu/course/cm.htm

6/29/2014

Manufacturing Cell Formation - Dr. Chao-Hsien Chu

Page 3 of 6

The case of bottleneck machines


When bottleneck machines exist, components similar in nature may be dispersed over more than one cluster; therefore, it is difficult to form the diagonal block
structure during the first iteration. The following procedure (King and Nakornchai 1982) can be followed to improve performance:
1.
2.
3.
4.

Identify the bottleneck machines.


Temporarily remove the bottleneck machines from the matrix.
Reapply the algorithm to generate a diagonal structure.
Duplicate the bottleneck machines into each group.

The case of exceptional elements


Exceptional elements may disrupt the clustering process and result in a poor formation. The following procedure (King and Nakornchai 1982) can be followed to
improve performance:
1.
2.
3.
4.

Identify the exceptional elements.


Temporarily remove the exceptional elements from the matrix.
Reapply the algorithm to the matrix.
Restore the previously ignored exceptional elements.

Measures of performance
The performance of clustering can be evaluated either according to computational efficiency or according to clustering effectiveness (Cunningham and Ogilvie 1971,
Kuiper and Fisher 1975, Rand 1971). Clustering efficiency is normally measured in terms of program execution time, the amount of memory needed, and the
complexity of the algorithm (Kusiak and Chow 1987). In this study, four effective measures were selected since they have been widely used in the literature (Chu
1989).
Total bond energy
Total bond energy (BE), also called measure of effectiveness (ME), is an index used to measure the relative clumsiness of a clustered matrix. Bond energy is defined
as the product of the two nearest-neighbor elements (McCormick et al. 1972). The total bond energy is equal to:

where Yij is the related cell value of a clustered matrix, i ranges from 1 to M (number of machines), and j ranges from to 1 to N (number of parts). Theoretically, a
matrix with denser clumps of elements will have a larger total bond energy; therefore, the larger BE that an algorithm can create, the better the clustering result
obtained. However, in some cases, even if the BE is not the highest, the clustering results may still be acceptable.
Percentage of exceptional elements
The quality of a clustering can also be measured by the number of part/machine cells that remain outside the diagonal blocks (Chan and Milner 1982, King 1980).
These off-diagonal part/machine cells are called exceptional elements. Per cent of exceptional elements (PE), obtained from dividing the number of exceptional
elements by the total number of elements with '1' in the entry, is used to include the possible effect of problem size. A better clustering algorithms results in a smaller
percentage of exceptional elements (PE).
Machine utilization
Machine utilization (MU) indicates the percentage of times the machines within the clusters are used in production. MU can be computed as (Chandrasekharan and
Rajagopalan 1986):

where N1 is the total number of 1's within the groups, K is the number of clusters, m, is the number of machines in the kth cluster, and n, is the number of parts in the
kth cluster. Generally speaking, the higher the value, the better the machines are being utilized. However, in some special cases, even if the clustering method
produces a smaller percentage of exceptional elements and a higher total bond energy, the machine utilization may be lower,
Grouping efficiency
Grouping efficiency (GE) is an aggregate measure which takes both the number of exceptional elements and machine utilization into consideration. A weight, p, must
be assigned to reveal the relative importance of each term, though a value of 0-5 is commonly used. GE can be defined as (Chandrasekharan and Rajagopalan 1986):

where MN is the size of the machine/part matrix, and NE is the number of exceptional elements. As a general rule, the higher the grouping efficiency, the better the
clustering results.
Examples for illustrating the computation of these measuring criteria are given in Table 2. Assume that (a) and (b) are the clustering results from two different
methods; although both (a) and (b) classify parts into two families, the results of (a) seem better because they have a lower percentage of exceptional elements, a
higher machine utilization and grouping efficiency, and a larger total bond energy.
(a) Example Using BEA:
-------------------------------------Components
-------------------------------------6
3
5
4
2
7
1
1
0
0
0
1
1
1
0

http://net1.ist.psu.edu/chu/course/cm.htm

6/29/2014

Manufacturing Cell Formation - Dr. Chao-Hsien Chu

Page 4 of 6

3
0
0
0
1
1
1
1
5
1
1
1
1
0
0
0
4
1
1
0
0
0
0
1
2
0
1
1
0
0
0
0
-------------------------------------BE = 17;
PE = 2/16 = 12.5%;
MU = 14/17 = 82.35%;
GE = 0.5x(14/17)+O.5x[1-2/(35-17)] = 85.62%
(b) Example Using DCA:
-------------------------------------Components
-------------------------------------3
4
5
6
1
2
7
5
1
1
1
1
0
0
0
4
1
0
0
1
1
0
0
2
1
0
1
0
0
0
0
3
0
1
0
0
1
1
1
1
0
1
0
0
0
1
1
-------------------------------------BE
PE
MU
GE

=
=
=
=

13;
3/16 = 18.8%;
13/18 =72.22%;
0.5x(13/18)+O.5x[1-3/(35-18)]=77.29%.

Table 2. Computational examples for measures


The problem (data) sets
Since clustering results are highly data dependent, the selection of data sets for testing is very important. The testing data can be either generated from a random number
generator via computer or collected from the literature (Chu 1989). In order to compare the performance of the tested algorithms with other clustering algorithms published
elsewhere, we elected to use the existing data sets. Eleven sets of data (problems) from the literature have been collected for the evaluation. We limit the problem size to 50
parts or machines because of the computer's internal limitation caused by the binary value of ROC algorithm. Table 3 summarizes the special features and the sources of
these data sets. The best clustering results from other studies are also included for comparison purposes. As shown, some of the data sets, such as I and I 1, can be
completely clustered and others cannot. To ensure maximum consistency and homogeneity, these data sets were separated from the program, and accessed when needed.
Back to Top of the Page

Computational results
Table 4 summarizes the computational results, where row (a) is the results that were obtained from the first iteration of computation, and row (b) is that which was arrived
at by using additional procedures to deal with bottleneck machines and exceptional elements.
=====================================================================
Data Size BM EE NG
PE
MU
GE
BE References
--------------------------------------------------------------------1
5x7
N N 2 0.0 82.35 91.18 14 King and Nakornchai (1982)
2
5x7
N Y 2 12.5 82.35 85.62 15 Waghodekar and Sahu (1984)
3
5x7
N Y 2 12.5 82.35 85.62 16 Waghodekar and Sahu (1984)
4
9x9
N Y 3 25.0 92.59 89.07 48 Gongaware and Ham (1981)
5 12x10 N Y 3 23.7 85.29 87.41 47 McAuley (1972)
6 15x10 N N 3 0.0 92.00 96.00 63 Chan and Milner (1982)
7
8x20 N Y 3 14.8 100.00 95.83 78 Chandrasekharan (1986b)
8 14x24 N Y 4 3.3 68.60 83.90 67 King (1980 a)
9 20x35 N Y 4 1.4 79.43 89.52 200 Burbridge (1971)
10 16x43 Y Y 5 2.4 61.81 80.70 144 King (1980 b)
4 0.8 50.59 75.18 135 King and Nakornchai (1982)
11 30x50 N N 4 0.0 36.67 68.33 82 Stanfel (1985)
=====================================================================
BM:
NG:
MU:
BE:

Bottleneck machines.
Number of groups.
Machine utilization.
Total bond energy.

EE: Exceptional elements.


PE: % exceptional elements.
GE: Grouping efficiency.

Table 3: Summary of Testing Data Sets


Relative performances
In the case of well-structured data sets, such as problems 1, 6 and I 1, both BEA and ROC performed very well. They completely clustered parts into groups without
producing any exceptional elements. Though DCA also performed well on data sets I and 6, it encountered difficulties in problem I 1. Since the bond energy analysis
algorithm always produces the largest total bond energies, it appears to be the better choice than the other two methods.
As to ill-structured data sets, the BEA algorithm still outperformed the other two methods in all criteria measured. During the first iteration, almost all of the ROC and DCA
matrices came out with at least one misplaced group, whereas, the BEA algorithm not only did not create misplaced groups but also produced higher total bond energy,
machine utilization and grouping efficiency. However, when additional procedures were applied, ROC and DCA performed as well as BEA. For example, in the case of
data set 7, the first iteration results of ROC, which had the highest percentage of exceptional elements, the least total bond energy and an unacceptable number of misplaced
groups, were the worst. But it performed much better during the second iteration, which had the same results as those from the first iteration of BEA. Similarly, in the case
of data set 5, DCA provided an ambiguous clustering matrix with a misplaced group. After an additional process was applied, the DCA resulted in a smaller percentage of
exceptional elements, higher machine utilization, and larger total bond energy. This indicates that DCA and ROC are quite sensitive to data sets with exceptional elements.
Effect of exceptional elements
The effect of exceptional elements is further illustrated in Table 5. As shown, the procedure for handling exceptional elements has almost no effect on the BEA algorithm.
The percentage of exceptional elements, machine utilization, and grouping efficiency of BEA did not change even when additional procedures were used. On the contrary,
the procedure by and large improved the performance of DCA and ROC. For instance, the average percentage of exceptional elements of ROC improved by 22.47%; the
mean machine utilization, by 6.18%; the average grouping efficiency, by 4.15%; and the mean total bond energy increased by 8.51%. Such changes are quite substantial.

http://net1.ist.psu.edu/chu/course/cm.htm

6/29/2014

Manufacturing Cell Formation - Dr. Chao-Hsien Chu

Page 5 of 6

===================================================================================
Data
ROC
DCA
BEA
------------------------- ------------------------ ------------------------Sets NG
PE MU(%) GE(%) BE NG
PE MU(%) GE(%) BE NG
PE MU(%) GE(%)
BE
----------------------------------------------------------------------------------1(a) 2
0.0 82.35 91.18 14
2
0.0 82.35 91.18 14
2 0.0 82.35 91.18
16
2(a) 2 12.5 82.35 85.62 15
2* 18.8 76.47 79.41 13
2 12.5 82.35 85.62
17
(b) 2 12.5 82.35 85.62 16
2 12.5 82.35 85.62 15
2 12.5 82.35 85.62
16
3(a) 2* 18.8 72.22 77.29 12
2 18.8 72.22 77.29 12
2 12.5 82.35 85.62
17
(b) 2 12.5 82.35 85.62 16
2 12.5 82.35 85.62 15
2 12.5 82.35 85.62
17
4(a) 2 15.6 67.50 77.65 35
2 9.4 70.73 81.62
39
3* 25.0 82.76 83.69 30
3 18.8 89.66 89.06
42
(b) 2
9.4 70.73 81.62 34
2 9.4 70.73 81.62
40
(b) 3 18.8 89.66 89.06 33
3 18.8 89.66 89.06
38
5(a) 3 13.2 78.57 86.08 49
3* 18.4 81.58 86.52 47
3 13.2 82.50 88.13
52
(b) 3 13.2 78.57 86.08 49
3 13.2 82.50 88.13 48
3 13.2 82.50 88.13
50
6(a) 3
0.0 92.00 96.00 63
3
0.0 92.00 96.00 63
3 0.0 92.00 96.00
65
7(a) 2
9.8 55.00 72.50 68
2
8.2 66.67 80.04 69
3* 21.3 85.71 86.61 68
3 14.8 100.00 95.83
82
(b) 2
9.8 55.00 72.50 80
2
8.2 66.67 80.00 76
3* 14.8 100.00 95.83 82
3 14.8 100.00 95.83
79
8(a) 3
3.3 50.86 74.98 66
4* 6.6 66.28 82.34 66
4 3.3 68.60 83.90
76
(b) 3
3.3 50.86 74.98 65
4
3.3 68.60 83.90 68
4
6.6 66.28 82.34 65
4 3.3 68.60 83.90
76
9(a) 4* 1.4 76.57 88.09 183
4* 4.3 75.58 87.22 186
4 1.4 76.57 88.09 206
(b) 4
1.4 76.57 88.09 195
4
2.2 76.00 87.71 190
4 1.4 76.57 88.09 209
10(a) (b) 4
0.8 50.20 75.02 138
4 0.8 47.17 73.51 162
5
2.4 61.81 80.70 138
5
2.4 59.90 79.75 131
5 2.4 61.81 80.70 162
11
4
0.0 36.43 68.21 116
4 14.6 37.58 67.59 118
4 0.0 36.43 68.21 163
----------------------------------------------------------------------------------Avg(a)
9.9 75.52 84.51 61.6
10.2 70.28 81.79
7.7 79.28 87.16 73.6
(b)
7.2 77.34 86.39 71.8
7.4 72.61 84.14 69.9
7.2 77.69 86.58 81.0
===================================================================================
* Some of the elements are misplaced and have been corrected manually.
Note: If two different clusterings existed, the one with better GE was selected for
computing the average.

Table 4. Summary of computational results.


====================================================================================
ROC
DCA
BEA
----------------------- ------------------------ ----------------------PE(%) MU(%) GE(%)
BE
PE(%) MU(%) GE(%)
BE
PE(%) MU(%) GE(%) BE
-----------------------------------------------------------------------------------1 Iter.
14.11 77.78 84.25 60.43 12.49 70.13 80.14 61.14 10.93 83.15 88.04 70.29
2 Iter.
10.93 82.59 87.74 65.57
9.23 75.27 84.43 63.29 10.93 83.15 88.04 69.29
Deviation -3.19 4.81 3.50 5.14 -3.26 5.14 3.99 2.14
0.00 0.00 0.00 -1.00
% Changed -22.47 6.18 4.15 8.51 -26.09 7.33 4.96 3.50
0.00 0.00 0.00 -1.42
====================================================================================
Data sets 1, 6. 10 and II were excluded from the statistics because no additional
procedure needed to be applied.

Table 5. Effect of exceptional elements.


Problem with bottleneck machines
Regarding the case of bottleneck machines, such as that of data set 10, none of the three methods can form natural groups at the first iteration because bottleneck machines
hinder the grouping process. However, after applying the additional procedure, the results from the three algorithms were quite similar and replicated the same as the best
results obtained from the literature.
Limitation of DCA algorithm
The original DCA algorithm has been criticized by Wemmerlov (1984) as deficient. Worth noting is that according to the results of data set number I 1, ROC and BEA can
completely cluster the input matrix into four groups without any exceptional elements. However, DCA resulted in 28 exceptional cells, a lower grouping efficiency and
machine utilization, and much less total bond energy. This indicates that the DCA algorithm have some deficiency; that is, it may not work well for some input matrices.
Competition with other algorithms
In general, the bond energy analysis algorithm is the best of the three methods tested. Comparing the results of BEA with the best results reported in the literature shows
that BEA can also compete with other more complicated methods. As seen from Tables 3 and 4, the BEA always results in a smaller percentage of exceptional elements
and in higher total bond energies. Although in some data sets, the BEA results in lower machine utilization and grouping efficiency, the difference is not substantial.
Back to Top of the Page

Conclusions
This paper has examined three well known array-based clustering techniques - rank order clustering, direct clustering analysis and bond energy analysis - for cell formation.
It also demonstrated a procedure for evaluating alternative clustering algorithms under different measuring criteria.
Based upon the eleven data sets that we tested, the following results can be obtained: (1) No matter which measure of performance or data set is tested, BEA is the best of
the three array-based clustering techniques under evaluation. (2) If a data set is well structured, almost all three methods can completely cluster parts into part families. (3)
If exceptional elements exist in the data set, it is much more efficient and effective to use BEA because the method does not require an additional procedure in order to
arrive at better results. (4) The BEA algorithm not only performs better than ROC and DCA, it can also compete with other more complicated methods presented in the
literature, especially if a company wants to reduce the percentage of exceptional elements and to increase the clumsiness of the clustering. (5) If bottleneck machines exist,
none of the three methods can produce acceptable clusters without additional processing; therefore, how to identify the bottleneck machines accurately remains an issue
into which further research is needed.
Back to Top of the Page

http://net1.ist.psu.edu/chu/course/cm.htm

6/29/2014

Manufacturing Cell Formation - Dr. Chao-Hsien Chu

Page 6 of 6

References
1. ANDERBERG, M. R., 1973, Cluster Analysis for Applications (New York: Academic Press).
2. BURBRIDGE, J. L., 1971, Production flow analysis. The Production Engineer, 50, 139-152.
3. BALLAKUR, A., and STEUDEL, H. J., 1987, A within-cell utilization based heuristic for designing cellular manufacturing systems. International Journal of
Production Research, 25, 639-665.
4. CARRIE, A. S., 1973, Numerical taxonomy applied to group technology and plant layout. International Journal of Production Research, 11, 399-416.
5. CHAN, H. M., and MILNER, D. A., 1982, Direct clustering algorithm for group formation in cellular manufacture. Journal of Manufacturing Systems, 1, 65-75.
6. CHANDRASEKHARAN, M. P., and RAJAGOPALAN, R., 1986a, An ideal seed non-hierarchical clustering algorithm for cellular manufacturing. International
Journal of Production Research, 24, 451-464.
7. CHANDRASEKHARAN, M. P., and RAJAGOPALAN, R., 1986 b, MODROC: An extension of rank order clustering for group technology. International Journal of
Production Research, 24, 1221-1233.
8. CHU, C. H., 1989, Clustering analysis in manufacturing cellular formation. OMEGA: International Journal of Management Sciences, 17, 289-295.
9. CUNNINGHAM, K. M., and OGILVIE, J. C., 197 1, Evaluation of hierarchical grouping techniques: a preliminary study. The Computer Journal, 15, 209-213.
10. GONGAWARE, T. A., and HAM, I., 1981, Cluster analysis applications for group technology manufacturing systems. Manufacturing Engineering Transactions,
503-508.
11. JACOBS, F. R., 1985, Computerized production flow analysis. Softcover Software: 28 Microcomputer Programs for IEs and Manager, edited by G. E. Whitehouse,
(Georgia: IE and Management Press), pp. 61-67.
12. KHATOR, S. K., and IRANI, S. A., 1987, Cell formation in group technology: a new approach. Computers & Industrial Engineering, 12, 131-142.
13. KING, J. R., 1980 a, Machine-component group formation in group technology, OMEGA, 8, 193 199.
14. KING, J. R., 1980 b, Machine-component grouping in production flow analysis: an approach using a rank order clustering algorithm. International Journal of
Production Research, 18, 213-232.
15. KING,, J. R., and NAKORNCHAI, V., 1982, Machine-component group formation in group technology: review and extension. International Journal of Production
Research, 20, 117-133.
16. KUIPER, F. K., and FISHER, L., 1975, A Monte Carlo comparison of six clustering procedures.Biometrics, 777-783.
17. KUSIAK, A., 1987, The generalized group technology concept. International Journal of Production Research, 25, 561-569.
18. KUSIAK, A., and CHOW, W. S., 1987, Efficient solving of the group technology problem. Journal of Manufacturing Systems, 6, 117-124.
19. McAuley, J., 1972, Machine grouping for efficient production. The Production Engineer, 53-57.
20. McCormick, JR., W. T., SCMWITZER, P. J., and WHITE, T. W., 1972, Problem decomposition and data reorganization by a clustering technique. Operations
Research, 20, 993-1009.
21. MOSIER, C. T., 1989, An experiment investigating the application of clustering procedures and similarity coefficients to the GT machine cell formation problem.
International Journal of Production Research, 27, 1811-1835.
22. PURCHECK, G. F. K., 1974, Combinatorial groupings lattice-theoretical method for the design of manufacturing systems. Journal of Cybernetics, 4, 27-60.
23. PURCHECK, G. F. K., 1975, A mathematical classification as a basis for the design of group technology production cells. The Production Engineer, 35-48.
24. RAJAGOPALAN, R., and BATRA, J. L., 1975, Design of cellular production systems: a graph theoretic approach. International Journal of Production Research, 13,
567-579.
25. RAND, W. M., 1977, Objective criteria for the evaluation of clustering methods. Journal of the American Statistical Association, 66, 846-850.
26. STANFEL, L. E., 1985, Machine clustering for economic production. Engineering Costs and Production Economics, 9, 73-81.
27. TARSUSLUGIL, M., and BLOOR, J., 1979, The use of similarity coefficients and cluster analysis in production flow analysis. Proceedings of the 20th International
Machine Tool Design and Research Conference, pp. 525-531.
28. WAGHODEKAR, P. H., and SAHU, S., 1984, Machine-component cell formation in group technology: MACE. International Journal of production Research, 22,
937-948.
29. WEMMERLOV, U., 1984, Comments on direct clustering algorithm for group formation in cellular manufacture. Journal of Manufacturing Systems, 3.
30. WEMMERLOV, U., and HYER, N. L., 1986, Procedures for the part family/machine group identification problem in cellular manufacturing. Journal of Operations
Management, 6, 125-147.
31. WEMMERLOV, U., and HYER, N. L., 1987, Research issues in cellular manufacturing. International Journal of Production Research, 25, 413-431.
Back to Top of the Page

http://net1.ist.psu.edu/chu/course/cm.htm

6/29/2014

S-ar putea să vă placă și