IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. 24, NO.

2 , FEBRUARY 1994 203

Group Technology and Cellular Manufacturing

Sunderesh S. Heragu

Abstract- A number of survey papers on group technology

and cellular manufacturing system design have been published.
Many of them focus primarily on clustering techniques that
manipulate rows and columns of the part-machine processing
indicator matrix to form a block diagonal structure. Since the last
survey paper was published, there have been some developments
in cellular manufacturing system design. A number of papers
that consider practical design constraints while designing cellular
manufacturing systems have been published. The purpose of
I lll." ........."

&I$&:
....,.l._l. .......,,l.l..l .-"I
I

.,.........._. ..... .......................

.,

this paper is to provide a thorough survey of papers on group

technology and cellular manufacturingsystem design. Its purpose
is also to state some important design factors that cannot be
ignored. BM
DM
TM
Broachng Machine
Dnllmg Machine
Turning Mxhme
VMM Vemcal hlrlling Machine
-
- - - JS-
Roumg of pans PI, P3, P9
Rouung of pans P2, P4, P7
Rouwlg of parts P5 P6 , P8 .
I. INTRODUCTION

G ROUP TECHNOLOGY (GT) is a management philoso-

phy that attempts to group products with similar design
and/or manufacturing characteristics. Cellular manufacturing
Fig. 1. Arrangement of cells in a jobshop environment

(CM) can be defined as an application of GT and involves

grouping machines or processes on the basis of parts or part
families they process [71]. It is a relatively recent concept
and has been applied in many manufacturing environments
successfully. There are significant benefits that can be achieved
as a result of implementing a CMS. A number of companies
surveyed in [70] have witnessed the following:
*Setup time reduction
@Work-in-processinventory reduction
*Material handling cost reduction
*Equipment cost and diredindirect labor cost reduction BM Broachmg Machme ___) RouungofpansP .P . P
1 3 9
*Improvement in quality DM Lhllmg Machine
Rouung of pans P2 , P 4 , P7
TM Turning Machine
*Improvement in material flow VMM Verucal hlrlhng Machine ?a- Rouung of pans P ,P
5 6
.P8
*Improvement in machine utilization
*Improvement in space utilization Fig 2 Arrangement of cells in a cellular manufactunng system
'Improvement in employee morale
The main difference between a traditional jobshop envi- The main objective in the design of a CMS is to create
ronment and a cellular manufacturing environment is in the machine cells, identify part families, and allocate part families
grouping and layout of machines. In a jobshop environment, to machine cells so that the inter-cellular movement of parts
machines are typically grouped based on their functional of part families is minimized. Of course, while doing so,
similarities (see Fig. 1). On the other hand, in a cellular a number of other constraints need to be considered. For
manufacturing environment, machines are grouped into cells example, the allocation of part families to machine cells must
SO that each cell is dedicated to the manufacture of a specific
be such that available capacity of machines in each cell is not
part family (Fig. 2). Typically, machines in each cell are exceeded; safety and technological requirements pertaining to
dissimilar in their functions. Such an arrangement in which the location of equipment and process must be met; size of
sets of machines are dedicated to specific part families allows a cell and number of cells must not exceed a user specified
easier control of a CMS. number; etc.
Manuscript received August 14, 1992; revised April 21, 1993. This work A majority of the papers on GT and CMS design ignore
was supported by grants from the New York State Science and Technology some or all the constraints listed above and focus only on
Foundation and Welch Allyn Medical Division, New York. the identification of machine cells and corresponding part
The author is with the Department of Decision Sciences and Engineering
Systems, Rensselaer Polytechnic Institute, Troy, NY 12180- 1390. families. Given a part-machine processing indicator matrix
IEEE Log Number 9214536. such as the one shown in Fig. 3, they basically attempt to find a
0018-9472/94$04 1.00 0 1994 IEEE

Authorized licensed use limited to: ULAKBIM UASL - GAZI UNIV. Downloaded on December 17,2023 at 13:07:23 UTC from IEEE Xplore. Restrictions apply.
 204 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. 24, NO. 2, FEBRUARY 1994

M a c h i n e M a c h i n e
M4 M, Me

i
M4 M6 M2

1 1
1 1
1 1
I 1
I 1
I
1 I 1

Fig. 3. Sample par-machine processing indicator matrix Fig. 5. Revised processing indicator matrix

M a c h i n e as is and machine cells are formed as indicated in Fig. 5 ,

it will be necessary for part PZ to visit both the cells MC1
and MC2. For such types of problems, there are a number
of methods that rearrange rows and columns (i.e., attempt to
identify part familiedmachine cells) so as to minimize the
intercellular movement of parts.

p5L
Just as some problems may have exceptional parts, other
1 1
I problems may have bottleneck machines (columns) which
p6 I ’ when removed will result in identification of the block diago-
nal structure. These bottleneck machines are so named because
Fig. 4. Rearranged processing indicator matrix. two or more part families share this resource. For example,
for the matrix shown in Fig. 5 , if the column corresponding to
rearrangement of the rows and columns of the matrix that will machine M I is removed, again two mutually separable clusters
provide a block diagonal form as shown in Fig. 4. Typically, of machine cells and part families can be identified. It is well
the part-machine processing indicator matrix consists of 0, 1 known that additional copies of exceptional machines can be
entries only; a 1 entry in row i, column j indicates that the purchased and placed in appropriate cells so as to eliminate
part corresponding to the ith row is processed by the machine or at least minimize the intercellular movement of parts; of
corresponding to the j t h column. A 0 entry means that the course, this must be done only if the cost of bolttleneck
part is not processed by the corresponding machine. Typically, machines is not high. A systematic procedure that uses cost
zeros are left blank in the matrix as in Figs. 3 and 4. information for evaluating the elimination of bottleneck ma-
The clusters of 1’s around the diagonal of the matrix chines and exceptional parts can be found in [32]. Shafer et
in Fig. 4 indicate three groups or machine cells MC1 = al. [58] also present a mathematical programming approach
{M1,M4rM~) and MCZ = {hfZ,M3,hfj,hf7);the cor- for dealing with exceptional parts and bottleneck machines.
responding part families are PF1 = { P I ,P3) and PF2 = Zhang and Wang [72] suggest that it is useful to use
{ P z ,P4, Pj,P6). It should be noted that not all part-machine nonbinary values between 0 and 1 to indicate the degree
processing indicator matrices can be rearranged to fit a block of match between each part-machine pair in the processing
diagonal form such as the one indicated by dashed lines in Fig. indicator matrix [a,,]. They show how the degree of match can
4. For many matrices, such a form may not even exist. For be calculated using fuzzy set theory. Compared to the binary
example, assume that part PZ requires processing on machine matrix approach, in which the only information available
M I in addition to MZ M3 and Mj. The corresponding part- is whether or not a part is processed on a machine, this
machine processing indicator matrix is shown in Fig. 5. For method is very flexible as it provides a mechanism to capture
such a problem, no rearrangement of the rows and columns a number of other relationships between each part-machine
will produce a block diagonal form such that no 1’s lie outside pair (e.g., cost of processing a part on a machine processing
the block indicated by dashed lines. time, etc). Ben-Arieh and Triantaphyllou [ 5 ] also use fuzzy
The rows (parts) corresponding to the 1’s that lie outside set theory and show how qualitative data and quantitative
the diagonal block are usually referred to as exceptional parts data that have a subjective meaning can be considered while
because when these parts are removed, a block diagonal struc- identifying machine cells and part families. They also provide
ture can be easily identified. Of course, some rearrangement of an algorithm that is capable of handling such fuzzy data.
the remaining rows and columns must still be done to uncover The part-machine processing indicator matrix may be mod-
the block diagonal structure. ified to handle other pieces of valuable information that
In the example considered (see Fig. 3,the exceptional part are used in identifying machine cells and part families. For
is Pz. If this part is removed from the matrix, two mutually example, two additional columns can be created to indicate
separable part familylmachine cell combinations can be easily the number of parts to be manufactured and the batch size for
identified. Some authors have suggested that exceptional parts each part [29]. Similarly, the sequence of machines visited by
be subcontracted out. On the other hand, if this part is left a part may also be recorded in the matrix [48]. This can be

Authorized licensed use limited to: ULAKBIM UASL - GAZI UNIV. Downloaded on December 17,2023 at 13:07:23 UTC from IEEE Xplore. Restrictions apply.
 ~

HERAGU: GROUP TECHNOLOGY AND CELLnAR MANUFACTURING 205

be focused. The focus of this paper is on GT and CMS

design. Consequently, techniques for GT and CMS design are
surveyed in detail with emphasis on papers pertaining to the
latter category. Also, a list of some of the important constraints
that msut be considered while designing a CMS are presented.
Each constraint is discussed in detail. Some general concluding
remarks are also provided.

OF GT TECHNIQUES
11. VARIOUSCLASSIFICATIONS
Fig. 6. Processing indicator matrix indicating sequence of operations.
Mitrofanov [46] first introduced the GT philosophy and the
machine grouping problem in the late 1950’s in the U.S.S.R.
accomplished simply by defining xzj to be the following: The other early pioneers in the field of GT are Burbidge [9] and
lcif part i visits machine j Ham [23]. Since then, a number of researchers have developed
xzJ=
i for the lcth operation
0 otherwise.
where lc is an integer representing the operation for which part
techniques for solving the GT and CM problems. These
techniques have been classified and reclassified a number of
times.
Burbidge [9] classified the techniques as follows:
i visits machine j. Such a modified matrix is shown in Fig. 6.
The advantage of such a representation is that available 1) Rule of Thumb techniques: These techniques, sometimes
clustering techniques such as the one in [27] can be used. The referred to as eyeball methods, use some simple rules
above representation enables us to capture routing sequence of thumb to identify part families and machine cells
information which is useful in determining material flow [6]. Clearly such techniques are not useful in solving
between machine pairs. As long as each machine is visited large-scale problems, but are relatively easy to use. In
by a part for any number of consecutive operations, the other words, they can be used to obtain “quick-and-dirty’’
representation shown in Fig. 6 is adequate. All the consecutive solutions to small-scale problems. Because there are a
operations on the same machine may be treated as a single number of efficient techniques available today, these are
operation. However, if the same machine is visited for two or not preferred.
more nonconsecutive operations, then the above representation 2) Classijication and coding techniques: Unlike the Produc-
is not adequate. In such a case, each entry in the above matrix tion Flow Analysis (PFA) techniques (discussed below)
will have to modified as a vector representing all the operations which group parts solely based on their processing char-
for which the corresponding machine is visited by the part. acteristics, classification and coding techniques group
For example, if part Ps visits machine Mz for the second parts on the basis of a number of attributes. Examples
and fourth operation, then this information will have to be of such attributes are shape of part; dimensions of part;
stored in the form of a vector-(2,4) in the 5th row and 2nd material composition of part; tolerance requirement and
column of the matrix. This obviously increases the memory operations requirement. Typically, each part is assigned
requirements for data storage. Hence, Heragu and Kakuturi a ten- to thirty-digit code with each code representing an
[29] suggest an alternate way of representing nonconsecutive attribute of the part. In some classification and coding
opeations on the same machine by use of dummy columns. In systems, a hierarchical relationship exists amongst the
general, it can be argued that it is advantageous to modify the digits; Le., information stored in digits with higher
part design or select alternate process plans so that the same numbers is related to the information stored in digits
machine is visited more than once for consecutive operations with lower numbers. For example, digit 2 may represent
only; otherwise there will be an unnecessary increase in set-up the part shape attribute, digit 3 may represent rotational
time. parts, digit 4 may respresent rotational parts with diam-
The above modifications suggested are useful in determining eter greater than 3 inches, and so on. In some other
the flow between each machine pair. The flow information is classification and coding systems, such a hierarchical
used in some algorithms for identifying machine cells (see relationship may exist partially or may not exist at all.
[28] for example). Thus, depending upon whether or not a hierarchical
To date, a number of survey papers on GT have been relationship exists, classification and coding systems
published. However, all of them have primarily surveyed have been classified as [41]
clustering techniques (i.e., techniques that are concerned with *hierarchical codes;
the rearrangement of rows and columns of the binary part- won-hierarchical codes; and
machine processing indicator matrix so as to identify a block *hybrid codes.
diagonal structure). Moreover, since the survey paper in 1988,
a number of developments have taken place. A number of Ham et al. [24] provided a survey of classification and
algorithms that are concerned with practical CMS design coding techniques.
problems have been published. It is therefore necessary to 3) Production Flow Analysis (PFA) techniques: PFA tech-
develop a new framework over which research attention can niques involve the systematic listing of information

Authorized licensed use limited to: ULAKBIM UASL - GAZI UNIV. Downloaded on December 17,2023 at 13:07:23 UTC from IEEE Xplore. Restrictions apply.
 206 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. 24, NO. 2, FEBRUARY 1994

contained in route cards and identification of part fam- to uncover clusters of machine cells and corresponding
ilies and machine cells by careful inspection. Some of part families.
the latter forms of PFA techniques use a part-machine Han and Ham [25] classified the GT algorithms as follows:
processing indicator matrix to specify the machining
1) Peripatetic and ocular techniques: These techniques are
requirements on parts and then attempt to manipulate
similar to the “rule-of-thumb” or “eyeballing” techniques
the rows and columns of this matrix to identify clusters.
described above. Knowledge concerning the parts and
While forming machine cells and part families, they do the manufacturing system are used in determining ma-
not take into consideration design, shape or other part char- chine cells and part families. These methods are also not
acteristics. Burbidge [SI argued that these characteristics are useful in practice as better techniques exist.
not useful in GT; processing requirement for each part is 2 ) PFA techniques.
the only information that is needed. He also argued that 3) Classification and coding techniques.
GT is suitable for all manufacturing systems. Citing personal 4) Mathematical programming techniques: These techniqu-
consulting experiences, Burbidge mentioned there is sufficient es use fuzzy mathematics, pattem recognition, cluster
flexibility in the part routing (or process planning) that allows analysis, etc. to identify part family/machine cell com-
PFA to be used to identify mutually separable machine cells binations.
and corresponding part families. However, a survey of many
Vakharia [65] used the following classification:
companies in the U.S. that use GT found that only in rare
circumstances were companies able to identify mutually sep- 1) Descriptive techniques: Descriptive techniques include
arable clusters of machine cells and part families [70]. Also, the PFA techniques discussed above and other compo-
Wei [68] pointed out that the studies by Morris and Tersine nent (or part) flow analysis (CFA) techniques discussed
[47] and Flynn and Jacobs [20] indicate that GT provides in [34]. King and Nakomchai [34] pointed out that there
operational advantages only under certain conditions. Through are a number of similarities between the PFA and CFA
a simulation study, Suresh [63] and [64] provided other techniques. Some minor differences between the two are
conditions under which functional systems may be preferable also noted in that paper.
to CMS’s. 2 ) Block diagonal techniques: These techniques are those
King and Nakornchai [34] used the following four cate- that were previously defined as clustering techniques.
gories: 3) Similarity coeficient techniques.
4) Other analytical techniques: The set-theoretic based
Similarity coeficient algorithms: These are algorithms techniques such as the one discussed in King and
that are derived from numerical taxonomy and attempt Nakomchai [34] come under this classification.
to measure the similarity coefficient between each pair of
machines or parts. For example, the similarity coefficient Wemmerlov and Hyer [71] did an extensive analysis of
for a pair of machines may be given by the number some of the above classifications and also various techniques.
of parts which visit both machines divided by the They pointed out that most other classifications do not help
number of parts which visit at least one machine. Most us in understanding the process of cell formation. Hence they
of the similarity coefficient algorithms use the Jaccard classified the GT techniques as follows:
similarity coefficient [59]. Examples of algorithms that 1) Techniques that identifi part families without the help of
measure the similarity between machines are, the single machine routing.
linkage clustering algorithm in [44], its modification in 2 ) Techniques that identifi part families using routing.
[53], the algorithm in [57], etc. Algorithms that measure 3) Techniques that identifi machine groups only.
the similarity between parts are presented in [lo], [17], 4) Techniques that identifi part families and machine
etc. groups simultaneously.
Set-theoretic algorithms: These techniques build super- The following categories were used to classify GT tech-
sets of machines and parts that can be represented as niques in [39]:
a path along the edges of a lattice diagram using the 1) PFA techniques.
union operation. These techniques were first suggested 2) Similarity coeficient techniques.
by Purcheck [50] and later refinements can be found 3) Sorting-based techniques.
in [51] and [52]. King and Nakomchai [34] mention 4) Bond-energy techniques.
that the lattice diagram grows exponentially and hence 5 ) Cost-based techniques.
such techniques will not be useful for practical problems 6) Cluster-identification techniques.
where the number of parts and machines are typically The sorting-based techniques, bond-energy techniques, and
large. cluster identification techniques are specific clustering tech-
Evaluative algorithms: These techniques are basically niques. Cost-based techniques attempt to identify part fam-
the same as PFA techniques described above. ily/machine cell combinations that minimize machining, set-
Other analytical techniques: These are techniques that up, inventory and material handling costs.
were defined as clustering techniques in Section I. They In addition to the above, there are a few other authors that
examine the part-machine processing indicator matrix have classified GT techniques. These are listed in [71] and the
and attempt to rearrange its rows and columns so as interested reader is referred to that source for more details.

Authorized licensed use limited to: ULAKBIM UASL - GAZI UNIV. Downloaded on December 17,2023 at 13:07:23 UTC from IEEE Xplore. Restrictions apply.
 HERAGU: GROUP TECHNOLOGY AND CELLULAR MANUFACTURING 207

111. DESIGNAND PLANNING ISSUES

CELLULARMANUFACTURING
IN
SYSTEMS .jl: xijktijk 5 rjNMj for each j
i=l k=l
As mentioned in the previous section, a number of tech- Xijk, N M j 2 0 and integer for each i , j , k .
niques have been developed for GT and CMS design. Most
of them are primarily concemed with the identification of part X i j k in the above model may be used to detemine the
families and machine cells and do not take into consideration number of times operation i is to be performed on part type k
other practical design constraints. These constraints are listed using machine type j . Note that it is implicitly assumed that
below in order of their importance. The list is an extension each machine type can perform all the operations on each part
of the one in [27]. type. If a machine type j cannot perform an operation i on
Available capacity of machines must not be exceeded. part type k , then that particular X i j k can be set to 0. This will
Safety and technological requirements must be met. decrease the number of Zijk variables. The objective function
Number of machines in a cell and number of cells must minimizes the machine purchase and operating costs. The first
not exceed an upper bound. constraint ensures that the required number of each operation is
Inter-cell and intra-cell cost of handling material between performed on the machines. The second constraint is a capacity
machines must be minimized. constraint and ensures that the time available on each machine
Machine utilization rate must be as high as possible. type is not exceeded. The last constraint is obvious and does
Machine purchase and operating costs and work-in-process not need any explanation.
inventory costs must be minimized. The above model only ensures that the required number
In addition to the above design constraints, there are other of each machine type is available. While allocating machines
planning issues that need to be addressed [65]. Two of the to cells, it is equally important to verify that there is adequate
more important ones include capacity in each cell to completely process all the part families
assigned to it. To explain this, assume that
a) scheduling of jobs in individual cells; and a) Two units of machine M5 are available and that these
b) throughput rate maximization. two machines process parts P I ,P4 and P5 only. (There is
just enough capacity on these two machines to process the
A. Machine Capacity three parts).
It is obvious that machine capacity is more important than b) High volume parts PI and P4 are in part family PF1,
the other constraints. It is therefore necessary to first ensure whereas low volume part Ps is in part family PF2.
that adequate capacity (in machine hours) is available to c) One unit of machine M5 is in machine cells MC1 and
process all the parts. Thus, the number of each type of machine MC2.
must be known apriori. Otherwise, it must be determined. For Since there is just enough capacity on the two units of
this purpose, we use the model presented in [28]. The model machine M5 to process all the three parts, the capacity
and notation used are provided below. constraint in cell MC1 will be violated as a result of the above
*type of operations to be performed on the various parts; grouping because the two high volume parts PI and P4 are
i = 1,. . . , P processed on only one unit of machine M5. Hence, because
*types of machines available; j = 1,. . . , m of the capacity limitation it may be necessary to process parts
*types of parts to be manufactured; k = 1, . . . , n PI and P4 on the other machine M5 in cell MC:!. This will
*number of units of part type k to be manufactured; obviously require some units of part PI or P4 to visit the
NPk,k = I , . . . , n two cells. This hypothetical example illustrates the fact that
*cost of performing operation i on part type k using machine while the number of pieces of each type of equipment may be
type j ; Cijk,z = 1, “ . , p , j = l , . ‘., m, IC = 1,. . . , n sufficient to process all the parts, the machine grouping may
*time required to perform operation i on part type k using be such that there is not enough capacity to process one or
machine type j ; tijk,z = 1, . . . , p , j = 1, . . . , m , k = more parts entirely within their corresponding cells. Thus, it
1,’ . . , n is necessary to ensure that the capacity constraint in each cell
*number of machines of type j selected; N M j , j = is not violated while allocating machines to cells and parts to
1,.’ . , m part families.
*purchase cost of machine type j prorated over the planning
period; C M j , j = l , . . . , m B. Safety and Technological Considerations
*time for which machine type j is available; r j , j = Safety or technological considerations may dictate two or
1,.. . , m more machines to be placed in the same cell, regardless of the
Model number of parts visiting the two machines. An example would
m be the forging and heat-treatment stations. Due to fire hazards,
these two stations must be placed in the same cell even if the
i = l j=1 k=l j=I number of parts visiting these two stations are relatively small
m in number. Conversely, two or more work-stations cannot be
subject to xijk = NPk for each i , IC placed in the same cell, even if a large number of parts are
j=1 processed by the two work-stations. The painting and welding

Authorized licensed use limited to: ULAKBIM UASL - GAZI UNIV. Downloaded on December 17,2023 at 13:07:23 UTC from IEEE Xplore. Restrictions apply.
 208 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. 24, NO. 2, FEBRUARY 1994

stations are two such stations. Because of the sparks generated technique does not determine the position of cells or machines
in the welding station, and the possibility of sparks igniting within each cell.
flammable solvents in the painting station, it is desirable to A number of researchers, for example, Co and Araar [16],
locate the welding and painting stations in different cells, as Harhalakis et al. [26] and Logendran [42] have recognized
far apart as possible, although there may be a high interaction that operations sequence for each part must be considered in
between the two stations. Such factors must be considered the formation of machine cells. For, if this is done, it will
while allocating each machine to a nonempty cell. enable identification of machine cells so that inter-cellular
movement of parts is minimized. For example, if a part has
one intermediate operation that is performed in a secondary
C. Upper Bound on the Number of Machine cell, and all the others in a primary cell, then this will
Cells and Size of a Cell necessitate two trips (for each batch) between the two cells.
It has been pointed out that in some CMS’s, a set of multi- On the other hand, if only the last operation is performed
functional workers may be assigned to oversee the operations in a secondary cell, then there is only one trip required. A
of a certain number of cells in order to improve the utilization technique to minimize inter-cellular movement of parts is
of employees [6]. Since the availability of such cross trained discussed in [27]. Since the product mix changes from time
workers is limited, it is necessary to impose an upper bound to time and is not known a priori, Seifoddini [56] suggests
on the number of cells. a probabilistic model be used to determine the expected
In many CMS’s, there is also an upper bound on the number intercellular material handling cost and outlines a mechanism
of machines that can be included in a cell. Askin and Chiu to capture this uncertainty. Representation of the part-machine
[I] mention that in practice, management can determine such processing data as shown in Fig. 6 is useful is recording
upper bounds based on their experience. There are a number of operation sequence information.
reasons why such upper limits may be necessary. For example,
for control purposes, it may be desirable to assign one operator E. Machine Utilization
to a machine cell. Since an operator can attend to a limited
Machine utilization is a very important issue that must
number of machines, it may be desirable to place an upper
be addressed especially when equipmentlprocess selection
bound on the number of machines in a cell. Furthermore, floor
decision is made. At later stages, the importance of this factor
plan dimensions may dictate the size of a cell in some GT
diminishes. This is because costs are already incurred as a
problems [29]. In addition to upper bounds on the number of
result of purchasing equipment in some previous stage and
machines in a cell, Askin and Chiu [ l ] also impose a lower
hence there is not much to gain by ensuring high equipment
bound. Such a consideration may be necessary if it is desired
utilization when making planning decisions.
to ensure that each operator is assigned a minimum workload.

F. Cost Minimization
D. Minimization of Inter-Cellular and Intra-Cellular In Fig. 5 , we saw how a bottleneck machine ( M I ) may
Material Handling Cost prevent the block diagonal form from occuring. For problems
In an ideal CMS, there will be no flow of material between with bottleneck machines, a number of authors have suggested
cells. But a number of researchers have pointed out that for duplication of bottleneck machines in the appropriate cells so
practical problems, it is difficult to form machine groups that parts belonging to a part family are processed entirely
between which there is no flow of material. Given this practical within the cell. For the data in Fig. 5, if an additional unit of
limitation, we often seek to form groups so that the inter- machine M I is place in cell MC1 (which consists of machines
cellular movement of parts in minimized. In fact, one of the M2, M3 Ms and M7), part P2 may be entirely processed
~

main objective of most GT techniques is to do just that (see within cell MC2. (This is illustrated in Fig. 7). However, the
Fig. 2). An equally important factor that needs to be addressed benefit of forming mutually separable clusters may be more
is that of minimizing intra-cellular cost of transporting parts. than offset by the expense incurred as a result of purchasing
Obviously, this requires the layout design to be done for additional units of bottleneck machines; or the addition of
each cell. Heragu and Kakuturi [29] have implemented a 3- duplicate machines may increase the cell size constraint.
stage approach in which machine cells (and corresponding part Under such circumstances, a physical arrangement in which
families) are identified first; layout of machines within each bottleneck machine(s) are placed between machine cells so
cell is determined next and layout of cells is determined in the that the corresponding part families may share the resource,
third stage. It should be noted that if the maximum number may be an attractive alternative. A physical arrangement that
of machines in a cell and maximum number of cells are less allows the sharing of an expensive bottleneck machine M I
than 15, optimal techniques that enable us to solve the two between two part families is shown in Fig. 8. For illustration
layout problems exist. purposes, it is assumed that the first set of machines is served
Instead of a sequential approach, a simultaneous solution of by a robot and the other set by an AGV in both figures.
the grouping and layout problems would be desirable but such The example considered above underscores the importance
problems are difficult to solve. A mathematical model that of considering machine procurement costs while designing
minimizes the inter-cell material handling costs in addition CMS’s. In addition to this cost, there are others that need
to other operating costs is presented in [l]. However, this to be considered. Examples are

Authorized licensed use limited to: ULAKBIM UASL - GAZI UNIV. Downloaded on December 17,2023 at 13:07:23 UTC from IEEE Xplore. Restrictions apply.
 HERAGU: GROUP TECHNOLOGY AND CELLULAR MANUFACTURING 209

Iv. LITERATURE
REVIEWOF RECENT
GT AND CMS DESIGN
TECHNIQUES FOR
In this section, we survey a number of papers published on

R GT and CM. Before surveying the papers, it is necessary to

provide a broad classification of the papers. For purposes of

H literature review, these techniques are classified as

*clustering techniques; and
*CMS design techniques.
;1,...,M7 Machines
Robot
AGV Automated Guided Vehicle
A. Clustering Techniques
Fig. 7. Duplication of machine M I results in the formation of two smaller In a broad sense, clustering techniques can be defined
cells. to be those techniques that, given an input object-object or
object-attribute data matrix, attempt to uncover and display
similar cluster or groups [45].As discussed in Section I,
“5
they often do so by rearranging rows and columns of the
input matrix. Clustering techniques have been applied in a
number of areas and are listed in [39]. Since this paper deals
with CMS’s, clustering techniques are defined rather narrowly.
They are defined as techniques that are concerned only with
the identification of machine cells and/or corresponding part
families; they neither take into account the design constraints
listed in Section 111 nor consider cost factors. Also, most
;I,...,M 7 Machines of the techniques discussed use process plan or part routing
Robot
AGV Automated Guided Vehicle information in forming machine cells andor part families.
Such techniques have been classified in [71] as discussed in
Fig. 8. Physical placement of machine M I allows two part families to share Section 11. We modify that classification as shown below.
the machine.
1) Techniques that identify part families only.
2) Techniques that identify machine cells only.
*work-in-process inventory; 3) Techniques that identify part families and machine cells
*machine depreciation; and simultaneously.
*machine setup. In CMS design, it is clear that techniques belonging to the
If the above costs are considered while identifying machine latter category are more useful than those belonging to the
cells and allocating part families to machine cells, the resulting other two. This is because knowledge concerning machine
solution will not only enable better control of the manufactur- grouping makes the task of the layout analyst easier. Part
ing system but also minimize operating costs significantly. family identification simplifies the planning process [27].
Techniques that Identify Part Families Only To our knowl-
edge, there are only a few techniques that come under this
G. Scheduling of Jobs in Individual Cells
category. The first one uses a similarity coefficient which
Vakharia [65] first recognized the need to consider sched- specifies the similarity between parts and was developed by
uling issues while designing CMS’s. He noted that since there Carrie [lo]. The similarity coefficient for a part pair { z ; j } is
are fewer jobs in each cell of CMS when compared to the determined by the ratio of the number of machines visited by
entire system, solving scheduling problems optimally in a parts i and j to the nubmer of machines visited by i or j .
CMS may not be impossible. This coefficient is somewhat similar to the one used in [44].
Carrie [lo] also developed a clustering technique to identify
part families. This technique adds a part to a part family
H. Throughput Rate Maximization
provided there is an acceptable level of similarity between the
The primary goal in any manufacturing system is to max- two. The technique incorporates an additional constraint that
imize throughput rate of parts. Maximization of throughput specifies the minimum number of parts allowed per family.
rate should be given more importance than maximization of Carrie [lo] demonstrated how the same principles can be
machine utilization, especially if the machines have already applied to develop plant layout.
been selected. Throughput rate can be maximized in a number Han and Ham [25] developed a technique that identifies
of ways. For example, set-up time for each operation can be part families only. However, the technique does not use part
reduced. In fact, as a result of implementing GT, set-up times routing information in generating part families. Rather, it uses
are reduced somewhat as similar parts are manufactured within a classification and coding system which specifies part design
a cell. If set-up times can be further reduced by redesigning and manufacturing characteristics. The part family identifica-
parts or redesigning fixtures, this will enable us to increase tion problem is formulated as an integer programming problem
the throughput rate even more. and the absolute Minkowski metric is used to specify the

Authorized licensed use limited to: ULAKBIM UASL - GAZI UNIV. Downloaded on December 17,2023 at 13:07:23 UTC from IEEE Xplore. Restrictions apply.
 210 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. 24, NO. 2, FEBRUARY 1994

distance between two parts. The objective function of the considers some additional design constraints, it is discussed
model lexicographically minimizes the distance between a part in a later section.
and part family. The model is then solved using the goal Since the methods in [44] and [53] do not always work well,
programming method. de Witte [18] proposed another approach in which machine
Offodile [49] applied the similarity coefficient method to types were divided into three categories:
parts coding and classification analysis. In his approach, a part-
a) primary (if only one machine of that type is available
part similarity matrix is constructed and used to identify part
or if all machines of that type have to be allocated to
families. First, the largest element in the matrix is identified
a single cell);
and the corresponding row, column indicate the parts to be
included in the first part family. Then the rows, columns b) secondary (if more than one machine of that type is
available); and
corresponding to the parts in the just formed part family are
c) tertiary (if enough number of machines of that type is
replaced by a single row and column indicating the maximum
available so this machine type can be allocated to all
similarity of every other part (that has not been included
cells).
in any part family) with the parts in existing part families.
The resulting matrix is used to determine whether a new part In order to analyze the relationship between the three
family is to be formed or whether a new part is to be added machine types, de Witt [ 181 used three similarity coefficients.
to an existing part family. This is repeated until each part He identified appropriate clusters (or cells). In the last two
is added to a part family. Kini et al. [35] also presented steps, he allocated corresponding parts to the cells and refined
a new approach to part numbering that has applications in the solution further.
GT. Faber and Carter [ 191 present a graph theoretical approach
Currie [17] presented a technique that uses a self-organiz- in which a machine-machine binary similarity matrix is con-
ing neural network model-the interactive activation and structed using a similarity measure. Then a similarity graph
competition (IAC) model, that identifies part families by con- (with nodes indicating machines and arcs between machine
sidering design and manufacturing similarities simultaneously. pairs only if the machine pair has a 1 value in the binary
A part-part similarity matrix is constructed and a bond energy matrix) is drawn. The densest subgraph of this graph is found
algorithm is used to identify part families. Machine cells are by solving a min-cut max-flow problem. Machine cells are
identified only after the part families are formed and they done identified using the densest subgraph by means of a polynomial
simply by examining the machines required to process parts in time algorithm. The technique consider upper bounds on cell
each family. Thus, although machine cells are identified, this size.
algorithm is discussed in this section because its emphasis is The single-linkage clustering algorithm in [44] was mod-
on identifying part families only. ified to overcome the chaining problem by Seifoddini and
Techniques that Identify Machine Cells Only Some of the Wolfe [57].They also used special data storage and analysis
techniques discussed below, for example, the algorithms in techniques to improve the algorithm’s computation perfor-
[18] and [57], identify machine cells first and then part mance. Further, some strategies dealing with the duplication
families based on the cells identified. Such techniques are also of bottleneck machines are incorporated.
discussed in this section because their primary focus is on Srinivasan et al. [61] proposed an assignment based algo-
identifying machine cells. rithm that identifies machine groups first and then part families.
McAuley [44] used a similarity coefficient applied in numer- An assignment problem is solved for a machine-machine
ical taxonomy and constructed a similarity coefficient matrix similarity coefficient matrix. This results in a number of
[Sjk], where Sjk is the similarity coefficient for a pair of closed loops each corresponding to a machine cell. Parts
machines { z , j } ; sjk is the ratio of the number parts that visit visiting these cells form a part family. If a disjoint set of
both the machines to the number of parts that visit at least part families can be identified, the algorithm stops. If not,
one of these two machines. Notice the similarity between this it checks whether two machine groups can be merged into
coefficient and the one used in [lo]. Whereas the coefficient one so as to identify a disjoint set of part families. If this is
above is defined for machine pairs, the similarity coefficient not possible, an assignment problem is solved for a part-part
used in [lo] deals with part pairs. similarity coefficient matrix. The resulting part families are
The single-linkage clustering technique developed by then assigned to machine cells so as to minimize the number
McAuley [44] adds a new machine to an existing cell if of parts visiting multiple cells. This algorithm was further
the similarity coefficient between the new machines and any improved in [60].
existing machine in the cell exceeds a certain threshold level. In addition to the above techniques there are some others
An obvious disadvantage of this method is that a machine that attempt to identify machine cells or groups only. Some
which has a high similarity coefficient with just one other of these are discussed in [71].
machine already in the cell will automatically be included Techniques that Identify Machine Cells and Part Families
(in it) even if the similarity coefficient between the new Simultaneously There are a number of a techniques that fall
machine and all other machines in the cell is very low. under this category. Iri [30] presented a clustering algorithm
This is referred to as the chaining problem. Some of the which starting from any row, masks all columns having a 1
deficiencies were remedied by Rajagopalan and Batra [53] entry in that row. It then masks all rows which have 1 entries in
who developed a graph theoretic method. Since their method the masked columns. This procedure is repeated until it is not

Authorized licensed use limited to: ULAKBIM UASL - GAZI UNIV. Downloaded on December 17,2023 at 13:07:23 UTC from IEEE Xplore. Restrictions apply.
 HERAGU GROUP TECHNOLOGY AND CELLULAR MANUFACTURING 211

possible to go to new unmasked rows or columns. When this Boe and Cheng [7] presented a close neighbor algorithm
condition occurs, a cluster of machines and corresponding part which rearranges a closeness matrix (the matrix indicates
families is formed. The above procedure is repeated to identify closeness between each pair of machines based on the parts
other clusters. Kusiak and Chow [40] presented essentially the they process) so as to identify clusters of machines. Then
same algorithm. Instead of masking rows and columns, they the part-machine processing indicator matrix is rearranged
draw horizontal and vertical lines. Modified versions of this such that adjacent rows correspond to machines in a cluster.
basic technique can be found in [39] and [37]. This gives a preliminary block diagonal form which is further
McCormick et al. [45] presented a heuristic bond-energy improved using another heuristic procedure. The algorithm’s
algorithm (BEA) that attempts to maximize the sum of the performance was compared with that of a number of other
bond energies for each element { z , j } in the part-machine clustering algorithms and found to be superior with respect to
processing indicator matrix [a,,]. The bond-energy for element solution quality and computation time.
{ i , j } is given by: Kaparthi and Suresh [31] and Kusiak and Chung [38] have
applied the Carpenter-Grossberg neural network to determine
a2&t3+1 + % - I + %+1, + &1J] the part family/machine cell combinations. This network has
The heuristic technique attempts (in a single pass for the row two layers of neurons with each neuron in the input layer
and column) to determine a permutation of rows and columns having a top-down and bottom-up connection to each neuron
of the part-machine processing indicator matrix [a,,] so as to in the output layer. The input layer serves as a comparison
maximize the sum of the bond energies. If a block diagonal layer while the output layer is a recognition layer. The basic
structure exists, it is immediately identified. However, if there procedure is an follows. When the first set of data, for example,
are one or more bottleneck elements that prevent the formation a row of the part-machine processing indicator matrix, is
of such a block diagonal structure. then the technique does not provided to the network, it is stored as a representative vector
perform well. Gongaware and Ham [21] developed a modified for the first cluster. The subsequent sets of data are also read
and improved version of the original BEA that eliminated a one row at a time, and compared to existing representative
tedious manual sorting procedure and outperformed the latter vectors. They are either treated as being part of existing
t711. clusters (in which case, the corresponding representative vector
King [33] and King and Nakornchai [34] developed rank- is suitably updated) or are stored as representative vectors of
order clustering (ROC) algorithms which determine a “binary new clusters. A key parameter called the vigilance parameter
value” for each row and column, rearrange rows and columns dictates whether the input vector is to be stored as a repre-
in descending order of their binary values and then identify sentative vector or whether it is to be treated as being part of
clusters. The ROC algorithms have certain deficiencies. For existing clusters. Different values of the vigilance parameter
example, the final, stable solution produced by the ROC generate different clusters.
algorithm is not necessarily the best one [34]. Chandrasekharan
and Rajagopalan 1131 developed a modified version of the
algorithm called MODROC which attempts to overcome some B. CMS Design Techniques
of the deficiencies of ROC. MODROC enables an “objective”
identification of machine cells. CMS design techniques are those that not only attempt
Chandrasekharan and Rajagopalan [ 141 presented a graph to find mutually separable clusters of machine cells (and
theoretic formulation of the clustering problem. They devel- corresponding part families) but also incorporate practical
oped expressions for a) the upper limit on the number of facility design constraints. Such techniques can be further
groups, and b) grouping efficiency that allows comparison of classified into
clustering algorithms. In addition, they presented a 3-stage a) cost based techniques; and
technique to determine clusters. This method was extended, b) design constraint based techniques.
improved and is presented in [ 121. Cost Based Techniques Cost based techniques explicitly
A graph theoretic approach in which the grouping problem consider the various costs listed in Section I11 while identifying
is modeled as a graph partitioning problem was presented in machine cells and corresponding part families. There are
[36]. The two-phase approach involves obtaining a good initial relatively fewer techniques in this category and are presented
solution (graph partition) by solving a transportation problem. in this section. Although some of the techniques consider
In phase 2, another algorithm is used to improve the initial design constraints also, they are discussed in this section
solution. Another network based method may be found in [67]. because their primary emphasis is on forming machine cells
Askin et al. [21 presented a 3-stage hamiltonian path ap- such that various costs are minimized.
proach for solving the clustering problem. In stage I , they One of the first models that explicitly considered costs
compute the distance matrix for parts and machines, respec- while forming machine cells was presented by Chakravarty
tively. Using each distance matrix, they suboptimally solve and Shtub [ 113. However, their model only considers work-
an associated traveling salesman problem (TSP). This is done in-process inventory cost; further, it assumes that only one
in stage 2. In stage 3, the tours of stage 2 are used to machine of each type is available.
find a suboptimal solution to the associated hamiltonian path The cost model presented [3] considers fixed machine costs,
problem. This method was shown to be superior to the ROC variable processing costs, set-up, cycle inventory, and intra-
algorithm and its extension using test problems. cellular material handling costs. The GT problem is solved

Authorized licensed use limited to: ULAKBIM UASL - GAZI UNIV. Downloaded on December 17,2023 at 13:07:23 UTC from IEEE Xplore. Restrictions apply.
 212 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. 24, NO. 2, FEBRUARY 1994

using a 3-stage approach. This approach is not suitable for work load and cell size constraints. The assignment of a part
large problems because of its complexity 111. to a part family is done so that a majority of the operations
Choobineh [15] also presented a two-stage approach in on the part are performed within the cell corresponding to the
which part families are determined in the first stage. In the part family to which the part is assigned. Thus the technique
second stage, a mathematical model is used to identify the cor- indirectly attempts to minimize the inter-cellular movement of
responding machine cells so that the equipment purchase, part parts. Its performance was compared with other well known
production, work-in-process and set-up costs are minimized. clustering techniques and in all cases produces solutions that
Machine capacity constraints are also included. Further, the were at least as good as the ones reported in the literature.
desired number of cells can be provided as input to the model. The technique is flexible as it allows the user to change two
The author suggests that this number be kept small; otherwise key parameters4ell admission factor (which controls the
the model cannot be solved easily. admission of a machine to a cell) and cell size factor. Thus a
Three mixed-integer programming models that assume number of alternative solutions can be generated for various
availability of alternate process plans are provided in [55]. values of these two parameters.
The first one provides information to construct a part- In addition to the above there are a number of other
machine processing indicator matrix which can then be solved techniques that address only one or two design constraints
using existing clustering techniques. The second forms cells listed in Section 111. Some of them are briefly reviewed here.
assuming part families are known whereas the third identifies Co and Araar [16] presented a hierarchical procedure for
part families and machine cells simultaneously. The models configuring CMS’s that involves three steps. The only design
minimize machine investment and operating costs. But as the constraint discussed in this paper is the capacity constraint. In
authors point out, the models are not suitable for industrial the first step, a mathematical model is used to determine the
problems because of the presence of integer variables. Another assignment of jobs to machines; the objective is to minimize
model that considers setup costs and setup time in addition deviation between workload assigned to machines and the
to investment costs is presented in [54]. Again, this model is available capacity. Solution of the model is used to construct a
appropriate only in environments where relatively few parts binary job-machine processing indicator matrix. In the second
are manufactured. stage, an extension of King’s ROC algorithm 1331 is used
Askin and Chiu [ l ] developed a model that considers to identify clusters of machines and corresponding jobs. In
machine depreciation, inventory, material handling and set- stage 3, a direct search algorithm is employed to determine
up costs. The cost model is first divided into two subproblems the number of cells as well as the composition of each. The
and solved suboptimally using graph partitioning techniques. objective of this algorithm is a) to minimize the number
The first subproblem determines the assignment of parts to of uncompleted jobs in each cell and b) to maximize the
machines, while the second one determines the assignment number of cells. Note that the first objective minimizes the
of machines to cells. This approach was shown to be not number of cells visited by each job and hence indirectly puts
only flexible but also successful for larger realistic prob- more machines in each cell, whereas the second objective
lems. attempts to do the opposite. Co and Araar [16] discussed
Design Constraint Based Techniques To improve the per- a simple procedure for achieving this; it involves applying
formance of the single-linkage clustering algorithm in [44], two key steps repeatedly until all machines are assigned to
Rajagopalan amd Batra [53] formulated the machine grouping cells.
problem as a graph-theoretic model and applied a partitioning Gupta and Seifoddini [22] presented a similarity coefficient
algorithm to identify clusters. First, a graph with vertices that considers part processing requirement, production volume,
representing machines is drawn. The vertices are initially routing sequence and operation time in determining the coef-
unconnected. The technique then connects the vertices { i ,j } ficient for each machine-pair. They also presented clustering
only if the similarity coefficient between machines i and j is techniques that identify machine cells and part families so that
greater than a threshold value. When this procedure is com- the design constraints mentioned in Section 111 are satisfied to
pleted, a clique consisting of several vertices, or alternately, the extent possible.
a machine cell consisting of several machines, is formed. A heuristic that attempts to minimize inter-cellular move-
Additional factors such as, imposing an upper limit on machine ment of parts is discussed in 1261. It is a two-stage algorithm in
cell size, maintaining an even and high utilization of the which a bottom-up aggregation procedure is used to minimize
machines in each cell, and determination of the number of inter-cellular movement of parts. The second step is a local
machines of a given type in each cell, are incorporated in the refinement procedure that attempts to improve the first stage
technique. solution. Nagi et al. [48] formulated the GT problem as a linear
Purcheck [50]-[52] presented a lattice theoretic method programming problem which addresses two subproblems: a)
which was previously discussed under set-theoretic methods. routing selection and b) cell formation. The algorithm in [26]
Their method maximized scheduling flexibility and minimized is used to solve the second subproblem. The solution approach
the total cost of forming machine cells while determining the iterates between the two subproblems.
machine-part grouping. As was previously mentioned, the sequence of operations
Ballakur and Steudel [4] presented a heuristic technique for is of utmost importance in determining the flow of material
identifying part family/machine cell combinations simultane- between machines and machine cells. Logendran [42] pre-
ously. The technique adds machines to a cell based on cell sented a four phase approach that was an improvement over his

Authorized licensed use limited to: ULAKBIM UASL - GAZI UNIV. Downloaded on December 17,2023 at 13:07:23 UTC from IEEE Xplore. Restrictions apply.
 HERAGU: GROUP TECHNOLOGY AND CELLULAR MANUFACTURING 213

previous algorithm in 1431 and incorporates machine utilization V. CONCLUSION

considerations. In the first phase, a certain number of key
Thus far, research on the GT and CMS design prob-
machines, one for each cell, is selected based on workload
lem has focused primarily on identifying mutually separable
consideration. The number of cells is predetermined. The
part family/machine cell combinations without considering
key machines are selected so that they are dissimilar with
practical design constraints that are frequently encountered.
respect to the parts they process. The remaining machines
However, there have been some efforts toward incorporating
are added, one at a time, to a cell that results in minimum
design constraints in the past few years. In this paper, some
total move. An improvement algorithm is used to improve the
of the more important ones were listed and discussed. In
solution in phase 2, if possible. Finally, parts are assigned
addition, some planning considerations were also discussed.
to their respective cells based on the cumulative processing
A number of papers on GT and CMS design were surveyed.
time.
A new classification scheme was also provided. Since the
Heragu [27] presented an algorithm that considers many
last survey paper was published, there have been a number
of the constraints listed in Section 111. This algorithm identi-
of algorithms that have been developed for the practical
fies a preliminary set of mutually separable machine groups
CMS design problem. It appears that research on clustering
and part families (simultaneously) in such a way that safety
algorithms (as defined in section I) has reached a saturation
and technological constraints are satisfied. Then it ascertains
point and further efforts must be focused on using existing
whether the number of cells formed is less than the upper
efficient clustering algorithms to solve CMS design problem
limit specified by the user. If this constraint is violated, it
that incorporate important design constraints. We feel that such
attempts to group two cells at a time so that cells having the
techniques will find more use in practice.
greatest interaction are combined into one. This is repeated
until the number of cells do not exceed the upper bound.
Next, the algorithm determines if the cells size constraint
has been violated for any of the cells formed. If so, it
attempts to identify bottleneck machines. If a bottleneck REFERENCES
machine (which costs less than a threshold value set by the R. G. Askin and K. S. Chiu, “A graph partitioning procedure for machine
user) cannot be identified, the algorithm determines if there assignment and cell formation in group technology,” Int. J. Production
are two bottleneck machines. If bottleneck machines cannot Res., vol. 28, no. 8, pp. 1555-1572, 1990.
be identified, it forms smaller cells of the required size by R. G. Askin, S. H. Creswell, J. B. Goldberg and A. J. Vakharia “A
Hamiltonian path approach to reordering the par-machine matrix for
retaining the required number of machines and placing the cellular manufacturing,” Int. J. Production Res., vol. 29, pp. 1081-1 100,
remaining in a temporary cell. However, the machines retained 1991.
are the ones that together have the maximum flow between R. G. Askin and S. P. Subramanian, “A cost based heuristic for group
technology configuration,” Int. J. Production Res., vol. 25, pp. 101-1 13,
them. The above procedure is repeated for machines in the 1987.
temporary cell also. In a final step, the algorithm determines A. Ballakur and H. J. Steudel, “A within-cell utilization based heuristic
if any part visits multiple cells. For such parts, it further for designing cellular manufacturing systems,” Int. J. Production Res.,
vol. 25, no. 5 , pp. 639-665, 1987.
determines if the processing is done on only one machine in
D. Ben-Arieh and E. Triantaphyllou, “Quantifying data for group
any of the cells. If so, it duplicates that machine in another technology with weighted fuzzy features,” Int. J. Production Res., vol.
cell visited by the part. Heragu and Gupta [28] provide 30, no. 6, pp. 1285-1299, 1992.
a modified version of this algorithm that also incorporates J. T. Black, “Cellular manufacturing systems reduce setup time, make
small lot production economical,” Indust. Eng., Nov., pp. 3 6 4 8 , 1983.
capacity considerations. W. J. Boe and C. H. Cheng, “A close neighbor algorithm for designing
A three stage approach aimed at identifying not only ma- cellular manufacturing systems,’’ Int. J. Production Res., vol. 29, no. 10,
chine cells (and corresponding part families) but also deter- pp. 2097-2116, 1991.
J. L. Burbidge, “Change to group technology: Process organization is
mining a layout of machines within each cell and the cells obsolete,” Int. J. Production Res., vol. 30, no. 5 , pp. 1209-1219, 1992.
themselves is presented in 1291. Using real world data, Heragu J. L. Burbidge, “Production flow analysis,” The Production Engineer,
and Kakuturi [29] demonstrate that the three-stage approach vol. 42, no. 12, pp. 742, 1963.
which considers practical design constraints can be used for A. S. Carrie, “Numerical taxonomy applied to group technology and
plant layout,” Int. J. Production Res., vol. 11, no. 4, pp. 399416, 1973.
industrial prolems. For performing the machine grouping, the A. K. Chakravarty and A. Shtub, “An integrated layout for group
technique in [28] is modified significantly and used. For technology with in-process inventory cost,” Int. J. Production Res., vol.
determining a layout of cells and machines within each cell, 22, pp. 4 3 1 4 4 2 , 1984.
simulated annealing based algorithms are used. The approach M. P. Chandrasekharan and R. Rajagopalan, “ZODIAC: An algorithm
for concurrent formation of part-families and machine-cells,” Int. J.
allows the user to change key data items and thereby perform Production Res., vol. 25, no. 6, pp. 835-850, 1987.
some sensitivity analysis. The algorithm’s solution quality and M. P. Chandrasekharan and R. Rajagopalan, “MODROC: An extension
performance are demonstrated using two grouping measures. of rank order clustering for group technology,” Int. J. Production Res.,
vol. 24, no. 5 , pp. 1221-1233, 1986.
It is shown to produce good quality solution without requiring M. P. Chandrasekharan and R. Rajagopalan, “An ideal seed non-
excessive computation time. hierarchical clustering algorithm for cellular manufacturing,’’ Int. J.
In addition to the above, there are other methods that Production Res., vol. 24, no. 2, pp. 451464, 1986.
explicitly consider some of the design constraints listed in F. Choobineh, “A framework for the design of cellular manufacturing
systems,” Int. J. Production Res., vol. 26, pp. 1161-1172. 1988.
section 111. Examples are Wei and Gaither 1691, Vakharia and H. C. Co and A. Araar, “Configuring cellular manufacturing systems,”
Wemmerlov [66], Sule [62]. Int. J. Produciton Res., vol. 26, no. 9, pp. 1511-1522, 1980.

Authorized licensed use limited to: ULAKBIM UASL - GAZI UNIV. Downloaded on December 17,2023 at 13:07:23 UTC from IEEE Xplore. Restrictions apply.
 214 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. 24, NO. 2, FEBRUARY 1994

1171 K. R. Currie, “An intelligent grouping algorithm for cellular manufac- [45] W. T. McCormick, P. J . Schweitzer, and T. W. White, “Problem
turing,” presented at 14th Ann. Computers and Industrial Engineering decomposition and data reorganization by a clustering technique,” Op.
Conf., Orlando, FL, March 9-11, 1992. Res., vol. 20, pp. 992-1009, 1972.
[18] J. de Witte, “The use of similarity coefficients in production flow [46] S. P. Mitrofanov, Group Technology in Industry, vols. 1 and 2,
analysis,” Int. J. Production Res., vol. 18, no. 4, pp. 503-514, 1980. Leningrad, USSR: Mashinostroienie, (in Russian) 1983.
[19] Z. Faber and M. W. Carter, “A new graph theoretic approach for [47] J. S. Morris and R. J. Tersine, “A simulation analysis of factors
forming machine cells in cellular production systems,” in Flexible influencing the attractiveness of group technology and cellular layouts,”
Manufacturing Systems: Methods and Studies, A. Kusiak, ed. New Management Science, vol. 36, no. 12, pp. 1567-1578, 1990.
York: North Holland, pp. 301-318, 1986. [48] R. Nagi, G. Harhalakis and J. M. Proth, “Multiple routeings and capacity
[20] B. B. Flynn and F. R. Jacobs, “A simulation comparison of group considerations in group technology applications,” Int. J. Production Res.,
technology with traditional jobshop manufacturing,” Int. J. Production vol. 28, no. 12, pp. 2243-2257, 1990.
Res., vol. 24, pp. 1171-1192, 1986. [49] 0. F. Offodile, “Application of similarity coefficient method to parts
[21] T. Gongaware and I. Ham, “Cluster analysis applications for group coding and classification analysis in group technology.” J. Manufactur-
technology manufacturing systems,” in Proc. Ninth Amer. Metal Working ing Systs., vol. 10, no. 6 , pp. 442L448, 1991.
Res. Coni, Society of Manufacturing Engineers, 198 I , [50] G. F. K. Purcheck, “Combinatorial grouping-a lattice-theoretic method
[22] T. Gupta and H. Seifoddini, “Production data based similarity coefficient for the design of manufacturing systems,” J. Cybern., vol. 4, no. 3, pp.
in the design of a cellular manufacturing system,” Int. J. Production 27-60, 1974.
Res., vol. 28, pp. 1247-1269, 1990. [Sl] G. F. K. Purcheck, “A mathematical classification as a basis for the
[23] I. Ham, “Introduction to Group Technology,” SME Tech. Paper MMR design of group technology -_production cells,” Production EnRineer, vol.
~

76-093, Society of Manufacturing Engineers, Dearborn, MI, 1976. 54, pp. 35-48: 1975.
1521. G. F. K. Purcheck, “A linear ~-
. programming method for the combinatoric
[24] I. Ham, K. Hitomi and T. Yoshida, Group Technology. Boston, MA:
grouping of an incomplete power set,” J . Cybern., vol. 5 , no. 4, pp.
Kluwer-Nijhoff, 1985.
51-76, 1975.
[25] C. Han and I. Ham, “Multiobjective cluster analysis for part family [53] R. Rajagopalan and J. L. Batra, “Design of cellular production systems-
formations,” J. Manufacturing Systs., vol. 5 , no. 4, pp. 223-230, 1986. A graph theoretic approach,” Int. J . Production Res., vol. 13, no. 6, pp.
[26] G. Harhalakis, R. Nagi and J. M. Proth, “An efficient heuristic in 567-579, 1975.
manufacturing cell formation for group technology applications,” Int. [54] D. Rajamani, N. Singh and Y. P. Aneja, “A model for cell formation in
J. Production Res., vol. 28, no. 1, pp. 185-198, 1990. manufacturing systems with sequence dependence,” /nr. J. Production
[27] S. S. Heragu, “A heuristic algorithm for identifying machine cells,” Res., vol. 30, no. 6, pp. 1227-1235, 1992.
Information and Decision Technologies, vol. 18, pp. 17 1-184, 1992. 1551 D. Rajamani, N. Singh and Y. P. Aneja, “Integrated design of cellular
[28] S. S. Heragu and Y. P. Gupta, “A heuristic for designing cellular manufacturing systems in the presence of alternative process plans,” Int.
manufacturing facilities,” Int. J. Production Res., vol. 32, no. 1, pp. J. Production Res., vol. 28, no. 8, pp. 1541-1554, 1990.
125-140, 1994. [56] H. Seifoddini, “A probabilistic model for machine cell formation,” J.
[29] S. S. Heragu and S. R. Kakuturi, “Grouping and placement of machine Manufacturing Systs., vol. 9, no. 1, pp. 69-75, 1990.
cells,” DSES Tech. Rep. #37-92-341, Rensselaer Polytechnic Inst., [57] H. Seifoddini and P. M. Wolfe, “Application of the similarity coefficient
Troy, NY 12180, 1992. method in group technology,” IIE Transactions, pp. 271-277, 1987.
1581 S. M. Shafer, G. M. Kern and J. C. Wei, “A mathematical programming
[30] M. hi, “On the synthesis of loop and cutset matrices and related approach for dealing with exceptional elements,” Int. J. Production Res.,
problems,” in RAAG Memoirs, K. Kondo, ed., Research Association
vol. 30, no. 5 , pp. 1029-1036, 1992.
of Applied Geometry, Japan, 4 (A-XII), pp. 376410, 1968.
[59] P. H. A. Sneath and R. R. Sokal, Numerical Taxonomy. San Francisco,
[31] S. Kaparthi and N. C. Suresh, “Machine-component cell formation in CA: Freeman and Co., 1973.
group technology: A neural network approach,” Int. J. Production Res., [60] G. Srinivasan and T. T. Narendran, “GRAFICS-A nonhierarchical
vol. 30, no. 6, pp. 1353-1367, 1990. clustering algorithm for group Technology,” Int. J. Production Res., vol.
[32] G. M. Kern and J. C. Wei, “The cost of eliminating exceptional elements 29, no. 3, pp. 4 6 3 4 7 8 , 1991.
in group technology cell formation,” Int. J. Production Res., vol. 29, no. [61] G. Srinivasan, T. T. Narendran and B. Mahadevan, “An assignment
8, pp. 1535-1547, 1980. model for the part-families problem in group technology,” Int. J.
[33] J. R. King, “Machine-component grouping in production flow analysis: Production Res., vol. 28, no. 2, pp. 145-152, 1990.
An approach using a rank order clustering algorithm,” Int. J. Production [62] D. R. Sule, “Machine capacity planning in group technology,” Int. J.
Res., vol. 18, pp, 213-219, 1980. Production Res., vol. 29, no. 9, pp. 1909-1922, 1991.
[34] J. R. King and V. Nakomchai, “Machine-component group formation in [63] N. C. Suresh, “Partitioning work centers for group technology: An-
group technology: Review and extension,” Int. J. Production Res., vol. alytical extension and shop-level simulation investigation,” Decision
20, no. 2, pp. 117-133, 1982. Sciences, vol. 23, pp. 267-290, 1992.
[35] R. B. Kini, L. R. Taube, and C. T. Mosier, “Part identification and group 1641 N. C. Suresh, “Partitioning work centers for group technology: Insights
technology: A new approach,” J. Manufacturing Systs., vol. 10, no. 2, from an analytical model,” Decision Sciences, vol. 22, pp. 772-791,
pp. 136145, 1991. 1991.
[65] A. J. Vakharia, “Methods of cell formation in group technology: A
1361 K. R. Kumar, A. Kusiak and A. Vannelli, “Grouping of parts and framework for evaluation,” J. Op. Management, vol. 6, no. 3, pp.
components in flexible manufacturing systems,” European J. Op. Res., 257-271, 1986.
vol. 24, pp. 387-397, 1986. [66] A. J. Vakharia and U. Wemmerlov, “Designing a cellular manufacturing
(371 A. Kusiak, “Branching algorithms for solving the group technology system: A materials flow approach based on operation sequence,” //E
problem,” J. Manufacturing Systs., vol. 10, no. 4, pp. 332-343, 1991. Transactions, vol. 22, no. 1, pp. 84-97, 1990.
[38] A. Kusiak and Y. K. Chung, “GT/ART: Using neural networks to form [67] T. Vohra, D. Chen, J. C. Chang and H. C. Chen, “A Network Approach
machine cells,” Manufacturing Review, vol. 4, no. 4, pp. 293-301, 1991. to Cell Formation in Cellular Manufacturing,” Int. J. Production Res.,
[39] A. Kusiak and W. S. Chow, “Decomposition of manufacturing systems,” vol. 28, no. 1 1 , pp. 2075-2084, 1990.
IEEE Trans. Robotics andAutomation, vol. 4, no. 5 , pp. 457471, 1988. 1681 J. C. Wei, “A note on ‘Change to group technology,’ Int. J. Production
”

[40] A. Kusiak and W. S. Chow, “An efficient cluster identification al- Res., vol. 30, no. 5, pp. 1221-1222, 1992.
gorithm,” IEEE Trans. Syst., Man, Cybem., vol SMC-17, no. 4, pp. [69] J. C. Wei and N. Gaither, “A capacity constrained multiobjective cell
696699, 1987. formation method,” J. Manufacturing Systs., vol. 9, no. 3, pp. 222-232,
1411 A. Kusiak and S. S. Heragu, “Group technology,” Computers in Industry, 1990.
vol. 9, no. 2, pp. 83-91, 1987. 1701 U. Wemmerlov and N. L. Hyer, “Cellular manufacturing in the U. S.
[42] R. Logendran, “Impact of sequence of operations and layout of cells industry: A survey of users,” Int. J. Production Res., vol. 27, no. 9, pp.
in cellular manufacturing,” Inr. J. Production Res., vol. 29, no. 2, pp. 1511-1530, 1989.
375-390, 1991. 1711 U. Wemmerlov and N. L. Hyer, “Procedures for the part-family/
[43] R. Logendran, “A workload based model for minimizing total intercell machine group identification problem in cellular manufacturing,” J. Op.
and intracell moves in cellular manufacturing,” 1nt. J. Production Res., Management, vol. 6, no. 2, pp. 125-147, 1986.
vol. 28, no. 5 , pp. 913-925, 1990. [72] C. Zhang and H. Wang, “Concurrent formation of part families and
[441 I. McAuley, “Machine grouping for efficient production,” The Produc- machine cells based on the fuzzy set theory,” J. Manufacturing Systs.,,
tion Engineer, vol. 51, no. 2, pp. 53-57, 1972. vol. 11, no. 1, pp. 61-67, 1992.

Authorized licensed use limited to: ULAKBIM UASL - GAZI UNIV. Downloaded on December 17,2023 at 13:07:23 UTC from IEEE Xplore. Restrictions apply.
 HERAGU: GROUP TECHNOLOGY AND CELLULAR MANUFACTURING 215

Sunderesh Heragu received the B.Eng. degree

in mechanical engineering from the University of
Mysore, India, in 1982; the M.B.A. degree from the
University of Saskatchewan, Canada, in 1985; and
the Ph.D. degree in industrial engineering from the
University of Manitoba, Canada, in 1988.
He is an Assistant Professor in the Department
of Decision Sciences and Engineering Systems at
Rensselaer Polytechnic Institute. Previously he was
Assistant Professor at the State University of New
York, Plattsburgh, and Adjunct Professor at the
University of Manitoba, Winnipeg, Canada. His current research interests are
in the integration of design and planning activities in manufacturing systems.
His previous research has focused on the application of knowledge-based
systems in automated manufacturing systems, cellular manufacturing, facility
location, layout, and network design. He has authored or coauthored over
fifty articles and technical reports, many of which have appeared in journals,
magazines, books, and conferences. He has edited a special issue of European
Joumal of Operational Research on facility layout and currently serves on
the editorial board of Intemational Joumal of Industrial Engineering. He is a
senior member of IIE and SME, and member of ORSA, POMS, and TIMS.

Authorized licensed use limited to: ULAKBIM UASL - GAZI UNIV. Downloaded on December 17,2023 at 13:07:23 UTC from IEEE Xplore. Restrictions apply.