Sunteți pe pagina 1din 39

Chapter II

Literature Survey on Cellular Manufacturing System

2.0 Preamble

The present thesis focuses on the development of visual clustering of machine-part cell
formation using artificial intelligence based self organizing map approach. The motivation
behind taking up the said topic is elaborated in Chapter I. In the present Chapter, a detailed
discussion on the cell formation approaches adopted in the literature would be presented. The
complexity of the cell formation problem is highlighted based on worked done in literature as it
is a criteria of applying specific cell formation approach. The present chapter proposes a
classification of the cell formation methods as it allows positioning the artificial neural network
among other methods and is followed by observation of the neural network applied to the cell
formation problem. This chapter completes with the necessity to apply neural network based
method to solve binary as well as non binary cell formation problem. A critical evaluation of the
neural network methods applied to the cell formation problem is presented to position this work.
In the preceding sections an extensive survey on the various above aspects would be presented
along with an exhaustive review of the plethora of literatures in these directions.

2.1 A Historical Context of Cellular Manufacturing

Group Technology (GT) has its underpinnings in the massive industrial production
environment of the former Soviet Union of the 1940s and 1950s in response to
alternatives sought for reducing inefficiencies and difficulties arising from the management
of large job shops (Petrov, 1968). Flanders (1925) was first to propose Group technology. In
the 1940’s S. R. Mitrofanov, A. P. Sokolovskii, and other leading Russian engineers made
inroads into job simplification and reducing set-up times. They believed that there was
commonality among parts, and from this commonality, basic part families could be created. In
19
the year 1966, the English translation of Mitrofanov’s book “The Scientific Principles of Group
Technology” in Russia (Mitrofanov, 1958) has paved the path to find the increasing
recognition, by both practitioners and researchers, as a philosophy that has broad
applicability in the manufacturing environment. The first well-documented Group Technology
operation in the free world was in France at Forges et Ateliers de Constructions Electrgnes de
Jenment (Phillips, 1978). In Germany, the concepts of Group Technology were advanced in the
early 1960’s by Professor H. Optiz at the Aachen Technical University. He developed the VDW
part classification and coding system which is widely accepted and led to the development of
many other part classification and coding systems (Opitz,1970; Phillips,1977). Of all industrial
countries, Great Britain is probably the most advanced in utilizing Group Technology concepts.
The government of England has supported a Group Technology Center since 1968 and the
British Institution of Production Engineers has formed a Group Technology Section. One of the
major proponents of GT was Burbidge (1975) who supported the relative performance
supremacy of CM over FL systems. His pioneering work on Production Flow Analysis
(PFA) involved manipulation of the part-machine matrix to form cells. Burbidge developed
methods suitable for hand computation in 1971. Burbidge, Gallagher and Knight, and others
propagated the technology (Burbidge 1963, 1971, 1975, and 1977; Gallagher and Knight 1973).
Manufacturing systems are complex because of the many elements and subsystems of a
manufacturing operation and their interactions (Zhang, 2011). Cellular manufacturing (CM) is a
production system in which similar parts are classified into part families and dissimilar machines
are assigned into machine cells in order to exploit the cost-effectiveness of mass production and
flexibility ofjob shop manufacturing.

To date the problem has been developed in several different mathematical forms, but the
optimisation of the cell formation problem, in general, is NP-complete (Cheng, et al. 1996),
which means that the problem is hard to solve for a computer. NP-complete (Garey and Johnson
1979) is a terminology in algorithm analysis (Sedgewick and Flajolet 1996), which studies the
efficiency of an algorithm. A brief overview on background of computation complexity and the
related issues involved in cell formation problem have is discussed below.

20
2.2 Complexity Issues in Cell formation Problem

Computational complexity is measured by the quantity of computational resources


(e.g., time, storage, program, communication) which is required for solving a particular
task (ElMaraghy et al. 2012). Computational complexity can be divided into two classes: P
problems and NP problems.

The class of P problems contains those that can be solved in polynomial time. This
complexity becomes problematic, when the number of elements (N) grows, because the
same algorithm that was able to solve a problem for a smaller N cannot solve one for a
larger NO in a reasonable time (or with using reasonable memory). For example, assume
every element in a system has a direct relationship with all other elements.

The computational complexity in terms of the number of relationships in this case


is 0(N(N-1)) = 0(N2), which is called polynomial complexity (0(Na), in which a is a
positive constant). Practically, it is well known that computational complexity grows very
quickly, when the complexity is 0(aN) (exponential), 0(N!) (factorial), or O (Nn) (double
exponential). These are called Non-deterministic Polynomial (NP) complexity as opposed to
polynomial cases (P).

The problems in the NP class are those that could be solved in polynomial time by an
algorithm that is not known so far, however the fact that they cannot be solved within
polynomial time has not been proved.

Polynomial time transfers can convert some of the problems in the NP class to another
form in the NP class. Thus, they are equally complex. If one of them can be solved in polynomial
time, they all can. These problems are known as NP-complete. A problem is called NP-hard if it
is at least as hard as an NP-complete problem. The details refer to the book by Garey and
Johnson (1979). The term "NP-complete" will be frequently used in this thesis, because most of
the decision problems involved in the current research are either NP-complete or NP-hard.

21 13289
Some NP complexity classes (NP-complete or NP-hard) are known to be difficult
or even impossible to manage and solve. In manufacturing, NP-complete problems can be
found, for instance, in production planning problems and logistics problems.

In most of these works, numerous assumptions have been made. The most recurrent and
important one is that cell size is specified a priori to maintain mathematical traetability. The
number of cells in the system is user-specified in order to reduce the computational difficulty of
assigning m machines to k non-empty cells. If the number of cells is not pre-specified the
problem is understood to be NP-complete (Garey and Johnson, 1979). However, when the
number of cells is pre-specified, Venugopal and Narendran (1992) have shown that the problem
of machine allocation to cells becomes simpler. The proposed model belongs to the class of NP-
complete problems as it includes the machine-part grouping problem. It is widely accepted in the
literature that the whole problem of designing a CMS, taking into account the numerous phases
and criteria involved, belongs to the class of NP-complete problems (King and Nakomchai,
1982; Ballakur, 1985). The consideration of cellular reconfiguration for a planning horizon is in
itself a problem that belongs to the class of NP-complete problems (Chen, 1998). He has
developed a model that considers system reconfiguration in terms of machine relocation,
showing that solving the model is NP-complete. Furthermore, the additional features of the
proposed model increase its complexity and combinatorial nature. Logendran et al. (1994) have
shown that the problem involved is NP-hard since it integrates the problem of cell formation,
system reconfiguration along with the consideration of multiple process routings, production
planning, machine capacities and availabilities.

Design of cellular manufacturing systems (CMSs) is a complex, multi-criteria and


multistep process. Ballakur and Steudel (1987) showed that this problem, even under
restrictive conditions, is NP-complete. Because cell formation problems are NP-hard, it is
difficult to obtain solutions that satisfy all constraints (Nair and Narendran, 1999). For that
reason, it is desirable to apply efficient computing techniques. Therefore, various parameters like
production factors, single or multi-objective criteria, implementation issues of cell formation
problem etc. and constraints like problem size which influence the selection of cell formation
approaches to be applied for efficient solution are studied in detail for a critical appraisal of those

22
approaches. A classification scheme is structured here to enable a holistic conceptual and
research methodological analysis of the cell formation solution.

2.3 Classification of Cell Formation Approaches

In the quest of efficient cell formation solution approaches a large number of literatures
have been identified during the last three decades. Based on these reviews the existing cell
formation approaches are classified in two categories (Joines et al. 1996):

• design oriented cell formation approaches


An overview of this type of approach have discussed by Askin and Vakharia (1990). It is
based on similarity in design features and includes the traditional and flow analysis methods.
In GT, the traditional methods are the applications of visual methods, rule of thumb,
Classification and Coding Systems of parts and machines. The Flow analyses are of two
types Production Flow and Component Flow analysis.

23
Cell Formation Approaches

' I
Design Oriented Production Oriented

X X
Model Cluster Analysis Algorithm Artificial
Traditional Methods Flow Analysis Intelligence
Methods
4 Graph Theory
Heuristic
Visual Classification Rule of
Mathematical
and Coding Thumb
Metaheuristic
Hierarchical .......
* Mono Code

Component Production Non- Trajectoiy Population


Poly Code Flow Analysis Flow Analysis Hierarchical
Ant Colony
Mired Array based Optimization

DCA
Linear Programming
1 Evolutionary
Simulated Tabu
ROC Algorithm
Annealing Search
Integer Programming

BEA
Fuzzy
Quadratic Programming Scatter Search
MODROC
Dynamic Programming Genetic Algorithm Neural Network
OV
Branch & Bound Particle Swarm Optimization

Legend:
DCA: Direct Cluster analysis MODROC: Modified ROC
BEA: Bond Energy Algorithm OV: Occupancy Value
ROC: Rank Order Clustering MST: Minimal spanning Tree

Figure 2.1. A taxonomic framework of cell formation approaches

• production oriented approaches


This is based on the basis of similar processing requirements consisting of Model based,
Cluster analysis based, Algorithm based and Artificial Intelligence based methods.
The above classification framework is shown in the Figure 2.1. The discussion on the
components of the framework is followed in the next sections.

24
2.3.1 Visual Methods

Eye-balling method that examines the data and performs clustering using the human eye
is one of the simplest approaches to the use of visual method in Group Technology. However it is
easier to perform but its success is highly dependent on human experience, knowledge and
preference. Due to the human limitation the maximum number of parts is limited. Burbidge
(1971) report of solution of cell formation problem with 2000 parts is overoptimistic and even it
is difficult to identify the part family with more than hundred parts (Carrie, 1973).

2.3.2 Classification and Coding

Parts Classification and Coding is concerned with identifying the similarities among parts
and relating these similarities to a coding system. A coding system is a series of alphanumeric
characters. Based on the presence or absence of similarities amongst the parts the Classification
is used to divide the parts into part families. In coding methodologies the design characteristic of
the parts has an important role in the formation of part families. These methodologies use a
coding system to assign numerical weights to part characteristics and identify families using
some classification scheme. Classification and Coding-based systems are traditionally design
oriented or shape-based; therefore they are ideal for component variety reduction. According to
Opitz (1970) “Classification and coding techniques have been used in practice but using features,
based on shape and then group parts accordingly is a very labor intensive job and in order to
solve this problem die idea of weighted codes is useful”. There are three types of coding systems
namely monocode, polycode and mixed type. In monocode, each number is qualified by the
preceding character and can therefore represent a large amount of information with very few
code positions. Polycode, also called chain code, represents every single digit despite the
precedent and is considered compact and easy to construct. The mixed or hybrid type is a
combination of the other two coding systems and encompasses the merits of both the systems.

2.3.3 Flow Analysis

2.3.3.1 Production Flow Analysis (PFA)

Burbidge (1969,1971) has introduced the concept of PFA which is the most widely used
job shop environment. It is a well-established methodology used for transforming a traditional
25
functional layout into a product-oriented layout The PFA concept was developed and detailed by
its inventor during late 70s and early 80s and is well documented (Burbidge, 1989). The cases
documented in detail report significant improvements in the general control of the previously
complex material flow system, reduced lead times and improved punctuality (Wemmerlov and
Johnson, 1997, 2000; Nahm et al. 2006). The PFA method uses part routings to find natural
clusters of work-stations forming production cells able to complete parts and components swiftly
with simple material flow, and simpler manufacturing systems have been seen to be more
efficient than more complex ones (Sarkis, 1997). The same approach can be used to analyze
production units as well as the line layouts within a cell. Once implemented, the scheduling
system is based on period batch control aiming to establish fixed planning, production and
delivery cycles for the whole production unit.

2.3.3.2 Component Flow Analysis

El-Essawy et al. (1972) developed this method to find groups and families which
considers both combinations and sequences for the best arrangement of plant layout. The
criticisms for this method are that the method is based more on intuition and subjective
judgments and also there is no explicitly easy determination of cell formation.

2.3.4 Model Oriented

2.3.4.1 Graph Theoretic

The relationships among the machines or the components to be manufactured can be


represented by a weighted graph where a node represents a machine or component and an arc
represents association between the nodes (Amirahmadi and Choobineh 1992). The first to
implement the graph partitioning approach to the cell formation problem were Rajagopalan and
Batra (1975) using combination of nodes and arcs. One thing that is considered as the major
weakness of graph partitioning approach is its non-simultaneous grouping of parts into families
and machines into respective cells. Also, being an analytical approach it cannot handle large size
of problems especially those which have ill-structured Machine-Part incidence matrices. Due to
these limitations graph partitioning approaches are not very popular and limited in use.

26
Applications in CMS:

Selim (2002) has applied a graph partitioning approach to demonstrate by comparing it to


two Machine-Part Cell Formation (MCFP) published solution methods using several problems
from the literature.

Manimaran et al. (2010) applied graph decomposition algorithm to the cell formation
problem with the objective of minimizing the exceptional elements using graph decomposition of
a bipartite graph into near complete bipartite graph. Here, the machines and parts are considered
as vertices and the links between machines and parts are edges. This approach is found to offer
reliable and better solutions when compared to earlier approaches available in the literature.

2.3.4.2 Mathematical

Approaches based on Mathematical Programming actually solve the Machine-Part


grouping problem as an optimization problem. Due to their ability to consider and incorporate a
number of critical system design information, Mathematical Programming based approaches
have been extensively used to solve the Machine-Part grouping problem. All the mathematical
optimization approaches applied to the cell formation are either linear or nonlinear integer
programming problems. In their work, Kusiak (1987 & 1988), and Boctor (1991), have shown
that these approaches have the ability to incorporate a lot of production related data, for example;
processing sequence, routing flexibility, setup and processing time etc. Being an optimization
technique, the objective; while clustering parts or machines; could be to maximize the total sum
of similarities between each pair of machines/ parts.

Though the mathematical optimization approaches are very attractive in terms of


incorporating more production level data into cell formation models but to find analytical
solution of medium and large size cell formation problems in polynomial time is not possible
which is a major limitation of mathematical optimization based techniques. “Obtaining optimal
solution for the mathematical programming approaches can be infeasible due to combinatorial
complexity of the CMS design problem” (Mungwattana, 2000).

27
Applications in CMS:

Won and Lee (2004) proposed a modified p-median approach for efficient GT cell
formation with the objective of maximizing the sum of similarities between machines in the
same cell.

Solimanpur & Jafari (2008) has proposed an algorithm based on branch-and-bound


approach to obtain the optimal cell formation solution and concluded that the proposed approach
performed inefficient for large-sized problems

2.3.5 Cluster Analysis

Cluster analysis is composed of various techniques to identify structure in a complex data


set. The fundamental objective of this cell formation approach is to group either objects or
entities or attributes into clusters so that individual elements within a cluster have a high degree
of “natural” association among themselves and very little “natural association” between clusters.
Clustering techniques are broadly three types: array-based clustering, hierarchical clustering and
non-hierarchical clustering techniques.

2.3.5.1 Array based Clustering

Array based clustering is one of the simplest as compared to other production oriented
techniques used to form machine cells and corresponding part families, simultaneously. The
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 box
are generated. This approach is easier to understand, but the results vary from algorithm to
algorithm.

Applications in CMS:

An overview on array based methods is presented by several authors (Chu, 1989; Shafer,
1998; Singh, 1993). There are at least eight array-based clustering algorithms as identified after
literature survey:

28
• Bond Energy Analysis (BEA) (McCormick, 1972)
• Rank Order Clustering (ROC) (King, 1980a)
• Improved ROC (ROC2) (King and Nakomchai, 1982),
• Modified Rank Order Clustering (MODROC) (Chandrasekharan and Rajagopalan, 1986)
• Direct Clustering Analysis (DCA) (Chan and Milner, 1982)
• Occupancy Value method (Khator and Irani, 1987)
• Cluster Identification method (Kusiak and Chow, 1988)
• Hamiltonian Path Heuristic (Askin et al. 1991).

2.3.S.2 Hierarchical Clustering

Hierarchical clustering in cell formation consists of two steps: initially, the set of
machines are partitioned into few large cells and then, each of these cells is separated into
smaller groups. The procedure is applied until the final group cannot be subdivided. Hierarchical
techniques may be categorized Into two methods: (a) Agglomerative methods where the process
starts with singleton clusters and proceeds to merge them into larger partitions until a partition
containing the whole set is obtained, and, (b) Divisive methods where the process starts with all
the data (machines or parts) in a single group and a series of partitions is created until each
machine (part) is in a singleton cluster. Hierarchical classifications are generally represented by
dendrograms or inverted tree structures. In the context of cell formation, only agglomerative
clustering techniques have been used. Most of the methods based on a similarity measure are
classified in this category. Hierarchical clustering methods can be implemented easily but these
methods tend to be very demanding of computational resources. They have the flexibility to
incorporate manufacturing data other than the binary machine/part incidence matrix. This is an
advantage compared to array-based clustering. The duplication of bottleneck machines is not
treated by most algorithms.

Applications in CMS:

Important work utilizing agglomerative methods can be found in Gupta and Seifoddini
(1990) and Gupta (1993). Also Vakharia and Wemmerlov (1990) proposed a methodology for the
CF problem, this time based on the identification of part families rather than machine cells.

29
2.3.53 Non-Hierarchical Clustering

Non-hierarchical clustering methods are iterative methods but they also employ a
measure of similarity or dissimilarity for grouping parts or machines. They begin with either an
initial partition of the data set or the choice of a few seed points. In either case, one has to decide
the number of clusters in advance. These techniques produce a single partition of the data.
Contrary to the hierarchical clustering, the non-hierarchical clustering does not need to compute
and store a similarity or distance matrix.

Non-hierarchical method can be divided in three main categories:

• Single-pass methods (such as Yu and Raghavan, 1977) produce clusters that are dependent
upon the order in which the machines or parts are treated.

• Relocation methods (such as k-means, Lloyd (1982) assign machines or parts to a defined
number of seed clusters and then iteratively reassign machines or parts to optimise the
result. Such methods are prone to reaching local optima rather than a global optimum, and
it is generally not possible to determine when or whether the global optimum solution has
been reached.

• Nearest neighbour methods (such as the Jarvis-Patrick method, (Jarvis and Patrick, 1973))
assign machines or parts to the same cluster as their nearest neighbours. A list with nearest
neighbours and the level of similarity between them needs to be defined.

Applications in CMS:

The most recent non-hierarchical procedures have been proposed by Nair and Narendran
(1998, 1999). Chandrasekharan and Rajagopalan proposed non-hierarchical technique ISNC
(Ideal Seed Non-hierarchical Clustering) based on the evaluation criterion “grouping efficiency”
(Chandrasekharan and Rajagopalan, 1986) which measures inter-cellular movement and within-
cell machine use. The same authors developed in 1987 a new algorithm: ZODIAC (Zero-One
Data-Ideal seed Algorithm for Clustering) which is a much more improved and expanded version
of ISNC (Chandrasekharan and Rajagopalan, 1987).

30
Srinivasan and Narendran developed the GRAFICS method that generates initial seeds
from an assignment problem, which maximizes the similarity between machines (Srinivasan and
Narendran, 1991). These techniques generates not proper clustering criterion for ill-structured
problems.

2.3.S.4 Similarity Coefficient

The approaches with similarity coefficient require a measure of similarity between each
pair of machines, parts, tools and design features. Fusions are based on these similarities.

Applications in CMS:

In the literature, a large diversity of similarity coefficients is defined between machines


(Logendran, 1990; Xu and Wang, 1989), parts (Burbidge, 1989; Vakharia and Selim, 1995), or
sequence of operations (Kusiak and Cheng, 1990; Shafer and Rogers, 1991). John et al. (2011)
has developed manufacturing cell successfully using weighted similarity coefficient method.

Yin et al. (2011a) compared the performance of twenty well-known similarity coefficients
based on two characteristics: discriminability and stability under different data conditions for
more than two hundred numerical cell formation problems. From the results, three similarity
coefficients are found to be more discriminable; Jaecard is found to be the most stable similarity
coefficient. Four similarity coefficients are not recommendable due to their poor performances.

Yin et al. (2011b) developed a similarity coefficient that incorporates alternative process
routing, operation sequence, and operation time and production volume factors. The method used
in stage 1 to generate machine cell and in stage 2 they refined the machine cell. The results show
that the approach is reliable and efficient for solving CFPs.

Unler and Gungor (2009) have effectively applied the K-harmonic means clustering
technique to form machine cells and part families simultaneously to a set of 20 test problems
with various sizes. It is found that the performance is improved in majority of the literature
problems.

31
2.3.6 Algorithm Oriented Methods

2.3.6.1 Heuristic Methods

"Heuristics stand for strategies using readily accessible though loosely applicable
information to control problem-solving processes in human beings and machines." (Pearl, 1984).
Heuristics, in more popular understanding, are rules of thumb, educated guesses, intuitive
judgements or simply common sense ideas to solve a particular problem. Heuristics are designed
to provide an alternative framework for solving a problem in contrast with a set of restricted
rules-constraints that cannot vary. Although heuristic approaches do not guarantee to provide
optimal solutions (usually sub-optimal results are derived) they are very useful in producing an
acceptable solution in reasonable time. A heuristic improves the efficiency of a search process,
possibly by sacrificing claims of completeness (Rich and Knight 1991). The heuristics in the
current section are restricted to those of rule-based algorithms, in which specific rules are used to
solve cell formation problems.

Applications in CMS:

Mukattash et al. (2002) proposed three heuristic procedures. Given a CF solution, the
heuristics were designed to assign parts to the cells in the presence of alternative process plans,
multiple alternative machines and processing times. The heuristics were tested using small
problem sizes only.

Chan et al. (2003) developed a heuristic algorithm that addressed problems of machine
allocation in cellular manufacturing only when the intra-cell materials flow was taken into
account. The objective function employed was to minimize the total travelling score within one
cell in which the total travelling distance was covered. The current algorithm outperformed other
approaches as it provided near optimum solutions.

Kim et al. (2004) considered a more comprehensive CF problem with a multi-objective


machine formulation and applied a two-phase heuristic algorithm. The authors concluded that the
two-phase heuristic algorithm was effective in minimizing intercell part movements and
maximum machine workload imbalance.

32
Mahdavi et al. (2008) reported the usefulness of a heuristic algorithm based on flow
matrix for cell formation and layout design in a simultaneous fashion using sequence data.

Liu et al. (2010) developed a heuristic algorithm with three stages: (a) form the
temporary machine group plan according to the alternative process routings of each part, (b)
select the appropriate process routing of each part with respect to the over-all material movement
cost, and (c) configure the regular manufacturing cells based on the appropriate process routing.
The results implied that it is useful for manufacturing cell design in both quality and speed.

2.3.6.2 Meta-Heuristic Methods

Over the past two decades, metaheuristics have been mainly developed for the solution of
NP-hard combinatorial optimization (CO) problems. The CF problem is considered to be a
complex and difficult optimisation problem. Many researchers in order to gain more benefits of
the CF problem have applied metaheuristic algorithms. Meta-heuristic techniques constitute
simulated annealing (SA), tabu search (TS), genetic algorithms (GAs), ant colony optimisation
(ACO) particle swarm optimisation (PSO), bees algorithm (BA), and water flow-like algorithm
(WFA). Since late 90s the applications of meta-heuristic techniques to GT/CF problems have
been encouraging.

• Simulated annealing (SA)


Simulated annealing (SA) and tabu search (TS) algorithms have a common characteristic as the
search process starts from one initial state (the initial solution) and describes a trajectory in the
state space.

Applications in CMS:

Souilah (1995) presents the general SA algorithm and also shows how it has been used to
group resources into manufacturing cells, to design the intra-cell layout, and to place the
manufacturing cells on the available shop-floor surface. Vakharia and Chang (1997) developed
two heuristic methods for the CF problem both based on simulated annealing and tabu search
algorithms. The results obtained indicated that simulated annealing outperformed tabu search in
terms of solution quality and computational time.
33
Ariafar & Ismail (2009) developed a mathematical model based on a variant of simulated
annealing algorithm to minimize both inter-cell and intra-cell material handling costs. The
algorithm produces solutions with better quality and less computation time in comparison with
the benchmarked algorithm even when the size of the problem increases.

Paydar et al. (2010) proposed a simulated annealing approach to develop a cellular


manufacturing system formulated as a multiple departures single destination multiple travelling
salesman problem (MDmTSP). The experiment has achieved optimal solutions and also
indicated that the SA approach performs well compared to some well-known cell formation
methods. Pailla et al. (2010) and Lin et al. (2010) has also reported better cell formation solution
after applying SA than the other existing evolutionary algorithms.

• Tabu Search

A description of the TS method and its concepts can be found in Glover and Laguna
(1997).

Applications in CMS:

Aljaber et al. (1997) proposed a TS heuristic algorithm for the solution of a pair of
shortest spanning path problems, one for the machines (rows) and one for the parts (columns).

Lozano et al. (1999) proposed a weighted sum of intra-cell voids and intercellular moves
to evaluate the quality of the solutions using TS algorithm. The performance of these TS was
benchmarked against two SA approaches, another TS approach and three existing heuristics.

Spiliopoulos and Sofianopoulou (2003) proposed a two stage heuristic approach for the
manufacturing cell design problem and a TS scheme for its solution. The first stage tackled parts
grouping whereas the second eliminated intercellular traffic flow. The TS algorithm, as the third
stage to be implemented, integrated proper short and long term memory structures and an overall
search strategy for their use.

Wu et al. (2004) considered a CF problem with production volume and cell size to
decompose the manufacturing shop into several manufacturing cells so that the total part flow

34
within the cells can be maximized. A comprehensive TS heuristic algorithm consisting of
dynamic tabu tenure and a long term memory structure was proposed and reported very good
result for small to medium sized problems.

Lei and Wu (2006) worked beyond a single objective for CF and presented a pareto-
optimality based multi-objective tabu search (MOTS) algorithm with objectives: minimization of
the weighted sum of intercell and intra-cell moves and minimization of the total cell load
variation. The computational results demonstrated the strong ability of MOTS in multi-objective
optimization.

Solimanpur and Elmi (2012) proposed nested tabu search (NTS) algorithm to evaluate 6
problems selected from the literature. The comparative results have shown effectiveness and
efficiency of NTS with SVS-algorithm.

Chung et al. (2011) proposed an algorithm where good initial solutions are first generated
and later on improved by a tabu search algorithm combining the mutation operator and an
effective neighborhood solution searching mechanism. The result after comparing with the
mathematical programming approach shows that the proposed algorithm is able to produce
optimal solutions in less than 2 seconds.

• evolutionary algorithms (EA)

Both evolutionary algorithms (EA) and ant colony optimization (ACO) could be
characterized as population-based searches. EAs are inspired by nature’s capability to evolve
living beings well adapted to their environment. EA prove to be particularly popular due to their
added characteristic of being able to search the solution space not from a single point but from a
population of points in parallel. There are several variants of EA but for the purpose of this work
two are identified: (a) genetic algorithms (GAs) (Holland, 1975), which has been used quite
extensively for CF, (b) scatter search (SS) (Marti et al. 2006), and (c) particle swarm
optimization (PSO) (Kennedy and Eberhart, 1995) which has been employed for CF only
recently.

35
• Genetic Algorithm(GA)
The core operation for GAs is based on evolution which usually starts from a population
of randomly generated individuals and happens in generations. In each generation, the
fitness of every individual in the population is evaluated, multiple individuals are
stochastically selected from the current population (based on their fitness), and modified
(recombined and possibly randomly mutated) to form a new population. The new
population is then used in the next iteration of the algorithm. Commonly, the algorithm
terminates when either a maximum number of generations has been produced, or a
satisfactory fitness level has been reached for the population.

Applications in CMS:

Venugopal and Narendran (1992) have shown the nature of GA is parallel when executed
on multi-processor system and efficiently reaches to optimality for CFP which deals with
multi-objectivity. Gupta et al. (1996) implemented GA as a solution methodology to CF
problem considering total movements of components and cell load variation while Joines
et al. (1996) developed an integer programming model using GA. Further Morad &
Zalzala (1996), Hwang & Sun (1996), Kazerooni et al. (1997) and Hsu & Su (1998)
successfully used GA in cell formation and proved efficient with multi-objective
problems. Moon & Gen (1999) considered production volume, machine capacity,
processing time, number of cells and cell size, Zhao and Wu (2000) solved multiple
objectives and part routing of CF problems with a modified GA. Chu and Tsai (2001)
proposed a GA based heuristic technique with a new similarity coefficient developed to
adjust the gene value of each part and heuristic mutation applied to fine tune the gene
value of machine and part. Dimopoulos and Mort (2001) developed genetic programming
(GP) based method to model Single Linkage Clustering (SLCA) problem. Mansouri et
al.(2003), Zolfaghari and Liang (2004) and Goncalves and Resende (2004) have shown
that GA can be more effective solving multi-objective problems with local heuristics in
this domain. Pai et al.(2005) and Chi and Yan (2004) attempted to test GA in fuzzy
environment considering the manufacturing factors of multi-process plan, fuzzy product
demands and fuzzy technical feasibility of machines. Muruganandam et al. (2005)
applied memetic algorithm (MA) which is a modified version of GA embedded with TS

36
on CFP. Filho and Tiberti (2006) introduced grouping genetic algorithm with new
crossover, mutation operators, correction scheme and a new codification scheme of
chromosomes based on machine groups rather than individual machine. Nsakanda et al.
(2006) modeled a CFP with multiple dimensions such as operations sequence, part
demands, machine capacities, multiple process plans and multiple routings and developed
a GA method combined with price-directed decomposition method. Defersha and Chen
(2006) developed a mathematical model, which incorporates dynamic cell configuration,
alternative routings, sequence of operations, multiple units of identical machines,
machine capacity, workload balancing among cells, operation cost, subcontracting cost,
tool consumption cost, set-up cost and other practical constraints. A two-phase GA based
heuristic technique was proposed to solve this CFP. Dimopoulos (2006) has shown the
earlier GP-SLCA model proposed by him is a single-objective technique and can be
clubbed with NSGA-II, a multi-objective technique, and this combination is a powerful
tool which is able to handle very large scale problem. Ponnambalam et al. (2007)
proposed a GA based technique in their work using non-binary real valued workload data
as an input matrix and developed a modified grouping efficiency. Boulif and Atif (2008)
considered dynamic production factors like input data, with realistic constraints.
Mahapatra and Pandian (2008) has taken into consideration the operational time and
sequence of operation of parts, to minimize cell load variation and exceptional elements
by applying GA methods. Chan et al.(2008) has introduced CFP with IAECLP with two
objectives, minimizing intracell and intercell part movement and total sum of intra-cell
and inter-cell part distance unit due to machine sequence and sequences of newly formed
cells and then applied GA on top of it for better result. Kao et al. (2008) presents a new
DE-based algorithm for solving cell formation problems, namely Differential Evolution
based Cell Formation Problem (DECF). Tariq et al. (2009) developed a local search
heuristic based GA as a methodology of CFP, which uses integer type representation,
multipoint crossover and roulette wheel selection procedure. The study shown by
Mahdavi et al. (2009) that cell formation with an objective of minimizing total number of
voids and EEs in part-machine incident matrix by using a GA embedded with a heuristic
inspired mutation is efficient and it yields significantly improved solution. Haleh et al.
(2009) developed new hybrid technique based on memetic algorithm and revised topsis
37
method called (HMA-RTM) and applied on multi objective CFP. Cao et al. (2009)
formulated a mathematical model for optimal lot splitting into alternative routes to
account for either positive or negative effects of production run length on product quality
in a Cellular Manufacturing environment. Noktehdan et al. (2010) has introduced a
Differential Evolution (DE) approach. A GA is used to solve simple numerical example
and it proves that the model is effective. Neto and Filho (2010) designed a multi­
objective optimization model using GA for CFP, where fitness evaluation was performed
via simulation of cellular system. A parallel hybrid grouping genetic algorithm has been
applied to team formation based on GT (Agustin-Blas et al. 2011). A hybrid GA method
has been used to improve the best-known solution for 31 of the 35 benchmark CFPs
(Elbenani et ah 2012). Baneijee and Das (2012) successfully applied a fine grain variant
of the predator-prey GA in CFPs.
• Scatter Search (SS)
In contrast to GAs, SS, which was first introduced by Glover (1977), is founded on the
premise that designs and methods for creating new solutions afford significant benefits
beyond of those derived from resource to randomization. Solutions are purposely
generated to take account of characteristics in various parts of the solution space. SS
orients its exploration to a set of reference points that typically consists of good solutions
obtained by prior problem solving efforts. The criteria for ‘good’ are not restricted to
objective function values and may apply to sub-collections of solutions rather than to a
single solution.
Applications in CMS:

Tang et al. (2010) proposed a scatter search approach considering two different methods
to generate diverse initial solutions and two improvement methods, and adopts the
roulette wheel selection in the combination method to further expand the conceptual
framework and implementation of the scatter search. The SS approach is compared with
the commercial solver CPLEX on a set of test problems, some of which are large
dimensions. The results have demonstrated the effectiveness of this approach.
Tavakkoli-Moghaddam et al. (2012) developed a meta-heuristic algorithm based on
scatter search to minimize the makespan, intracellular movement, tardiness, and
sequence-dependent setup costs, simultaneously. This algorithm has been evaluated and
38
the related results confirmed the efficiency and the effectiveness of the method to provide
good solutions, especially for medium and large-sized problems.

• Particle Swarm Optimization (PSO)


PSO is inspired by flocking birds and it is initialized with a population of random
solutions evolving over generations to find optima. Generally speaking, the set of rules
that govern PSO are: evaluate, compare and imitate. The evaluation phase measures how
well each particle (candidate solution) solves the problem at hand. The comparison phase
identifies the best particles. The imitation phase produces new particles based on some of
the best particles previously found. These three phases are repeated until a given stopping
criterion is met. The objective is to find the particle that best solves the target problem.

Applications in CMS:

Andres and Lozano (2006) were the first authors to consider a PSO algorithm for the CF
problem. The objective involved was the minimisation of inter-cell movements. A
number of published results were used to assess the proposed algorithm. The
computational results showed that the proposed algorithm can generate optimal or near
optimal solutions but only for small data sets.
Duran et al. (2008, 2010) proposed a modified PSO algorithm. The main modification
made to the original PSO algorithm is that the current algorithm did not use the vector of
velocities as the standard PSO algorithm does. The proposed algorithm used the concept
of proportional likelihood with modifications, a technique that is used in data mining
techniques. Some simulations were presented and compared. The criterion used to group
the machines in cells was based on the minimization of inter-cell movements. The
computational results showed that the PSO algorithm is able to find the optimal solutions
in almost all instances.

• Ant Colony Optimization (ACO)

ACO is one of the newest metaheuristic for the application to CF problems. The basic
ideas of ACO were introduced in Dorigo et al. (1996, 1999). ACO was inspired by the foraging

39
behavior of real ants (Deneubourg et al. 1999) and its search process can be described as the
evolution of a probability distribution over the search process.

Applications in CMS:

Prabhaharan et al. (2005) also proposed an ACO approach for grouping the machines,
with the objective of minimizing total cell load variation and total intercellular moves. A number
of parameters were also considered in this study, such as demands for numbers of parts, routing
sequences, processing time, machine capacities and machine workload status. The results of the
ACO approach were compared with a GA taken from the literature. The former proved to have
better performance.

Kao and Li (2008) presented a part clustering algorithm for the CF problem that used the
concept of the recognition system of artificial ants. The proposed algorithm mimicked the
random meetings of real ants to build up the ability of object recognition and then to form many
initial part clusters with high similarities. These initial part clusters were further merged into
larger and larger clusters in an agglomerative way until the designated number of part families
was reached. The effectiveness of this algorithm was tested with a variety of data sets collected
from the literature.

Spiliopoulos and Sofianopoulou (2008) proposed an ACO algorithm, which used a tight
eigenvalue-based bound to guide and accelerate the search, when minimization of inter-cellular
moves was considered. The resulting algorithm produced most promising results for medium to
large scale problems.

Megala et al. (2008) considered the problem of cell formation with the objective of
maximizing the grouping efficacy and developed an ACO algorithm to obtain machine cells and
part families. Their proposed algorithm was tested by using many benchmark data sets. The
grouping efficacy obtained was compared with grouping efficacies of existing approaches.

Islier (2005) developed an ACO algorithm for the CF problem in order to get block
diagonalised structures for the part/machine incidence matrices. The grouping problem was first
represented as an artificial ant system, via which better and better groupings were obtained as
semi-blind ants could find their way by a communication-supported random search process. The
40
proposed technique was compared with other approaches such as GA, SA and TS. The most
remarkable outcome was that ant systems performed better than the other techniques as far as an
equal number of solution alternatives was concerned.

Farahani and Hosseini (2011) developed an ACO algorithm to solve 35 cell formation
problems and it is found to improve the best known values of the grouping efficacy for 5
problems.

Agrawal et al. (2011) incorporated more generalized framework of ants’ behavior, has
been applied to the parts and machines grouping problems taken from the literature. The results
has shown that the average performance of Tabu search with multiple ants was found to be the
best and thus the parameter values for this approach were also determined using design of
experiments methodology.

Survey also revealed that other metaheuristics are strongly competitive with EA, SA, TS
and ACO such as PSO, BA, WFA are used less frequently. Most of them are used to solve single
objective CFPs and only few are developed to solve multi-objective problems. According to the
survey till date done usage and winning possibility of EA is higher.

Table 2.1. Summary of cell formation methods with references

SI Cell Formation Implementation ofMethods with References


No Approaches
1 Visual Methods It is not a systematic approach and cannot be applied if the number of parts is huge
(Kusiak, 1987)
2 Classification and Development of classification and coding systems including Brisch system
Coding (Gombinski, 1963), Optiz system (Optiz, 1970),
MICLASS/MULTICLASS(Houtzeel, 1975). Singh and Rajamani (1996) described
the coding systems. Ghosh et. al. (2010)
3 Production Flow PFA concept was introduced by Burbidge (1971).
Analysis
4 Component Flow CFA was designed by El-Essawy et al. (1972)
Analysis
5 Graph Theoretic Rajagopalan and Batra (1975) developed Graph theoretic approach to machine cell
formation. Few other applications included are: Vannelli and Kumar (1986), Kumar
and Vannelli (1987), Deutsch et al.(1998), Mukhopadhyay et al. (2000), Ashayeri et
al. (2005), Won and Lee (2004), Won and Currie (2006), Veeramani and Mani
(1996), Selim (2002), Manimaran et al. (2010)
6 Mathematical Integer programming: Han and Ham (1986), Kusiak (1987), Shtub (1989),
Choobineh (1988), Wei and Gaither (1990), Boctor (1991), Selim et al. (1998),
Albadawi et al. (2005), Bulgak & Bektas (2009). Non-linear goal programming:
Purcheck (1975), Bilic & Veza (2001). Branch and Bound technique: Kusiak et al.
41
(1993), Ramabhatta and Nagi (1998), Solimanpur & Jafari (2008), Boulif and Atif
(2006). P-median approach: Kusiak (1987), Won and Lee (2004), Won & Currie
(2006). Mixed-integer programming: Foulds et al. (2006). Dynamic Programming:
Ballakur and Steudel (1987). Goal programming: Shafer and Rogers (1993). Linear
programming: Boctor 1991, Wang (2003). Eigen value matrix: Albadawi et al.
(2005), Hachicha et al. (2008). Spectral Clustering: Oliveira et al. (2009).
Lagrangean: Kattan (2007). Data Envelope Analysis: Sofianopoulou (2006).
Quadratic Assignment Problem: Schaller et al. (2000), Albadawi (2003), Defersha
& Chen (2006), Tsai & Lee (2006), Mahdavi et al. (2007), Kaku & Tang (2008),
Sangwan & Kodali (2009), Jerbi et al. (2010), Paydar et al. (2011), Mahdavi et al.
(2012)
7 Cluster Analysis Agglomerative clustering: Gupta and Seifoddini (1990) and Gupta (1993), Vakharia
and Wemmerlov (1990). Non-hierarchical clustering algorithms:
Chandrashekharan and Rajagopalan (1987),Jayakrishnan and Narendran (1998,
1999). ROC developed by King (1980a,b), ROC2: King and Nakomchai(1982).
DCA : Chan and Milner (1982). Similarity Coefficient: McAuley (1972), Mosier and
Taube (1985), Seifoddini and Wolfe (1987) and Gupta and Seifoddini (1990), Islam
& Sarker (2000), Garbie et al. (2005), Yin & Yasuda (2006), Oliveira et al. (2008),
John et al. (2011), Yin et al. (201 lb)MODROC: Chandrasekharan and Rajagopalan
(1986a). BEA: McCormick et al. (1972). Occupancy Value: Khator and Irani,
(1987). K-means: Unler & Giingor (2009)
8 Heuristic Waghodekar and Sahu (1984), Panneerselvam and Balasubramanian (1985), Chen
(2001), Mukattash etal. (2002), Chan et al. (2003), Kim et al. (2004), Yin et al.
(2005), Mahdavi et al. (2008), Suer et al. (2009), Liu et al. (2010), Elbenani &
Ferland (2012). close neighbour algorithm: Boe and Cheng (1991). GRAFICS:
Srinivasan and Narendran (1991). CASE: Nair and Narendran (1998). ACCORD:
Nair and Narendran (1999)
9 Meta-Heuristic Simulated Annealing: Boctor (1991), Venugopal and Narendran (1992), Liu and
Wu (1993), Vakharia and Chang (1997), Zolfaghari and Liang (1998), Caux et al.
(2000), Abduelmola & Taboun (2000), Jayaswal and Adil (2004), Arkat et al.
(2007), Tavakkoli-Moghaddam et al. (2007), Defersha & Chen (2009), Tavakkoli-
Moghaddam et al. (2009), Paydar et al. (2010), Defersha & Chen (2008a), Pailla et
al. (2010), Lin et al. (2010). Ant Colony Optimization: Islier (2005), Giri et al.
(2007), Ateme-Nguema & Dao (2007), Solimanpur et al. (2010), Jiahan et al. (2010).
Tabu Search: Logendran et al. (1994), Lozano et al. (1999), Onwubolu & Songore
(2000), Adenso-Diaz et al. (2001), Spiliopoulos and Sofianopoulou (2003), Cao &
Chen (2004), Chen & Cao (2004), Lei & Wu (2005), Lei & Wu (2006), Rodrigues
and Weller (2008), Ateme-Nguema and Dao (2007, 2009). Memetic Algorithm:
Muruganandam et al. (2005). Evolutionary algorithm: Venugopal and Narendran
(1992), Joines et al. (1996), Kazerooni et al. (1997), Pierreval and Plaquin (1998),
Moon and Gen (1999), Zhao and Wu (2000), Chu and Tsai (2001), Chi and Lin
(2002), Suer et al. (2003), Zolfaghari and Liang (2003), Gon?alves & Resende
(2004), Solimanpur et al. (2004a), Yasuda et al. (2005), Car & Mikac (2006),
Pierreval & Plaquin (2006), Rajagopalan and Fonseca (2006), Nsakanda et al.
(2006), Filho and Tiberti (2006), Noktehdan et al. (2010), Mahapatra and Pandian
(2008), Kao et al. (2008), Pailla et al. (2010), Neto and Filho (2010), Nouri & Hong
(2012). Particle Swarm Optimization: Andres & Lozano (2006), Duran et al.
(2008), Duran et al. (2010),. Scatter Search: Wang & Tang (2009), Tavakkoli-
Moghaddam et al. (2010), Gholipour-Kanani et al. (2011). Bees Algorithm: Pham et
al. (2007). Water Flow like Algorithm: Wu et al. (2010)

42
2.3.7 Artificial Intelligence

2.3.7.1 Background of Artificial intelligence (AI)

Artificial intelligence (AI) is simply a way of providing a computer or a machine to think


intelligently like human beings (Laha & Mandal, 2008). Since human intelligence is a complex
abstraction, scientists have only recently begun to understand and make certain assumptions on
how people think and to apply these assumptions in order to design AI programs. It is a vast
knowledge base discipline that covers reasoning, machine learning, planning, intelligent search,
and perception building. The phrase “artificial intelligence” had been first coined at a famous
Dartmouth College conference in 1956 (http://en.wikipedia.org/wiki/History_of_artificial_intelligence).

Artificial intelligence has been claimed to yield revolutionary advances in manufacturing


(Zhang & Huang, 1995). According to Wong et al. (1997) 53.5% of the reviewed literature in
artificial intelligence refers to applications in production and operations management.
Applications of AI in manufacturing includes: condition monitoring in different manufacturing
operations such as metal forming (Selles et al. 2012), drilling (Kamik, 2008), turning (Asilturk &
£unka§, 2011), and tool wearing and breaking (Chen et al. 2011), cost estimation (Duran et al.
2012), fault diagnosis (Demetgul et al. 2009), parameter selection (Li et al. 2003), production
scheduling (Feng et al. 2003), manufacturing cell formation (Chen & Cheng, 1995), and quality
control (Du et al. 2012), operations (Wong et al. 2011) etc.

Artificial intelligence techniques consist of neural networks, case base reasoning, genetic
algorithms, genetic programming, and many others (Chen & Leung, 2005). For example, the
field of artificial neural networks (ANN) is per se an extensive area of artificial intelligence.
Haykin (1994) provides a comprehensive foundation on neural networks. In this work the major
interest area is Artificial Neural Network.

23.1.2 Description of Artificial Neural networks (ANN)

Artificial Neural networks (ANN) are massively parallel computational devices whose
design is inspired by the pattern of thinking in the human brain (Simpson, 1990; Wasserman,
43
1989; Rojas, 1996; Saberi & Yusuff, 2012). ANNs process information in parallel and are robust
to data errors. These biological systems encompass many processing elements, the neurons that
are interconnected according to some structure (Hanggi & Mosehytz, 2000). They have the
capability to generalize, adapt, approximate given new information, and provide reliable
classifications of data. While the possible architectures vary widely, all neural networks consist
of the same basic elements—an abstract structure composed of nodes and arcs (Zhang et al.
1999; Hu et al. 1999). Artificial neural networks are characterized by the network architecture
(Figure 2.2), such as the number of layers, the number of nodes in each layer, and how the nodes
are connected. The network has an input layer, an output layer and any number of hidden layers.
A neuron is linked to all neurons in the next layer (Hammerstrom, 1993), as shown in Figure.
2.2. They are highly interconnected by links (synapses) with weights (Yao, 1999; Rodvold et al.
2001).

ANNs may be feed forward (the commonest) or feedback. Information flow is


unidirectional in feed forward ANNs, with no cycles, but in both directions in feedback ANNs so
they have cycles, by which their state evolves to equilibrium (Stergiou and Siganos, 1996).
ANNs require few prior assumptions, learning from examples (Rodvold et al. 2001) by adjusting
the connection weights to produce the desire output from inputs and is called training
(Rojas,1995; Patterson, 1996). Learning may be supervised or unsupervised. Supervised learning
gives the ANN the correct output for every input pattern (Caudill & Butler, 1994). The weights
are varied to minimise error between ANN output and the given output. One form of supervised
learning, reinforcement learning, tells the ANN if its output is right rather than providing the
correct value (Yao, 1999). Unsupervised learning gives the ANN several input patterns. The
ANN then explores the relations between the patterns and learns to categorise the input (Jain et
al. 1996). Some ANNs combine supervised and unsupervised learning. The data structure and
non-linear computations of ANNs allow good fits to complex, multivariable data.

44
2.3.7.3 Basic Structure of ANN

Input #1

Input #2 a — purelin(n)
Output
Input #3 Figure 2.3a. Linear Activation Function

Input #4 a
*+l

-1
a = logsig(n)

Figure 2.2. Arehitechture of Neural Network Figure 2.3b. Sigmoid Activation Function
Neuron jc has n inputs and one output
j(x)=g^_Zw/^j (2.1)

Where (wo,........ ,wn) are the input weights and g is the non-linear activation function (Nissen,
2003), usually a step (threshold) function or sigmoid (Figure 2.3a and Figure 2.3b). The step-
function output is y=\ if x>0, 0 if x<0. The sigmoid function, more commonly used, is
asymptotic to 0 and 1 (Jain et al. 1996) and antisymmetric about (0,0.5):
^W=--Jzx-^>0 (2.2)
l + e px

23.1 A Application Domains of ANN

ANNs can be applied to seven categories of problems (Jain et al. 1996): pattern
classification, clustering, function approximation, prediction, optimisation, retrieval by content
and process control. Pattern classification assigns an input pattern to one of the pre-determined
classes, e.g. food classification (D^bska & Guzowska-Swider, 2011), power systems engineering
(Yap et al.2011), Control chart pattern recognition (Kim et al. 2012), Multivariate Quality
Control (Masood & Hassan, 2012), land classification from satellite imagery (Barros and
Rodrigues, 1994) or sewage odour classification (Onkal-Engin et al. 2005). Clustering is

45
unsupervised pattern classification, e.g. load forecasting in power systems (Jain & Satish, 2009),
input patterns to predict ecological status of streams (Vellido et al.2007). Function
approximation, also called regression, generates a function from a given set of training patterns,
e.g. control the amount of feed during the constant-force grinding (Chen et al.2011), central
composite rotatable design (Xiong et al. 2012), cost estimation of sheet metal parts (Verlinden et
al. 2008), modelling river sediment yield (Cigizoglu and Alp, 2006) or predicting daily
maximum ozone concentration (Yi & Prybutok, 1996). Prediction estimates output from
previous samples in a time series, e.g. of prediction of various stock indices (Zhang & Wu,
2009), predicting chaotic time series (Zhang & Xiao, 2008), Prediction of rainfall time series(Wu
et al. 2010 ), or water quality management in river basin(Wen and Lee, 1998). Optimisation
maximises or minimises a cost function subject to constraints, e.g. Maximizing the recoverable
strain by optimizing process parameters in Ti rich Ni Ti (Sinha et.al., 2013). Content based
retrieval of multimedia image (Dowe, 1993), probabilistic information retrieval (Kwok, 1989)
and retrieval by process control (Dominguez & Bolle, 1998).

2.3.7.5 ANN Applications in Cellular Manufacturing

Though neural networks have been successfully applied in a variety of fields, their use in
Cellular Manufacturing problems started in the late 80s and early 90s. Recognizing ANN’s
pattern recognition ability, several researchers began to investigate neural network methods for
the part-machine grouping problem. Since the last two decades, researchers had applied various
models of ANN to solve the CF problems and from the pie chart in Figure 2.4 obtained from the
literature survey, it could clearly be seen that among the major ANN approaches, ART (nearly
47%) and SOM (nearly 21%) (including the modifications) are the most common approaches
with significant results while other models occupy only a small space in the CF literature.

46
21%

Figure 2.4. Usage percentage of major neural network


approaches in Cellular Manufacturing

Based on the observation, the ANN approaches are classified into three broad categories.
• ART (Adaptive Resonance Theory)
ART is a family of algorithms for unsupervised learning developed by Carpenter and
Grossberg (1987). ART is similar to many iterative clustering algorithms where each pattern
is processed firstly by finding the "nearest" cluster to that exemplar and then updating that
cluster to be "closer" to the exemplar. While ART1 is designed for binary features, ART2 is
designed for continuous (analog) features. Figure 2.5 and Figure 2.6 shows the architecture
of ART1 and ART2 models.

Figure 2.6. Basic Architecture of ART2 Figure 2.7. Architecture of KSOM

Kao and Moon (1991), Dagli and Huggahali (1991), Kusiak and Chung (1991), Dagli and
Sen (1992) and Liao and Chen (1993) used ART1 to group parts or machines while Burke

47
and Kamal (1992), Suresh & Kaparthi (1994), Burke & Kamal (1995) and Suresh et al.
(1999) applied Fuzzy ART neural network for cellular manufacturing. In an attempt to
improve performance, Dagli and Huggahalli (1995) proposed modifications to the basic
ART1 algorithm while (Chen & Cheng, 1995) proposed a set of supplementary procedures to
handle the ill structured data. Chen et al. (1996) proposed an improved ART neural network
for machine cell formation to reduce the effects of the input order of the vectors in the
machine part matrix to yield better solution. Kamal and Burke (1996) further proposed a
new clustering algorithm named FACT (Fuzzy ART with Add clustering technique) for
Group Technology. Enke et al. (1998) presented a new ARTI paradigm which involved
reordering of the input vectors with a modified procedure for storing a group's representation
vectors. Lee and Fischer (1999) proposed a new part family classification system (IPFACS:
Image Processing and Fuzzy ART based Clustering System). Enke et al. (2000) modified
ARTI paradigm which reordered the input vectors, along with a modified procedure for
storing a group's representation vectors. Chen et al. (2002) presented integrated approach of
ARTI and Tabu Search to solve CF problems. Further Park and Suresh (2003), and extended
Fuzzy ART for better results while Won and Currie (2007) used Fuzzy ART Neural
Network/ RRR-RSS: a two-phase neural network algorithm to solve the comprehensive part
machine grouping (PMG) problem in Cellular Manufacturing (CM) where the operation
sequences with multiple visits to the same machine, production volumes of parts and multiple
identical machines were incorporated simultaneously. Modification in ARTI made by
(Venkumar and H&q, 2005) and Venkumar and Haq (2006a) showed improved results while
Kuo et al. (2006) applied Fuzzy ART2 neural network approach for part family formation.
Miljkovic and Babic (2005) used ARTI Simulator and FLEXY in Machine-Part Family
Formation problem. Yang and Yang (2008) identified the existing limitations of the modified
ART approach by (Dagli and Huggahalli, 1995) and tried to correct them. Pandian &
Mahapatra (2009) also proposed some modifications to ARTI considering production
sequence data and operation time of the parts.

• Self Organizing Map:


The Self-Organizing Map (SOM) or Self-Organizing Feature Map (SOFM) was proposed by
Kohonen (1997). This model implements a clustering algorithm that is very similar to the
classical K-means clustering algorithm. The SOFM network typically has two layers of
nodes, the input layer and the two-dimensional Kohonen layer. SOFM networks are good at
48
accepting multi-dimensional non-binary input and transforming it into a map of fewer
dimensions. Figure 2.7 shows the architecture of KSOM model.
Lee et al. (1992) and Rao and Gu (1994) applied self-organizing neural network part family
formation, machine cell identification, bottleneck machine detection and the natural cluster
generation. Chakraborty and Roy (1993) and Pilot and Knosala (1998) used Kohonen's self­
organizing feature maps for part family and machine cell formation. Kulkami and Kiang
(1995) and Kiang et al. (1995) applied self-organizing feature map (SOM) in CF problem.
Inho & Jongtae (1997) proposed Self Organizing Feature Maps (SOFM) network where the
manufacturing environment factors such as the operational sequences and lot sizes were
considered while Rao et al. (2001) applied SOFM for Cell formation utilizing a syntactic
pattern recognition approach. Guerrero et al (2002) used self organizing neural network and
Kuo et al.(2001) applied Fuzzy self-organizing feature map neural network for clustering the
parts into several families. Further Kiang (2001) and Venkumar & Haq (2006b) extended the
Kohonen self-organizing map (KSOM) networks for better clustering results.

• Other traditional Neural Networks:


Apart from the two major approaches as discussed above, various other traditional neural
network approaches were applied in CMS over the years. The other traditional neural
network approaches are briefly described below in Table 2.2.

2.4 Analysis of Reviewed Literature

Each cell formation design approach considers a different number of design objectives
and constraints, to varying extents, depending on the scope and interest of each approach
(Prabhaharan et al. 2005). Each of them has its own advantages and limitations. Some of them
are easy to implement to solve cell formation problem in reasonable time. Some of them have
considered a number of objectives but require a substantial amount of time to generate solutions.
The above observation for the major types of cell formation approaches are discussed below:

49
Table 2.2. Other traditional neural networks with application in cell formation of Cellular Manufacturing Systems
SI Approach Brief description In CMS
No
1 Competitive learning It one of the simplest paradigms in unsupervised Grossberg (1976),
(CL) learning which was first proposed in the mid 1970s. Malave and
Models and algorithms based on the principle of Ramachandran (1991),
competitive learning include vector Venugopal and
quantization and self-organising maps (Kohonen Narendran (1992, 1994),
maps). Chu (1993),
Malakooti and Yang
(1995)
Interactive Activation The processing units are organized into several Moon (1992), Moon and
and Competition (IAC) pools (layers). Each pool represents specific Chi (1992), Chu (1997)
characteristics of the problem. In each pool, all the
units are mutually inhibitory.
Back-propagation It is based on the error correction learning rule. Kao and Moon (1991),
Network (BPN) Error propagation consists of two passes through Moon and Roy (1992),
the different layers of the network, a forward pass Kao and moon (1998),
and a backward pass. Onwabolu (1999)
Hopfield Neural Network single layer network architecture. The output of Zolfaghari and Liang
each neuron feeds into every other neuron and (1997), Ateme-Nguema
causes the network to be recursive. and Dao(2009)
Transiently Chaotic It has a transiently chaotic phase to locate global Soleymanpour et
Neural Network (TCNN) optima followed by a convergent phase. The al.(2002), Solimanpur et
chaotic behavior helps the network to explore al.(2004b)
different states and successive bifurcations lead the
output to strange attractors. As a self-feedback
connection parameter decreases, TCNN gradually
converges to an equilibrium solution.
FuzzyMin-Max (FMM) FMM uses hyperbox fuzzy sets for the different Dobado et al.(2002)
categories. Each hyperbox is defined by a min and
a max point. A hyperbox defines a region in the M-
dimensional input pattern space and all input
patterns contained within the hyperbox have full
cluster membership.
Graph Neural Network In this method, information is derived from the Mahdavi et al.(2001)
machine-part incidence matrix. One converts this
matrix to a multi-graph matrix, whose vertices
correspond to the machines and whose edges
represent those common parts that are processed by
corresponding machines.
Stochastic Neural stochastic neural network models are influenced by Arizono et al.(1995)
Network the value of the temperature parameter in the
Boltzmann machine model and the values of the
temperature and reference activation level
parameters in the Gaussian machine model.
Adaptive Hamming Net It has a similar structure to the traditional Lee et al.(1997)
Hamming net. adaptive Hamming net is an adaptive
version of the Hamming net, a two-layer neural
network that consists of a matching score net with a
threshold 8 and a MAXNET

50
2.4.1 Clustering Analysis

Clustering analysis in cell formation literature considers only one objective, the
minimization of intercellular moves. Only the operations of parts and the processing for
machines are considered in cluster analysis. During the design process the product factors like
operational sequences and processing time and the rate of production as production factor are not
included for which only partial solutions might achieve. Still, clustering produce solution in
legitimate time period and easy to implement.

2.4.2 Mathematical Programming

Most of the formulations for CF propose a mathematical programming model with the
main objective of minimizing the total number of intercellular movements. Other mathematical
programming formulations involve cost related objective functions and only a few consider
machine/load utilization as a goal parameter for CF. They are hard to implement due to
computational limitations for large scale problems. Hence, the majority of these formulations
serve for setting the problem up and for providing a lower bound or a sub optimum value against
which computational results of additional methodologies proposed within each study can be
compared. Obtaining optimal solutions from mathematical programming approaches can be
unfeasible due to the combinatorial complexity of the CM design problem (Singh and Rajamaani
(1996), Selim etal. (1998)).

2.4.3 Heuristic and Meta-heuristic approaches

Due to the NP-hard nature of the CF problem, methodologies such as heuristic and
metaheuristic strategies have been employed to integrate larger scale systems meeting today’s
requirements. The most popular metaheuristics employed in the last decade are GAs, TS and SA.
They have thus been used as alternatives to obtain reasonably good solutions within an
acceptable amount of time. The problem specific heuristic works only for one problem; it cannot
be used to solve a different problem. For instance, a specific heuristic developed to solve a
traveling salesman problem is unlikely to be applied to solve a general assignment problem.
Another type of heuristic, meta-heuristics, e.g., genetic algorithm and simulated annealing, tabu

51
search, ant algorithms, etc., which are more general and can be used for different types of
problems.

The survey presents an in-depth review of recent CF based meta-heuristic methodologies.


Since mid 90s EAs, SA, TS have evolved as powerful optimization techniques in CFP and a
substantial volume of research works are available which are focused on these techniques.
Amongst these, GA has been the most adaptable technique for the researchers when it is
combined with other algorithms and dominating as a solution methodology in CM since the last
two decades. ACO, PSO, BA and WFA are lately developed and able to compete with GA.
Therefore, it seems that the trend is on using population based methodologies as well. According
to literature survey, it seems that the new methodologies are built combined with GA to solve
CFP as hybrid technique. As a future research, it is possible to use hybrid techniques to solve
more realistic and complex GT/CM problems to handle large scale industrial optimization
problems with due focus on higher efficiency, less computational time.

The work has also identified the trend of application of genetic algorithm in the area of
cell formation in cellular manufacturing with respect to computational time, numerical result and
the crossover, mutation and population size of the genetic operation have been improved over the
years.

PSO and ACO have only recently received attention for CF. The former has been
employed for clustering purposes for CF when a part/machine incidence matrix was considered,
whereas the latter has been proposed for addressing in most of the studies the intercellular layout
of the CF problem.

Most of the research in the CFP literature focus on single meta-heuristic approach, which
is compared either to the variants of the same technique, or to previously available methods such
as similarity coefficient method, mathematical programming method, or to simple heuristics such
as random search, greedy search, or to exact methods when these are available. Few works
perform comparisons among different meta-heuristics.

This is also an emerging research area where new techniques based on meta-heuristic can
be formed. In the frequency of the usage of meta-heuristic in CF domain where EA is winner in

52
most cases and SA and TS are believed to be effective methods. For many large scale problems,
computational time is a major concern of many researchers, and hence better EAs are being
proposed. From the late 90s GA, TS, SA are mostly considered as optimization tool in CFP.

New population based tools like PSO, ACO are attracting more research interest as they
are computationally more attractive and less complex. Many other tools such as BA, WFA, SS
and other hybrid tools are evolving with time as CF solution methodologies.

CF Problems can be formulated in more complex way by means of multiple objectives,


such as inter-cell or intra-cell part movement, within cell load variation, count of EEs and voids,
machine utilization, machine investment, machine duplicity, WIP level, part subcontracting, part
cycle time, part routing, operational time, operational sequence of parts. The analysis of the
present work and the trend of genetic algorithm on various observation criteria in cell formation
revealed that

• CFPs can be formulated realistically in more complex way by means of multiple


objectives.
• GAs is very complex in nature, therefore in order to minimize the computational time
advanced machines are required.
• Due to the complexity of the industrial data improved genetic techniques are more useful
while solving the CFPs.
• ZODIAC, GRAFICS, MST and LINGO solutions are the benchmarks in CFP literature
while computing the goodness of the proposed methods.
• Selection or formulation of appropriate performance measure depends on the type of
problem undertaken.

2.4.4 Artificial Neural Network

From the review as presented above it could be seen that ART1 and SOM models have a
more efficient and significant role in the part machine grouping problem compared to other ANN
approaches as found in the literature. ART1 occupies around 47% of the total analyzed literature
while SOM is applied on another 21% of the works. Other approaches cover the remaining 32%.
Hybrid approaches such as incorporation of Fuzzy logic in the ANN models gave some brilliant

53
solutions with real time data. A detailed survey is presented below to analyze the solution trend
in the cell formation problem using neural network since the last two decades. The survey
highlights the major contributions, NN models used, their clustering efficiencies based on group
efficiency, group efficacy, Group Technology efficiency, computational time and weighted
group capability index (WGCI), the computer configurations and tools used to analyze the
overall solution’s performance.

Table 2.3. Observations based on neural net approaches applied in Cellular Manufacturing systems

ObsavaiunAxca Cbsemlon&faib Conclusion


Percentage oftotal ART 47% SOM 21% Other lladiticQud Network 32% ART SOM are more suited to the CMS problem
analyzed literature
Most Common Fuzzy Neural Covers 25% of the total analyzed literature Hybrid Aproaches often produce better results. Hybrid
Approach Fuzzy Neural network proves to be an efficient approach to
tackle nan-b mazy machine part incidence matrix
Common group efficiency Most Common Covers around 53% of the total Selection / formulation of appropriate Performance measure
Badbxmanoe group efficacy; group efficiency, analyzed literature depends on the type of problem undertaken. Grouping
measures group technology group efficacy efficiency and efficacy are primarily designed to tackle
efficiency, binary problems without considering production factors
computational where as several other measures such as group technology
tune and efficiency , weighted group capability index, Modified
weighted group grouping efficiency are used to tackle several other complex
capability index problems with production factors such as processing time,
(WGCI) sequence, machine capacity, production volume etc
Clustering Number of Ntimber of Performance Computational To reach optimality, a balance between first three
Efficiency voidsClesserthe Exceptional measure(greater Time(ksser interrelated efficiency measures is necessary. Time
Measures letter) elements (lesser the better) tiie better). complexity is also an important consideration while
the better) Depends on seleciing/famDjlaiing the algorithm
tiie
complexity of
the selected
algorithm
Most onunoa C i3l C++ Covers 37% ofthe total analyzed literature simple and verypopular
programming
language
Analysing Trend oompatalbnal Data Set 4CK10Q Early90's, raid 2000^ Proves the improvement inalgorithms over the years
time Chandrasekharan Nearly 120 0.05s eeonds
and Rajagopalan seconds (less
(1987) titan 2 minutes)

exceptional DataSet 40X100 Early 90’s, raid 2000, Proves the improvement in algorithms over the years
dements Chandrajekharan 36 exceptional 2 exceptional
and Rajagopalan elements elements
(1987) (fractional cell
formation).

In order to summarize the above discussions, elaborate the observations and analyze the research
trend, Table 2.3 is presented.
Among the different neural network approaches the present review work concluded the
followings based on various observation criteria from the Table 2.3.

54
• ART and SOM are more suited to the CMS problem

• Hybrid approaches often produce better results. Hybrid Fuzzy Neural network proves to
be an efficient approach to tackle non-binary machine part incidence matrix

• Little work has been done on formal visualization of Machine-Part clustering and
measuring the success or failure of visual clustering approaches

• Selection or formulation of appropriate performance measure depends on the type of


problem undertaken. Grouping efficiency and efficacy are primarily designed to tackle
binary problems without considering production factors where as several other measures
such as group technology efficiency , weighted group capability index, Modified
grouping efficiency are used to tackle several other complex problems with production
factors such as processing time, sequence, machine capacity, production volume etc

• To reach optimality, a balance between first three interrelated efficiency measures is


necessary. Time complexity is also an important consideration while
selecting/formulating the algorithm

• Computer languages like C and C++ are simple and very popular languages to develop
neural network based application to solve cell formation problem

• The work has also identified the trend of neural network based research in the area of cell
formation in cellular manufacturing in the area of computational time and exceptional
elements where there are evidence of the improvement in algorithms over the years.

• Only a few hybrid formulations, where a combination of methodologies is examined,


have been produced for CF.

2.5 Summary

Although the majority of the research efforts in the field of cellular manufacturing have
been focused on traditional algorithm for generating machine-part cell formation, very little has
been reported on the context of Artificial Intelligence in cell formation. In Table 2.1, a summary
of cell formation methods has been shown based on literature review.

55
The purpose of the present chapter is to present an in-depth and critical review of the cell
formation approaches in the design of Cellular Manufacturing Systems. A detail analysis has
been carried out to identify the research trend till date that captures the chronological progress
and continuous improvement in the cell formation methods.
Although this review cannot be exhaustive, the analysis carried out is based on a large
number of literature references relating to the above areas. Although some of the simple
production factors such as processing time, processing sequence etc are considered by
researchers mostly from early 2000, to our view, the change isn’t as prominent and presentable
yet. Surely with present research focusing on demand, production volume etc would create a
space for such trend analysis in near future. In future, new meta heuristic approaches (eg: BEAS,
ANT, PSO etc) and their hybrid forms may be studied in depth and a comparative analysis
between the latest and traditional approaches may be presented to portray the relevance of
present research and fixture prospects in cellular manufacturing.
CF problems can be formulated using single objective or multiple objectives, such as
intercell or intracell part movement, within cell load variation, count of EEs and voids, machine
utilization, machine investment, machine duplicacy, WIP level etc. by considering operational
time and operational sequence of parts. The review work reveals that around 80% of the research
studied are bi-objective, and around 50% amongst them comprised total cell movements and cell
load variations.
• Most of the proposed methodologies in the last decade focus on a single criterion for CF
and only a few studies deal with multiple objectives.

• Only half of the studies in the literature have their results compared with other existing
methodologies. The reason for this is that both objective and constraint specification of
the methods differ.

• Most of the proposed models involve two or three of the identified major constraints but
only very few comprehensive formulations have been proposed.

• The largest size of problem solved varies among authors. The number of machines and
number of cells are principal determinants of the complexity of the CF problem. The
number of parts is a secondary factor influencing problem difficulty and this number will
generally rise in proportion to the number of machines.
56
• The dynamic nature of the CF problem reflecting today’s market requirements has
received attention in the last five years with a few formulations being proposed. A multi­
period time horizon constraint is assumed where reconfiguration of cells is required as
the best cell design for one period may not be the same for subsequent periods.

• Fuzzy theory has been employed mainly for clustering purposes and within mathematical
programming formulations for addressing uncertainty in certain model parameters.

• Neural network algorithms have been employed quite extensively over the last decade for
the CF problem; however only a few of them addressed a more comprehensive CF
problem with additional constraints within.

All the CMS design approaches discussed during this chapter have certain advantages
and disadvantages. Some approaches are very simple as far as their practical application is
concerned, for example ROC. Whereas, some have the ability to formulate the CMS design
problem more precisely by considering different objectives and constraints, but the problem with
such approaches is that they need substantially long computational time to find solutions, for
example Mathematical programming. Al-based approaches including ANN, FL, SA and GA
have been applied to CMS design, because of their ability to find solutions in comparatively less
computational time, capture and employ design knowledge, handle a number of constraints,
utilize several nonlinear performance measures, and simultaneously form machine groups and
part families with a lot of ease. Both Heuristic Search and AI based approaches are relatively
new in this area and therefore most of the recent research is utilizing these techniques to handle
the cell formation problem.

Thesis work reported in this Chapter has been published in the following reference:

Manojit Chattopadhyay, Saurav Sengupta, Tamal Ghosh, Pranab K. Dan, & Sitanath
Mazumdar (2012). Neuro-Genetic Impact on Cell Formation Methods of Cellular Manufacturing
System Design: A Quantitative Review and Analysis. Computers & Industrial Engineering, 64,
256-272, DOI: 10.1016/j.cie.2012.09.016.

57

S-ar putea să vă placă și