Int. J. Intelligent Systems Technologies and Applications, Vol. x, No. x, xxxx
Copyright © 20xx Inderscience Enterprises Ltd.
A hybrid neural network approach to cell formation in
cellular manufacturing
Sourav Sengupta, Tamal Ghosh* and
Pranab K. Dan
Department of Industrial Engineering & Management,
West Bengal University of Technology,
BF 142, Salt Lake City, Kolkata 700064 India
Email: sengupta.sourav86@gmail.com
Email: tamal.31@gmail.com
Email: danpk.wbut@gmail.com
*Corresponding author
Abstract: The design of Cellular Manufacturing Systems (CMS) has attained
the significant interest of academicians, researchers and practitioners over the
last three decades. CMS is regarded as an efficient production strategy for
batch type of production. Literature suggests that since the last two decades
neural network based methods have been intensively used in cell formation
problems while production factor such as operation time is merely considered.
This paper presents a new hybrid neural network approach, Fuzzy ART-
Centroid Linkage Clustering Technique (FACLCT), to solve the part machine
grouping problems in cellular manufacturing systems considering operation
time. The performance of the proposed technique is tested with problems from
open literature and the results are compared with the existing clustering models
such as simple C-Linkage, K-Means, modified ART1 and genetic algorithm
and achieved better performance. The novelty of this study lies in the simple
and efficient methodology to produce quick solutions with least computational
efforts.
Keywords: cell formation; group technology; cellular manufacturing; artificial
neural
network; fuzzy adaptive resonance theory; centroid linkage; agglomerative
clustering.
Reference to this paper should be made as follows: Sengupta, S., Ghosh, T.
and Dan, P.K. (xxxx) ‘A hybrid neural network approach to cell formation
in cellular manufacturing’, Int. J. Intelligent Systems Technologies and
Applications, Vol. x, No. y, pp.xxx–xxx.
Biographical notes: Sourav Sengupta is a graduate student in the Department
of Industrial Engineering & Management, West Bengal University of
Technology, India. He obtained his Bachelor of Technology (BTech) degree in
Information Technology from West Bengal University of Technology, India.
His current research area is application of artificial neural network in cellular
manufacturing.
Tamal Ghosh is a graduate student in the Department of Industrial Engineering
& Management, West Bengal University of Technology, India. He obtained his
Bachelor of Technology (BTech) degree in Computer Science and Engineering
from National Institute of Technology Calicut, India. His current research area
is application of soft-computing techniques in cellular manufacturing.
S. Sengupta, T. Ghosh and P.K. Dan
Pranab K. Dan is a Reader in the Department of Industrial Engineering &
Management, West Bengal University of Technology, India. He obtained his
PhD in Production Engineering in 1996 from Jadavpur University, India, and
Bachelor of Engineering and Master of Engineering degrees in Mechanical
Engineering from Bengal Engineering College, Shibpore affiliated to Calcutta
University, India, in 1980 and 1982, respectively. His research and professional
experience is in the area of industrial engineering.
1 Introduction
Over the past three decades, in response to the competitive markets need for increased
industrial automation, product diversification and the trend towards shorter product life
cycles, new manufacturing philosophies have been adopted by many of the established
manufacturing firms. Among those new manufacturing philosophies, group technology
(GT) has been used to reduce throughput and material handling times, to decrease work-
in-progress and finished goods inventories and to increase the ability to handle forecast
errors (Won and Currie, 2007). Group technology can be defined as a manufacturing
philosophy identifying similar parts and grouping them together to take advantage of
their similarities in manufacturing and design (Selim et al., 1998). Cellular manufacturing
(CM) is an application of GT and has emerged as a promising alternative manufacturing
system. CM could be characterised as a hybrid system linking the advantages of both the
jobbing (flexibility) and mass (efficient flow and high production rate) production
approaches. CM entails the creation and operation of manufacturing cells. Parts are
grouped into part families and machines into cells. As reported by Wemmerlöv and Hyer
(1989), the aim of CM is to reduce set-up and flow times and therefore to reduce
inventory and market response times. Set-up times are reduced by using part-family
tooling and sequencing, whereas flow times are reduced by minimising set-up and move
times, wait times for moves and by using small transfer batches. Group technology
addresses issues such as average lot size decreasing, part variety increasing and increased
variety of materials with diverse properties and requirements for closer tolerances. As
described in a review (Venugopal, 1999), the basic idea behind GT/CM is to decompose
a manufacturing system into sub-systems by identifying and exploiting the similarities
amongst part and machines. The very first step in this process is to solve the complex
Part Machine Grouping (PMG) problem and the problem being quite challenging under
real-time scenario, various approaches have been developed, and among which soft
computing approach has an eminent role in the GT/CM literature. Soft computing is
the state-of-the-art approach to artificial intelligence which mostly comprises fuzzy
logic, artificial neural network and evolutionary computing. This paper presents a new
hybrid neural network approach, Fuzzy ART Centroid Linkage Clustering Technique
(FACLCT), to solve the PMG problem in cellular manufacturing systems considering
operation time. In light of the literature survey, it is well understood that very few studies
focus on cell formation considering production factors such as operational time,
operational sequence, batch size, production volume and other factors. In this work, it is
attempted to form the cells considering operation time, a real-time production factor. To
solve such problem the zero-one Machine Part Incidence Matrix (MPIM) is converted
into real valued workload data. The workload represents the operational time required by
A hybrid neural network approach to cell formation
the parts in the machines. The proposed hybrid model has been tested on wide variety of
problems from literature and compared to the solutions obtained from simple C-Linkage,
K-Means, modified ART1 and genetic algorithm in the recent literature.
2 Literature review
Burbidge (1977) viewed group technology as a change from an organisation of people
mainly on process, to an organisation based on completed products, components and
major completed tasks. Since 1960, various approaches were presented to solve the
machine part grouping problem. Initially the methods like Similarity Coefficient Methods
(SCM) (Seifoddini and Wolfe, 1986), graph theory (Rajagopalan and Batra, 1975) and
Rank Order Clustering (ROC) (King, 1980) methods were developed only to group the
similar machines into machine cells while the grouping of parts into part families was
done in the supplementary step of the procedure. Later clustering methods such as the
MODROC (Chandrasekharan and Rajagopalan, 1986), ZODIAC (Chandrasekharan and
Rajagopalan, 1987) MACE (Waghodekar and Sahu, 1984) are reported for solving the
cell formation problems. Since late 1980s soft-computing approaches began to gain
popularity (Venugopal, 1999; Papaioannou and Wilson, 2009) which included artificial
neural network, fuzzy logic and meta-heuristics like simulated annealing (SA) algorithm,
genetic algorithm (GA), tabu search (TS).
2.1 Artificial neural network
Neural networks are massively parallel computer algorithms (Wasserman, 1989) with
an ability to learn from experience. They have the capability to generalise, adapt,
approximate given new information, and provide reliable classifications of data. These
algorithms involve numerous computational nodes that have a high connectivity. Each
of the nodes operates in a similar manner which makes them ideal for a parallel
implementation. During the execution, each node receives an input, processes this
information, and produces an output which is provided as an input to other nodes in the
network. The connections between the nodes, and in particular the learning rules that
modify the strength between the connections, give neural networks their power and
flexibility (Enke et al., 2000). The neural network approach has been the subject of
intensive study by interdisciplinary researchers for a long time. Though neural networks
have been successfully applied in a variety of fields, their use in cellular manufacturing
problems started in the late 1980s and early 1990s. Recognising ANN’s pattern
recognition ability, several researchers began to investigate neural network methods for
the part-machine grouping problem. Neural network is of major interest because when it
is connected to computer, it mimics the brain and bombard people with much more
information.
2.2 Fuzzy adaptive resonance theory
Fuzzy ART proposed by Grossberg (Carpenter et al., 1991) belongs to the class of
unsupervised, adaptive neural networks. Adaptive neural networks always had an
important role in cellular manufacturing beginning in the early 1990s in the works of Kao
and Moon (1991), Malave and Ramachandran (1991), Dagli and Huggahalli (1991) and
S. Sengupta, T. Ghosh and P.K. Dan
Moon and Chi (1992). Dagli and Huggahalli used ART1 in such problems while Malave
and Ramachandran used competitive learning. Fuzzy ART was another common adaptive
resonance framework as presented in the works of Suresh and Kaparthi (1994), Burke
and Kamal (1995), Kamal and Burke (1996), Suresh et al. (1999), Peker and Kara (2004),
Won and Currie (2007) and Ozdemir et al. (2007) which provided a unified architecture
for both binary and continuous valued inputs. Although fuzzy ART does not require a
completely binary representation of the parts to be grouped, it possesses the same
desirable stability properties as ART1 and a simpler architecture than that of ART2.
Figure 1 shows the architecture of the fuzzy ART network (Chang et al., 2005). It
consists two layers of computing cells or neurons, and a vigilance sub-system controlled
by an adjustable vigilance parameter. The input vectors are applied to the fuzzy ART
network one by one. The network seeks for the ‘nearest’ cluster that ‘resonates’ with the
input pattern according to a ‘winner-take-all’ strategy and updates the cluster to become
‘closer’ to the input vector. In the process, the vigilance parameter determines the
similarity of the inputs belonging to the same cluster. For the same set of inputs, the
similarity of elements in one cluster grows as the vigilance parameter increases, leading
to a larger number of trained clusters. The choice parameter and the learning rate are two
other factors that influence the quality of the clustering results. In this paper, fuzzy ART
is used to form the part families while agglomerative centroid linkage-hierarchical
clustering algorithm is used to form the machine groups. The detailed description of the
hybrid algorithm is discussed in the next section.
Figure 1 Topological structure of the fuzzy ART architecture
Y
1
Y
2
Y
3
X
1
X
2
X
3
Y
M
X
N
td
j
i
Z
bu
ij
Z
F1
F2 (WTA)
T
j
X
ρ l
RESET
l
1
l
2
l
3
l
N
Comparator
3 The proposed hybrid approach
This study presents a hybrid FACLCT, a new pattern recognition neural network
approach, for clustering problems, and illustrates its use for machine cell design in group
technology. FACLCT is a bimodal clustering model. While mode1 is concerned with the
identification of part families using the fuzzy ART architecture, mode2 is concerned with
the formation of machine groups using the centroid linkage-agglomerative hierarchical
clustering algorithm. The fuzzy ART neural network was introduced by Carpenter et al.
A hybrid neural network approach to cell formation
(1991) and Suresh and Kaparthi (1994) implemented it to the CF problem. The latter
found that in terms of bond energy recovery, fuzzy ART outperformed ART1 and
ART1/KS. The execution time of fuzzy ART was higher than ART1 and ART1/KS, but
for larger datasets, execution times were significantly lower than DCA and ROC2. The
fuzzy ART neural network involves several changes to ART 1: (a) non-binary input
vectors can be processed; (b) there is a single weight vector connection (w
ij
); and (c) in
addition to vigilance threshold (ρ), two other parameters have to be specified; a choice
parameter (α) and a learning rate (β). The step-by-step illustration of fuzzy ART network
is as follows (Suresh and Kaparthi, 1994):
Step 1: Initialisation
Connection weights: w
ij
(0) = 1.
0 ≤ I ≤ N – 1, 0 ≤ j ≤ (M – 1)
Select values for: α > 0, β ε (0, 1), ρε (0, 1)
Step 2: Read a new input vector I consisting of binary or analogue elements
Step 3: Compute choice function (T
j
) for every input node
T
j
= ||I ^ wj|| / [α + ||wj||], 0 ≤ j ≤ (M – 1),
where ^ is the fuzzy AND operator, defined as: (x^y) = min (x
i
, y
i
)
Step 4: Select the best matching exemplar.
T
Ө
= max {T
j
}
Step 5: Resonance test:
If ||I^w
Ө
|| / ||I|| ≥ ρ then go to step 7 otherwise go to step 6
Step 6: Mismatch reset: set T
Ө
= –1 and go to step 4
Step 7: update best matching exemplar (learning law)
w
Ө
new
= [β × (I^w
Ө
old
)] + [(1 – β) × w
Ө
old
]
Step 8: Repeat: go to step 2.
The above algorithm although could produce efficient clustering solutions, the literature
suggests that hybrid approaches often produced better clusters. To establish the fact,
centroid linkage-hierarchical clustering algorithm is integrated to the fuzzy ART neural
network to form the machine group based on the part families formed by the neural
model.
Hierarchical Agglomerative Clustering (HAC) is conceptually and mathematically
simple algorithm practised in clustering analysis of data (Anderberg, 1973).[AQ1] It
delivers informative descriptions and visualisation of potential data clustering structures.
When there exists hierarchical relationship in data this approach can be more competent.
The algorithm in contrast to machine grouping is presented below.
3.1 Step 1: formation of input dataset
An input dataset for HAC is a machine–part incidence matrix. Machines are the items
that should be grouped based on their similarities. Parts are the components that contain
routing information. The type of input dataset can be classified into binary data (contains
only 0 or 1, i.e. the routing information) and ratio data (contains information about
production volume, operation time). Figure 2 shows a 5 × 7 binary dataset.
AQ1: The references
flagged with [AQ1]
are not included in the
reference list. Please
provide the complete
reference details to
include in the
reference list.
