Computers and Industrial Engineering Vol. 23, Nos 1-4, pp. 97-100, 1992 0360-8352/92 $5.00+0.

00
Printed in Great Britain. All rights reserved Copyright© 1992 PergamonPress Lid

NEURAL NETWORKS APPLICATIONS IN MANUFACTURING PROCESSES

Godwin J. Udo, Ph.D.

Department of Decision Sciences
Tennessee Technological University
Cookeville, TN 38505

ABSTRACT Although several studies have reported success in

applications of neural nets on manufacturin~ no
Advanced manufacturing systems are becoming literature survey exists to help researchers and
too complex and dynamic for traditional rule-based practitioners in this emerging technology. Materna's
decision support systems. Neural network technology [22] statement highlights this problem:
application is gathering momentum in all aspects of Over the last 15 years, a tremendous amount
manufacturing processes. This paper overviews the of literature has been generated on neural
technology and surveys the literature on its application networks, their capabilities, and (more
in manufacturing processes. recently) on how to implement them.
Unfortunately, most of this literature is
INTRODUCTION widely scattered over many different
disciplines and scientific publications, and
Manufacturing systems in industrialized countries there is no single source for locating all.
have dramatically changed as a result of advanced The aim of this study is to present an overview of
manufacturing technologies employed in today's neural networks and to examine the literature which
factory. Factories are now striving to attend and deals with their applications in several aspects of
maintain a worid-class status through automation that manufacturing processes. The study discusses the
is made possible by sophisticated computer programs. characteristics of a neural network that make it
Since the decision-making process in an advanced suitable for different aspects of advanced
manufacturing systems environment is becoming manufacturing systems environments. The rest of this
increasingly difficult and overwhelming to humans, paper is divided into three sections. The first section
artificial intelligence (AI) is becoming widely adopted gives an overview of neural networks. The second
to assist human efforts. There is a wide use of several section discusses the literature, and the third section
types of AI in all aspects of manufacturing processes provides the conclusions and identifies areas for future
but the oldest and the most common type of AI in research.
manufacturing is expert systems (ES). ES is widely
used in design, planning process, scheduling, material OVERVIEW OF NEURAL NETWORKS
handling, quality control, and other operations. Some
authors have identified the strengths and weaknesses Neural computing is an attempt to model the
of ES [1]. ESs are less effective in the ever-changing, functions of biological neural systems. The major
complex, and open system environment of today's purpose of biological neural systems is centralized
manufacturing systems. control of biological functions: energy supply,
The type of AI capable of responding to the rapid metabolism, cardiovascular control, respiration, etc.
changes and ill-structured nature of the automated [18]. Kohonen [18] defines artificial neural networks
manufacturing environment has to be able to learn, as "massively parallel interconnected networks of
adapt to changes, recognize trends, and mimic human simple (usually adaptive) elements and their
thought processes with little human interventions. It hierarchical organizations which are intended to
must be able to recognize patterns and relationships interact with objects of the real world in the same way
among prevailing factors which are not apparent to as biological nervous systems do."
human experts. Neural net technology fits the above A neural network technology, also known as
description to a large extent. Neural net's ability to neurocomputing~ is based on the massive parallelism
recognize patterns and trends in an historic database (a large number of neurons working simultaneously)
is one of the reasons the technology is effectively used and high connectivity of the neurons which
in manufacturing processes, as reported by several differentiate it from° conventional computing.
recent studies [2, 4, 40]. In the last four years, many Conventional computers are made up of circuits that
studies have reported successful applications of neural are programmed for specific functions based on a step-
network technology in many areas including strategic by-step process. These circuits work separately,
planning, financial management, banking and causing commands to be fetched and executed one at
insurance services, engineering, and manufacturing. a time. The sequential processing of conventional

97
98 Proceedings of the 14th Annual Conference on Computers and Industrial Engineering

computers is time-consuming, but the output is exact. and other optimization problems prevalent in
In contrast, the neurons of a neural network work manufacturing systems [14]. Hopfield and Tank [13]
together and simultaneously, producing a good solution used the traditional salesman problem (TSP) as an
within a short time. But a neural network's solution is example to demonstrate how neural networks can
not necessarily the best (optimal). A neuron receives provide a rapid solution to a hard optimization
input signals from several other neurons and, in turn, problem. They mapped TSP on an analog circuit to
passes a single signal to other neurons through demonstrate the computational power and speed of a
weighted connections feeding into a summation box. parallel and massively connected network of neurons.
The output signal passes through several pathways to Dahl [5] applied Hopfield net to solve a map and
provide input signals to other neurons. The graph coloring problem, which concerns deciding
connection between any two neurons is called a whether a K-coloring of the map exists. They first
synapse, through which neurons send and receive proved that a 100-region map problem is NP-Complete
messages from each other. Some of the connections and then transformed it into a TSP before using a
are stronger than others based on the adaptive neural network to obtain a good solution. Page and
coefficients often referred to as weights. Operating Tangliarini [26] used Hopfield net to solve a resource
alone, each neuron has a weak computational strength allocation problem. Neural network provided optimal
but the high connectivity among neurons provides the solutions in some of the problem instances, but in
strength and speed that make neural net second to no other instances, suboptimal solutions were obtained.
other algorithm. In another study, Tangliarini and Page [38] used
The two forms of learning in neural networks are neural network to solve queens' constraints problems.
supervised and unsupervised [3]. In supervised Neural network provided ninety feasible solutions out
learning, a trainer is needed to specify the desired of ninety-two problem instances.
output for a given input. The training process involves Marcus [21] presents an approach considered to be
providing the network with sets of input and the an alternative to Hopfield's net. The approach is a
corresponding outputs. A network is said to have connectionist algorithm, which is capable of
learned if it can produce the desired outputs when a automatically generating analog versions of any
sequence of inputs are provided. Unsupervised problem for Hopfield net. Levy and Adams [20]
learning involves the trainer informing the network of proposed a combinatorial optimization technique that
the input data and of the number of output combines simulated annealing procedure and Hopfield
parameters, and the system then figures out the net. The study shows the hybrid approach to be more
solutions. Neural networks are especially good for efficient than Hopfield. In a different type of
those tasks that conventional computers perform application, Saylor and Stork [36] used a neural
poorly or not at all. Such tasks include recognizing network to model a decision making process. Their
human faces or other patterns, diagnosing medical results indicate that a neural network converges at a
problems, sensor processing, and so forth. shorter time than a conventional method. Recently,
Neural networks possess certain strengths not Udo [39] mapped warehouse problems on neural
found in conventional computers [37]. These strengths network to generate initial solutions for branch and
lead many believe that the technology will lead to a bound algorithm. His results indicated that the hybrid
drastic change in the way we process information. The algorithm (neural network/branch and bound)
advantages of a neural network over conventional outperformed the traditional branch and bound
computing are listed below. technique by 66 percent in terms of computational
time and memory storage requirements.
1. It does not need an expert to represent knowledge
and does not require much programming Scheduling/Maintenance and Repairing
knowledge. Foo [8] used a stochastic neural network approach
2. It has a simpler structure, since only the input and to solve a 4-job 3-machine job-shop scheduling
the output of the system are simulated. problem. Optimal solutions were obtained in some of
3. It has a good ability to tolerate faults and missing the problem instances; a "good" solution was found in
data, making it useful where all the rules and data other problem instances. The integer linear
are not known. programming neural network (ILPNN) was developed
4. It is self-organizing and can learn. These by Foo and Takefuji [9] and used in solving a job-shop
characteristics make it suitable for those tasks that scheduling problem. The results indicated optimal or
were originally done by humans. near-optimal solutions, though optimal solution was
5. The massive number of processing elements not guaranteed by the ILPNN. Rabelo and Alptekin
(neurons) makes neurocomputing faster than [32] developed an intelligent scheduling systems (ISS)
conventional computing. for flexible manufacturing systems in which neural
6. When fully developed, neural hardware is expected networks and expert systems are used to generate good
to be 20 to 50 times smaller in size than schedules. This system ensures proper integration
conventional computers. between ,~cheduling and control, and so enhances
quality decision making, which leads to an efficient
AREAS OF APPLICATIONS and quality manufacturing process.
Noaker [25] had proposed that the technology be
Resource Allocation/Constraints Satisfaction employed in sensing five functions on the factory floor:
The pioneer works of Hopfield and Tank [12, 13] pressure/absence detection, positioning, inspection,
have precipitated several recent studies that apply condition measurement, and condition identification,
neural networks in resource allocation, transportation, with the aim of readily sensing when something is
 UDO: Neural Networks 99

wrong and fixing it before damages are done. that of the traditional method in that temporal
Yamashina et al. [41] employed a neurai network to nonstationary state transition probabilities could be
diagnose failures on a pneumatic servovalve. The revised by a parameter learning paradigm. The neural
study concluded that neural networks utilized majority network simulator (NNSIM) provides a simulation
voting techniques to decrease both the failed-safe and environment that easily interfaces with other
failed-dangerous chances concurrently, and so improve simulators [24]. With NNSIM, it is possible to
the effectiveness of the diagnosis. Ray [33], reports on evaluate and determine how neural network would
a project in an integrated steel industry that applied behave inside a given simulator and how the
neural network in engine fault diagnosis in India. The simulation outputs would be affected.
results obtained from the project indicated that neural McClelland and Rumelhart [23] is a presentation
networks provided the same diagnosis as human of a simulation-based tutorial system used in exploring
experts about 75 percent of the time. In earlier study, parallel distributed processing (PDP) models. The
Ray and Murty [34] reported success in application of PDP model is good for large problems with no known
neural networks on maintenance management solutions. Johnstone [14] cites a number of places that
problem. Charles Bauer [2], reports on a study that use neural networks in simulating the whole social
introduced an unsupervised neural network designed system including the economy, crime, and population
to detect tool failure in milling operations. growth, among other factors. Williams and Brosofsky
[40] applied a neural network in the simulation of an
Process Control and Planning autonomous vehicle, which is capable of sensing its
At Lockheed Missiles and Space Research Center, environment and controlling its movement.
a probabilistic neural network (PNN) is applied to
control active or "rubber" mirrors [31]. Pugh [31] Robotics Control
describes the application of a neural network on Pan and Sobajic [27] developed a neural network
manufacturing process control. Patrick [29] reports on system for controlling robots. They reported a
the application of a neural network to brownstock robotics control algorithm that is computationally
washer operations at a pulp and paper mill in simple and robust because of the application of
Arkansas by Automation Technology. A preliminary massively parallel information processing capabilities
results indicated a 25 percent improvement in solid of the multi-layer neural network they used. All et. al.
variability which led to improved washing efficiency, [1] outlined the features that make neural networks
higher quality, and increased profits. Chryssoloris et more suitable for robotics control than conventional
al. [4] used a neural network to determine operational techniques. Ingiesby [16] presents a proprietary neural
policies for manufacturing systems. network system, GLN (graded learning network), used
in controlling robots' movement and action.
Database Management
Neural network technology is used to manage Quality Control and Machine Vision
database systems [7, 10]. Conventionally, key words Motorola Incorporated uses artificial vision based
are used to retrieve information from databases, on neural network technology for the quality
making it possible for too much or too little infor- inspection of their chips [30]. Port [30] also reports on
mation to be retrieved. Ersoy et al. [7] developed how Anderson Company intends to use a neural
schemes for training neural networks to optimally network in its quality inspection with the aim of
store, recall and process information in the presence improving its performance from 350 per million to
of noisy and incomplete knowledge. Hillman [11] about 100 per million.
describes a database advisor in which neural networks CTS Electronics of Texas has used a neural
and expert systems are used to advise a user on the network system to detect defective loudspeakers
requirements of developing a database. Hillman assembly lines [17]. c r S ' s inspection is more precise,
reported an increase in productivity without a decline faster, and more accurate with neural network
in the quality of the final product. Lelu [19] discusses application. Ford Motor Company uses neural
neural network application in database management network technology to check car paint finishes [17].
and navigation. This application allows a full-text or Ford also hopes to use the technology to diagnose
multimedia browser to be placed in the database in engine problems. English et al. [6] presents a new
such a way that the data can be queried in any method of quality control in a continuous flow process
contexts. using neural network technology. The results of their
Parsaye [28] describes an intelligent database as an initial experiments with the system were said to be
evolution and a merger of many technologies, such as promising.
automatic recovery, hypermedia, high-level user
interface, and intelligent database engine. Kohonen CONCLUSIONS
[18] contains an application of a neural network on
stochastic search of information in knowledge As this literature survey indicates, some of these
databases. Although all the possible or best solutions manufacturing processes have attracted more research
were not guaranteed, neural network provided an attention than others. Some manufacturing processes
acceptable solution for practical purposes. such as design, tool selection, plant layout, and
inventory management, are yet to be explored. There
Simulation is still more studies and investigations to be done to
Sastri et. al. [35] modeled the Markovian decision enhance the application of the technology and to
process with neural network. The neural network achieve the numerous benefits thereof. The
application in this simulation problem was better than conclusions of this survey are:
100 Proceedings of the I4th Annual Conference on Computers' and Industrial Engineering

• There is an intense research interest in neural Neural Networks, 1, pp. 3-16 (1988).
networks. [19] A. Lelu, From data analysis to neural networks: New
• Some of the theoretical claims of neural networks' prospects for efficient knowledge through databases,
abilities have been tested in real-world Journal of Information Science Principles and
manufacturing problems and found to be true. Practice (Netherlands) 17, No. 1, pp. 1-12 (1991).
• Many of the applications reported in the literature [20] B. C. Levy and M. B. Adams, Global optim~tion
are either laboratory experiments or preliminary with stochastic neural networks, IEEE 1st
applications. International Conference on Neural Networks, San
• Some of the studies were prescription of what Diego, CA, pp. III 681-689 (1987).
neural network is capable of doing. [21] R. Marcus, A connectionist algorithm for minimum
cost network flows, IEEE 1st International
• A major limiting factor for an all-out application
Conference on Neural Networks, San Diego, CA, pp.
of the technology is the hardware.
III 735-739 (1987).
• One of the major research needs is that of [22] T. Materna, Neural networks enter high speed market
developing a generalized network for solving place, Teeh. Mgmt's Newsmonthly, 7, No. 7 (1987).
different types of large-scale problems. [23] J.L. MeClelland and D. E. Rumelhart, A simulation-
REFERENCES based tutorial system for exploring parallel distributed
[1] A. L. All, D. L. Aft, and S. Kamal, The processing, Behavior Research Methods Instruments
undeterministic manipulation of solid models for & Computers, 20, No. 2, pp. 263-275 (1988).
robot program synthesis. Computer & Industrial [24] J. Nijhuis, L. Spaanenburg, and F. Warkowski,
Engineering 19, pp. 465-468 (1990). Structure and application of NNSIM: a general
[ 2] C. M. Bauer, NAMRC XIX Report, Manufacturing purpose neural network simulator. Mieroprocessing
Engineering, 107, pp. 55-58 (1991). & Microprogramming, 27, pp. 189-194 (Aug. 1989).
[ 3] E. B. Baum and F. Wilezek, Supervised learning of [25] P. M. Noaker, Smart sensing on the factory floor.
probability distributions by neural networks, Neural Production 100, pp. 42-48 (1988).
Information Processing Systems, D. Z. Anderson, Ed. [26] E. W. Page and G. A. Tagliarini, Algorithm
American Institute of Physics, pp. 52-61 (1988). development for neural networks. Proc. SPIE on
[ 4] G. Chryssolouris, M. Lee, and M. Domroese, The use neural networks, pp. 11-19 (1988).
of neural networks in determining operational policies [27] Y. Pao and D. J. Sobajic, Artificial neural net based
for manufacturing systems. Journal of Manufacturing intelligent robotics control. SPIE Robotics/IECON
Systems, 10, pp. 166-175 (1991). '87 Meeting, Cambridge, MA, November 1987.
[ 5] E. D. Dahl, Neural network algorithm for an NP- [28] K. Parsaye, M. Chignell, S. Khoshafian, and H. Wong,
complete problem: Map and graph coloring, IEEE Intelligent databases. AI Expert, 5, pp. 38-47 (1990).
1st International Conference on Neural Networks, [29] K. L. Patrick, Neural network keeps BSW filtrate
San Diego, CA pp. HI 113-120 (1985). solide at maximum uniform levels. Pulp and Paper,
[ 6] J. R. English, R. John, and T. Sastri, Enhanced 65, pp. 55-58 (March 1991).
quality control in continuous flow processes. [30] O. Port, Adding eagle eyes to those lighning fast
Computers & Industrial Engr. 19, pp. 258-262 (1990). computers. BusinessWeek, 3194, pp. 60B (Dec, 1990).
[ 7] O. K. Ersoy, C. H. Chen, P-T Yu, and P. Yegani, [31] G. A. Pugh, Synthetic neural networks for process
Information processing and storage based on control. Computers and industrial Engineering 17,
transforms and neural networks. IEEE 1st Int. Conf. pp. 24-26 (1989).
on Neural Networks, San Diego, CA. pp. III 777-784 [32] L.C. Rabelo and S. Alptekin, Integrating scheduling
(1988). and control functions in computer-integrated
[ 8] Y. S. Foo, Neural network optimization techniques manufacturing using artificial intelligence. Computers
for job-shop scheduling, Ph.D. dissertation, Dept. of and Industrial Engineering 14, pp. 101-106 (1989).
Electrical and Computer Engineering, University of [33] A. K. Ray, Equipment fault diagnosis--a neural
South Carolina (1988). network approach. Computers in Industry 16, pp.
[9] Y. S. Foo and Y. Takefuji, Integer linear 169-177 (1991).
programming neural networks for job-shop [34] A. K. Ray and M. S. S. N. Murty, Application of
schedulings, IEEE 1st International Conf. on Neural artificial intelligence in maintenance management.
Networks, San Diego, CA, pp. II-341-348 (1987). Proc. 2nd Inf. Conf. on Industrial and Engineering
[10] B.W. Henderson, Neural network computers finding Applications of AI and Expert Systems, ACM
practical applications at Lockheed. Aviation Week & Tullahoma, USA, pp. 112-118 (1989).
Space Technology, 132, pp. 53-55 (1990). [35] T. Sastri, J. English, and M. Krishnamurthi, Modeling
[11] D. V. Hillman, Integrating neural networks and a Markovian decision process by neural network.
expert systems. AI Expert 5, pp. 54-58 (June 1990). Computer & IE, 17, pp. 464-468 (1989).
[12] J.J. Hopfield, Neural networks and physical systems [36] J. Saylor, and D. G. Stork, Parallel analog neural
with emergent collectivecomputational abilities, Proc. networks for tree searching, American Institute of
of National Sciences USA, 79, pp. 2554-2558 (1982). Physics, pp. 392-397 (1986).
[13] J.J. Hopfield and D. W. Tank, Neural computation of [37] P.K. Simpson, Artificial neural systems: Fundations,
decisions in optimization problems, Biological paradigms, applications, and implementations.
Cybernetics, 5, pp. 141-252 (1985). Pergamon Press, New York (1990).
[14] S. J. Johnston, Modelling the future. Far Eastern [38] G. A. Tagliarini and E. W. Page, Solving constraint
Economic Review, 152, No. 32, pp. 70 (1991). satisfaction problems with neural networks, IEEE 1st
[15] S.J. Johnston, Neural networks are making inroads, International Conference on Neural Networks, San
Infoworld, 13 No. 27 pp. 13, 16 (1991). Diego, CA, pp. III 741-747 (1987).
[16] T. Inglesby, No downing, neural networks can help [39] G. J. Udo, Neural network applications to
manufacturing. Mfg. Systems 6, pp. 26-31 (1988). combinatorial optimization problems, Proc. Decision
[17] D. E. Kirrane, Machine learning. Training and Sciences Institute, pp. 491-493 (1991).
Development Journal, 40, pp. 24-29 (1990). [40] B. Williams and R. Brosofsky, An autonomous vehicle
[18] T. Kohonen, An introduction to neural computing, controlled by a neural network simulation.