Anda di halaman 1dari 17

212

IEEE TRANSACTIONS ON COMPONENTS, PACKAGING, AND MANUFACTURING TECHNOLOGY-PART

A, VOL. 17, NO. 2, JUNE 1994

Artificial Neural Networks in Manufacturing: Concepts, Applications, and Perspectives


Samuel H. Huang and Hong-Chao Zhang
Abstract- New approaches and techniques are continuously and rapidly introduced and adopted in todays manufacturing environment. Recently, there is an explosion of interest in applying artificial neural networks to manufacturing. Artificial neural networks have several advantages that are desired in manufacturing practice, including learning and adapting ability, parallel distributed computation, robustness, etc. There is an expectation that neural network techniques can lead to the realization of truly intelligent manufacturing systems. This paper introduces the basic concepts of neural networks and reviews the current application of neural networks in manufacturing. The problems with neural networks are also identified and some possible solutions are suggested. We hope that the material presented in this paper can provide useful guidelines and references for the research and implementation of artificial neural networks in the field of manufacturing.

I. INTRODUCTION ANUFACTURING is a set of correlated operations and activities, which includes product design, material selection, planning, production, inspection, management, and marketing of the products, for the manufacturing industries [ 11. Computers are commonly used in modern manufacturing practice. In the late 1960s, computers were first used for the direct control of groups of machine tools. In the 1970s, the concept of flexible manufacturing system (FMS) was introduced. An FMS is a computer-controlled grouping of semi-independent work stations linked by automated materialhandling systems [2]. Systems of this type are capable of producing a variety of products automatically. They are now widely used in the manufacturing industry. Currently, the development of computer-aided manufacturing systems is advanced to the stage where computer integrated manufacturing (CIM) systems are in a rapid, world-wide proliferation. A CIM system results when the design effort includes the use of computers to achieve an integrated flow of manufacturing activities, based on integrated information flow that links together all organizational activities [3]. It is becoming clear that the next stage will be that of intelligent manufacturing systems (IMS). The systems belonging to this stage may be characterized by their ability to solve problems without either a detailed, explicit algorithm available for each solution procedure, or all the facts, mathematical relationships and
Manuscript received November 15, 1993; revised February 5, 1994. This paper was presented at the International Electronics Manufacturing Symposium, Santa Clara, CA, October &6, 1993. The authors are with the Department of Industrial Engineering, Texas Tech University, Lubbock, TX 79409-3061 USA. IEEE Log Number 9401776.

models available in perfect arrangement and complete form for a deterministic (and unique) answer to be found [4]. Since the decision-making process in an advanced manufacturing system environment is becoming increasingly difficult and overwhelming to humans, artificial intelligence (AI) is widely adopted to assist human efforts. Artificial intelligence is an unfortunate choice of title for an ill-defined technology that may, in the long run, be the single most important and most pervasive ingredient for the realization of true computerintegrated manufacturing (CIM) [5]. AI has provided several techniques with applications in manufacturing. Knowledgebased expert systems are the most popular AI techniques in the 1980s. Recently, there is an explosion of interest in applying artificial neural networks to manufacturing. Artificial neural networks are a type of AI that have the potential of increasing the product quality, reducing the reaction time of a manufacturing system, improving system reliability and enhancing its intelligence [6]. Since the late 1980s, hundreds of papers have been published pertaining to neural network applications in manufacturing. Most of them are widely scattered over many different disciplines and publications. This makes it very difficult for one to locate all the information necessary for using artificial neural networks in manufacturing. A paper that can help researchers and practitioners in applying this emerging technology is highly desired. The objective of this paper is to present the concepts, applications, and perspectives of artificial neural networks in manufacturing, thus providing some useful guidelines and references for the research and implementation. Section 2 of this paper reviews the development of neural network techniques and introduces the basic concepts of artificial neural networks. Section 3 gives a comprehensive survey of neural network applications in manufacturing. Section 4 identifies the problems with artificial neural networks and suggests that neural network techniques should be combined with knowledge-based expert system techniques in order to develop intelligent manufacturing systems. This is followed by a conclusion section.
11. ARTIFICIAL NEURAL NETWORKS A. Historical Overview

The development of artificial neural networks has an interesting history. In the 1940s, scientists found out that the physiology of the brain was similar to the processing mode used by electronic computing devices (computers). In both cases, large amounts of data are manipulated. In computers, the

0162-8828/94$04.00 0 1994 IEEE

HUANG AND ZHANG: ARTIFICIAL NEURAL NETWORKS IN MANUFACTURING


inputs OUtDUf

213

processing element is the bit, while in the brain, the processing element is the neuron. This phenomenon suggests that research in the design of brain-like processing might be interesting. In 1943, McCulloch and Pitts [7] outlined the first formal model of an elementary computing neuron. Their neuron model laid the groundwork for future development. In 1949, Hebb [8] stated that information can be stored in connections and proposed a learning scheme for updating a neurons connections. His learning rule (known as the Hebbian leaming rule) made a primary contribution to the development of artificial neural networks. In 1954, Farley and Clark [9] set up models for adaptive stimulus-response relations in random networks. Their theories were further elaborated by several researchers. Rosenblatt [ 101 invented the perceptron, which is a trainable machine capable of learning to classify certain patterns by modifying connections to the threshold elements. Caianiello [ l l ] outlined a theory for thinking machines. Widrow and Hoff [ 121 introduced a device called ADALZNE (for ADAptive LINEar combiner). A learning rule (the popular Widrow-Hoff leaming rule) was then devised for neural networks with multiple adaptive elements [ 131. Other early works include Starks mode-seeking technique [ 141 and Steinbuchs Learning Matrix [15]. A book on learning machines by Nilsson [16] clearly summarizes many of these early developments. The book also formulates inherent limitations of learning machines with modifiable connections. During the 1950s, the first neurocomputer was built and tested [ 171. Many implementations of neural computers were realized in the 1960s. However, the existing machine learning theorems of that time were too weak to support more complex computational problems. The computational resources available were also relatively modest at that time. As a result, neural network research entered a stagnation phase. In 1969, Minsky and Papert [ 181 criticized the existing neural network research as being worthless in their book entitled Perceptrons. It has been claimed that the pessimistic views presented in this book discouraged further funding for neural network research for several years. Instead, funding was diverted to further research of knowledge-based expert systems, which Minsky and Papert favored. It is only after the mid-1980s that neural networks have begun to make a strong comeback. The renaissance of neural networks is the result of rigorous works by a handful of researchers. In Japan, Amari [19], [20] studied the learning in networks of threshold elements and the mathematical theory of neural networks. Also in Japan, Fukushima [21], [22] developed a class of neural network architectures known as Cognitrons and Neocognitrons. In Finland, Kohonen [23], [24], [25] pursued the research on associative memory. He also developed unsupervised learning networks for feature mapping into regular arrays of neurons [26]. Carpenter and Grossberg [27], [28], [29], [30] introduced a number of neural network architectures and theories and developed the Adaptive Resonance Theory (ART). Hopfield [31], [32] developed a recurrent neural network architecture for associative memories in the early 1980s. More recently, Kosko [33] extended some of the ideas of Hopfield and Grossberg to develop his adaptive bidirectional associative memory (BAM) network.

X.

y naptic cLnnections

processing element

Fig. 1.

An artificial neuron.

Hinton [34], [35] developed a probabilistic neural network model known as the Boltzmann Machine. Other significant contributors include Anderson [36], Werbos [37], Rumelhart [38], etc. The overview provided above is by no means complete. Readers who are interested in more details may refer to [39], [40], [41]. One more thing worth mentioning here is the publication of a two-volume set of books on parallel distributed processing. It is believed that these two books, edited by Rumelhart and McClelland [42], revitalized the field of neural network research. Beginning in 1986-87, many new research programs in neural networks were initiated. The intensity of research in neural networks can be measured by a quickly growing number of conferences and journals devoted to this field.

B. Basic Concepts

Artificial neural networks have been defined by Kohonen as massively parallel interconnected networks of simple (usually adaptive) elements and their hierarchical organizations which are intended to interact with the objects of the real world in the same way as biological nervous system do 1431. Neural networks attempt to achieve good performance via dense mesh of computing nodes and connections. They also go by many names such as connectionist models, parallel distributed processing models, and neuromorphic systems. We will use the term neural networks or simply networks in the following discussion. Neural networks have the following main benefits [61: 1) processing speed through massive parallelism; 2) learning and adapting ability by means of efficient knowledge acquisition; 3) robustness with respect to fabrication defects and different failures; 4) compact processors for space- and power-constrained applications, etc. The basic components of a neural network are nodes (neurons) and weights (connections). The adjustable weights correspond to biological synapses. A positive weight represents an excitatory connection. A negative weight represents an inhibitory connection. The weighted inputs to a neuron are accumulated and then passed to an activation function which determines the neurons response. Fig. 1 shows an artificial neuron with weighted connections.

214

IEEE TRANSACTIONS ON COMPONENTS, PACKAGING, AND MANUFACTURING TECHNOLOGY-PART A, VOL. 17, NO. 2, JUNE 1994

t
output layer

i'i'
hidden layer

Y1

YM

input layer

Fig. 2.

Neural networks: (a) a back-propagation network, (b) a Hopfield network, and (c) an ART network.

Neural Netwrok Modek

Unsupe&ised Hopfield Network Network

Associative Memory CounterPropagation 1 Network BackNetwork Propagation Self-organizing Network Feature Maps

Memory

Temporal Associative

Multidirectional Associative Memory

Autoassociative Memory

ART network

Fig. 3.

A taxonomy of neural network models.

The output of a neuron is given as o = f(net) in which


(1)

f ( n e t ) is referred to as an activation function. Typical activation functions used are 2 f ( n e t ) = 1 e--X.net - 1 (3)
+

or

recurrent networks. In a feedforward network (Fig. 2(a)), signals propagate in only one direction from an input stage through intermediate neurons to an output stage. While in a recurrent network (Fig. 2(b) and (c)), signals may propagate from the output of any neuron to the input of any neuron. Neural networks are able to store experiential knowledge through learning from examples. They can also be classified in terms of the amount of guidance that the learning process receives from an outside agent. An Unsupervised learning network learns to classify the input into sets without being told anything. A Supervised learning network adjusts weights on the basis of the difference between the values of output units and the desired values given by the teacher, given an input pattern [45].

in which the parameter X (A > 0) is used to control the steepness of the continuous function f ( n e t ) near n e t = 0. There are also other types of activation functions that can be used. Readers may refer to [44], [45] for more details about model neurons and activation rules. A neural network is a collection of interconnected neurons (Fig. 2). A neural network may be distinguished on the basis of the directions in which signals flow. Basically, there are two types of networks, i.e., feedforward networks and

C. Network Models
Today, more than a dozen neural network models have been developed. For the purpose of study, analysis, understanding, and utilization, it is necessary to provide an overview and taxonomy of neural network models. Fig. 3 summarizes the taxonomy of the most important network models. Since this paper is focused on the application of neural networks in manufacturing, we have no intention of discussing every network model in detail. Instead, we will briefly in-

HUANG AND ZHANG. ARTIF'ICIAL NEURAL NETWORKS IN MANUFACTURING

215

evolves to an equilibrium state. Therefore, one may design a neural network for function minimization by associating variables in an optimization problem with variables in the energy function. Hopfield and Tank illustrated the use of the energy function to configure networks for several optimization applications including the traveling salesman problem [47], a signal processing problem [48], and a linear programming problem [48]. In manufacturing applications, Hopfield networks are typically used for solving combinatorial optimization problems such as scheduling problems. ART networks are neural networks based on the Adaptive Resonance Theory (ART) provided by Carpenter and Grossberg [29]. They belong to the class of unsupervised 8E learning networks. An ART network selects its first input as Wij(t 1) = W i j ( t ) - 7 ) 8Wij the exemplar for the first cluster. The next input is compared in which 17 is a user-selected, positive constant called learning to the first cluster exemplar. It is clustered with the first if the rate. The error back-propagation is a typical supervised learn- distance to the first cluster is less than a threshold. Otherwise ing procedure. The learning procedure of a back-propagation it is the exemplar for a new cluster. This procedure is repeated for all following inputs. If an input is clustered with the jth network is as follows: cluster, the weights of the network are updated according to 1) Initialize the weights of the network at small random the following formulae values. 2) Start the learning cycle by exposing the network to a certain input pattern paired with the desired output. (7) 3) Compute the network's output and compare it with the Vij(t 1) = X i V i j ( t ) desired output so that the error can be calculated. 4) Adjust the weights of the network using the error backpropagation algorithm so that a certain amount of the where i = 1, 2,. . . ,M. Carpenter and Grossberg's ART2 detected error is removed. architecture, which works for analog and binary input pattern 5 ) Repeat step (2), (3), and (4) with all the input pat- as well, embodies solutions to a number of design principles terns and their correspondent desired outputs (training such as the stability-plasticity tradeoff and the search-direct examples), compute the cumulative error. access tradeoff. ART networks use an unsupervised learning 6) If the cumulative error is within a tolerable range, procedure. They are stable because new input patterns do terminate the training process; otherwise, go back to not wash away previously learned information. They are also step (2). adaptive because new information can be incorporated until Back-propagation networks can be applied to almost all the full capacity of the architectures is utilized. ART networks applications in the manufacturing domain. In fact, they are provide a fertile ground for gaining a new understanding of the most popular neural network models in manufacturing biological intelligence. They also suggest novel computational theories and real-time adaptive architectures with promising applications. Hopjield networks are recurrent networks introduced by properties for tackling some of the outstanding problems Hopfield [31], [32]. The earliest Hopfield network, which in computer science and technology [29]. In manufacturing employs two-state (odoff) neurons, is used for the design of applications, ART networks have been used in solving group neural content-addressablememories [311. Hopfield later intro- technology (GT) and engineering design problems. It seems duced a modified version of his earlier model that employed a the family of the ART networks can act as very valuable continuous nonlinear function to describe the output behavior modules in intelligent manufacturing systems, where real-time of the neurons. The neurons represent an approximation to classification abilities with self-learning and self-organizing biological neurons in which a simplified set of important com- capabilities are required [6]. putational properties is retained. This neural network model, which consists of nonlinear graded-response model neurons organized into networks with effectively symmetric synaptic 111. MANUFACTURING APPLICATIONS connections, can be easily implemented with electronic devices The development of computer-aided manufacturing systems [46]. A Hopfield network can be characterized by its energy has evolved to the phase of computer integrated manufacfunction, as shown below: turing (CIh4). It has been proposed that the next phase will be that of intelligent manufacturing systems (IMS). As a I n E= ~ w i j v i v - ki i g t i v i . j iv (6) trend, manufacturing systems are demanding more and more 2 i = l j=1 i=l i=l flexibility in product design, process planning, scheduling, i#j process control, and quality assurance. This may be achieved The network will seek to minimize the energy function as it by building intelligent systems that can adapt to changes in

troduce three types of network models that are very popular in manufacturing applications. There are back-propagation networks, Hopfield networks, and ART networks. Readers may refer to [41], [44] for detailed information about different network models and learning algorithms. Back-propagation networks are multi-layered feedforward neural networks that apply the error back-propagation procedure [38] for learning. The back-propagation procedure uses a gradient descent method which adjusts the weight in its original and simplest form by an amount proportional to the partial derivative of the error function with respect to the given weight. The formula used for the adjustment is

--r

216

IEEE TRANSACTIONS ON COMPONENTS. PACKAGING, AND MANUFACTURING TECHNOLOGY-PART A, VOL. 17, NO. 2, JUNE 1994

their environment. The first effort in this endeavor is the application of knowledge-based expert systems. Since the late 1970s, knowledge-based systems have been used in various manufacturing domains [49]. However, they are less effective in the ever-changing, complex, and open system environment of today's manufacturing systems [50]. On the other hand, neural networks emerged as a revolutionary technology for solving problems that are intractable or difficult for traditional computation. Successful applications of this technology in pattern matching and speech recognition generated enormous interest in the manufacturing arena. Since the late 1980s, neural network applications in manufacturing have been reported. The application area of neural networks in manufacturing is surprisingly broad. It covers nearly all of the fields spreading from the design phase through process planning, scheduling and process monitoring to quality assurance. This section provides a comprehensive survey of neural network applications in manufacturing.
A. Design
Fig. 4.

1
1

Needs

I
I

Functional Requirements

KnownElemental Structures

Designer'sStored Patterns of Known Designs

Physical Device

Mutation

I
Physical Device

Design model based on associative memory [54].

Retrieval of old product designs that meet current requirements on geometrical and/or technical information is a problem that is often encountered in batch manufacturing systems. Venugopal and Narendran [51] modeled the design retrieval system as a human associative memory and used a Hopfield network to develop a design retrieval system. The system was verified with test cases on rotational as well as non-rotational parts. The results show that neural network methodology is a promising tool for the development of practical design retrieval systems. The use of neural networks for design data retrieval was also studied by Kamarthi et al. [52]. Instead of Hopfield networks, a back-propagation network was used. The result is also promising. Kumara [53], [54] suggested an associative memory based modeling procedure for conceptual design. The motivation for their research stems from the following discussion [54]: It is possible that the designer may be aware of the structures that satisfy a particular set of functions. In his memory the designer may have stored the representations of a number of physical devices (design solutions). Given a (a set of) functional requirement(s) the designer will be able to identify a structure or a set of structures that will satisfy the required function(s) and hence by associating these structures with physical devices stored in his memory, he can selectively retrieve those designs. It must be noted that the physical devices could be from different domains. After having retrieved the designs, the designer mutates them to come up with new and creative design solutions. The author used a back-propagation network for modeling the associative memory. A design model based on associative memory is also proposed to capture the conceptual design process (Fig. 4). The application of neural networks in design also has been studied by quite a few other researchers. Among them, Coyne and Postmus [55] explored the application of neural networks to simple spatial reasoning in computer-aided design.

Arai and Iwata [56] suggested a four-layer neural network to connect lower level items to upper level ones on the design specification step of the conceptual design phase. Kim et al. [57] applied a neural network approach for engineering drawing with geometrical constraints. Dhingra and Rao [58] examined a new conceptual framework for solving design optimization problems based on a neural computing paradigm. Another interesting study was conducted by Chovan and Waldron [59]. The authors proposed a cognitive model of the transformation from perceived form to function based on findings from a behavioral study of expert mechanical designers when they were reading two dimensional mechanical drawings. The model was simulated using an ART network and was exercised and compared with findings from the behavioral study. The top-down and bottom-up reasonings exhibited by the subjects can be easily represented in the ART network. Their results show that ART networks might be useful for representing the behavioral system. The evidence demonstrated in their paper may provide useful information for the application of neural networks as the research community works toward the development of intelligent computer-aided systems for engineering design.
B. Process Planning

The first attempt to use neural network techniques in process planning might be that of Osakada et al. 1601, 1611, [621. The authors applied neural network techniques to an expert system for the process planning of cold forging in order to increase the consultation speed and to provide more reliable results. A three-layer neural network is constructed to relate the shapes of rotationally symmetric products to their forming methods. The shapes of the products are transformed into 16x 16 black and white pixels and are given to the input layer of the neural network. The back-propagation algorithm is employed. After training, the network is able to determine the forming methods for the products which are exactly the same or slightly

HUANG AND WANG: ARTIF'ICIAL NEURAL NETWORKS I MANUFACTURING N

217

different from those used as training examples. To exploit the self-learning ability, the authors further applied neural network techniques to the prediction of the most probable number of forming steps by considering the shape complexity and material property, the prediction of the die fracture and surface defect in the formed product, and the generation of rules from the knowledge acquired from an FEM (Finite Element Modeling) simulation. It is found that the prediction of the most probable number of forming steps can be made successfully and the FEM results are represented better by the neural network than by statistical methods. Hwang and Henderson [63] applied a perceptron network in feature recognition, which is the first step in automated process planning (i.e., to interpret the design data from a CAD model). The goal of feature recognition is to convert a low-level representation such as face, edge, vertex to a semantically higher feature-based model. The network training is accomplished by manually presenting exemplars of features the user considers important in an engineering analysis (for example, manufacturing-relatedfeatures for process planning). Their results show that the neural network approach took less time in feature recognition than other traditional approaches. Knapp and Wang [U], [651 applied neural network techniques for the automatic acquisition of process planning knowledge. In their approach, two cooperating neural networks are utilized. The primary network is a three-layer backpropagation network. The second fixed-weight network utilizes the MAXNET architecture. Parts to be planned are decomposed into machining features such as slots, holes, and planes. Each feature type is associated with a set of characterizing attributes such as dimensions and tolerances. Each feature is represented by a vector whose elements identify the feature type and its attribute values. This vector forms the input pattern to the primary network. The network responds to the presentation of a feature vector by activating certain output nodes, corresponding to the proposal of particular machining operations. The response of the network is trained using example process plans and the back-propagation learning algorithm. The second network is used to force a decision between competing operation alternatives. Its output is then fed back to the input layer of the primary network to provide a context for deciding the next operation in the machining sequence. The part is presented to the neural network feature by feature, the network then generates a sequence of operations for machining each feature of the part independently (global sequencing of operations across all features is not considered). Process planning is knowledge-intensive in nature. Neural networks, being a useful knowledge acquisition tool, are expected to play an important role in process planning. However, process planning is a challenging task due to the interdependencies among the steps in a plan. It is not clear how neural networks can reason about the causal relationships among the entities in a process plan. Compared with symbolic systems (e.g., knowledge-based expert systems), neural networks are less effective in representing structured, contextual knowledge [66]. Therefore, neural networks should be integrated with symbolic systems in order to solve process planning problems.

C. Scheduling
Foo and Takefuji [67] developed the Integral Linear Programming Neural Network (ILPNN) and used it to solve jobshop scheduling problems. In a job-shop scheduling problem, the resources are typically machines and the jobs are the basic tasks that need to be accomplished using the machines. Each task may consist of several subtasks related by certain precedence restrictions. This problem can be formulated as a linear programming problem. The cost function to be minimized is defined as a sum of the starting times of all jobs subject to compliance with precedence constraints. The problem is solved using a linear programming network. Vaithyanathan and Ignizio [68] investigated the use of neural networks for solving certain types of large-scale, resource constrained scheduling problems. Their work was focused on dynamic resource constrained scheduling problems. Such problems, so characteristic of real world situations, involve the determination of a schedule subject not only to limited resources but also to sudden, unforeseen changes. They first decomposed the problem to be solved into a series of multidimensional knapsack models and established an equivalent neural network model for each particular representation. Then, they developed an approach that ultimately served to solve the original problem by extending the work of Hopfield and Tank [47]. Their approach, to a great extent, avoided common neural network difficulties such as instability and local minima. In addition to Hopfield networks, feedforward backpropagation networks can also be used in solving scheduling problems. Yih er al. [69] provided a hybrid method that combines back-propagation neural network, simulation, and semi-Markov optimization to solve the crane scheduling problem. The crane scheduling problem occurs in a circuit board production line where one overhead crane is used to transport jobs through a line of sequential chemical process tanks. Because chemical processes are involved in this production system, any mistiming or misplacing will result in defective jobs. The proposed method consists of three phases: data collection, optimization, and generalization. Training data are purified using an operations research method (semi-Markov optimization). The neural network is used in the building of the decision making model. The resulted system performed better than the human scheduler from whom the models were formulated. Other neural network approaches to scheduling problems reported in the literature include: time table scheduling [70], real-time scheduling [7 11, multiple-job scheduling [72], assembly scheduling [73], robot scheduling [74], a stochastic neural network (Gaussian machine) for scheduling [75], and an intelligent scheduling system (ISS) for flexible manufacturing systems where neural networks and expert systems are used to generate good schedule [76], [77]. The application of neural networks in scheduling has been studied by many researchers. In some cases, the neural network approach holds significant advantages. In others, the usefulness and effectiveness of the neural network approach is debatable. However, at the very least, neural networks present a legitimate alternative to the more conventional methods for

218

IEEE TRANSACTIONS ON COMPONENTS. PACKAGING, AND MANUFACTURING TECHNOLOGY-PART A, VOL. 17, NO. 2, JUNE 1994

I
Error

m
f d v
(Current Optimal Inputs) Current Net Parameters

LEARNING MODE

Current Optimal Inputs

Current

SYNTHESIS MODE
Fig. 5. Incremental learning and synthesis approach for process modeling and control [78].

scheduling. Since the scheduling problem is one of immense importance, it is certain that considerable efforts will continue in applying neural networks in scheduling.
D. Process Modeling and Control

For the control of machining processes, perhaps the greatest difficulty is that reliable models of the processes do not exist. Rangwala and Dornfeld [78] suggested neural networks as learning structures for intelligent controllers that can model machining processes and use adaptive control. They also proposed an incremental learning and synthesis approach as shown in Fig. 5. This approach relies only on sensor data generated by the manufacturing process and does not require a priori process model. Training samples are composed from the inputs and measured variables during operation and are continually used to enhance the network performance; while the knowledge embedded in the partially learned system is used for the determination of the local optimum of the input variables. In the process industry, usually human operators rather than mathematically based advanced algorithms are used to achieve product control. Gingrich [79] discussed a methodology that uses neural networks for capturing the knowledge of human process operators. The results show that it is possible to learn the control actions of a human operator using neural networks, at least for simple systems. The neural network approach is preferred over heuristic approaches for several reasons including: (1) there is no need for the operator to formulate his knowledge as rules, and (2) to train a neural network is easier than to design, build, and maintain an expert system. Anderson et al. 1801, 1811 applied the neural network approach in arc welding. Any arc welding process is controlled by a number of parameters, and the ultimate objectives of the

process are specified in terms of numerous parameters as well. As a result, any arc welding process can generally be viewed as a multiple-inputlmultiple-output system. The lack of reliable, general, and yet computationally fast physical models of this multi-variable system, makes the design of a generalized realtime controller for arc welding nontrivial. One of the basic issues to be considered regarding the generalized control for arc welding is to determine what is to be controlled and which parameters are accessible to enact control actions on the process. The back-propagation networks were used for the modeling and control of the Gas Tungsten Arc Welding (GTAW). Various configurations, in terms of the number of layers and the number of network nodes, were tested. For the application presented, two-layer (not counting the input layer) networks consisting of a single hidden layer and an output layer have been proved to be adequate. Smartt et al. [82] also applied the neural network approach in arc welding. Instead of using neural network to model the arc welding processes, they developed a new approach to quantify conditional logic rules and represent them in a neural network. The network was then used in place of a PI controller for a gas metal arc welding process. Process modeling and control are a manufacturing domain where neural networks can play a very important role. Chryssolouris and Guillot 1831 combined process modeling and AI techniques for the determination of the operational range of the process parameters. The process models were determined by neural network learning techniques. Jalel et al. 1841 applied neural networks in imitating a skilled process operator. h g h [85] described the application of a neural network on manufacturing process control. Levin et al. [86] discussed the use of neural network architecture for adaptive system modeling and control. Patrick 1871 reported the application of a neural network to brownstock washer operations at a pulp and paper mill. Karsai et al. [88] applied neural network methods for the modeling and control of welding processes. Guha [89] presented some neural network-based adaptive control strategies that can be used for designing controllers for continuous process control problems. Madey et al. 1901 used neural networks and general-purpose simulation to model a continuous improvement system. The advantages of using neural network approaches in process modeling and control are indicated by Chryssolouris et al. [91]. The authors pointed out that while synthesis of multiple sensor information would provide better result, neural networks excelled in dealing with situations in which process models do not adequately reflect the process complexity. A neural networks ability to learn a control algorithm without the benefit of a priori analysis or modeling can be of great help for difficult, complex, and nonlinear control applications. The same conclusion also can be found in 1921, in which Bozich and Mackay applied a neurocontroller in vibration cancellation.
E. Monitoring and Diagnosis

The first application of neural networks in monitoring reported in the literature could be that of Govekar et al. [93]. The authors applied a back-propagation network for

HUANG AND W A N G : ARTIFICIAL NEURAL NETWORKS IN MANUFACTURING


PROCESS OUTPUTS

219

PROCESS INPUTS

b)

A
SENSORY SYSTEM SKILLED OPERATOR

c)

Sensory inputs that include data from a variety of sensors such as force, acoustic emission, accelerometers, etc. The operator modifications to feed and cutting

I
Fig. 6.

1
ARTIFICIAL NEURAL

speed.
2) Operation Mode: Upon convergence, the network is fed with the process and the sensory input and according to the mapping (model) established during the training session, it automatically modifies the cutting parameters. Many research efforts have been dedicated to the application of neural networks in monitoring and diagnosis. Sorsa et al. [99] reported the use of neural networks in process fault diagnosis. Burke and Rangwala [lo01 discussed a neural network approach for tool condition monitoring in metal cutting. Ray [loll reported on a project in an integrated steel industry that applied neural network in engine fault diagnosis. Bauer [lo21 reported on a study of unsupervised neural network for tool failure detection in milling operations. Elanayar and Shin [ 1031 applied neural network techniques to estimate the evolution of flank and crater wears using feed and cutting force measurements. Wasserman et al. [lo41 used the neural network approach to detect and measure small cracks in the shafts of rotating machines. Guillot and El Ouafi [lo51 applied a three-layer feedforward neural network in the identification of tool breakage in metal cutting processes. Wu et al. [lo61 presented a neural network approach to diagnose processing damages in injection molding. Lin et al. [ 1071 used neural network techniques for sensor failure detection and data recovery. Chow et al. [lo81 discussed a neural network approach to real-time condition monitoring of induction motors. Knapp and Wang [1091 applied a back-propagation neural network for machine fault diagnosis. Upadhyaya and Eryurek [110] used neural networks for sensor validation and plant monitoring. Murphy and Kagle [ 1 1 11 applied neural networks for the recognition of electronic malfunctions. As a logical consequence of these and similar investigations, some authors reported on the development of neuro monitoring system using accelerator cards for neural network computations [94], [112], [113]. Other authors [97] argued that combining an expert system with neural network based pattem recognition might lead to potential applications in manufacturing process monitoring and fault diagnosis.
F. Quality Assurance
Quality is the single most important factor in determining market share [114]. There are two different approaches to quality assurance, namely, reactive quality assurance and proactive quality assurance. Reactive tools include sampling plans, lot acceptance determination, scrap or rework analysis, and so forth. Proactive strategy requires. an emphasis on physical cause-effect knowledge, risk analysis, experience, and judgment to justify action. Neural networks have been used for both reactive and proactive quality assurances. Reactive quality assurance is strongly related to monitoring and diagnostics. Therefore, neural networks can also play an important role in reactive quality assurance, especially where high processing and classification capabilities are required.

Proposed monitoring scheme for machining processes [98].

the monitoring of a drilling process. Their results show that Acoustic Emission (AE) signals detected by a single sensor are applicable to the recognition of operations like free run, drilling at different positions, and drilling with a wom drill. However, they also discovered that the recognition ability was influenced not only by the neural network but also by the properties of the detected system. Later applications [94], [95] show that neural networks can advantageously be used in real-time monitoring of manufacturing processes and other technical prodesses. Kamarthi et al. [96] investigated a new method for on-line measurement of flank wear through the synergy of vibration and force sensing. Their method used a Kohonens selforganizing feature map for correlating force and vibration signals with t e tool wear levels. In the particular experiment, h it was found that the trained feature map learned to detect the tool wear levels with a reliability close to 95%. Thkr results show that methods based on sensor data fusion using neural networks have the potential to offer reliable and robust methods for correlating tool wear levels with the signatures from force, vibration, and acoustic emission sensors for a wide range of process conditions. Hou and Lin [97] designed a monitoring system for identifying both periodic and aperiodic process signals using neural networks. First, digital signal processing techniques are used to convert collected manufacturing signals into frequency domain. Then, a neural network based program is used to identify the signals by examining their characteristic frequencies. Two examples (periodic signals and aperiodic signals) were presented to demonstrate the feasibility of the monitoring system and its recognition ability. The results are quite promising and show that the neural network based system seems to have a good potential in monitoring automatic manufacturing processes. Masory [98] proposed a scheme for machining processes monitoring using multi-sensor readings fused by neural networks. His scheme, shown in Fig. 6, functions in two modes as the following: 1) Training Mode: In this mode a skilled operator is monitoring the process and the ANN is used to capture his knowledge. The ANN is fed with three groups of data:

a)

Process inputs that include feed, cutting speed, process geometry and workpiece/tool materials.

220

IEEE TRANSACTIONS ON COMPONENTS, PACKAGING, AND MANUFACTURING TECHNOLOGY-PART A, VOL. 17, NO. 2, JUNE 1994

Barschdorff [ 1 151 discussed the application of neural network techniques for the quality control of electric drive motors. Sixteen spectral features of the motors vibration were input to a three-layer neural network. The network was able to recognize production faults, such as unbalance of the rotor, nonconducting winding connections, magnetic field failures, failures on commutators or bearings, loose parts, etc. The classifying ability of the network was compared with the results obtained by different pattern recognition algorithms. It was found that the performance of the neural network was superior. Smith [ 1161 reported the use of back-propagation neural networks in quality control in an injection molding corporation (Orscheln Industries, Moberly, Missouri). Injection molding is a process with many variations in raw materials, machinery conditions and ambient conditions. It also has a temporal aspect where line conditions change during operation, affecting the end product. Neural networks are especially applicable when the data considered do not follow a known distribution or pattern, and, hence, are well suited for the quality control of injection molding. The results show that the neural network approach is comparable to other quality control methods, including control charts and statistical techniques, in goodness of output for quality control. An advantage of the neural network approach is the convenience of learning to establish the relationships directly, rather than through analysis and assumptions. Using a single network to monitor multiple products and/or quality parameters is an additional advantage. Neural network techniques also can be used in proactive quality assurance. Schmerr et al. [ 1171 provided an innovative approach for planning robust design experiments through the use of neural networks. Robust design is a cost-effective technique for achieving high quality and reliability. The basis of the approach was to train a neural network on a set of tuples where each tuple corresponds to a Taguchi experiment along with its observed product response. Once trained, the neural network can probe the entire parameter space of design parameter settings, equivalent to performing a full factorial experimental design. With this information, a designer can identify alternative settings and search for optimal designs. The neural network approach was compared with analytical modelbased approach. It was found that the neural network had remarkable capabilities for generalization when trained on the same sparse array of experiments as used in complementary Taguchi analysis. The research of neural networks for quality assurance has actually be implemented in manufacturing practice. CTS Electronics of Texas has used a neural network system to detect defective loudspeaker assembly lines [ 1181. Ford Motor Company used neural network techniques to check car paint finishes [118]. Motorola Incorporated used artificial vision based on neural network techniques for the quality inspection of their chips [ 1 191.

G. Group Technology
Two engineering problems associated with implementing GT are part classification and part family formation. Kaparthi

and Suresh [120] proposed a neural network system for shape-based classification and coding of rotational parts. A back-propagation neural network was trained to generate part geometry-related digits of the Optiz code from the bitmaps of part drawings. Their principal study shows that the network can generate codes accurately and promises to be a useful tool for the automatic generation of shape-based classes and codes. The use of neural networks for part family formation was studied by a couple of researchers [121], [122], [123]. In their approach, a three-layer feedforward neural network was trained with the back-propagation algorithm. Each input unit of the network represents a part feature, and each output unit represents a part family. An operator plays the role of a teacher to the network by presenting each part in terms of part features, and by telling which family it belongs to. The trained network stores implicit discrimination rules through a set of connection weights, and exhibits consistent classification practice. The research demonstrates that the neural network approach is capable of dealing with large part family formation problems efficiently. A problem in dealing with part classification and part family formation is how to maintain consistency. The consistency problem can be addressed most effectively if the classification and formation are a single procedure rather than two separate procedures. Kao and Moon [124] presented an approach for using the learning ability of neural networks to automate the process of part family formation during the part classification process. They divided the task into four phases: ( 1 ) seeding phase, (2) mapping phase, (3) training phase, and (4) assigning phase. In the seeding phase, three to five apparently distinctive parts are chosen from the part pool as part family representatives (seed parts). In the mapping phase, a code is assigned to each seed part. A neural network is also constructed. The number of input units is equivalent to the number of features contained in a classification code system. In the training phase, a set of training pattern pairs is presented to the network. The network learns based on the back-propagation algorithm. In the assigning phase, the network compares the presented part to the training set. If the features of the part are similar to the generalized features of any existing part family representative, the output of the network will be the same part family number. If there are no significant similarities between the part and any of the part family representative, the output of the network will not indicate any existing family code. That means a new part family should be set up for this part. Another GT problem is the machine-part cell formation problem. The assignment of a group of similar parts to a cell of machines having common processing characteristics greatly improves the efficiency of batch manufacturing. However, approaches to the problem of machine-part cell formation have been computationally inefficient for large machine-part matrices. Dagli and Huggahalli [ 1251 applied the binary ART network (ART-1) for machine-part cell formation. The columns and rows of the machine-part matrix are binary vectors that can be directly applied as inputs to an ART-1 network. Dagli and Huggahalli analyzed the performance of the neural network classifier and found that direct application of the basic ART-I network did not provide very satisfactory

HUANG AND ZHANG: ARTMCIAL NEURAL NETWORKS IN MANUFACTURING

22 1

propagation networks and counter-propagationnetworks) have the ability to mathematically solve the kinematic problem within acceptable error rates. Nguyen et al. [133] have designed several neural network configurations for solving the forward kinematic problem for robots with two degrees of freedom. Each of the networks used for evaluation has two input nodes, 81 and 8 2 , and two output nodes, 2 and y. The network architectures studied have 6een: 2-10-10-2 for the network trained with the error back-propagation (BP), 2- 10-10-2 for the network trained with the back-propagation algorithm but with output splitting (BPOS), 2-8-2 for the functional link network (FL),and a counter-propagationnetwork (CP) for trajectory mapping. The networks were trained using a set of data (64 input/output pairs) obtained from measurements taken in the robot work area. The training of all four network architectures resulted in quick learning convergence to the desired trajectory. The basic BP network with two hidden layers gave a reasonably accurate solution. The BPOS network, which involved considerably fewer connections than the BP network, yielded comparable accuracy. The best solution of the trajectory learning task was achieved with the FL network. The CP network resulted in a relatively fast approximation of the trajectorjr, but the endeffector position error was rather large and uneven. Therefore, the CP network in this application can be recommended only H. Robotics if a short development time is required for low-precision Robotics constitute the study of a finite number of rigid trajectory tracking. Iberall [ 1341, [ 135J applied cooperativekompetitive neural mechanical chains that represent a multi-variable non-linear coupled system. The solution of this problem is difficult networks to compute the inverse kinematic solution for fingers because even the simplest desired movement requires sophis- of a simulated robot hand during grasping. Guez and Ahmad ticated and computationally intensive mathematics [ 1291. The [ 1361 investigated the inverse kinematics solution for twoproblem areas involved in robotics are kinematics, dynamics, and three-degrees-of-freedom manipulators using multiplelayer perceptrons. They suggested that neural networks might trajectory planning, control, sensing, and intelligence. Kinematics refers to the study of robot joint motions without be best for providing a good initial estimate for a manipulator considering the causes of motion. Two distinct subproblems that required iterative methods for its solution, because of are involved, namely, forward kinematics and inverse kine- the relative numerical accuracy of neural networks compared matics. The forward kinematics problem involves a non-linear with the closed form solutions. Aylor et aZ. [137] provided matrix mapping from the joint space (a description in terms of proof that back-propagation neural networks can provide acjoint angles and positions) of the robot to its Cartesian space ceptable solutions to the inverse kinematic problem. The (Cartesian coordinates of the robot end-effector). The forward authors developed a network that is capable of taking into kinematic solution computes the Cartesian location of the end- account the unknown number of faults, errors, and limitations effector given the joint space description. On the other hand, of the poorly constructed Fishertechnik robot and provides the inverse kinematic solution computes the joint space coordi- solutions that are as accurate as the robots mechanical and nates given the end-effector location in Cartesian coordinates. electrical systems can support. Also, they argued that backGiven the angular movement of each joint and the length of propagation neural networks were ready to leave the laboratory each arm it is relatively easy to find the Cartesian location and enter the real working world to solve the inverse kinematic (coordinates) of the end-effector, except that several non- problem. Pa0 and Sobajic [138], and Sobajic et al. [139] studied linear trigonometric and matrix multiplications are needed. r The inverse kinematics problem can yield multiple number of the use of neurocontrollers in robot a m movement with two solutions when the robot has multiple degrees of freedom. It degrees of freedom and with a desired final end-effector posiis more computationally intensive than the forward kinematic tion. The results that were achieved demonstrated the unusual problem and is the most difficult problem in robotics [130]. efficiency and potential of neurocontrollers for use in robot Neural networks can be used to reduce the computational kinematics control systems. Albus [140], [141], [ 1421 precomplexity of the kinematic problem. The trade-off is that sented his Cerebellar Model Articulation Controller (CMAC) the results yielded by the neural network approach may not for general robotic control. The basic idea of the CMAC be as accurate as those yielded by the conventional analytic approach is to learn an approximation of the system characapproach. Hecht-Nielsen [ 1311, Josin [ 1301, and Saxon and teristics and then to use it to generate the appropriate control Mukerjee [ 1321 have proven that neural networks (back- signal. The approximation of the system characteristics is

result. They then reordered the input vectors derived from the machine-part matrix, and modified the original ART1 paradigm. This produced drastic improvement over the performance of the basic ART-1. The result obtained compared favorably with popular algorithms proposed in the literature, such as the ROC2 algorithm provided by King and Nakomchai [126]. The application of neural network approaches in the design of cellular manufacturing systems, which involves the machine-part family formation problem, has been studied by MalavC and Ramachandran [1271 and Lee et al. [128]. Back-propagation networks and ART networks both can be used in solving GT problems. However, the learning results of a back-propagation network depend on the frequency of presenting a part. For example, if a network is trained three more times with part A than with part B, the network becomes biased toward part A [124]. This problem is not so obvious when an ART network, which applies the unsupervised learning algorithm, is used. Since ART networks are both adaptive and stable, they are more suitable for solving GT problems than back-propagation networks. The use of ART networks to solve GT problems is a research topic that is worthy of further studies.

222

IEEE TRANSACTIONS ON COMPONENTS, PACKAGING, AND MANUFACTURING TECHNOLOGY-PART A, VOL. 17, NO. 2, JUNE 1994

understood as gradual learning based on the observations of the plant input-output data in real time [143], [144]. Miller et al. [ 1451 used CMAC neural networks to control a robot whose kinematics is unknown. After training, the neural network controller performed very well. The average error was always below the error of a fixed gain controller without learning. The applications of neural networks in robot control have been addressed by many authors. Readers may refer to the following partial list of projects: 1) Multiple-layer perceptrons for an inverse Jacobian controller [ 1461. 2) The use of a proprietary neural network system, GLN (Graded Learning Network), for controlling robots movement and action [147]. 3) A neural network based controller for a single degree freedom robot arm [148]. 4) A neural network based controller for tracking moving objects with a two-joint robot arm [149]. 5 ) Hierarchical neurocontroller for robotics manipulation [150]. 6) Learning optimization in robot positioning control [ 1511. 7) Visual control of robotics manipulator based on neural networks [152]. 8) Neural network compensator for uncertainties of robotics manipulators [ 1531. Neural networks have also been used for robot dynamics due to their nonlinear mapping property. The basic idea is that the neural network learns the inverse dynamical relationship of the robot directly, which can be used as an inverse dynamic controller. Kawato [ 1541 implemented a novel neural network for inverse dynamics control based on neurophysiology. The research of neural network approaches in trajectory planning includes the following: 1) Eckmiller [ 1551 developed a novel neural network called Neural Triangular Lattice (NTL) for storing and retrieving trajectories. 2 ) Jorgenson [ 1561 investigated the use of neural networks for mobile robot path planning. 3) Tsutsumi and Matsumoto [ 1571 implemented obstacle avoidance in an unknown environment using Hopfield networks for multi-joint robots and truss structures. 4) Seshadri [ 1581 investigated the use of Hopfield networks for mobile robot path planning. 5 ) Lee and Park [159] used a neural network approach for collision-free path planning. The objective of the research in neural network approaches to robotics is to investigate the feasibility of creating an intelligent robot control scheme in an unmanned working environment. The research is quite diverse and rudimentary, leaving much room for improvement and new areas for development. Much of the preliminary research discussed here is quite promising. However, some of the works yield poor performance relative to conventional techniques partially because neural networks are not yet fully understood or developed. Hopefully many of these problems will disappear as we develop and understand neural networks more fully.

I. Other Applications
The ill-structured characteristics of manufacturing problems provided fruitful grounds for neural network applications. Due to the abundance and diversity of technical publications, we cannot cover all the manufacturing applications in this paper. The following is a list of specific applications that may be of interest to the readers. 1) Multisensor fusion and integration [ 1601. 2) Sensing on the factory floor [161]. 3) Maintenance management problem [ 1621. 4) Plant layout determination [ 1631. 5 ) Operational policy determination [ 1641. 6) Assembly system selection [165]. 7) Sampling inspection plans determination [ 1661. 8) Design of manufacturing systems [167]. 9) Design of assembly system [168]. 10) General-purpose simulation [ 1691. IV. PROBLEMS WITH NEURAL NETWORKS AND SUGGESTED SOLUTIONS Neural networks have several advantages over knowledgebased expert systems. Their learning ability and parallel distributed structure are very attractive to researchers who want to develop intelligent systems. Currently, there is an explosion of interests in neural networks. However, some of the scientific and economic expectations on neural networks (e.g., replacing conventional computers and eliminating programming) are unreasonable [ 1701. This is because neural networks have the following disadvantages: 1) A neural network cannot explain its results explicitly, which implies that the user interface of a neural network may not be as friendly or productive as that of an expert system. 2 ) The knowledge representation of a neural network is vague and not easily understood. 3) The configuration of a neural network is usually time consuming, as one needs to use a trial-and-error method to find the proper neural network architecture for a given problem. 4) The current neural network learning algorithms are not efficient enough and cannot guarantee network convergence. 5 ) How to derive some type of optimal training set for a neural network application still remains a question. Experience with expert systems has shown that the ability to generate explanations is absolutely crucial for the user acceptance of AI systems [171]. However, neural networks have no explicit, declarative knowledge representation and therefore have considerable difficulty in explaining their results. In order to solve this problem, Caudill [172] suggested that one can construct an additional small rule-based system and use it for explanation. The explanation system would be neither as complete nor as detailed as a full-blown rule-based solution; its function is solely to explain why an answer is correct. Another solution to the problem is to introduce structure to a neural network, e.g., the explicit encoding of relations and modular network architectures [171].

HUANG AND ZHANG ARTIFICIAL NEURAL NETWORKS I MANUFACTURING N

223

To build a neural network based system, one has to make decisions regarding the type of neural network to be used, its architecture, topology, type of nonlinearity, and associated parameters. However, there is still not an efficient method that exists to make the decisions. Cios and Liu [173] proposed a machine learning method to automatically generate a neural network architecture for a given problem. Their approach also allows interpretation of the knowledge embedded in the generated connections and weights. However, neural networks generated using the approach are usually more complicated than those generated using the trial-and-error method. While very large rule-based systems have been built, extremely large neural network based systems are not feasible with todays technology [172]. Such networks can be constructed, but they cannot be effectively trained. The current neural network learning algorithms available are heuristic in nature. For small problems those algorithms are sufficient enough. However, problems with moderate size usually require long learning time; and network convergence is not guaranteed. One solution to this problem is to break big problems into smaller pieces. This method also facilitates the selection of a proper set of training examples. One can expect that expert systems can help overcome the problems inherent in existing neural networks and enable them to do what they cannot accomplish alone in the following [174]: 1) to deal with knowledge representation, symbolic reasoning, non-numerical operations, which conventional neural networks cannot cope with; 2) to determine the number of nodes, number of layers, connections, and initial weights of a neural network, based on the degree of nonlinearity, complexity, and other data that are available in the design process; 3) to improve the learning ability of neural networks by controlling on-line the minimal number of nodes, weights, and layers (i.e., determine the optimal network architecture while training); and 4) to help train neural networks using the key data classified by expert systems from a great number of the collected data. Quite a few researchers are interested in building hybrid systems that merge the structures and functions of expert systems with those of neural networks. There are some manufacturing applications that exemplify the functionality and power of combining expert systems and neural networks. We will provide a brief review in the following. In order to overcome the knowledge acquisition bottleneck in the building of expert systems, neural networks have been used as a learning mechanism to transfer engineering experience into formulated knowledge. Yeh et al. [175] employed a neural network to build an expert system for debugging FEM program input data. The back-propagation learning algorithm was used to train the network for extracting knowledge from training examples. The authors examined the influences of various control parameters including learning rate and momentum factor, and various network architecture factors (e.g., the number of hidden nodes and the number of

hidden layers). To apply the knowledge in the trained network, a reasoning strategy that hybridizes forward-reasoning and backward-reasoning schemes was proposed to realize the inference mechanism. The results proved that neural networks can work sufficiently for classification problem-related expert systems. Ben-David and Pao [176] presented an architecture, which is based upon a hybrid of neural networks and rule-based models, for the design of self-improving expert systems. The architecture was tested on an expert system which assists users during the planning stage of a chemical product that has given properties and composition. In the application, 15 chemical substances are components that will be mixed and processed in order to produce a product that exhibits certain required properties. The designers goal is to find a composition that meets these properties. A neural network can be used to predict a composition given the required properties. Any prediction is naturally susceptible to errors. Laboratory tests must be carried out in order to verify the actual properties before any mass production takes place. The results of the laboratory tests (i.e., the actual properties) can be stored for further network training. When a prediction is not accurate, the system invokes another source of knowledge, a rule-based component that iteratively suggests possible modifications to the current formula. All the suggestions made by the rule-base are fed into a second neural network. The network estimates the effects of each suggestion before resorting to a costly laboratory test. The designer then selects a plan based on the estimation provided by the network for laboratory testing. The results are fed back into the system for further training of the neural network. At intervals, all the neural networks are trained using feedbacks. The rule-based component, on the other hand, is static. It can only be modified explicitly. The consultation process repeats iteratively until an acceptable formula is found. AI Ware has developed a commercial Integrated Technology Platform [177], [178], [179], [180]. The workstation-based software platform combines neural networks, expert systems, and optimization. The platform system attempts to provide all the capabilities of the individual technologies, while minimizing the limitations of each. The individual technologies within the platform are essentially transparent to the user. This leaves the user free to concentrate on the problem to be solved. The user does not need to address the application issues and technical quirks typically associated with each technology. AI expertise is not mandatory when using the platform system. Therefore, domain experts can both develop and apply systems based on the platform. The platform has been incorporated in a new applicationspecific product designed for chemical engineers. The product, called CAD/Chem Custom Formulation System, helps designers rapidly optimize the formulation and processing of products such as plastics, rubber, alloys, composite materials, and pharmaceuticals [177], [ 1781, [179]. The functionalities of CAD/Chem and the underlying Integrated Technology Platform include Best Match, Estimate, and Optimize. Best Match enables a product designer to find in the database the current or past product that best meets a given set of new product requirements. This capability is based on a neural

224

IEEE TRANSACTIONS ON COMPONENTS, PACKAGING, AND MANUFACTURING TECHNOLOGY-PART A, VOL. 17, NO. 2, JUNE 1994

networks pattern matching ability. The Estimate capability basically is also a neural network function. It tells the user what product properties to be expected for a given change in formulation. This estimation capability is based on a system model that is automatically learned by the neural network. An automatic neural network configuration capability is provided by an expert system embedded in the platform. The rules in the expert system are used to select an initial network configuration and a set of parameters, based on the nature and dimension of the task. Additional rules adjust the configuration and parameters in real-time, as required, to enhance the learning rate of the neural network. The Optimize function finds the best way to formulate or process a product. The platform system has also been applied to tasks such as quality control, product design, process control, power system security monitoring, material selection, and finite element modeling. A Functional Link Network (FLN) Combustion Optimizer, which is also built upon the Integrated Technology Platform, as discussed in [ 1SO]. Other significant applications reported in the literature include the following. Holden and Suddarth [181] used neural networkhowledge-based hybrid systems for large scale dynamic control. Bulsari et al. [ 1821, [ 1831 applied the so-called productive neural networks for continuous steel casting diagnosis. Hall and Lu [184] used a hybrid methodology to develop an adaptive controller for optimizing the operation of a harvester. Spelt el al. [185] presented a hybrid architecture for the diagnosis and decision-making in manufacturing. When we are dealing with a manufacturing problem, most often we have a certain amount of apriori knowledge about the problem domain. However, the a priori knowledge is neither complete nor necessarily accurate. While expert systems can handle a priori knowledge easily, neural networks can be used to refine the knowledge base. Due to the recent successes of connectionism, one may argue that the a priori knowledge can also be learned by neural networks. This is possible but may be inappropriate. First, the task may not be well suited to inductive learning. For example, in his connectionist model of beam-balance physics, McClelland [ 1861 suggested that the highest level of knowledge, i.e., using the concept of torque to predict which way the beam will fall, cannot be learned from examples and required explicit transfer of knowledge. Second, many learning algorithms are often designed to extract previously unknown regularities from their environment. Thus the representations they developed may not be comprehensible to the user. Third, it makes little sense to train a neural network when the a priori knowledge can be explicitly embedded in the system architecture. Therefore, it is desirable to build a hybrid system which consists of expert system as well as neural network components. Although hybrid systems can exploit the advantages and neutralize the disadvantages of their components, they also bring some new problems. Hybrid systems face two problems not found (or found to a lesser extent) in their single-paradigm counterparts [ 1871. First, the expert system and neural network components in a hybrid system must be able to communicate. Second, if learning is to be supported, credit must be assigned to each component. The communication problem arises be-

cause the two paradigms (expert systems and neural networks) use different representations. One solution to this problem is to subdivide the problem, delegating one set of tasks to each component. However, this approach restricts the effectiveness of the hybrid system; the components cannot benefit from each others information when solving their individual tasks. An alternative is to convert the representations of each component into a common language, or simply use one of the original representations (e.g., connectionist representation). The credit assignment problem also can be largely solved by using a homogeneous representation and adjusting all learning parameters with a single mechanism. This approach seems to be a better solution. It is a very interesting research topic which requires tremendous amounts of effort. V. CONCLUSIONS The idea of building manufacturing systems that can function automatically has attracted a lot of attention and created continuous research activities. Recently, artificial neural networks have emerged as a revolutionary AI approach and generated enormous interests in the manufacturing arena. This paper introduces the basic concepts of neural networks and reviews the current application of neural networks in manufacturing. The problems with neural networks are also identified and some possible solutions are suggested. We hope that the materials presented in this paper can provide some useful guidelines and references for the research and implementation. Our conclusions are summarized as follows: 1) There is an intense research interest in applying neural network techniques in manufacturing. 2) Some manufacturing problems have attracted more research attention than others. 3) Many of the applications reported in the literature are either laboratory experiments or preliminary applications. 4) While Hopfield networks are used in solving optimization problems and ART networks are used in group technology, back-propagation networks have been applied in a variety of problems (including optimization and group technology). 5 ) Some of the theoretical claims of neural network abilities have been tested in real manufacturing practices and found to be true. 6) Most of the applications are simulated in conventional computers rather than implemented in hardware. 7) The development of VLSI (Very-Large Scale Integrated) neural chips will further accelerate the computation speed of neural networks. 8) The development of a generalized neural network for solving different types of large-scale problems is desired. 9) There is a trend to integrate neural network and expert system techniques to solve manufacturing problems. Artificial neural network techniques are still very new and are developing quickly. We are witnessing fast expansion of neural network-based intelligent machines. Neural networks are an enhancing rather than replacing technology. As most researchers agree, artificial neural networks are not going to replace conventional computer and eliminate programming.

HUANG AND ZHANG: ARTIFICIAL NEURAL NETWORKS IN MANUFACTURING

225

Instead, they Will be integrated with conventional computing techniques and other AI schemes (e.g., knowledge-based expert systems) in order to develop truly intelligent manufacturing systems. Hopefully, the era of Intelligent Manufacturing Systems will soon be realized via the help of artificial neural networks.
&FEFENCES [ l ] T.-C. Chang, R. A. Wysk, and H.-P. Wang, Computer-Aided Manufacturing. Englewood Cliffs, NJ: Prentice-Hall, 1991, p. 3. [2] R. Jaikumar, Postindustrial manufacturing, Harvard Business Review, vol. 86, no. 6, p. 70, 1974. [3] F. H. Mitchell, Jr., CIM Systems: An Introduction to ComputerIntegrated Manufacturing. Englewood Cliffs, NJ: Prentice-Hall, 1991, p. 67. [4] WAC, Artificial intelligence in manufacturing, Computers in Industry, vol. 5, p. 159, 1984. [5] G. H. Schaffer, Artificial intelligence: A tool for smart manufacturing, American Machinist and Automared Manufacturing, vol. 130, no. 8, p. 83, 1986. [6] D. Barschdorff and L. Monostori, Neural networks: Their applications and perspectives in intelligent machining, Computers in Industry, vol. 17, pp. 101-119, 1991. [7] W. S. McCulloch and W. A. Pitts, A logical calculus of the ideas imminent in nervous activity, Bulletin of Mathematics and Biophysics, vol. 5, pp. 115-133, 1943. [8] D. 0. Hebb, The Organization of Behavior: A Neuropsychological Theory, New York John Wiley, 1949. [9] B. G. Farley and W. A. Clark, Simulation of self-organizing systems by digital computer, Institute of Radio Engineers-Transactions of Professional Group of Information Theory, vol. PGIT-4, pp. 76-84, 1954. [IO] F. Rosenblatt, The perceptron: A probabilistic model for information storage and organization in the brain, Psychoanalytic Review, vol. 65, pp. 386-408, 1958. [ l l ] E. R. Caianiello, Outline of a theory of thought-processes and thinking machines, J. Theoretical Biolom vol. 2. DD. 204-235. 1961. Adapti;; switchingcircuits, 1960 B. Widrow and M. E. Hoff, WESCON Convention, Record Part IV,pp. 96104, 1969. B. Widrow, Generalization and information storage in networks of adaline neurons, in Self-organizing Systems, Washington, D.C.: Spartan Books, 1962, pp. 435461. L. Stark, M. Okajima, and G. H. Whipple, Computer pattem recognition techniques: Electrocardiographic diagnosis, Communications of the ACM, vol. 5, pp. 527-532, 1962. K. Steinbuch and V. A. W. Piske, Learning matrices and their applications, IEEE Trans. Electronic Comput., vol. EC- 12, pp. 846-862, 1963. N. J. Nilsson, karning Machines: Foundations of Trainable Pattern Classifers, New York McGraw-Hill, 1965. M. Minsky, Neural nets and the brain, Ph.D dissertation, Princeton Univ., NJ, 1954. M. Minsky and S. Papert, Perceptrons, Cambridge, MA: MIT Press, 1969. S. I. Amari, Learning pattems and pattem sequences by self-organizing nets of threshold elements, IEEE Trans. Comput., vol. C-21, pp. 1197-1206, 1972. S. I. Amari, Neural theory of association and concept formation, Biological Cybernetics, vol. 26, pp. 175-185, 1977. K. Fukushima, Cognitron: A self-organizing multilayered neural network, Biological Cybernetics, vol. 20, pp. 121-136, 1975. K. Fukushima and S. Miyaka, Neocognitron: A self-organizing neural network model for a mechanism of pattern recognition unaffected by shift in position, Biological Cybernetics, vol. 36, pp. 193-202, 1980. T. Kohonen, Associative Memory: A System-Theoretical Approach, Berlin: Springer-Verlag, 1977. T. Kohonen, Self-organization and Associative Menwry, Berlin: Springer-Verlag, 1984. T. Kohonen, Adaptive, associative, and self-organization functions in neural computing, Applied Optics, vol. 26, pp. 49104918, 1987. T. Kohonen, Self-organized formation of topologically correct feature maps, Biological Cybernetics, vol. 43, pp. 59-69, 1982. G. A. Carpenter and S. Grossberg, A massively parallel architecture for a self-organizing neural pattem recognition machine, Computer Vision, Graphics, and Image Processing, vol. 37, pp. 54-115, 1983.

JE

1281 G. A. Carpenter and S. Grossberg, ART 2: Self-organization of stable category recognition codes for Lalog output pattems, Applied Optics, vol. 26, pp. 49194930, 1987. [29] G. A. Carpenter and S. Grossberg, The ART of adaptive pattem recognition by a self-organizing neural network, Computer, pp. 77-88, Mar. 1988. [30] G. A. Carpenter and S. Grossberg, ART 3 hierarchical search Chemical transmitters in self-organizing pattem recognition architectures, in Proc. Int. Joint Con5 on Neural Networks, vol. 2, 1990, pp. 30-33. [31] J. J. Hopfield, Neural networks and physical systems with emergent collective computational abilities, Proc. Natl. Acad. Sci.. vol. 79, pp. 2554-2558, 1982. [32] J. J. Hopfield, Neurons with graded response have collective computational properties like those of two state neurons, Proc. Natl. Acad. Sci., vol. 81, pp. 3088-3092, 1984. [33] B. Kosko, Adaptive bidirectional associative memories, Applied Optics, vol. 26, pp. 49474960, 1987. [34] G. E. Hinton, T. J. Sejnowski, and D. H. AcMey, Boltzmann machines: Constraint satisfaction networks that learn, Tech. Rep. CMU-CS-84119, Dept. Comp. Sci., Carnegie-Mellon Univ. [35] G. E. Hinton and T. J. Sejnowski, Learning and relearning in Boltzmann machines, in Parallel Distributed Processing: Explorations in the Microstructure of Cognition, Vol. I : Foundations, Cambridge, MA: MIT Press, 1986, pp. 282-317. [36] J. A. Anderson, J. W. Silverstein, S. A. Rite, and R. S. Jones, Distinctive features, categorical perception, and probability learning: Some applications of a neural model, Psych. Rev., vol. 84, pp. 413-451, 1977. [37] P. J. Werhos, Beyond regression: New tools for prediction and analysis in the behavioral sciences, Ph.D dissertation, Harvard Univ., MA, 1974. [38] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, Learning intemal representations by error propagation, in Parallel Distributed Processing: Explorations in the Microstructure of Cognition, Vol. I : Foundations, Cambridge, MA: MIT Press, 1986, pp. 318-362. [39] B. Widrow and M. A. Lehr, 30 years of adaptive neural networks: Perceptron, madaline, and backpropagation, Proceedings of the IEEE, vol. 78, no. 9, pp. 1415-1442, 1990. [40] R. Hecht-Nielson, Neurocomputing, Reading, MA: Addison-Wesley, 1990. [41] J. M. Zurada, Introduction to Artificial Neural Systems, St. Paul, MN: West, 1992. [42] D. E. Rumelhart and J. L. McClelland, eds., Parallel Distributed Processing: Explorations in the Microstructure of Cognition, Vol. I : Foundations: Vol. 2: Psychological and Biological Models, Cambridge, MA: MIT Press, 1986. [43] T. Kohonen, An introduction to neural computing, Neural Nehvorks, vol. 1, p. 4, 1988. [44] R. P. Lippmann, An introduction to computing with neural nets, Acoustics, Speech, and Signal Processing Mag., vol. 4, no. 2, pp. 4-22, 1987. [45] M. Zeidenberg, Neural Networks in Artificial Intelligence, New York: Ellis Honvood, 1990. [46] J. J. Hopfield and D. W. Tank, Computing with neural circuits: A model, Science, vol. 233, pp. 625-633, Aug. 1986. [47] J. J. Hopfield and D. W. Tank, Neural computation of decisions in optimization problem, Biological Cybernetics, vol. 52, pp. 141-152, 1985. [48] D. W. Tank and J. J. Hopfield, Simple neural optimization networks: An A/D converter, signal decision circuit and a linear programming circuit, IEEE Trans. Circ. Syst., vol. CAS-33, no. 5, pp. 533-541, 1986. [49] A. B. Badiru, Expen Systems Applications in Engineering and Manufacturing, Englewood Cliffs, NJ: Prentice-Hall, 1992. [50] G. J. Udo, Neural networks applications in manufacturing processes, Computers and Industrial Engineering, vol. 23, no. 1 4 , pp. 97-100, 1992. [51] V. Venugopal and T. T. Narendran, Neural network model for design retrieval in manufacturing systems, Computers in Industry, vol. 20, pp. 11-23, 1992. [52] S . V. Kamarthi, S. T. Kunara, F. T. S. Yu, and I. Ham, Neural networks and their applications in component design data retrieval, J. Intelligent Manufacturing, vol. 1, no. 2, pp. 125-140, 1990. [53] S. R. T. Kumara and I. Ham, Use of associative memory and selforganization in conceptual design, Annals ofthe CIRP, vol. 39, no. 1, pp. 117-120, 1990. [54] S. R. T. Kumara and S. V. Kamarthi, Function-to-structure transformation in conceptual design: An associative memory-based paradigm, J. Intelligent Manufacturing, vol. 2, no. 5, pp. 281-292, 1991. [55] R. D. Coyne and A. G. Postmus, Spatial application of neural networks in computer-aided design, Artificial Intelligence in Engineering, vol. 5, no. 1, pp. 9-22, 1990.

226

IEEE TRANSACTIONS ON COMPONENTS, PACKAGING, AND MANUFACTURING TECHNOLOGY-PART A, VOL. 17, NO. 2, JUNE 1994

[56] E. Arai and K. Iwata, CAD with design specification decomposition and its applications, Annals of the CIRP, vol. 39, no. 1, pp. 121-124, 1990. 1571 N. Kim, Y. Takai, and T. L. Kunii, Geometrical constraint solving based on the extended Boltzmann machine, Computers in Industry, vol. 19, pp. 239-250, 1992. [58] A. K. Dhingra and S. S. Rao, A neural network based approach to mechanical design optimization, Engineering Optimization, vol. 20, pp. 187-203, 1992. [591 J. D. Chovan and M. B. Waldron, Towards intelligent CAD: A neural network model of the transformation from perception to function in mechanical engineering design, in Intelligent Engineering Systems Through Artificial Neural Networks, New York: ASME Press, 1991, pp. 75 1-756. [60] K. Osakada, G. B. Yang, T. Nakamura, and K. Mori, Expert system for cold forging process based on FEM simulation, Annals of the CIRP, vol. 39, no. 1, pp. 249-252, 1990. [61] K. Osakada and G.-B. Yang, Neural networks for process planning of cold forging, Annals of the CIRP, vol. 40, no. 1, pp. 243-246, 1991. [62] K. Osakada and G. Yang, Application of neural networks to an expert system for cold forging, Int. J. Machine Tools and Manufacture, vol. 31, no. 4, pp. 577-587, 1991. [63] J.-L. Hwang and M. R. Henderson, Applying the perceptron to threedimensional feature recognition, J. Design and Manufacturing, vol. 2, no. 4, pp. 187-198, 1992. [64] G. M. Knapp and H.-P. Wang, Neural networks in acquisition of manufacturing knowledge, in Intelligent Design and Manufacturing, New York: John Wiley & Sons, 1992, pp. 723-744. [65] G . M. Knapp and H.-P. Wang, Acquiring, storing and utilizing process planning knowledge using neural networks, J. Intelligent Manufacturing, vol. 3, no. 5, pp. 333-344, 1992. [66] TAI, Are neural networks a tool for AI, in Proc. I992 IEEE Int. Conf on Tools for AI, 1991, pp. 5-6. (671 Y. P. S. Foo and Y. Takefuji, Integer linear programming neural networks for job-shop scheduling, in Proc. 1988 Int. IEEE Conf Neural Networks, vol. 2, 1988, pp. 341-348. [68] S. Vathyanathan and J. P. Ignizio, A stochastic neural network for resource constrained scheduling, Computers & Operations Research, vol. 19, no. 3 4 , pp. 241-254, 1992. [69] Y. Yih, T.-P. Liang, and H. Moskowitz, A hybrid approach for crane scheduling problems, in Intelligent Engineering Systems Through Artijicial Neural Networks. New York: ASME Press, 1991, pp. 867-872. 1701 T. L. Yu, Time table scheduling using neural network algorithm, in Proc. Int. Joint Conf on Neural Networks, vol. 1, 1990, pp. 279-284. [71] T. P. Liang, H. Moskowitz, and Y. Yih, Integrating neural networks and semi-Markov processes for automated knowledge acquisition: An application to real-time scheduling, Decision Sciences, vol. 23, no. 6, p. 1297, 1992. [72] Z.-P. Lo and B. Bavarian, Multiple job scheduling with artificial neural networks, Computer & Electrical Engineering, vol. 19, no. 2, pp. 87-102, 1993. [73] C. L. P. Chen, Design of a real-time AND/OR assembly scheduler on an optimization neural network, J. Intelligent Manufacturing, vol. 3, no. 4, pp. 251-262, 1992. [74] Y. Yih, T.-P. Liang, and H. Moskowitz, Robot scheduling in a circuit board production line: A hybrid O W A approach, IIE Transactions, vol. 25, no. 2, pp. 26-34, 1993. [75] I. Arizono, A. Yamamoto, and H. Ohta, Scheduling for minimizing total actual flow time by neural networks, Int. J. Production Research, vol. 30, no. 3, pp. 503-511, 1992. [76] L. C. Rabelo and S. Alptekin, Integrating scheduling and control functions in computer-integrated manufacturing using artificial intelligence, Computers and Industrial Engineering, vol. 14, pp. 101-106, 1989. (771 L. C. Rabelo, S. Alptekin, and A. Kiran, Synergy of artificial neural networks and knowledge-based systems for intelligent Fh4S scheduling, Int. Joint Conf on Neural Networks, San Diego, CA, 1990. 1781 S. S. Rangwala and D. A. Dornfeld, Learning and optimization of machining operations using computing abilities of neural networks, IEEE Trans. Syst., Man, and Cybem., vol. 19, no. 2, pp. 299-314, 1989. [79] C. G. Gingrich, Modeling human operators using neural networks, /SA Transactions, vol. 31, no. 3, pp. 81-90, 1992. [80] K. Anderson, G. E. Cook, and K. Gabor, Artificial neural networks applied to arc welding process modeling and control, IEEE Trans. Industry Applications, vol. 26, pp. 824-830, 1990. [81] K. Anderson, G. E. Cook, J. F. Springfield, and R. J. Bamett, Applications of artificial neural networks for arc welding, in Intelligent Engineering Systems Through Artif cia1 Neural Networks, New York: ASME Press, 1991, pp, 717-728.

[82] H. B. Smartt, J. A. Johnson, C. J. Einerson, and G. A. Cordes, Development of a connectionist fuzzy logic system for control of gas metal arc welding, in Intelligent Engineering Systems Through Artifcial Neural Networks, New York ASME Press, 1991, pp. 711-716. [83] G. Chryssolouris and M. Guillot, An AI approach to the selection of process parameters in intelligent machining, ASME on Sensors and Controlsfor Manufacturing, WAM, Chicago, IL, Nov. 27-Dec. 2, 1988. [84] N. A. Jalel, A. R. Mirzai, J. R. Leigh, and H. Nicholson, Application of neural network in process control, in Neural Network AppIications, London: Springer-Verlag, 1991, pp. 101-1 13. [85] G. A. Pugh, Synthetic neural networks for process control, Computers & Industrial Engineering, vol. 17, pp. 24-26, 1989. [86] E. Levin, R. Gewirtzman, and I. Gideon, Neural network architecture for adaptive system modeling and control, Neural Networks, vol. 4, no. 2, p. 185, 1991. [87] K. L. Patrick, Neural network keeps BSW filtrate solide at maximum uniform levels, Pulp and Paper, vol. 65, pp. 55-58, Mar. 1991. [88] G. Karsai, K. Anderson, G. E. Cook, and R. J. Barnett, Neural network methods for the modeling and control of welding processes, J. Intelligent Manufacturing, vol. 3, no. 4, pp. 229-236, 1992. [89] A. Guha, Continuous process control using neural networks, J. Intelligent Manufacturing, vol. 3, no. 4, pp. 217-228, 1992. [90] G. R. Madey, J. Weinroth, and S. Vijay, Integration of neurocomputing and system simulation for modeling continuous improvement systems in manufacturing, J. Intelligent Manufacturing, vol. 3, no. 4, pp. 193-204, 1992. [91] G. Chryssolouris, M. Domroese, and P. Beaullieu, Sensor synthesis for control of manufacturing processes, J. Engineering for Industry, vol. 114, pp. 158-174, May 1992. [92] D. J. Bozich and H. B. Mackay, Vibration cancellation using neural controllers, in Intelligent Engineering Systems Through Artifcial Neural Networks, New York: ASME Press, 1991, pp. 771-776. [93] E. Govekar, I. Grabec, and J. Peklenik, Monitoring of a drilling process by a neural network, The 21st CIRP Int. Seminar on Manufacturing Systems, Stockholm, Sweden, June 5-6, 1989. [94] J. Nacsa and L. Monostori, Real-time monitoring of machining processes, in Proc. Int. Conf on Automatic Supervision, Monitoring, and Adaptive Control in Manufacturing, 1990, pp. 197-233. [95] L. Monostori and J. Nacsa, On the application of neural nets in realtime monitoring of machining processes, The 22nd CIRP Int. Seminar on Manufacturing Systems, Enschede, Netherlands, June 11-12, 1990. [96] S. V. Kamarthi, G. S. Cohen, and S. R. T. Kumara, On-line tool wear monitoring using a Kohonens feature map, in Intelligent Engineering Systems Through Artifcial Neural Networks, New York: ASME Press, 1991, pp. 639-644. [97] T.-H. Hou and L. Lin, Manufacturing process monitoring using neural networks, Computers & Electrical Engineering, vol. 19, no. 2, pp. 129-141, 1993. [98] 0. Masory, Monitoring machining processes using multi-sensor readings fused by artificial neural network, J. Materials Processing Technology, vol. 20, no. 1&2, pp. 231-240, 1991. [99] T. Sorsa, H. N. Koivo, H. Koivisto, Neural networks in process fault diagnosis, IEEE Trans. Syst., Man, and Cybem., vol. 21, no. 4, pp. 815-825, 1991. [lo01 L. 1. Burke and S . Rangwala, Tool condition monitoring in metal cutting: A neural network approach, J. Intelligent Manufacturing, vol. 2, no. 5, pp. 269-280, 1991. [loll A. K. Ray, Equipment fault diagnosis: A neural network approach, Computers in Industry, vol. 16, pp. 169-178, 1991. [ 1021 C. M. Bauer, NAMRC XIX report, Manufacturing Engineering, vol. 107, pp. 55-58, 1991. [lo31 S. Elanayar and Y. C. Shin, Tool wear estimation in turning operations based on radial basis functions, in Intelligent Engineering Systems Through Artificial Neural Networks. New York: ASME Press, 1991, pp. 685-692. [lo41 P. D. Wasserman, A. Unal, and S. Haddad, Neural networks for online machine condition monitoring, in Intelligent Engineering Systems Through Artificial Neural Nemorks. New York ASME Press, 1991, pp. 693-700. [lo51 M. Guillot and A. El Ouafi, On-line identification of tool breakage in metal cutting processes by use of neural networks, in Intelligent Engineering Systems Through Artifcial Neural Networks. New York ASME Press, 1991, pp. 701-710. [IO61 H.-J. Wu, C.-S. Liou, and H.-H. Pi, Fault diagnosis of processing damage in injection molding via neural network approach, in Intelligent Engineering Systems Through Artifcial Neural Networks. New York: ASME Press, 1991, pp. 645-650. I1071 C . 4 . Lin, I.-C. Wu, anb T.-H. Guo, Neural networks for sensor failure

HUANG AND W A N G ARTIFICIAL NEURAL NETWORKS IN MANUFACTURING

227

detection and data recovery, in Intelligent Engineering Systems Through Art$cial Neural Networks, New York ASME Press, 1991, pp. 735-740. [lo81 M.-V. Chow, P. M. Mangum, and S. 0. Yee, A neural network approach to real-time condition monitoring of induction motors, IEEE Trans. Industrial Electronics, vol. 38, no. 6, pp. 449-454, 1991. [lo91 G.M. Knapp and H.-P. Wang, Machine fault classification: A neural network approach, Int. J. Productions Research, vol. 30, no. 4, pp. 81 1-823, 1992. ill01 B. R. Upadhyaya and E. Eryurek, Application of neural networks for sensor validation and plant monitoring, Nuclear Technology, vol. 97, no. 2, pp. 170-176, 1992. [ l 111 J. H. Murphy and B. J. Kagle, Neural network recognition of electronic malfunctions, J. Intelligent Manufacturing, vol. 3, no. 4, pp. 205-216, 1992. [I121 D. A. h m f e l d , Unconventional sensors and signal conditioning for automatic supervision, in Proc. Int. Con5 on Automatic Supervision, Monitoring and Adaptive Control in Manufacturing, 1990, pp. 197-233. [ 1131 D. Barschdorff, L. Monostori, A. F. Ndenge, and G.W. Wostenkiihler, Multiprocessor systems for connectionist diagnosis of technical processes, Computers in Industty, vol. 17, no. 2&3, pp. 131-145, 1991. [114] T. Hill, Manufacturing Strategy: Text and Cases. Homewood, IL: h i n , 1989. [115] D. Barschdorff, Case studies in adaptive fault diagnosis using neural networks, The IMACS Annals on Computing and Applied Mathematics, Brussels, Sept. 3-7, 1990. [ 1161 A. E. Smith, Quality control using backpropagation: An injection molding application, in Intelligent Engineering Systems Through Artijcial Neural Networks. New York ASME Press, 1991, pp. 729-734. [117] L. W. Schmerr, S. M. Nugen, and B. Forouraghi, Planning robust design experiments using neural networks and Taguchi methods, in Intelligent Engineering Systems Through Artijcial Neural Networks, New York: ASME Press, 1991, pp. 829-834. [I 181 D. E. b a n e , Machine learning, Training and Development Journal, vol. 40, pp. 24-29, 1990. [I 191 0. Port, Adding eagle eyes to those lightning fast computers, Business Week, vol. 3194, pp. 60B, Dec. 1990. [120] S. Kaparthi and N. C. Suresh, A neural network system for shapebased classification and coding of rotational parts, Inc. J. Production Research, vol. 29, no. 9, pp. 1771-1784, 1991. [121] Y. Kao and Y. B. Moon, Learning part families by the backpropagation rule of neural networks, in Pmc. Ist Int. ConJ: on Automation Technology, 1990, pp. 819-824. [122] Y. B. Moon and U. Roy, Learning group technology part families from solid models by parallel distributed processing, Tech. Rep. MFE902, Dept. Mech. and Aerosp. Eng., Syracuse Univ., 1991. [I231 Y. B. Moon and S. C. Chi, Generalized part family formation using neural network techniques, J. Manufacturing System vol. 11, no. 3, pp. 149-160, 1992. [I241 Y. Kao and Y. B. Moon, A unified group technology implementation using the backpropagation learning rule of neural networks, Computers and Industrial Engineering, vol. 20, no. 4, pp. 425437, 1991. [125] C. Dagli and R. Huggahalli, Neural network approach to group technology, in Knowledge-Based Systems and Neural Networks: Techniques and Applications, New York: Elsevier, 1991, pp. 213-228. [126] J. R. King and V. Nakomchai, Machine-component group formation in group technology: Review and extension, Int. J. Production Research, vol. 20, no. 2, pp. 117-133, 1982. [127] C. 0. MalavC and S. Ramachandran, Neural network-based design of cellular manufacturing systems, J. Intelligent Manufacturing, vol. 2, no. 5, pp. 305-314, 1991. [128] H. Lee, C. 0. MalavC, and S. Ramachandran, A self-organizing neural network approach for the design of cellular manufacturing systems, Journal of Intelligent Manufacturing, vol. 3, no. 5, pp. 325-332, 1992. [129] B. Home, M. Jamshidi, and N. Vadiee, Neural networks in robotics: A survey, J. Intelligent and Robotics Systems, vol. 3, pp. 51-66, 1990. [130] G. Josin, Intepting neural networks with robots, AI Expert, pp. 50-58, Aug. 1988. [ 1311 R. Hecht-Nielsen, Applications of counterpropagation networks, Neural Networks, vol. 1, pp. 131-139, 1988. [132] J. B. Saxon and A. Mukerjee, Learning the motion map of a robot arm with neural networks, in Proc. Int. Joint ConJ: on Neural Networks, vol. 2, 1990, pp. 777-781. [133] L. Nguyen, R. V. Patel, and K. Knorasani, Neural network architectures for the forward kinematics problem in robotics, in Proc. Joint IEEE Int. Neural Networks Conf., vol. 3, 1990, pp. 393-399. [134] T. Iberall, A ballpark approach to modeling human prehension, IEEE ConJ Neural Networks, vol. 4, pp. 535-544, 1987.

[135] T. Iberall, A neural network for planning hand shapes in human prehension, IEEE Conference on Decision and Control, 1987, pp. 2288-2293. [136] A. Guez and Z. Ahman, Solution to the inverse kinematics problem in robotics by neural networks, IEEE Conference on Neural Networks, vol. 11, pp. 617-624, 1988. [137] S. Aylor, L. Rabelo, and S. Alptekin, Artificial neural networks for robotics coordinate transformation, Computers & Industrial Engineering, vol. 22, no. 4, pp. 481493, 1992. [138] Y.-H. Pao, and D. J. Sobajic, Artificial neural-net based intelligent robotics control, in Proc. SPIE-IntelligentRobots and Computer Vision, vol. 848, 1987, pp. 542-549. [139] D. J. Sobajic, J. J. Lu, and Y.-H. Pao, Intelligent control for the Intelledex 605 T Robot Manipulator, in Proc. 1988 IEEE Int. Neural Networks Conf., vol. 2, 1988, pp. 613-640. [140] J. Albus, A new approach to manipulator control: The cerebellar model articulation controller (CMAC), J. Dynamic Systems, Measurement, and Control, pp. 220-227, Sept. 1975. [141] J. Albus, Data storage in the cerebellar model articulation controller (CMAC), J. Dynamic Systems, Measurement, and Control, pp. 228-233, Sept. 1975. [142] J. Albus, Mechanisms of planning and problem solving in the brain, Mathematical Biosciences, vol. 45, pp. 247-293, 1979. [143] L. G.Kraft and D. S. Campagna, A comparison between CMAC neural network control and two traditional adaptive control systems, IEEE Control Systems Mag., pp. 36-43, Apr. 1990. [144] L. G.Kraft and D. S. Campagna, A summary comparison of CMAC neural network and traditional adaptive control systems, in Neural Networks for Control, Cambridge, M A MIT Press. [145] W. T. Miller, F. H. Glanz, and G.Kraft, CMAC: An associative neural network altemative to backpropagation, Proc. IEEE, vol. 78, no. 10, pp. 1581-1587, 1990. [146] R. Elsley, A learning architecture for control based on back-propagation neural networks, IEEE Con$ Neural Networks, vol. 4, pp. 545-550, 1987. [I471 T. Inglesby, No clowning, neural networks can help manufacturing, Manufacturing Systems, vol. 6, pp. 26-31, 1988. 1481 K. Wilhelmsen and N. Cotter, Neural network based controllers for a single-degree-of-freedom robotics arm,in Proc. 1990 Int. Joint Con5 on Neural Networks, vol. 3, 1990, pp. 407413. 1491 D. Shumsheruddin, Neural network control of robot arm tracking movements, in Neural Network Applications, London: Springer-Verlag. 1991, pp. 129-139. I501 L. C. Rabelo and X. Avula, Jr., Hierarchical neurocontroller architecture for robotics manipulation, IEEE Control Systems Mag., vol. 12, no. 2, pp. 3 7 4 1 , 1992. 1511 J. C. C. Chen, B. C. Jiang, and C.-H. Wu, Learning optimization for CPN-based training in robot positioning control, J. Intelligent Manufacturing, vol. 3, no. 4, pp. 237-250, 1992. 1521 H. Hashimoto, T. Kubota, and M. Sato, Visual control of robotics manipulator based on neural networks, IEEE Trans. Industrial Electronics, vol. 39, no. 6, pp. 490-496, 1992. 1531 A. Ishiguro, T. Furuhashi, and Y. Uchikawa, A neural network compensator for uncertainties of robotics manipulators, IEEE Trans. Industrial Electronics, vol. 39, no. 6, pp. 565-570, 1992. 1541 M. Kawato, Hierarchical neural network model for voluntary movement with application to robotics, IEEE Control Systems Mag.. vol. 8, no. 2, pp. 8-16. 1988. [155] R. Eckmiller, Neural network mechanisms for generation and learning of motor programs, IEEE Conference on Neural Networks, vol. 4, pp. 545-550, 1987. [156] C. C. Jorgenson, Neural network representation of sensor graphs in autonomous robot path planning, IEEE Conference on Neural Networks, vol. 4, pp. 573-582, 1987. [157] K. Tsutsumi and H. Matsumoto, Neural computation and learning strategy for manipulator position control, IEEE Conference on Neural Networks, vol. 4, pp. 525-534, 1987. [158] V. Seshadri, A neural network architecture for robot path planning, in Second Int. Symposium on Robotics and Manufacturing: Research, Foundation, and Applications, New York ASME Press, 1988, pp. 249-256. [159] S. Lee and J. Park, Neural computation for collision-free path planning, J. Intelligent Manufacturing, vol. 2, no. 5, pp. 315-326, 1991. [160] G. Chryssolouris and M. Domoroese, Sensor integration for tool wear estimation in machining, in Proc. ASME on Sensors and Controls for Manufacturing, 1988, pp. 115-123. [161] P. M. Noaker, Smart sensing on the factory floor, production, vol. 100, pp. 4 2 4 8 , 1988.

228

IEEE TRANSACTIONS ON COMPONENTS, PACKAGING, AND MANUFACTURING TECHNOLOGY-PART

A, VOL. 17, NO. 2, JUNE 1994

[I621 A. K. Ray and M. S. S. N. Murty, Application of artificial intelligence [I821 A. B. Bulsari, A. Kraslawski, and H. Sax&, Implementing a fuzzy expert system in an artificial neural network, Computers & Chemical in maintenance management, in Proc. 2nd Int. Conf on Industrial and Engineering, vol. 17, pp. S405-S410, 1993. Engineering Applications of AI and Expert Systems, 1989, pp. 1 12-1 18. 1831 A. B. Bulsari, A. Kraslawski, and H. Sax&, Continuous steel cast[I631 B. L. Foote, T. Tillinghast, S. Tretheway, J. Y. Cheung, and C.-H. ing diagnostic fuzzy expert system in an artificial neural network, Chang, Application of neural networks to optimize plant layout probScandinavian J. Metallurgy, vol. 21, no. 4, pp. 146150, 1992. lems, in Knowledge-Based Systems and Neural Networks: Techniques 1841 J. W. Hall and S. C.-Y. Lu, An adaptive machine controller utilizing and Applications. New York Elsevier, 1990, pp. 229-235. domain knowledge and quantitative data, Knowledge-Based Engineer[I641 G. Chryssolouris, M. Lee, and M. Domoroese, The use of neural neting Systems Research Laboratory Annual Report, Univ. of Illinois at works in determining operational policies for manufacturing systems, Urbana-Champaign, 1991, pp. 4 1 4 8 . J. Manufacturing Systems, vol. IO, no. 2, pp. 166175, 1991. 1851 P. F. Spelt, H. E. Knee, and C. W. Glover, Hybrid artificial intelli[165] A. Shtub and Y. Zimerman, Estimating the cost of robotics and manual gence architecture for diagnosis and decision-making in manufacturing, assembly systems by a neural network, in Intelligent Engineering Journal o Intelligent Manufacturing, vol. 2, no. 5, pp. 261-268, 1991. f Svstems Through Artificial Neural Networks. New York ASME Press. [ 1861 J. L. McClelland, Parallel distributed processing: Implications for 1991, pp. 801-806. cognition and development, in Parallel Distributed Processing: Impli11661 J. Wang and V. Chankong, Neurally-inspired stochastic algorithm ~. cations for Psychology and Neurobiology. Oxford: Clarendon Press, for determining multi-stage multi-attribute sampling inspection plans, 1989. Journal of Intelligent Manufacturing, vol. 2, no. 5, pp. 327-336, 1991. [I671 G. Chryssolouris, M. Lee, and J. Pierce, Use of neural networks for [ 1871 D. E. Rose, Appropriate uses of hybrid systems, in Connectionist Modeh: Proc. 1990 Summer School, San Mateo, CA: Morgan Kaufmann, the design of manufacturing systems, Manufacturing Review, vol. 3, 1991, pp. 277-286. no. 3, pp. 187-194, 1990. 11681 C. L. P. Chen and Q.-W. Yan, Design of a case associative assembly planning system, in Intelligent Engineering Systems Through Artijcial Neural Networks. New York: ASME Press, 1991, pp. 757-762. [I691 J. Nijhuis, L. Spaanenburg, and F. Warkowski, Structure and application of NNSIM: A general purpose neural network simulator, Samuel H. Huang is a Ph.D. Student in the DepartMicroprocessing & Microprogramming, vol. 27, pp. 189-1 94, Aug. ment of Industrial Engineering, Texas Tech Univer1989. sity, Lubbock, TX. [ 1701 J. A. Feldman, Neural networks, artificial intelligence and compuMr. Huang is a member of SME, ASME, and tational reality, Computers in Industry, vol. 14, pp. 145-148, May IIE. He is also the vice-president of SME student 1990. chapter at Texas Tech University. His research in[I711 J. Diederich, Explanation and artificial neural networks, Int. J . Manterests are intelligent manufacturing, CIM, CAPP, Machine Studies, vol. 37, pp. 335-355, 1992. PDEWSTEP, CAD/CAM, and manufacturing sys[I721 M. Caudill, Using neural nets: Hybrid expert networks, AI Expert, tems. He has published several technical papers in vol. 5, no. 11, pp. 49-54, 1990. these areas. [I731 K. J. Cios and N. Liu, A machine learning method for generation of a neural network architecture: A continuous ID3 algorithm, IEEE Trans. Neural Networks, vol. 3, no. 2, pp. 280-291, 1992. [I741 H. Y. Xu and C. R. Baird, Synergism of neural networks and expert systems for system identification, Expert Systems With Applications, vol. 5, pp. 25-33, 1992. [I751 Y.-C. Yeh, Y.-H. Kuo, and D . 4 . Hsu, Building an expert system Hong-Chao Zhang received the Ph.D. degree in for debugging FEM input data with artificial neural networks, Expert manufacturing from the Technical University of Systems With Applications, vol. 5, pp. 59-70, 1992. Denmark in 1989 and the M.Sc. degree in mechan[I761 A. Ben-David and Y.-H. Pao, Self-improving expert systems: An ical engineering from the University of Aalborg, architecture and implementation, Information & Management, vol. 22, Denmark, in 1986. pp. 323-331, 1992. He is currently an Assistant Professor in the [I771 W. H. VerDuin, Neural network software assists expert system, Department of Industrial Engineering, Texas Tech Control Engineering, pp. 3 7 4 0 , July 1991. University, Lubbock, TX. His research and teaching [ 1781 W. H. VerDuin, Solving manufacturing problems with neural nets, interests are in the areas of concurrent engineering Automation, pp. 54+, July 1990. and computer integrated manufacturing (CIM), in1791 W. H. VerDuin, The role of integrated AI technologies in product cluding computer aided process planning (CAPP), formulation, ISA Transactions, vol. 3 I , no. 2, pp. I5 1-1 57, 1992. CAD/CAM, application of artificial intelligence (AI) in manufacturing, auto1801 W. H. VerDuin, Optimizing combustion with integrated neural netmated tolerancing analysis, as well as manufacturing processes and systems. works and AI technologies, Control Engineering, pp. 3 8 4 0 , July Dr. Zhang has published more than 40 technical articles in a variety 1992. of journals and conferences. One of his recent books, titled Computerized 1811 A. D. C. Holden and S. C. Suddarth, Combined neural-netknowledgeManufacturing Process Planning Systems, was published by Chapman and based adaptive systems for large scale dynamic control, Int. J. Pattern Hall in 1993. He has received a research initiation awards from the NSF. Recognition, vol. 5, no. 4, pp. 503-522, 1991.

~~

Anda mungkin juga menyukai