You are on page 1of 24

Keynote Papers

Intelligent Computing Methods for Manufacturing Systems


R. Teti (2), Department of Materials b Production Engineering, University of Naples 'Federico II', Italy S. R.T. Kumara ( l ) ,Department of Industrial b Management Systems Engineering, Pennsylvania State University, USA

Abstract

Intelligent computation is taken to include the development and application of artificial intelligence (Al) methods i.e. tools that exhibit characteristics associated with intelligence in human behaviour. Many approaches have been proposed to apply Al methods, techniques and paradigms to the solution of manufacturing problems. This paper discusses current trends in applications of intelligent computing tools to manufacturing and reviews the motivation and basis for the utilisation of these systems. The topics of the paper were confined to four main issues of manufacturing systems: design, planning, production and system level activities. A discussion about intelligent manufacturing systems from these four basic functional view points was introduced, the relevant intelligent computing methods and their use in manufacturing were surveyed, and a number of significant research issues and applications were illustrated. The main developments that were observed comprise the integration of Al methods into CAD, CAPP, etc.; the improvement of the performance of present Al techniques; the development of hybrid Al systems; the elaboration and application of new Al paradigms in manufacturing. Intelligent systems in the future are expected to be integrated, modular, and hybrid in nature, and they may well include all the techniques described in this paper and further more.
Kevwords: Artificial intelligence, Manufacturing systems. Acknowledaements

Acknowledgements are due for papers, contributions and correspondence received from messrs ['CIRP member]: M. Balazinsky', Ecole Polytechnique de Montreal, Canada; R. Bueno', Tekniker Research Centre, Eibar, Spain; G. Dini', University of Pisa, Italy; W. Eversheim', RWTH Aachen, Germany; M. Geiger', University of Erlangen-Nuremberg, Germany; I . S Jawahir', University of Kentucky, Lexington, USA; J. Jedrzejewski", Wroclaw Technical University, Poland; H.J.J. Kals', University of Twente, Enschede, NL; F.-L. Krause', Institute for Production Systems and Design Technology, Berlin, Germany; P. Martin, University of Nancy, France; J. Milberg', Technical University of Munich, Germany; L. Monostori', Hungarian Academy of Sciences, Budapest, Hungary; A.Y.C. Nee', National University of Singapore, Singapore; T. Pfeifer', RWTH Aachen, Germany; D.T. Pham, University of Wales, Cardiff, UK; G. Reinhart, Technical University of Munich, Germany; W.B. Rowe', Liverpool John Moores University, UK; G. Rzewsky, The Open University, Milton Keynes, UK; H. Sato', Chuo University, Tokyo, Japan; K. Ueda', Kobe University, Japan; G. Warnecke', University of Kauserslautern, Germany; M. Weck*, RTWH Aachen, Germany; E. Westkamper', Technical University of Braunschweig, Germany; H.-P. Wiendhal', University of Hannover, Germany. Graduate student F. Parretta is gratefully thanked for his help and assistance in the preparation of the text. Support for preparing this paper was provided by the lstituto Nazionale per la Fisica della Materia (INFM), Genova, Italy, within Progetto Sud.
1. Introduction 1.1 Backaround

The Task Group focused on the identification of fields of successfuVpossible applications of intelligent computing tools in the different sectors of manufacturing systems and the definition of new and future research thrusts in Al applications for manufacturingengineering. This paper was prepared as part of the activities of the Task Group on Applications of Al and is the result of 6 meetings held between Aug. 1994 and Jan. 1997, as well as contributions, correspondence and papers received from Task Group participants and ClRP members who are gratefully acknowledged for generously providing their time and input.
1.2 Reviews on A l ADDlications in Manufacturinq

A framework for investigating the international state-ofthe-art, technological challenges and development trends in applications of Artificial Intelligence (Al) techniques in the manufacturing environment was established by the ClRP within the context of the STC " 0 Task Group on Applications of Al Methods in Manufacturing Engineering which was active between 1994 and 1997 (Teti 1 9 9 5 ~ 1996a, 1996b).

A significant amount of information on this subject is available, mostly associated with general discussion and detail elements of Intelligent Manufacturing Systems (IMS). The last ClRP Keynote Paper on this topic area was by Rowe 1994b, covering issues of application of Al in grinding. A recent ClRP technical report by Monostori 1996b, prepared within the activities of the Task Group on Applications of Al, presents a detailed review on machine learning in manufacturing methods. An extensive list of references to literature on Al applications in manufacturing engineering has been provided by G. Dini within the activities of the Task Group and will be published in a separate paper (Dini 1997). Prior to these, other reviews have been published on application of Al in manufacturing. Hatvany 1983 categorises IMS as the fourth phase in manufacturing research and indicates its main stresses. Merchant 1984 analyses existing technological forecasts in the application of Al for manufacturing engineering and identifies the main utilisation areas. The 1986 ClRP Round Table focused on the use and development of knowledge based and expert systems in manufacturing with contributions from Chryssolouris, Davies, Eversheim, Hatvany, Milacic, Preiss, and Sohlenius. The authors of this paper were assigned the responsibilityof compiling, with the help and contribution of ClRP colleagues, a state-of-the-art in the use and application of Al tools in manufacturing systems. Though at the outset the task looked reasonable, over a period of time it proved extremely difficult to put together an entire field and its development which spans over a

Annals of the ClRP Vol. 46/2/1997

629

decade. After several iterations, it was decided to confine the topics of the paper to four main issues of manufacturing systems, namely: Design, Planning, Production, and System Level Activities. A discussion about IMS from these four basic functional view points is introduced. Then, the relevant intelligent computing tools are described, their use in manufacturing is surveyed, and a number of significant research issues and applications are presented.
1.3 Functional View of Manufacturino Svstems

COMMON SYSTEM DATABASE INTELLIGENT DATABASE


and control

A manufacturing system can be conceptually thought of as being an integrated whole of complex interacting subsystems, organised in such a way as to synergically endeavour towards a common set of goals (Merchant 1984). Due to the inherent complexities associated with the manufacturing sub-systems, modelling them using common analytical and mathematical tools has proved to be very difficult, if not futile. In this paper, based on literature survey as well as inputs and contributions from ClRP colleagues, an intelligence based architecture to model manufacturing systems is proposed. The overall objectives of a manufacturing system are presented and, accordingly, the sub-system development is carried out. Then, the architecture of the sub-systems is presented, with special emphasis on intelligent computing. The functional scheme of a typical manufacturing system is shown in Fig. 1. A manufacturing system is a multi-objective and multiple goals seeking system. At the top level, a manufacturing system takes in the customer needs, feedback (responses), and part of society's total energy (be it raw materials, humanpower, resources, etc.), transforms them in such a way as to produce the outputs (products) more efficiently. From a life cycle point of view, it has to deal with scrap, waste disposal, personnel issues, governmental and environmental issues. Looking at systems "one level closet", they may be thought of as an organisation of many sub-systems (components) cooperating and coordinating together to function as an integrated whole. The integration of different system components can be achieved from a common intelligent system database. This integration is different from a sub-system interfacing in that the latter only involves information transfer and local analyses. For a fully integrated system, information transfer is bidirectional. This can be accomplished through: (a) system database for those sub-systems that do not directly affect each other very much: (b) direct connection when the sub-systems have high level of interaction; (c) treating the manufacturing system as a distributed system and developing and implementing an open system architecture for integration. Fig. 2 gives an overall view of manufacturing systems from a functional view point. For the sake of this paper, all the issues related to these sub-systems have been combined into four major themes: Design, Planning, Production and System Level Activities (Kumara 1995). "rules" from the society customer
needs

I SOCIETY

Needs Costumer feed back More efficiently produced products Energy Utility Scrap-recycled Information & technology Social activities

1 &I

Fig. 2 - Functional view of manufacturingsystems.


1.3.1 Design

In manufacturing systems, the customer needs are fed in, along with the feedback from the system and/or surroundings. Using the available technology related to materials, properties, etc., and to some extent human intuition, the product is conceived (conceptual design). Subsequently, the product is given a configuration and is parametrised. While performing this function, the product manufacturability concerns are also borne in mind. The design is tested using simulation analysis for its manufacturability. The corrections are made iteratively and finally the component is manufactured. Rule-based, case-based, graph-based expert systems and artificial neural networks are among the Al tools most widely utilised for conceptual, configuration, and parametric design.
1.3.2 Planning

1-

Feedback Energy Information

+
Manufacturing system

More efficiently produced products

Recycled scrap Technology Social activities/srrvices

Three major issues are dealt with in the planning subsystem. These are Process Planning, Scheduling and Manufacturing Resource Planning. Product design is made accessible to the process planning sub-system that transforms manufacturable design into its process plan with a schedule. As the process planning sulj-system does not need the overall design data such as safety factors, etc., they are not directly transferred to that sub-system, but if a need arises, any piece of information can be accessed from any sub-system. This method of structuring the database will render the system knowledge intensive rather than merely information intensive. In addition, the database will also aid in faster development of the process plan. Fuzzy logic reasoning and hybrid techniques combining expert systems, neural network and fuzzy logic appear to be the preferred Al approaches in planning. Scheduling and Manufacturing Resource Planning are in general carried out as service functions. In the ClRP community, not many references to these problems are found. Thus, it was decided not to deal with these issues in detail. However, intelligent computing tools that are used in these areas are taken into consideration.
1.3.3 Production

Means

Fig. 1 - A SADT view of manufacturing systems - level 0

With the process plan thus prepared, the product may be manufactured. The production system could be

630

Keynote Papers

considered as completely autonomous or semiautonomous. In such cases, to ensure that the process is under control, it is necessary to monitor the process, obtain status information about the different process and product variables, diagnose the problems, and control the process so that the diagnosed problems can be rectified. The principal functionalities are: Process Modelling, Monitoring, Diagnosis, Control, Inspection, and Assembly. Knowledge based systems, neural networks and fuzzy logic, often combined in hybridsystems, are most frequently applied to the production functionalities mentioned above.
1.3.4 System Level Activities

14. Heuristic Search 15. Holonic Manufacturing Systems 16. Immune Networks 17. Knowledge Based Systems 18. Knowledge Representation 19. Machine Learning 20. Multi Agent Systems 21. Natural Language Processing 22. Neural Networks 23. Petri Nets 24. Qualitative Reasoning 25. Reasoning techniques 26. Reinforcement Learning 27. Similarity Theory 28. Simulated Annealing 29. Syntactic and Statistical Pattern Recognition For the sake of this paper, the above have been grouped into the following categories: Knowledge Based Systems (KBS) and Expert Systems (ES), Neural Networks (NN), Fuzzy Logic (FL), Multi Agents (MA), and others techniques such as Genetic Algorithms (GA), Simulated Annealing (SA), etc. Features of Al applied in manufacturing engineering are summirised in Table 1.
Artificial Intelligence in Ma1 ifacturing I A/ fechniaues Acronym I Mno. Sector Advice Expert Systems ESAs-sembly Fuuy Logic Communication Control FL Control Genetic Algorithms Design GA Hybrid Systems Inspection Decision Making HS KB systems GoaCSeeking Maintenance K BS Neural Networks Knowledge NN Monitoring Neuro-Fuzzy Planning Learning NF Multi Agents Production Pattern Recognition MA Simulated Annealing Scheduling Reasoning SA Self-Improvement Systems Self-Maintenance Self-Organisation
A/ functions

The system level activities sub-system considered includes the following functionalities: System Modelling, System Control, and Integration.
System Modelling At a higher level, the activities of the production subsystem have to be controlled. This requires a model of the system activities in order to develop a control software that can be applied to a flexible manufacturing system. Following the NlST control hierarchy, an FMS at the facility level is broken down into shops, and at the shop level, the system is decomposed into cells, and so on, building a tree-like structure. System Control The control actions proceed in the forward direction (direction of decreasing hierarchy) and the feedback in the reverse. The physical entities of components at each level are represented by computer systems and control devices, connected through a communication networks. Multi agents, holonic systems, automata theory, Petri nets, process algebra, process calculus and blackboard architectures are being used for system modelling and control. One of the main objectives here is to automatically generate control software. These architectures are hierarchical in nature. Integration Manufacturing systems integration is a particularly important issue. Distributed Al techniques, multi agents, holonic systems and open system architectures are the current research focus. In order to make the system concurrent it is of paramount importance to pay particular attention to integration.

Table 1 - Al functions and techniques in manufacturing. One of the issues of the activities of the Task Group on Applications of Al in Manufacturing Engineering was represented by the preparation of a mapping between intelligent computing tools and manufacturing issues. Table 2 reports this mapping, which was arrived at after an extensive literature survey (Kumara 1995).

2.2 Relevant A / Tools


Artificial Intelligence (Al) is the branch of computer science dealing with the design of computer systems that exhibit characteristics associated with intelligence in human behaviour, including reasoning, learning, selfimprovement, goal-seeking, self-maintenance, problem solving and adaptability (Barr 1981). In a broad sense, it is the discipline concerned with the development and application of computational tools that mimic or are inspired by natural intelligence to execute tasks with performances similar to or higher than those of natural systems. Since its inception in the 1950s, the field of Al has produced a variety of tools. This section focuses on the most relevant Al tools applied to the solution of manufacturing engineering problems in design, planning, production, process modelling, monitoring, inspection, diagnosis, maintenance, assembly, system modelling, control, and integration. A summary of the application of these methods in recent papers contributed by ClRP members is given in Table 3, where manufacturing issues are grouped into the main categories of Design, Planning, Production and Systems.

2. lntellipent ComDutina Tools and Techniaues 2.7 MaDDinU of A1 Tools to Manufacturing


Numerous intelligent computing tools, techniques and paradigms have been applied to manufacturing related issues. A list is given below: 1. Artificial Life 2. Associative Memory 3. Automata Theory 4. Blackboard Architecture 5. Biological Manufacturing Systems 6. Chaos Theory 7. Computer Vision 8. Conceptual Dependency Graphs 9. Constraint Based search 10. Expert Systems 11. Fuzzy Logic 12. Genetic Algorithms 13. Graph Theory

63 1

SEARCH CONCEPTUAL

REASONING

LEARNING

Al tools, techniques and paradigms I WOV!XDXAEpR I NEURALNETS I FUUYLOGIC I

DlSTRlBUTEDAl

OTHERS

DESIGN

CONFIGURATION PARAMETRIC PROCESS PLANNING

PLANNING

SCHEDULING INVENTORY SYSTEMS PROCESS ANALYSIS MONITORING DIAGNOSIS

*Yes'

Neuro luzm 'Yes' New0 luzzy


'Yes'

pRowcTm

PROCESS CONTROL INSPECTION ASSEMBLY FMS MODELLING

'Yes'

Neuro luzm 'Yes'

I
SYSTEMS
FMSCONTROL
case-based

casebased

DlSlnDuled KB

Mull0 agents Holanr svslema

Pmceucakuku Biol menu1 sys~ Aulomala theory Pelnnets Process algebra Process cakulus
Blackboard

INTEGRATION

Asxlative memory

Cooperative Conslrawll propag

Frames Holon~ syslems

Table 2 - Mapping of A1 tools onto manufacturingissues.


Contributors Balazinsky BourdellMartin Bueno Din1 Eversheim Geiger Jawahir Jedrzejewski Kals Krause KNth MilbergIReinhart Monostori Nee Pteiler Rowe Design FL. SA Planning 3 FL 3 MA KBS 2 KBS. 2 NN, FL KBS. BA KBS,NN.
YE
I I*.
Q8"

Production 2 FL. NF KBS,3NN

Systems

KBS NN KBS KBS, NN HS KBS, SN

ELI

FL.2HS KBS KBS

I
KBS

I I I

GA

3KBS 2KBS

I KBS I2NN.2NF.

I KBS.NF I

KBS 2 HMS

KBS,FL. 2 HS. SA KBS,"

I KBS. BA.

2HS
KBS, NN NN

Table 3 - Applications of Al in manufacturing engineering. KBS = Knowledge Based Systems, NN = Neural Networks, FL = Fuzzy Logic, MA = Multi Agents, GA = Genetic Algorithms, SA = Simulated Annealing, Neuro-Fuzzy = NF, HS = Hybrid Systems, SN = Semantic Nets, BA = Blackboard Architecture, HMS = Holonic Manufacturing Systems, BMS = Biological Manufacturing Systems.

2.2.1 Knowledge Based and Expert Systems


Knowledge Based Systems (KBSs) are computer programs embodying knowledge about a narrow domain for solving problems related to that domain (Pham

1994). The first KBSs appeared in the 70s and because they contained high-level, domain specific knowledge elicited from human experts, they were called Expert Systems (ESs). Today, the terms KBS and ES are often used synonymously, the only distinction being the context in which the terms are utilised. A KBS or ES usually comprises three main elements: a knowledge base, an inference mechanism and a user interface. The knowledge base contains domain knowledge which may be expressed as any combination of IF-THEN rules, factual statements, frames, objects, procedures and cases. The inference mechanism allows to manipulate the stored knowledge for solving problems. Knowledge manipulation methods include the use of inheritance and constraints (in a frame-based or object-oriented KBS), the retrieval and adaptation of case examples (in a case-based KBS) and the application of inference rules (in a rule-based KBS) according to some control procedure ("forward chaining" or data-driven and "backward chaining" or goal-driven) and search strategy procedure ("depth first" and "breadth first"). Most KBSs are nowadays developed using programs known as "shells". These are essentially ready-made KBSs complete with inferencing and knowledge storage facilities but without domain knowledge. Some sophisticated KBSs are constructed with the help of "development environments". The latter are more flexible than shells in that they also provide means for users to implement their own inferencing and knowledge representation methods. The process of building a KBS or ES is called knowledge engineering. It typically involves a special form of interaction between the ES builder, called the knowledge engineer, and one or more human experts in some problem area. The main players in the ES gam'e are the ES itself, the domain expert, the knowledge engineer, the ES-building, and the user (Fig. 3). For more information on the technology of KBSs and ESs see Price 1990 and Durkin 1994.

632

Keynote Papers

provided during training: the NN learns automatically to cluster them in groups with similar features. Interviews
I

Extends, and tests lines,

Supervised Learning In supervised learning paradigms, backpropagation networks, which are multiple-layered feedforward networks (Rumelhart 1986), have been extremely popular for their performance.
4 4

du &
Clerical Fig. 3 - The players in the ES game. Among the Al tools considered in this paper, KBSs and ESs are probably the most mature, with many commercial shells and development tools available to facilitate their construction. Consequently, once the domain knowledge to be incorporated in a KBS has been extracted, the process of building the system is relatively simple. The ease with which KBSs can be developed has led to a large number of applications of the tool. In manufacturing engineering, applications can be found for a variety of tasks including selection of materials, machine elements, tools, equipment and processes, signal interpreting, condition monitoring, fault diagnosis, machine and process control, design, process planning, production scheduling, and system configuring. 2.2.2 Neural Networks An artificial Neural Network (NN) is a computational model of the human brain that assumes that computation is distributed over several simple interconnected processing elements, called neurons or nodes, which operate in parallel (Rumelhart 1986) (Fig. 4). The outputs of nodes in one layer are transmitted to nodes in another layer through connections that amplify or attenuate the outputs through weight factors. Except for input layer nodes, the net input to each node is the sum of the weighted outputs of the nodes in the prior layer. Each node is activated in accordance with the input to the node, the activation function of the node, and the threshold of the node; accordingly, the node fires an output. In this way, a NN provides a mapping through which points in the input space are associated with corresponding points in an output space on the basis of designated attribute values, of which class membership might be one. NNs can capture domain knowledge from examples, do not archive knowledge in an explicit form such as rules or databases, can readily handle both continuous and discrete data, and have a good generalisation capability. N N s can be employed as mapping devices, pattern classifiers or patterns completers (auto-associative content addressable memories and pattern associators). For more information on NNs, see Hertz 1991 and Haykin 1994. Applications in manufacturing engineering range from modelling, prediction, control, classification and pattern recognition, to data association, clustering, signal processing, and optimisation. Knowledge is built into a NN by training. Some NNs can be trained by feeding them with typical input patterns and expected output patterns. The error between actual and expected outputs is used to modify the weight of the connections between neurons. This method is known as supervised training. Some NNs are trained in an unsupervised mode where only the input patterns are

Fig. 4 - Multilayeredfeedforward backpropagation NN. In such a network, the input arrives on the neurons of the entry layer and spreads f.rom layer to layer to the exit layer (Fig. 4). The output of the network is dependent on the inputs and on the whole set of weights of the connections between neurons which can be indicated in vector form as w. The entire capability of adaptation is in the possibility of adapting the NN weight vector w. This adaptation operates as follows: suppose a set of examples are available for which the answer that the network must give (desired answer) is known. The network training problem is stated in the following manner: find the configuration of the networks weight vector w such that the error committed by the network on the whole set of examples would be minimal. That functions minimisation is achieved by a lowering of gradient according to w, which leads to a very simple algorithmic expression. Everything happens as if an error alert ran through the network in reverse (outputs to inputs), modifying the weights as it goes. The algorithm thus obtained is called the gradients backpropagation.
Unsupervised Learning In unsupervised learning, only input stimuli are shown to the network and the network organises itself internally so that each hidden processing element responds strong to a different set o input stimuli or closely related f group of stimuli. These sets of input stimuli represent clusters in the input space which typically represent distinct real word concepts. Among the numerous unsupervised NNs which have been proposed, the SelfOrganising Map (SOM) networks have been largely employed for their performance (Kohonen 1984). The SOM creates a two-dimensionalfeature map of the input data in such a way that order is preserved. In other words, if two input vectors are close, they will be mapped to processing elements that are close together in the two-dimensional layer that represents the features or clusters of the input data. The SOM network has typically two layers (Fig. 5). The input layer is fully connected to a two-dimensional Kohonen layer. In the Kohonen layer each neuron measures the Euclidean distance of its weights to the incoming input values. During the operation phase, the Kohonen processing element with the minimum distance is called the winner and has an output of 1, while the other Kohonen neurons have an output of 0. During the training, the

633

Kohonen processing element with the smallest distance adjusts it weights to be closer to the values of the input data. The neighbours of the winning neuron also adjust their weights to be closer to the same input data vector. The adjustment of neighbouring neurons is instrumental in preserving the order of the input space which is one of the most useful and intriguing aspects of the SOM.

V ar iabI es

Fig. 6 - Fuzzy logic processing.

--

000 000

Kohonen layer

Fig. 5 - Self-Organising Map (SOM) architecture.


2.2.3 Fuzzy Logic

Fuzzy Logic (FL) has two different meanings. In a narrow sense, FL is a logical system, which is an extension of multivalued logic. But in a wider sense, which is in predominant use today, FL is almost synonymous with the theory of fuzzy set (Klir 1988). A fuzzy set is a set without a crisp, clearly defined boundary. It can contain elements with only a partial degree of membership. A fuzzy set defines a mapping between elements in the input space (sometimes referred to as the universe of discourse) and values in the interval [0, 11. A membership function is a curve that defines how each point in the input space is mapped to a membership value (degree of membership or truth degree) between 0 and 1. The membership function can be any arbitrary curve, the shape of which can be defined as a function suitable from the point of view of simplicity, convenience, speed and efficiency. Typical membership function shapes are triangular, rectangular, trapezoidal, gaussian, sigmoidal, etc. The processing core of a FL is based on a collection of IF-THEN rules, where the IF part is called the "antecedent" and the THEN part is called the "consequent". Fuzzy rule sets usually have several antecedents that are combined using fuzzy operators. The combination is called a "premise" and it generates a single truth value that determines the rule's outcome. In general, one rule by itself does not do much good. What's needed are two or more rules that can play off one another. The output of each rule is a fuzzy set, but in general the output for an entire collection of rules should be a single number. Thus, first the output fuzzy sets for each rule must be aggregated into a single fuzzy set. The most common aggregation methods are: MAX (maximum), PROBOR (probabilistic or) and SUM (simply the sum of each rule's output set). Then, the resulting set is defuzzified or resolved to a single number. The most popular defuzzification methods are: centre of area (CoA), bisector, middle of maximum (MOM), largest of maximum, and smallest of maximum. The actual process of mapping from a given input to an output using FL is called fuzzy inference, which involves all the pieces that were discussed above. Thus, a global fuzzy inference system calculation is composed by the following five steps illustrated in Fig. 6. This inference method, due to Mamdani 1974, is the most popular one.

Neuro-Fuzzy Fuzzy systems are considered to be a natural link between symbolic and subsymbolic approaches in Al. On the one hand they can work in real time circumstances and handle uncertainties as NN, on the other hand they can manage both symbolic and numerical information. However, fuzzy systems usually do not incorporate automatic learning abilities and adaptive features. It seems that a very high performance can prospectively be obtained by combining NN and FL approaches and integrating their benefits. The resulting neuro-system is a hybrid system, where the architecture remains fuzzy, but using neural learning techniques, it can be trained automatically. Thus, the Neuro-Fuzzy (NF) approach (Halgamuge 1994) can comply with the requirements enumerated earlier, i.e. real-time nature, uncertainty handling and learning ability, but moreover expert knowledge can be easily incorporated in the system and the transparency of rule-based ESs is preserved.
2.2.4 Multi Agents and Holonic Systems

Multi Agents An alternative view of human intelligence has recently emerged which postulates that intelligence is the capability of a system to interact with its environment without clearly defined goals, to learn from this interaction and, in an incremental fashion, to both articulate and achieve its goals. The system may attempt to impose controls on its environment with a view to changing it to suit its goals, or to adapt itself to the environment if it comes to the conclusion that the environment will not change. Multi Agents (MA) represent a promising approach to design an intelligent system as a society of intelligent agents capable to deal with distributed problems (Rzevsky 1995a). The idea, inspired by Minski's seminal work (Minski 1985), is that a centralised, hierarchical control could be replaced by a group of loosely connected agents which are computer programs that are capable of communicating with each other, reasoning about received messages and collectively learning from experience. The overall system performance is not globally planned but develops through the dynamic interaction of agents. A plausible specification for such approach is as follows. A team of agents is given responsibility for planning and controlling the behaviour of a particular system (such as a machine tool). An agent within this team is given responsibility for initiating and controlling a particular machine behaviour (e.g. machining a workpiece, avoiding collision). Another agent is given responsibility for scheduling the machine activities with a view to maximising effectiveness of the higher system of which the given system is a component (a manufacturingsystem). Yet another agent is given responsibility for recording, keeping accounts and reporting on all system activities. All decisions are made by negotiation among agents stakeholders (those agents whose work may be affected by consequences of a particular decision). A protocol is established regulating negotiations and specifying nonnegotiable categories (such as safety);

634

Keynote Papers negotiation protocols impose constraints on the freedom of agent actions. In such a scheme a manufacturing system is seen as a society of intelligent agents negotiating with each other how best to achieve specified goals (Fig. 7). Each intelligent machine, in turn, is controlled by a team of intelligent agents rather then a centralised control system. The novelty of this approach is in replacing hierarchical architectures with network configurations in which nodes with advanced communication capacities are capable of negotiating how to achieve specified goals without any centralised control. Two main communication architectures can be distinguished: blackboard and message passing architectures (Kadar 1997). The blackboard is a shared memory space (common database) on which agents write messages, post partial results and obtain informatiop; the main characteristics of a blackboard architecture are: (a) independence of agents, (b) centralised control of communicating agents, and (c). event-driven behaviour. The classic message passing architecture involves a defined communication protocol and precise content; in the Contract Net Protocol (Davis 1983) the approach envisages negotiation consisting in task announcement, bid process, and award of a contract to accomplish the task. Other scheduling agents This technique can solve complex problems by imitating Darwinian theories of evolution on a computer (Holland 1975). The first step in the use of a GA is building a computer model to represent a given problem (Goldberg 1989). Interacting variables in the problem are first combined and encoded into a series of binary strings (rows of ones and zeros) to form numerical "chromosomes". The computer randomly generates an entire "population" of these chromosomes, and ranks them based on a "fitness function" which determines how well they solve the problem. Those strings which are deemed the "fittest" are allowed to "survive", and "reproduce" with other chromosome strings, through genetic operators such as "crossovet' and "mutation", to create "offspring" chromosomes. This population of strings evolves by continuously cycling the genetic operators (Fig. 8). A powerful search engine is thus available which inherently preserves the critical balance needed with any search: the balance between exploitation (taking advantage of information already obtained) and exploration (searching new areas). The result is then decoded back to its original value to reveal the solution. The output to be maximised (or minimised) is often a straightforward result, but special attention must be given to the design of the fitness function which is 'solely responsible for determining whether each chromosome is fit enough to survive and reproduce in the genetic population. Although simplistic from a biologist's viewpoint, these algorithms are sufficiently complex to provide robust and powerful search mechanisms.
Parent I: 100010~011110 Parent 2:01 101 110001 10 New string 1: 100011~000110 New string 2:011010~011110

agent/
Fitness Seeded Random Ievaluationi selectionH selection

Fig. 7 - A society of agents. Holonic Systems Some researchers working with the IMS concept refer to as the use of Holonic Manufacturing Systems (HMS) the solution for distribution of tasks in complex systems where a single problem solver, single machine or locus of computation is inappropriate (Davis 1980). In HMS, an emerging concept in the manufacturing arena, the idea is to provide a dynamic and decentralised manufacturing environment, where humans are effectively integrated, so that changes can be made dynamically and continuously. HMS are based on the notion of "holon" (Koestler 1967), a term that means at the same time a whole and a part of a whole. A holon is an autonomous, co-operative, and sometimes intelligent entity: a holon can be made of other holons (Fig. 43). A system of holons co-operating to achieve a goal or objective forms a holarchy; the holarchy defines the basic rules for coloperation among holons and thereby limits their autonomy. A HMS is a holarchy that integrates the entire range of manufacturing activities from order booking through design, production and marketing to realise the agile manufacturing enterprise (Valkenaers 1994). 2.2.5 Genetic Algorithms Genetic Algorithms (GAS) belong to a branch of computer science called "natural computation" where programmers, inspired by phenomena in the biological world, create models of these systems on a computer.

Fig. 8 - Structure of GA and genetically based operators. Genetic Programming Genetic Programming (GP) is a branch of GA. The main difference between GP and GA is the representation of the solution. GP creates computer programs in the lisp or scheme languages as the solution, whereas a GA creates a string of numbers that represent the solution. GP uses four steps to solve problems (Koza 1992): (a) generate an initial population of random compositions of the functions and terminals of the problem (computer programs): (b) execute each program in the population and assign it a fitness value according to how well it solves the problem: (c) create a new population of computer programs (copy the best existing programs, create new computer programs by mutation, create new computer programs by crossover); (d) the best computer program that appeared in any generation, the best-sofar solution, is designated as the result of GP. 2.2.6 Simulated Annealing The Simulated Annealing (SA) algorithm comes from an algorithmic analogy with the annealing of materials, where the purpose is to lead the material to a state corresponding to a global minimum of its internal

635

energy. This process is attained by searching the global minimum of an objective function in a space of solutions. SA uses a control parameter called temperature that is decreased by iterations until it becomes close to zero (Kirkpatrick 1983). This method allows the algorithm to generate random points for, which the objective function has a greater value (these points are accepted according to the Botzman probability) as defined with the Metropolis test. This worse situation may give a better minimum. New positions are randomly chosen on a feasible domain to satisfy the constraints. For improvement purposes, parameters are not modified simultaneously. One technique consists in projecting the variables on a lower dimensional subspace and moving only the selected parameters. A uniform distribution of the randomised choice is maintained to give each variable the same probability of being selected. SA algorithms are usually used to solve discrete variable problems; however, optimising continuous variables requires some adaptations of the algorithm (Brooks 1988).

Fig. 9 - Configuration of a hydrostatic journal bearing.

Design of recessed cylindrical Lm liquid fL journal bearing


(Click required icons, buttons and character titles)

1.

3. ADDlication of lntelliqent ComDutinq Methods in Manufacturinq Enqineerinq


(Onhographic Projection Applications of intelligent computation to manufacturing engineering are reviewed with reference to the four main issues of manufacturing systems: Design, Planning, Production, and System Level Activities.
3.1 Desian

Singleconstraint optiinisation based on minimum power dissipation Multi-constraint optimisation based on minimum power, maximum stiffness and minimum temperature rise.

) 2.

Technical Drawing
I

Fig. 10 - The system menu card.

Several researchers have utilised a variety of intelligent computing methods in the solution of problems and issues in Design. KBS were elaborated for representing knowledge interdependencies between sub-systems in parametrised product models to enable the designer to concentrate on creative tasks (Milberg 1994); designing AGV based on a conceptual scheme consisting of sensing the external world, defining motion parameters, taking judgmental decisions and travelling from source to destination (Furukawa 1991); integrating engineering design tools based on a variety of software applications of different types such as CAD, CAE, optimisation, etc., through a hypermedia tool (Rowe 1994a); developing a representation concept of technical objects and their characteristics by semantic modelling based on object oriented formalisms (Specht 1993); designing effective grooved chip breakers for minimum power consumption under varying machining conditions (Jawahir 1994); finding optimal thermal properties in the design of machine tools (Jedrzejewski 1994); accomplishing computer aided design of fixtures (Dini 1994). A computer aided fixture design was developed by Nee 1991 into a full 3D system using Nexpert Expert system shell. A powerful feature extraction software based on the principle of attributed adjacency graph (AGG) approach was implemented by Senthil 1996 which is capable of extracting some 35 different machining features. A case-based reasoning technique was devised by Senthil 1995 to index, retrieve and modify fixture designs which have been carried out previously to produce new designs.

START

Determine bearing parameters: W,D, L,N

NO
Select c
I

Determine bearing parameters:

Determine one of following restrictors: (a) Capillary tube; (b) Slot (c) Orifice; (d) Diaphragm valve Calculate bearing parameters: P,, q Modify bearing size parameters: a, b, c, D, L

Calculate bearing performance based on the bearing

Calculate restrictor parameters Modification

Fig. 11 - Hydrostatic bearing general design procedure.

636

Keynote Papers

Knowledge-basedsystem for the bearing selection

,/

....................

(Bearing selecting : instruction

'1

m,

i
f
a

h ae -tivertise b s 1I
Knowledge
r

:............................

To expert system for the bearing design

Fig. 12 - The basic framework of the KBS. An intelligent design system for hydrostatic journal bearing (Fig. 9) was developed by Rowe 1992. The incorporation of KBS into conventional numerical design techniques in the system makes it possible for the bearing design to take into account both numerical and qualitative issues. The system, implemented on Hypercard, includes three intelligent design modules (Fig. 10): the basic design module, based on the general design procedure of hydrostatic journal bearing reported in Fig. 1 1, the single-constraint optimal bearing design module, and the multi-constraint constraint optimal bearing design module. The design modules take the maximum stiffens, the minimum power dissipation and the minimum temperature rise as the bearing optimisation objectives. In each module, A1 in the form of KBS is incorporated into the procedure for bearing design defined by Fig. 1 1. The basic framework of the KBS in the modules is illustrated in Fig. 12. It consists of three parts, these being the KB, the inference engine and the user interface. The knowledge is collected from experts and built into the system as a KB structured by the "production rules" technique. For a KBS to generate a solution, an inference engine must be used to search the KB and logically arrive at a conclusion. As the starting point of the bearing design system reasoning and searching is given by the user requirements, the system employs forward chaining, the mechanism of data-driven reasoning, as its inference engine. Finally, the user interface is made in the form of dialogue between system and user. NN systems were devised for recognising complicated intersecting features for computer aided fixture design (Gu 1995) and for associating functions and structures for conceptual design using a feedforward backpropagation network (Kumara 1990), or an adaptive resonance network (ART2) (Kumara 1992). In Franke 1996, an automated construction method of press brake tools for metal sheet bending is developed. Construction methods utilised by a company are implied in previously realised constructions. Using a self learning NN, this implied knowledge is made usable for the development of new tools. The tool is designed in a way that no collision occurs in a specific bending stage, starting from predefined tool variants. First the right tool variant is derived and then the tool parameters from the shape of the workpiece are identified. A NN system finds the tool type class and the tool parameters on the basis of a given workpiece profile. Fig. 13 shows the tool variants and tool parameters for a gooseneck tool. The structure of the NN system is shown in Fig. 14. A two phase architecture is used: in the first phase, the workpiece is classified and the best tool variant is selected. In the second phase, the NN associated with the selected variant indicates the tool parameters. The NN input is the geometrical projection of the bending stage. The output of the first phase NN gives the classification of tool variants. The output of the second phase NN yields the parameters of the selected tool.

4
Fig. 13 - Tool variants and parameters.
1st

stogc c 1 ~ s s ~ 1 ~ c o t01 o n ~ tool r m o n t


8"D"t

0"tP"l

Fig. 14 NN application for variant construction. An NN system was used by Kowal 1996 for selecting the type of spindle bearing sets based on sample solutions quoted in catalogues of machine tool and bearings manufacturers. A NN trained on such examples is able to generalise the solutions over the entire search area and indicate the most suitable solution. As a result, factors such as bearing diameter, maximal rotational speed, bearing capacity strength, and bearing width are made available at the initial stage of design. Employing available information, data for various solutions of bearing sets with low power losses were collected. In the database, several types of design solutions of various sizes were stored together with characteristic parameters of the bearings. Care has been taken to ensure that examples for every bearing set covered the most often used part of its dimensional range.
1

Fig. 15 Diagram of designer/"

co-operation.

637

1000 2600 3000 4000 5000 6000

rotation speed Lrpni]

Fig. 16 - NN-based indication of the optimal value of the a/b parameter for a two-layer housing. A backpropagation NN was tested on a number of design and testing cases and it could correctly recognise all types of bearings. Fig. 15 depicts the NN application at the particular stage of design to indicate an optimal value of the a/b parameter for the two-layer spindle bearing housing. Having been trained by an adequately numerous set of representative solutions employing two-layer housings, the NN determines the sleeve thickness based on few parameters. The favourable outcome of the decision to apply the NN indicated sleeve thickness is illustrated by the bearing temperature diagram in Fig. 16. FL decision support systems were used by Dupinet 1,996 in order to deal with tolerance allocation problems encountered in the design of mechanical parts; the technique is a combination of FL to evaluate manufacturing difficulties and a SA algorithm to optimise tolerance allocation over the complete dimensioning chain through minimisation of the objective function. A hybrid Al system including rule based, graph based and NN based techniques was used by Henderson 1994 for the automated recognition of form features from boundary representations of solid models. Finally, a tool for automatic shape optimisation in the design of mechanical components based on mathematical programming algorithms is proposed by Weck 1995.
3.2 Planninq

An intelligent knowledge-based object-oriented process planning (IKOOP) approach was used by Lee 1993 to plan the manufacture of progressive die plates. The object-oriented scheme, production rules and heuristic representation are effectively integrated to represent the complex relationships in die design, process planning, heat treatment and fixture design. A case-based reasoning (CBR) approach was illustrated by Rowe 1996 for the selection of grinding conditions. CBR is an approach which seeks to identify a close match between a new operation to be performed and the characteristics of a previously successful case stored in a data base. It solves a current problem by studying the solutions to previously experienced problems and where necessary modifying the solutions to achieve appropriate conditions for grinding. The presented CBR system may be used for the initial selection of conditions which are then optimised by adaptive control or by an operator. The new optimised case may be learned and in this way the database is constantly developed so that the system may be considered to form a learning system. The modules of a generative CAPP system, named PART (Fig. 17), dealing with machining methods and cutting tools selection was developed by Van Houten 1990 utilising KB selection mechanisms embedded in the procedural structure of the PART system. Interactive graphics were used to support the maintenance. The inference engine was not fixed as in many ESs, but strategy and selection mechanisms could be specified and altered. The machining method selection module (Fig. 18) determines which sequence of machining operations is the most favourable to produce the manufacturing features considering the objective function and the constrains; it determines the type and range of cutter to be used. The cutting tool selection module combines the tool requirements over the different features and composes the complete set of unique tool-holder combinations required to machine the part. The design and implementation of the cutting tool database, shown in Fig. 19, is also described.
User

4 + User Interface
I
I
Supervisor

A number of researchers have employed Al appraches in dealing with various aspects and issues in Planning. KBSs were developed for determining the optimised bending sequences for complex die-bending parts (Hoffman 1992); selecting tool holder and insert in turning processes and minimising the number of tools for operations in the same set up (Pina 1990); establishing the optimal sequence of turning operations (Giusti 1989); classifying of the rough shapes of input part model based on arbitrary criteria specified by the user (Iwata 1987). ESs were designed for representing know-how and its application of machining reference surface in CAPP (Iwata 1986); automating tool selection in turning (Giusti 1986); developing an expert inspection task planning system for coordinate measuring machines (ElMaraghy 1987); determining drawing stages in cold drawing processes and specifying initial starting profiles of round and square bars (Brucker 1988); designing group technology part family database (Ham 1988).

Common Database Interface

Database

Fig. 17 - Schematic representationof the PART system.

638

Keynote Papers

: Features & Tools

: v
Database

Fig. 18 - Sub-modules of the Method Selection Module.

Fig. 20 - Planning system architecture for arc welding. The system allows an almost automated performance of the planning process, but the planner still has the opportunity to overrule the system. Therefore, results of the planning process are displayed by using the functionality of the CAE system to allow user validation. NN architectures were developed for grinding wheel selection (Li 1994); bending sequence determination (Geiger 1995a); modelling for planning of manufacturing cells (Westkamper 1997); computer aided process planning in multiple-blow cold forging (Alberti 1997a); automated selection of tools in turning (Dini 1995). NNs were positively experimented by Santochi 1996 to perform a fully automatised tool selection in turning. The particular procedure of knowledge acquisition and the capability to face unknown situations, without having the explicit rule for solution, make the NN an effective tool for this kind of problem. Using this approach, the technological knowledge for selecting tools was characterised through a set of representative examples concerning the various technological parameters (cutting edge angles, corner radius, insert grade, etc.). Fig. 21 reports, as an example, a 3-layer feed-forward NN used for the selection of the included angle e,.
HIDOEN LAYER

7, *

CunER TABLES

SHANK TABLES

IMAGE TABLES

Fig. 19 - Database layout for cutting tools.

In Eversheim 1995 a prototype task-level programming system for arc welding applications was developed which uses KB modules and planning modules embedded in a commercial CAE system. Input to the system is a product model that contains welding features generated during the design process. The planning functionality covered by the system starts with set-up planning and ends with the planning of sensor usage, including also simulation capabilities. Domain specific knowledge is used to control the planning process within the system more efficiently. A dedicated architecture of the planning system based on the blackboard principle, able to improve data integration and to provide a new approach to the optimisation problem in small batch production, is presented. Advantages of this principle are the support of the hybrid approach and potential parallel execution of the planning task. The architecture of the planning system for arc welding is shown in Fig. 20. Different categories of planning modules communicate by using a blackboard containing product and process data. A control module, realised using an ES shell, directs the planning process. This module analyses the blackboard and activates one of the planning modules after reasoning about the suitable proceeding. The major difference between the approach chosen and conventional blackboard-based system lies in the fact that domain specific knowledge is used to control more efficiently the planning process and to carry out the optimisation. Since some of the planning modules are implemented by using a CAE system, inter-process communication is used within the planning system.

OUTPUT LAYER

Machmng Type I 1 Cutting Condhons 12 ClompNng Type 1 3 Workplece Moterqol 14 Workpiece Slenderness 1 5

a .C _
+ ..

v)

0.5

0.5 0

.-

C m

U
C
v) Q

Q
E

I1

12 1 I5 3

: :

z
55 60 80

Fig. 21 - NN system for the selection of the included angle e, of a cutting tool.

639

FL reasoning systems were designed for allocating jobs


to the most suitable machines and assigning load

I
I

~ ~ ~ ~ ~ y s t e m
I

priorities (Balazinsky 1995b); selecting optimal cutting tools (Balazinsky 1995a); selecting object surfaces to be grasped by a gripper (Dini 1993); automating the tool path planning (Balazinsky 1994b); implementing a chipform classification technique and mapping methodology for use in CAPP systems (Fei 1993); developing group technology classification of mechanical parts (Zhu 1987); assessing feasibility of cold forged products in single blow operations (Alberti 1997b). An integrated application of ES and NN approach was used by Shan 1990 to solve the complex machining operation sequencing problem of single-spindle Swiss type automatics, the most common mode of producing small, precision parts typically found in watches and other precision instruments. In the hybrid system, instead of deciding the sequence for the selected operations in a pure algorithmic search manner, an ES is used to produce partial orders and to satisfy the constraints according to expert rules. A NN is designed to determine the final operation sequence. An effort to integrate computer-aided process planning and computer-aided fixture planning was attempted by Ong 1994 and 1996 using the fuzzy set theory. Set-up planning is proposed as an effective means to bridge this gap. The system developed uses the fuzzy set representation, together with production rules and object representation. This hybrid representation is able to depict the inexact relationship among features. Set-up plans produced are comparable to those practised by experienced planners. An intelligent technique for prediction of chip formkhip breakability in CAPP systems was developed and implemented by Ghosh 1995 based on the reconstruction and modification of the chip-groove profile, chip breakability classification, database management of machining data, and rule-based prediction of chip breakability, as shown in Fig. 22. 3-D data for the active tool face can be obtained from optical measurements by using a mapping algorithm, from which profiles at any cross-section on the tool face can be determined. Chip breakability is classified into five fuzzy sets: Very Poor (VP), Poor (P), Fair (F), Good (G), and Excellent (E) by using a fuzzy classification approach based on the size and shape of a chip. A relational DBMS model manages machining data, which is capable of searching information concerning insert type, insert profile, cutting conditions, chip-form, cutting forces, surface finish and material property, and is used as a tool in constructing rules for a predictive system for chip breakability. The rule-based predictive system for chip breakability is developed based on the analysis of effective chipgroove profile and is extended to include predictions for varying cutting conditions such as side-cutting edge angle, nose radius, cutting speed, with limited experiments. The rule-base is of the following form: IF Groove width is (GW), Backwall height is (BH} Backwall slope is (BS}, Rake angle is (RA}, ... AND Other conditions hold THEN Chip breakability is (CB} AND Chip breakability certainty-level is (certainty} The results of this rule-base operation are then postprocessed using a fuzzy inference engine to obtain the final prediction. A typical mapped chip breakability prediction chart is shown in Fig. 23.

Rule Base

tMachining theory

:
,

DBMS

niodule

module

L . . . . . . - . . - . . - . . - - - - - - - - . . . - . . - . . - - . - . . . . . - - . - . . . - .

Fig. 22 Modules developed and used in the rule-based predictive system for chip breakability.
7
ck.
Pnknm..hrm91.aMdd.lns

1 M

1.1

Fig. 23 - A mapped chip breakability prediction chart. Several Al tools were used by Greska 1993 for the planning of manufacturing operations for complex die bending parts: semantic networks for classifying metal sheet parts, FL for recognising different shapes of parts, and KBS for determining manufacturing operations based on the results of the classification process. A semantic net consists of a pattern of nodes and linking edges. In such a net, there are two fundamental types of objects: Concepts and Instances. Concepts are nodes of KB and contain definitions, algorithms for analysis, description of permissible attributes and explanations. Instances, on the other hand, relate. to the physical object. One Concept may have different Instances. The example in Fig. 24 contains two Instances for the Concept "Angle". Concepts and their relations make up the so called Expert Net. Evaluation and results of calculations, represented as Instances of the Concepts, are built up in the Results Net.
I n p u t Data
I

Processing

Output Dat
Angle I

li n f e r e n c e

Knowledae-Base
Angle 2

...

Result N e t Expert Net

Fig. 24 Data representation in a Semantic Network.

640

Keynote Papers

The recognition of the basic form of a complex die bending part is done by a simplified description of the geometry of the object: e.g. the shape of bending wings is approximated by simple shapes, such as a rectangle. A U-Form, shown in Fig. 25, is defined as follows: a U-Form is represented by 3 faces which have an outer contour similar to a rectangle. The value of the rectangle membership function must be at least 0.8; bending angles have to be about 90"; the breadth b l and b2 of two connected bending wings must have a ratio between 0.7 and 1.4; the I1 and 1 of two bending wings must have a ratio 2 between 0.5 and 2.0.

breakability in machining, and by Geiger 1995b for automating the classification of sheet parts to determine manufacturingplans and select appropriate tooling. An SA based approach was used by Shan 1992 to solve a typical TSP-type problem where there could be local optimal values to carry out assignment of tools on single-spindle Swiss type automatics. A GA based system was realised by Rossgoderer 1996 to optimise positions of production resources, positions of robots and movement sequences of cooperative handling tasks. The definition for a general optimisation task is given in the form: "minimise an objective function by varying certain optimisation variables and take into account that there are some constraints that must be adhered to." In the considered case, the objective function consists of a weighted combination of the following criteria: (a) kinematics stability of the robot arm configuration, (b) required cycle time, and (c) required driving torque. Examples for optimisation variables are: (a) position of production resources, (b) position of robots, (c) position of tool or workpiece, and (d) movement sequence of coordinated handling tasks. Possible constraints within a production system are: (a) access to all frames of each single robot program, (b) no collisions while robot movements, and (c) position constraints (e.g. perpendicular placement).

//
2
35 c

0.7 < b l / b 2 t l

. / 2 <20

Individual Genetic Algorithm (GA) Objective function value Arrangement of components


m

Fig. 25 - Determination of the rough geometry of a part. The topology of a U-Form, that consists of three connected rectangular faces, is represented directly in the Expert Net. The other features are gathered in Slots and the Test Function of a Concept. If the calculations of the Test Function are successful, an Instance of the tested Concept is generated in the Results Net. The definition of a basic form can even include rough and uncertain conditions. In the above example the bending angles have to be about 90". In this case, FL offers convenient methods to describe and process this condition very concisely. In this way, many variants with insignificant deviations can be described with few rules. In Fig. 26, the Linguistic Variable "Angle" has the two values: 90" and 120". Depending upon the given Base Variable verified on the real object, the appearance of a feature is evaluated differently. The relation between a Base Variable and the values of a Linguistic Variable is described by the membership functions.
ANGLE
Linguistic Variable
Values

I
I

Automatic Assessment Optimisation Criteria

Fig. 27 - System for automatic layout optimisation. Fig. 27 shows the whole system. The planner does not have to interact with the system while the optimisation runs. He only has to define the optimisation variables and the objective function with some simple menu functions that are integrated in the user interface. After activating the start button, the GA creates an initial population, which represents a set of alternative production system layouts. For each individual of the population, the layout is built up in the simulation system. Calculation of the criteria is done by simulating the robot programs while concurrently , watching constraints by checking collision and access of robot frames. By weighting and adding each single criterion, the current value of the objective function to be minimised is determined. Back in the GA, the fitness of the current individual is calculated by scaling the objective function value to a range between zero and a certain maximum value. After each individual of the current population has been evaluated, the next s population i produced by applying evolutionary operators such as cross-over and mutation. The results of the optimisation run are the optimised arrangement of components, optimised robot programs and the value of each single criterion.

120' Angle

1 )

Base Variable
Fig. 26 - Definition of a FL linguistic variable. Further hybrid Al systems combining NN and FL were employed by Osakada 1991 for process planning of cold forging, by Fang 1996 for predicting chip formkhip

64 1

3.3 Production
A significant amount of research work has been carried out in applying intelligent computing tools to the various aspects, topics and issues in Production. KBSs were developed for lifetime prediction of machine tool elements (Laucirica 1992); integration of deep knowledge about production with shallow knowledge about fault-cause relationships (Pfeifer 1993a); accommodation of unscheduled disruption, such as the breakdown of machines, high priority jobs, and power failures (Iwata 1988); an all-round solution to a number of problems related to the wire-EDM process (Snoeys 1988). ESs have been realised for conceptual diagnosis and maintenance of FMS working stations (Majstorovic 1989); keeping the knowledge on fault behaviour of automated assembly systems and making it available for diagnosis in the shop floor (Seliger 1990); tool condition assessment in order to avoid unnecessary interference with machining process (Villa 1985). In a number of papers on adaptive control systems for grinding, KBS and model based Al approaches were utilised to maximise feed rate, optimise dwell period, determine target position for size, test for impending thermal damage, and achieve close manufacturing tolerances (Rowe 1991); optimise grinding cycles by parameter adjustment using a simulation developed from a model of the deflection and force relationships of the grinding machine/wheel/workpiece system (Chen 1994); identify system time constant in real time by correlating measured data with a model (Thomas 1995); coping with variability of the grinding process with different dressing conditions and the effects of dressing .tool wear (Rowe 1995). In Kadar 1996, a reactive KBS for the management of manufacturing cells is proposed in order to cope with internal disturbances. Tool failure represents a critical type of internal disturbance and proper tool management can significantly increase production effectiveness. Using the information stored in the production database and supplied by monitoring systems, KBS techniques determine the control policy to be followed in a given situation. A hierarchical structured monitoring system was developed by combining an ES and a NN in a hybrid approach (Fig. 28).
*

wear is indicated by the monitoring module, the KBS task is to take into consideration the following factors: (a) estimated residual tool life, (b) available spare tools, (c) number of workpieces to be machined, (d) cost of tools, scraps, delays, etc. As a consequence, the KBS cell management can: (a) lower the cutting parameters, (b) replace the tool, (c) reschedule the task, etc., with proper messages to other parts of the system. NN paradigms were employed for noise analysis of machine tools (Pfeifer 1993b); identification of cutting conditions (Warnecke 1994); identification of tool wear state (Barschdorff 1993 and Monostori 1993); prediction of tool wear development (Wang 1992); monitoring progressive tool wear in single point turning operations (Dornfeld 1990); estimating the remaining cutting tool life under given cutting conditions (Teshima 1993); tool wear sensing in quasi-orthogonal cutting (Teti 1995b); monitoring of manufacturing processes (Grabec 1994); monitoring of turning processes (Etxeberria 1996); developing thermal actuators to actively compensate for thermal deformation (Hatamura 1993); developing a decision making process model for grinding operations with multi-stage structure consisting of feedfoward and brain-state-in-a-box NNs (Sakakura 1992); detecting the onset of chatter vibrations in grinding by removing the subjectivity of operator decision making (Chen 1996). In Revilla 1996, the deformations in a large milling machine (Fig. 29) due to 'temperature variations were studied and a NN method to compensate for them was provided. The algorithm utilises a FIR multilayer perceptron, i.e. an algorithm providing the advantages of a backpropagation NN and adding the possibility to include the temporal dimension of the phenomena.

11 II I
MOTOR

Y
HEADSTOCK

..........................................

1 D;g;seI #:
I
Production
Oatabase

:-.....

xo
.........
U " U " U ,

1
Fig. 29 - Milling machine. The main internal and external heat sources in the milling machine were identified in the ram. Thermal deformations depend not only on the temperature fields of the machine, but also on the position of the X, Y, Z axes. Heating tests were carried out to determine the relationships between temperatures and deformations. A NN trained with data from heating tests could provide fairly accurate estimates of deformations in subsequent tests. The main advantages of the NN approach are: no need to know a previous model to relate temperatures and deformations; the NN can learn linear and nonlinear models; most of the operations can be automated and the quality of results is independent of operator skill. Some disadvantages are present: it is not easy to program a NN in a simple PLC, which may imply the need of an external computer.

[Manufacturina

I y ..............: if, ANN !

Fig. 28 - Reactive, KB cell management system. The ES part receives information from the NN part, which performs the monitoring of the tool conditions, and from a production database, which includes the data on the machine, workpieces and cutting tools. The NN outputs are conveyed to an ES which provides process control information. On the basis of accumulated knowledge, the ES generates optimal process parameters and informs the user on the actual process state. The production database contains all the information necessary to control and manage the cell. The major part of the production database is the tool database, where tool wear and tool life data are stored as provided by analytical techniques and the on-line tool monitoring system. As an example, when extensive tool

642

Keynote Papers

A x e s Nndependonl I

aeformallan
I

4l
Tff

---OX

--oY

-Z
I

Axes position
Fig. 30 - Coupling between compensation algorithms. Thus, a two stage algorithm was realised (Fig. 30): first a FIR multilayer perceptron provides an estimate of the deformation of the machine in one point of its workspace. This deformation depends only on the temperature field and not on the position of the machine axes. Then, a' multivariable linear regression algorithm expands this deformation to any other point of the workspace. In this way, both algorithms are suitable to be implemented in a common CNC/PLC. FL decision support systems were used for the evaluation and selection of cutting conditions (Balazinsky 1993a); monitoring work material heat treatment conditions during machining (Teti 1995a); pattern recognition of metal cutting states (Wang 1985); control of machine tools to obtain the desired surface finish by controlling the machining process (Zhu 1982); nondestructive inspection of composite material electrical insulators (Teti 1997). In Balazinsky 1994a, a FL system based on machinist experience is utilised to select machining procedures and cutting tool size in the manufacturingof steel molds through milling. A fuzzy decision support system (FDSS) is particularly suitable to capture this type of expert knowledge because much of the information describing the machining process is imprecise. The molds are modelled as cavities (Fig. 31a): the goal is to machine the material in the cavity precisely and efficiently. The general machining procedure is a three step one: rough cutting, semi-finishing and finishing (Fig. 31b, c, d). Two types of tool are used: ball-end and flat-end mill cutters.

section to be machined, cavity depth and relative height of the islalid in the cavity. This information is entered as the K 6 in the FDSS which comprises the mold cavity fuzzy description and the results presentation as fuzzy indication of the best machining procedure (Fig. 32). The tool diameter is also selected on the basis of the cavity shape. In the case of rough cutting with flat end mill cutter, the cavity is described by four conditions (Fig. 33): top view area, average slope of cross section, complexity of cavity contour and contour shape. The conclusion is the diameter of the cutting tool to be used.
small medium larqe

0
72

144 288 432 576

864100811521296144

Area shallow deeo

0.3 0 . 6 0.9 1.2 1.5 1.8 Dewth no island small island

2.1

2.4

2.7

bia island

0.1

0.2

03

0.4

0.5 0.6

0.7

08

0.9

Relative island heiaht

Machininq Procedure

Fig. 32 - Machining procedure decision system.


?I, 3 1 1
I ' l K r C 'I

I I,.;;.

. . . ................
...........................

., .....
.......

18

27

36

45

54

.... 63

72

81

90

10

20

30

40

50

60

70

80

90

100

__

10

20

30

40

50

60

70

80

90

100
.

______.______-._

__

Fig. 31 - (a) Mold cavity; (b) rough cutting; (c) semifinishing: (d) finishing. The machining procedure is chosen on the basis of the cavity shape described by the area of the top view

Fig: 33 - Tool selection decision system. Black polygons in first 4 rows are the input from user: top view area 45*5 in2, slope 35"*5, complexity 40%&5, shape 30%. The conclusion is given in the last row: 1.5 in tool.

643

NF systems were applied to the monitoring of turning processes (Balazinsky 1993b); automatic control of EDM (Kruth 1995); tool wear sensing in turning (Monostori 1996a); tool wear detection in milling (Monostori 1994); provision of recommended grinding wheel specifications for particular grinding problems (Li 1994); decision making for the identification of vibration onset in grinding by removing the subjectivity of human operators (Chen 1996). In Westkamper 1996, artificial NNs are applied to the learning of the interdependencies between process conditions and quality parameters. The objective is to obtain a model to describe the grinding process more suitable than the present conventional methods. The authors argue that, on principle, the method can be applied also to other metal cutting processes. Process modelling for profile grinding was attempted using a multilayer feedfoward backpropagation NN and a product approach. In order to create the neural process model, the variable grinding parameters Vc, Vn, frd, and qd are used as input to the trained NN, and the process characteristics (FI, F,) and quality parameters (AS*, Ay, Ra, Rz) are obtained as output of the network (Fig. 34).
11. EGerimental ihase
Grinding

longitudinal form deviation .Ayon cutting speed Vc and speed rate VICis for example: "If V, is high and Vn is small then Ay is small". The second step is carried out in the adaptation phase (Fig. 35, middle). The rules formulated through linguistic terms (small, medium, high) are associated to ranges of values of input grinding parameters and output quality quantities through the help of the fuzzy set membership functions. The adaptation of the ranges of value is carried out through the partial neural model previously created with real experimental data. The experimental database is thus enhanced to cover the entire input parameter range and train a complete neural process model. This is done in the third step, the generic phase (Fig. 35, below).

CD grinding process
= I

1
Vr
VC

.,(=-

parameters
vc, VIb
frds q d

vn

Quality quantities Ash. Ay. Rz. Ra

Partially trained network

Membership functions

2. Training phase

Neural network Input


Grinding parameters

Vtl

Aswk

-+ frd
qd

Quality quantities

I.

13. Generic phase

phase

Fuzzy rules as in the

Ra

3. Simulation

Membership functions

Neural network Input


Grinding parameters

=+
Quality quantities
vc

training database

Input:
parameters

Vtl

Aswk

-+ ffd
qd

output
quality

Fuzzy rules as in the expert phase

Rr Ra

Fig. 34 Modelling the grinding process with NNs. Vc = cutting speed, V I ~ table feed rate, frd = feed, qd = = dressing rate ratio, AS& = wheel wear, Ay = longitudinal form deviation, R a and Rr = surface roughness parameters. This is carried out in 3 phases: experimental, training and simulation. A training database is set up in the experimental phase by varying input quantities and measuring the resulting process and quality parameters in grinding tests (Fig. 34, above). Then, the NN is trained using the previously determined training set (Fig. 34, middle). Finally, in the simulation phase the NN is evaluated using unforeseen input cases and verifying the NN output (Fig. 34, below). When modelling with NNs, it is possible to integrate previous available knowledge into the process model: this was carried out by representing expert knowledge in the form of fuzzy rules. The integration of previous available knowledge is carried out in 3 phases. The knowledge available on the process, i.e. the functional relations between grinding conditions and quality parameters, is gathered verbally and transformed into fuzzy rules (expert phase) (Fig. 35, above). A typical fuzzy rule for the dependence of the

Fig. 35 Integration of expert knowledge into neural process models by means of FL. Finally, a hybrid system making use of NN, FL and GA was developed by Egresits 1996 for tool state classification in milling on the basis of measured cutting force and vibration signals. 3.4 Svstem Level Activities A number of research workers have illustrated applications of intelligent computing techniques to topics, problems and issues regarding System Level Activities. An overview of application areas for KBSs in development, design and systems is presented by Krause 1995. KBSs make possible an increased utility potential for software in design and process planning, taking account of constantly increasing production system complexity. The authors identify activities that can be supported consistently with the present state-ofthe-art and undertake a categorisation of industrial applications and research activities in this area. To reduce performance losses in shop operations, Szelke 1995 considers reactive scheduling in the

644

Keynote Papers

reactive operation management of IMSs with two options concerning reactive repair and proactive adjustment to compensate fodprevent from disturbances during schedule execution. With a supervisory control perspective of IMSs' operational control, reactive scheduling with the above options is included as one major intelligent supervisor function of a Supervisory Reactive Scheduler (SUPREACT) and complemented with a learning function for improving search space/control knowledge to guide iterative schedule repairs. Cooperating rule and case-based reasoning and learning agents contribute to the opportunistic problem solving during iterative schedule repair via a blackboard architecture (Fig. 36). The casebased reasoning component's inherent learning ability allows to capture new schedule repair knowledge to be used for future decisions and improve system efficiency.
Clustering Rules Operator Shop Floor Statistics &Update Menlory

(Memory)

Find Similar Case Class (or Episode)


I

I.(

Similarity Metrics

)I I

Prior

"

Adapt One/Sonie If Necessary

Solition Evaluate (Select) Executable Solution Install Repair Solution

- - isodic Case
I

-I

Learned: Human Preferred

Fig. 36 Casebased reasoning and learning in SUPREACT.


#
~~

quality cssur-once

diagnosis

Fig. 37 - Structure of the knowledge presented for quality assurance and diagnosis. Reinhart 1996 verifies that existing concepts for control and malfunction handling are mostly centrally structured. Accordingly, the existing KBSs for diagnosis and quality assurance have a central structure (Fig. 37). These approaches do not fulfil future requirements towards increased availability and malfunction tolerance. The authors developed and implemented a new

hierarchic control concept for autonomous manufacturing cells. The approach enables an independent handling of occurring malfunctions, relieving the operators from routine interventions. The efficient use of decentral degrees of freedom to react autonomously to occurring malfunctions requires KB approaches. An analysis of the distributed KBs and different representations of knowledge showed that an integrated design of the KBS is vital for the successful development of future autonomous manufacturingcells. MA system approaches were adopted for manufacturing scheduling and control (Shafer 1995) and real time scheduling o production cells using the client-server f paradigm (Sohier 1993). According to Rzevski 1995b, the planning and control of the operation of a machine tool may be managed by a MA system made of five autonomous intelligent agents. One agent is charged with a goal of achieving the optimal speed for cutting various workpieces and another with a goal of maintaining the machine in the best possible working order. Under conditions of normal operation, the first agent will monitor speed of cutting and maintain it at the appropriate level, whilst the second agent will be monitoring the tool wear. When the tool wear reaches a critical limit, the second agent may decide that, unless the cutting speed is reduced, the tool will fail in the middle of the current operation and will send a request for slow-down. Since the first agent has a preference for continuing at the same speed until the end of the current operation, there will be a conflict. The conflict is resolved by negotiation in accordance with relevant protocols. Over a period of time, agents may be allowed to modify these protocols with the aim of improving the overall system effectiveness. The third agent, quite independently from the first two, schedules the work load and is given a task to minimise idle time for the machine under its responsibility without reducing overall system effectiveness. This agent negotiates the schedule with agents responsible for other machines in the shop. The fourth agent records all machine activities and commitments, and alerts other agents if the need arises. The fifth agent is monitoring the immediate environment of the machine and is responsible for avoiding collision with human operators, robots or vehicles transporting workpieces. A MA system was developed by Patriti 1995 to increase the robustness and reactivity of a machine tool with parallel architecture by using a self-organising distributed control system. The machine, named Shiva, is equipped with two axes capable of machining the same workpiece in parallel. A MA system for the control of the production activities of Shiva is designed with reference to the operative control of the machine: axes, grippers, tool magazine, pallets, as well as to elements not directly controllable: workpieces and tools. Through a simplified analysis of the system structure, the components of the machine and their connections are identified (Fig. 38). Three possibilities for decomposing the structure into agents are presented in Table 4. Each agent has knowledge only about its direct environment. The base for the exchange of information between agents is the machining "entity" defined as a geometric form and a set of specifications through which a machining process is known. The tool agents know what entities they can machine, how long it will take and in what way they can do it, whereas the workpiece agents know what entities they are made of.

645

Fig. 38 - Structural analysis of Shiva with identification of component agents.


4 Connection agents

--

Fig. 40 - General model of the MA system. Wiendhal 1995 developed a model describing manufacturing companies as systems where a redistribution of decision making responsibilities onto local distributed units is realised. In such systems, the distributed activities still need co-ordination. The authors demonstrate the application of market mechanisms also within a single enterprise. Company units which are autonomous and specialised in solving definite tasks have to co-ordinate their co-operation continuously by negotiating the conditions between each other. The model is based on the MA theory which permits describing the behaviour of such sub-systems in separate programs (agents) and realising the communication between them in the form of exchange of messages. In order to be able to solve an overall task, several agents in the system must cooperate with each other. The proposed market model and the division of tasks within a company are very closely oriented towards the roles demanders and suppliers play on the macroeconomicalmarket. The suppliers in this case are the resources, which are distinguished into machines and transport agents. Additionally, there is an agent for each customer order which represents the demand.

2 Axes Table 2 Spindles 2 Grippers X Pallets Tool Magazine X Worpieces X Tools

I 3 Connection aqents I 2 Connection agents 1 2 Axes+Spindles+Grippers I Machine agent


Table X Pallets Tool Magazine (comprising dl other components) X Workpieces

X Tools X Workpieces X Tools

Table 4 - Different structural decompositions of Shiva. Agents can have two types of functions: distributed and local functions. For example, the manufacturing of a workpiece implies the realisation of the following functions: (a) select the workpiece to machine, (b) select the cutting tool, (c) select the pallet, (d) select the machining operation to perform, (e) select the workpiece position. Functions (a), (b) and (c) are distributed in the system, and function (d) and (e) are kept local in the agent. To realise the distribution of the functions, a MA system model based on the negotiation network is set up (Fig. 39). In Table 5, the structure of distributed functions is illustrated for a simplified case which considers only three main functions.
MANAGER

c ..request .. ..... ,.!en+?-. ... Information system


n c e ment
4

'

- ...... - b
4
........

BIDDER
Fig. 39 - Negotiation network.
Function Select the workpiece Select the tool

+
I I

request _..- ' ..... - - - tender.- ... .....- .contract- b ?xcepttons ! a Yn! . e d . !e - ......

Implementation system

Bidders Workpiece agents Tool agents

Manager Free pallet agent Workpiece agent asking for machining Workpiece agent under machining

Select the pallet

Pallet agents

Bid Number of operations on that pallet Described by Time necessary to operations to be performed perform the operations Described by Number of remaining operations that operations to the pallet can be performed take care of

Task Described by type of pallet

I'

Order agent

Fig. 41 - Structure of an order agent. Fig. 41 depicts the structure of an order agent which includes three sub-systems: the implementation system that combines special abilities; the information system that carries out the communication with other agents by sending and receiving messages; the objectives system that characterises the autonomy of the agent in its relations to other agents. All agents have the task to care for nothing but the interests of their respective clients and to optimise their individual profits. This leads to competition situations among all demanders and among all suppliers.

646

Keynote Papers

agent. The latter develops a workplan with optimal prices and dates and informs the salesperson who finally proposes a tender to the customer. If the customer accepts, the order agent makes contracts with the machine agents which do the same with the transport agents. When the start date comes, the order agent notifies the respective transport agent, etc., until the entire order has been processed, i.e. the last machine reports completion (Fig. 42c). In Markus 1996, the concept of holonic manufacturing systems (HMS), based on the co-operation of autonomous, functionality complete entities with diverse, often conflicting goals, is elaborated. A market mechanism for co-ordinating the activities of intelligent holons that pursue their own interest by operating under bounded rationality in a changing, hardly predictable environment is used to solve dynamic order processing and scheduling problems. A new architecture and negotiation protocol for dynamic scheduling of manufacturing systems based on HMS is proposed by Sousa 1997. The main contribution in the architecture of the scheduling holon is the co-existence of holons representing tasks together with holons representing resources (Fig. 43). The well known Contract Net Protocol (Davis 1983) was adapted to handle temporal constraints and to deal with conflicts.

Production planning Holon

Fig. 43 - Holonic architecture. In Ueda 1997, Biological Manufacturing Systems (BMS) are proposed as a next generation manufacturing system concept. EMS aims at dealing with dynamic changes in manufacturing environments based on biologically inspired ideas such as self-growth, selforganisation and evolution (Fig. 44).

machine agent

transport vehicle agent messages

customer order agent material flow

> -

, +
(4

Fig. 42 - (a) Inviting tenders for operations required by a customer order; (b) inviting tenders for transport jobs required by operations; (c) processing a customer order. In Fig. 42, the scenario of order control is illustrated. When a customer asks a company about the possibility of getting a certain number of products by a given date, the salesperson creates on the computer a new order agent and informs it about the inquiry. The order agent breaks down the product into a list of parts and sends iniviations for tender to all machine agents (Fig. 42a). The latter find out which parts they can manufacture and, before submitting a tender to the order agent, send invitations for tender to the transport agents to verify under which conditions they can obtain delivery of necessary materials (Fig. 42b). Transport agents find out what they can transport depending on geometry and weight and submit tenders (price and date) for all feasible transports. The machine agents, taking into account transport prices, submit tenders to the order

gironment Genetic Information

Fig. 44 - An overview of self-organisationand evolution.

647

It also covers whole product life cycle from planning to disposal. Fig. 45 compares the life cycles of living organisms and artefacts: in common they have similar processes of birth-growth-deterioration-death, and morphogenesis (morphology generation in artefacts) and activities (use and consumption in artefacts) formed by energy and information. There are also differences: information in morphogenesis of living organisms (DNA) is inherited while in artefacts design information is supplied externally, DNA continues to exist throughout all the periods and in all the cells, and living organisms possess both DNA and BN-type information (acquired information dependent on the brain). DeatJ

4. Conclusions and Future Research Thrusts

yposal

DNA DNA
Repair

(a) organisms

The users motivations for developing and implementing intelligent computation in manufacturing systems relate to improving decision making, feature selection and classification; increasing performance, flexibility, efficiency and productivity; ameliorating modelling, control strategies, accuracy and optimisation; enhancing real-time operation, automation and integration. This puts stringent demands on future intelligent computing tools for manufacturing engineering. To fullfil such demands the following achievements are required: improving the available knowledge bases of manufacturing issues improving the performance of present Al tools and methodologies reducing computing time reducing costs for industrial applications identifying new applications for experimented/ successful Al tools Besides the improvements listed above, the future intelligent systems will be characterised by: integration of Al methods into CAD, CAPP, CAE, etc. development of hybrid Al systems development and implementationof IMSs newly elaborated Al methods and paradigms for manufacturing In conclusion, intelligent computing systems in the future will most likely be integrated, modular and hybrid in nature. The particular method utilised will be the method most appropriate for the particular module of the integrated system. Future systems may well include all the techniques described in this paper and further more.
References

(b) artefacts
Fig. 45 - Comparison of life cycles between organisms and artefacts. In order to realise BMS, there are at least two issues to be examined. One is to develop the embedding technology of biological information in artefacts, and the other is to understand the general behaviour of BMS, i.e. a theoretical aspect of the concept. Modelling of the BMS is related to the latter. The modelling theory is closely related to methods in Artificial Life such as GA, GP, Evolutionary Strategy, L-System, Cellular Automata, Reinforcement Learning and Immune Networks rather than symbolic approaches in Al. The paper focuses on theoretical aspects of the BMS adaptation, its application to manufacturing engineering is described and a floor level architecture for dynamic reconfiguration is modelled (Fig. 46).

Fig. 46 - Field model of the manufacturing floor. Finally, several Al tools were devised by Sat0 1993 for applying intelligence to machines in order to develop advanced machining in Japan.

Alberti, N., Di Lorenzo, R., Micari, F., Teti, R. and Manzoni, A. (1997a). Intelligent Computation Technique for Process Planning of Cold Forging, Pd World Conf. On Intelligent Manufacturing process and Systems, June 10-13, Budapest: 229-235. Alberti, N., Di Lorenzo, R., Teti, R., Manzoni, A. and Russo, G. (1997b). Computer Aided Process Planning of Cold Forging Based on a Fuzzy Logic Approach, 111 AITEM Conf., 18-20 Sept., Salerno. Balazinsky, M., Bellerose, M. and Czogala, E. (1993a). Application of Fuzzy Logic Techniques to the Selection of Cutting Parameters in Machining Processes, Int. J. for Fuuy Sets and Systems, 61: 307-317. Balazinsky, M., Czogala, E. and Gravelle, S. (1994a). Fuzzy Logic Techniques in the Automatic Manufacturing of Molds, Int. Fuzzy Systems and Intelligent Control Conference 94, Louisville, KY: 316-325. Balazinsky, M., Czogala, E. and Gravelle, S. (1995a). Automatic Tool Selection Using Fuzzy Decision Support System, FUZZ-IEEWFES 95, Yokohama: 615-620. Balazinsky, M., Czogala, E. and Sadowski, T. (1993b). Modelling of a Neural Fuzzy Controller with Application to the Control of Selected Machining Processes, Int. J. for Fuzzy Sets and Systems, 56, 3: 273-280. Balazinsky, M., Gravelle, S. and Trochu, F. (1994b). Automatic Process Planning Using a Fuzzy Logic Decision System, 10th ISP//FAC Int. Conf. on CAD/CAM, Robotics and Factories of the Future CARS & FOF 94, Ottawa: 325-330. Balazinsky, M. and Kops, L. (1995b). Intelligent Job Allocation on Machine Tools Using Fuzzy Logic,

648

Keynote Papers The First World Congress on Intelligent Manufacturing Processes & Systems, 13-17 Feb., MayaguezEan Juan: 950-961. Barr, A. and Feigenbaum, E.A. (1981). The Handbook of Artificial Intelligence, 1, Kaufmann, Los Altos, CA. Barschdorff, D., Monostori, L., Kottenstede, T., Warnecke, G. and Miiller, M. (1993). Wear Estimation and Stale Classification of Cutting Tools in Cutting via Artificial Neural Networks, Tooldiag '93, Int. Conf. On Fault Diagnosis, 5-7 April, Toulouse. Brooks, D.G. and Verdini, W.A. (1988). Computational Experience with Generalised Simulated Annealing over Continuous Variables, American J. of Mathematical and Management Sciences, 8: 425449. Bucker, M., Keller, D. and Reissner, J. (1988). CAD of Profiles from Round and Square Bars, Annals of CIRP, 37, 1: 247-250. Chen, X., Allanson, D.R., Thomas, D.A., Moruzzi, J.L. and Rowe, W.B. (1994). Simulation of Feed Cycles for Grinding between Centres, Int. J. Mach. Tools Manufact., 34, 5: 603-616. Chen, X., Rowe, W.B., Li, Y. and Mills, B. (1996). Grinding Vibration Detection Using a Neural Network, J. of Engineering Manufacture, 210: 349352. Chryssolouris, G. and Wrigth, K. (1986). KnowledgeBased System in Manufacturing, Annals of CIRP, 35, 2: 437-440. Davies, B.J. (1986). Knowledge-Based Systems in Production Engineering, Annals of CIRP, 35, 2: 423424. Davis, R. (1980). Report on the Workshop on Distributed Artificial Intelligence, SIGART Newsletter, 73: 42-52 Davis, R. and Smith, R., (1983). Negotiation as a metaphor for distributed problem solving, Ahificial Intelligence, 20, 1: 63-109. Dini, G. (1993). A Module for the Automated Planning of Grasps in Robotised Assembly Operations, Annals Of CIRP, 42, 1 1-4. Dini, G. (1995). A Neural Approach to the Automated Selection of Tools in Turning, I1 AITEM Conf., 18-20 Sept., Padua: 1-10. Dini, G. (1997). Literature Database on Applications of A Methods in Manufacturing Engineering, Annals of 1 CIRP, 47,2. Dini, G. and Santochi, M. (1994). Computer Aided Design of Fixtures Using a Rule-Based Selection Method, Idh Int. Conf. On Computer Aided Production Eng., 7-9 June, Palermo: 273-282. Dornfeld, D. (1990). Neural Network Sensor Fusion for Tool Condition Monitoring, Annals of CIRP, 39, 1: 101-105. Dupinet, E., Balazinsky, M. and Czogala, E. (1996). Tolerance Allocation Based on Fuzzy and Simulated Annealing, J. of Intelligent Manufacturing, 7, 6: 1-11. Durkin, J. (1994). Expert Systems Design and Development, Macmillan, Ney York. Egresits, C., Hornyak, J. and Monostori, L. (1996). f h Int. DAAAM Symp., 17-19 Oct., Vienna: 115-116. EIMaraghy, H.A. and Gu, P.H. (1987). Expert System for Inspection Planning, Annals of CIRP, 36, 1: 85-89. Etxeberria, J., Lopez, P., Arnaiz, A. and Alberdi, G. (1996). Toward Turning Monitoring by the Processing of Indirect Measures Using Neural Networks, XI Congress of Machine Tools and Manufacturing Technologies, San Sebastian. Eversheim, W. and Humburger, R. (1995). Dedicated Task-Level Robot Progamming System for Small Batch Production in Arc Welding, Annals of the German Academic Society for Production Engineering, 2 173-176. Eversheim, W. and Neitzel, R. (1986). Present Restictions in Industrial Application of AlTechniques, Annals of CIRP, 35, 2: 441-442. Fang, X.D., Fei, J. and Jawahir, IS. (1996). A Hybrid Algorithm for Predicting Chip Form/Chip Breakability in Machining, Int. J. Mach. Tools Manufact., 36, 10: 1093-1107. Fei, J. and Jawahir, I.S. (1993). A Fuzzy Classification Technique for Predictive Assesment of Chip Breakabilty for Use in Intelligent Machining Systems, PdI Conf., March 28 - April 1, San Francisco: 1275-1280. Franke, V., Backes, F. and Geiger, M. (1996). Automated Construction Methods for Press Brake Tools, 4th Int. Conf. on Sheet Metal, Enschede: 427434. Furukawa, Y., and Moronuky, A. (1991). Design Procedure of Intelligent Machines and Its Application to AGV, Annals of CIRP, 40, 1: 119-122. Cser, L., Lenard, J.G., Farkas, K., Backes, F. Geiger, M., and Hwu, Y.J. (1995a). Neural Networks in Metal Forming, PdInt. Workshop on Learning in lntelligent Manufacturing Systems, Budapest: 527-548. Geiger, M., Greska, W., Franke, V. (1995b). Classification Problems in Manufacturing of Sheet Metal Parts, 2 Int. Workshop on Learning in " ' Intelligent Manufacturing Systems, Budapest: 437461. Giusti, F., Santochi, M. and Dini, G. (1986). COAST: an Expert Module for Optimal Tool Selection, Annals of CIRP, 35, 1: 337-340. Giusti, F., Santochi, M. and Dini, G. (1989). KAPLAN: a Knowledge-Based Approach to Process Planning for Rotational Parts, Annals of CIRP, 38, 1: 481-484. Goldberg, D.E. (1989). Genetic Algorithms in Search, Optimisation, and Machine Learning, AddisonWesley, NY. Ghosh, R., Lin, M., Fei, J., Jawahir, IS., Khetan, R.P. and Bandyopadhyay, P. (1995). Chip Breakability Assessment Using a Chip-Groove Classification System in Finish Turning, Manufacturing Science in Engineering, ASME, MED-Vol. 2-1/MH-Vol. 3-1: 679-701. Greska, W., Hoffman, M. and Geiger, M. (1993). A Feature Based Approach for Simultaneous Engineering of Sheet Metal Components, Trans. of the North American Manufacturing Research Institution, XXI: 271-277. Grabec, I. and Kuljanic, E. (1994). Characterisation of Manufacturing Process Based Upon Acoustic Emission Analysis by Neural Networks, Annals of CIRP, 43, 1: 77-80. Gu, Z., Zhang, Y.F. and Nee, A.Y.C. (1995). Generic Form Feature Recognition and Operation Selection Using Connectionist Modelling, J. of Intelligent Manufacturing, 6: 263-273. Halgamuge, S.K. and Glesner, M. (1994). Neural Networks in Designing Fuzzy Systems for Real World Applications, Fuzzy Sets and Systems, 65: 112. Ham, I., Goncalves, E.V. and Han, C.P. (1988). An Integrated Approach to Group Technology Part Family Database Design Based on Artificial Intelligence Techniques, Annals of CIRP, 37, 1: 433438. Hatamura, Y., Nagao, T., Mitsuishi, M., Kato, K., Taguchi, S., Okumura, T. Nakagawa, G. and Sugishita, H. (1993). Development of an Intelligent Machining Centre Incorporating Active

649

Compensation for Thermal Distortion, Annals of CIRP, 42,l: 549-552. Hatvany, J. (1 986). Available and Missing Al-Tools, Annals of CIRP, 35,2:433-435. Hatvany, J. and Lettner, F.J. (1983). The Efficient Use of Deficient Knowledge, Annals of CIRP, 32, 1: 423-

Koza, J.R. (1992). Genetic Programming: on the Programming of Computers by Means of Natural Selection, MIT Press, Cambridge, MA. 1 Krause, F.L. and Schlingheider, J. ( 995).Development and Design with Knowledge Based Software Tools An Overview, Expert Systems with Applications, 8,

425.
Haykin, S. (1 994). Neural Networks: A Comprehensive Foundation, Macmillan, NY. Henderson, M.R., Srinath, G., Stage, R., Walker, K. and Regli, W. (1 994). Boundary Representation-Based Feature Identification, in Shah, J.J., Mantyla, M. and Nan, D.S., eds., Advances in Feature Based Manufacturing, Manufacturing Research and Technology, Vol. 20,Elsevier Science, Amsterdam. Herz, J., Krogh, A. and Palmer, R.G. (1991). Introduction to the Theory of Neural Computation, Addison Wesley, NY. Hoffman, M., Geissler, U. and Geiger, M. (1992). Computer Aided Generation of Bending Sequences for Die-Bending Machines, J. of Materials Processing Technology, 30: 1-1 2. Holland, J.H. ( 975).Adaptation in Natural and Artificial 1 Systems, The University of Michigan Press, Ann Arbor, MI. Iwata, K. and Fukuda, Y. (1986).Representation of Know-How and its Application of Machining Reference Surface in Computer Aided Process Planning, Annals of CIRP, 35,1 : 321-324. Iwata, K., Nemoto, T. and Sugimura, N. (1988). A Knowledge Based Production Control System to Accommodate Unscheduled Disruptions in the Manufacturing Process, Annals of CIRP, 37,1: 439-

2:233-248.
Kruth, J.P. and Van Coppenolle, B. (1995). New Trends in Automatic Control of Electro-Discharge Machining, 4'h Int. Conf. on Monitoring and Supervision in Manufacturing, 28-29Aug., Warsaw. 1 Kumara, S.R.T. ( 995). Intelligent Manufacturing Systems: Mapping of Al Tools to Manufacturing, Proc. Pd Meeting on Applications of A 1 Manufacturing Engineering, CIRP, Aug. 1995, Enschede: 1 19-1 45. Kumara, S.R.T. and Ham, I. (1990). Use of Associative Memory and Self Organisation in Conceptual Design, Annals of CIRP, 40,1 : 1 17-1 20. Kumara, S.R.T. and Kamarthi, V.S., ( 992).Application 1 of Adaptive Resonance Theory to Conceptual Design, Annals of CIRP, 4 : 2 3-21 1 1 6. Laucirica, J., Arnaiz, A. and Igartua, A. (1992).Expert System for Wear Control and Life Time Prediction of Machine Tools, COMADEM 92, 4" Int. Congress on Condition Monitoring and Diagnostic Eng. 7 Management, CETIM, 15-1 July, Senlis: 448-453. Lee, I.B.H., Lim, B.S. .and Nee, A.Y.C. (1993). Knowledge-Based Process Planning System for Manufacture of Progressive Die, Int. J. of Prod. Res., 31,2:251-278. Li, Y., Mills, B., Moruzi, J.L. and Rowe, W.B. (1994). Grinding Wheel Selection Using a Neural Network, 1 dh Nat. Manufacturing Research Conf., 13-1 5 Sept., Loughborough: 594-601. Mamdani, E.H. (1974). Application of Fuzzy Algorithms for Control of Simple Dynamic Plant, IEEE, 121,12:

442.
Iwata, K., Sugimura, N. and Fukuda, Y. (1987). Knowledge Based Flexible Parts Classification System for CAD/CAM, Annals of CIRP, 36, 1 : 3 71

320.
Jawahir, I.S. and Fang, X.D. (1994).A KnowledgeBased Approach for Designing Effective Groove Chip Breakers - 2D and 3D Chip Flow, Chip Curl and Chip Breaking, Int. J. Adv. Manuf. Technol., 9. Jedrzejewski, J. and Kowal, Z. (1994). Use of Knowledge Bases in Optimising Thermal Behaviour of Machine Tools, CIRP Int. Workshop on Design Theory and Methodology, Stockholm. Kadar, B., Markos, S. and Monostori, L (1996). Knowledge Based Reactive Management of Manufacturing Cells, Conf. on Integration in Manufacturing, 2-4Oct., Galway: 197-205. Kadar, B., Monostori, L. and Szelke, E., (1997).An Object Oriented Framework for Developing Distributed Manufacturing Architectures, 2nd World Congress on Intelligent Manufacturing Processes & Systems, 10-1 Budapest: 548-554 3 Kirkpatrick, S., Gelatt, C.D. and Vecchi, M.P. (1983). Optimisation by Simulated Annealing, Science, 220:

1585-1 588.
Majstorovic, V.D. and Milacic, V.R. (1989).An Expert System for Diagnosis and Maintenance in FMS, Annals of CIRP, 38,1 : 489-492. Markus, A., Kis Vancza, T. and Monostori, L., (1996). A Market Approach to Holonic Manufacturing, Annals Of CIRP, 45,1 : 433-436. Merchant,. M.E. (1984). Analysis of Existing Technological Forecasts Pertinent to the Utilisation of Artificial Intelligence and Pattern Recognition Techniques in Manufacturing Engineering, 1s'" CIRP Int. Sem. on Manufacturing Systems, 14, 1:

11-16.
Milacic, V.R. (1986).How to Build Expert Systems, Annals of CIRP, 35,2:445-450. Milberg, J., Reinhart, G. and Kaiser, J. (1994). Knowledge Representation through Constraints in Parametrised Product Models, Annals of the WGP,

2 173-176. :
Minsky, M. (1 985). The Society of the Mind, Heinemann. Monostori, L. (1993). A Step towards Intelligent Manufacturing: Modelling and Monitoring of Manufacturing Processes through Artificial Neural Networks, Annals of CIRP, 42,1 : 485-488. Monostori, L. and Egresits, Cs. (1994).Modelling and Monitoring of Milling Through Neuro-Fuzzy Techniques, Fd IFAC/IFlP/IFORS Workshop on Intelligent Manufacturing Systems, 13-1 June, 5 Vienna: 381 -386. Monostori, L., Egresits, Cs. and Kadar, B. (1996a). Hybrid Al Solutions and their Application in Manufacturing, dh Int. Conf. on Industrial and Engineering Applications of Artificial Intelligence and Expert Systems, 4-7June, Fukuoka: 469-478.

671-680.
Klir, G.J. and Folger, T.A. (1988). Fuzzy Sets Uncertainty and Information, Prentice-Hall, Engelwood-Cliffs, NJ. Koestler, A., (1967). The Ghost in the Machine, Hutchinson & Co., London. Kohonen, T. (1 984). Self organisation and associative memory, Springer-Verlag, Berlin. Kowal, Z., Winiarski, Z. and Jedrzejewski, J. (1996). Artificial Intelligence in the Shaping of Thermal Properties of Machine Tools, VII Workshop on Thermal Behaviour, Intelligent Diagnostic and Supervising of Machining Systems, W roclaw Technical University.

650

Keynote Papers
Monostori, L., Markus, A., Van Brussel, H. and Westkamper, E., (1996b). Machine Learning Approaches to Manufacturing, Annals of CIRP, 45, 2: 685-712. Nee, A.Y.C. and Senthil kumar, A. (1991). A Framework for ObjecVRule-Based Automated Fixture Design System, Annals of CIRP, 40, 1: 147-151. Ong, S.K. and Nee, A.Y.C. (1994). Application of Fuzzy Set Theory to Set-up Planning, Annals of CIRP, 43, 1: 137-144. Ong, S.K. and Nee, A.Y.C. (1996). An Intelligent Fuzzy Set-Up Planner for Manufacturing and Fixturability Evaluations, lnt. J. Prod. Res., 34, 3: 665-686. Osakada, K. and Yang, G.B. (1991). Neural Networks for Process Planning of Cold Forging, Annals of CIRP, 40, 1: 243-246. Patriti, V. and Charpentier, P. (1995). Application des SMA a la Fabrication. Cas de Shiva: Machine a Architecture Parallele, 3eme J. Francophones sur I'lntelligence Artificielle Distribuee et les Systemes Multi-Agents, St Bandolf: 66-78. Pfeifer, T. and Grob, R. (1993a). Using Deep Knowledge Acquisition in a Knowledge-Based Fault Analysis System, TdInt. Workshop on Learning in Intelligent Manufacturing Systems, 20-21 April, Budapest: 782-802. Pfeifer, T., and Ritschel, W. (1993b). Pattern Classification with Multi-Layer Perceptrons in Quality Assurance, Proc. of the EUFIT '93, Aachen. Pham, D.T. (1994). Artificial Intelligence in Engineering, 1st Conf. on Applications of A1 Techniques in Engineering, 5-7 Oct., Naples: 4-37. Pina, A., Bengoa, A. and Yaiez, J.I. (1990). Seleccion de Herramientas de Torneado Mediante un Sistema Basado en el Conocimiento, VIII Congreso de lnvestigacion Diseiio y Utilizacion de MaquinasHerramienta. San Sebastian: 1-15. Preiss, K. (1986). Artificial Intelligence in Manufacturing Systems, Annals of CIRP, 35, 2: 443-444. Price, C.J. (1990). Knowledge Engineering Toolkits, Ellis Horwood, Chichester. Reinhart, G., Diesch, R. and Koch, M.R. (1996). Knowledge-Based Information Processing in Manufacturing Cells - the Present and the Future, AISMC 3, Sept., Linz. Revilla, J. and Arana, R. (1996). Thermal Deformations Compensation by Neural Networks, VII Workshop on Supervising and Diagnostics of Machining Systems, 10-15 March, Karpacz: 1-14. Rossgoderer, U., Reinhart, G., Milberg, J (1996). Optimisation of Production Systems Using Genetic Algorithms, p d World Automation Congress, Montpellier. Rowe, W.B., Allanson, D.R., Pettit, J:A., Moruzzi, J.L. and Kelly, S. (1991). Intelligent CNC for Grinding, J. of Engineering Manufacture, 205, B: 233-239. Rowe, W.B. and Cheng, K. (1994a). Development of a Prototype Hypermedia-Based Engineering Design Environment-Hypercad, ESDA '94 Conf., ASME, London, 64, 5 : 123-129. Rowe, W.B., Cheng, K. and Ives, D. (1992). An Intelligent Design System for Recessed Hydrostatic Journal Bearings, Wear, 159, 95-105. Rowe, W.B., Cheng, X. and Mills, B. (1995). Towards an Adaptive Strategy for Dressing in Grinding Operations, 31" Int. Matador Conf., 20-21 April: 41 5-420. Rowe, W.B., Li, Y., Cheng, X. and Mills, B. (1996). Case-based reasoning for selection of grinding conditions, Computer Integrated Manufacturing, 9, 4: 197-205.
,

Rowe, W.B., Li, Y., Inasaki, I. and Malkin, S. (1994b). Applications of Artificial Intelligence in Grinding, Annals of CIRP, 43, 2: 521-531. Rumelhart, D.E., Hinton, G.E.and Williams, R.J. (1986). Learning Internal Representations by Error in Propagation, McClelland, J.L., Rumelhart, D.E. and the PDP Research Group, eds., Parallel Distributed Processing: Exploration in the microstructures of cognition, Foundations: 1. Rzevsky, G. (1995a). Artificial Intelligence in Engineering: Past, Present and Future, 7d* Int. Conf. AIENG 95, 4-6 July, Udine: 2-16. Rzevsky, G. (1995b). Mechatronics: Designing Intelligent Machines, Butterworth-Heinemann. Sakakura, M. and Inasaki, I (1992). A Neural Network Approach to the Decision Making Process for Grinding Operations, Annals of CIRP, 41, 1: 353356. Santochi, M. and Dini, G., (1996). Use of Neural Networks in Automated Selection of Geometrical and Technological Parameters of Cutting Tools, Computer Integrated Manufacturing Systems, 9, 3: 137-148. Sato, H. (1993). Basis and Future for Intelligent Machine, 1 Symp. on Advanced Intelligence Machine Complex, 1 1 June, Tokyo: 21-32. Seliger, G., Bollmann, 0. and Kaase, W. (1990). Knowledge Based Diagnosis in Flexible Automated Assembly, Annals of CIRP, 39, 1: 9-14. Senthil kumar, A. and Nee, A.Y.C. (1995). A Framework for a Variant Fixture Design System Using CaseBased Reasoning Technique, Manufacturing Science in Engineering, MED-Vol. 2-1/MH-Vol. 3-1: 763-775. Senthil kumar, A., Salim, F. and Nee, A.Y.C. (1996). Automatic Recognition of Design and Machining Features from Prismatic Parts, J. of Adv. Manuf. Tech., 11: 136-145. Shafer, K. (1995). L'Approache Multi-Agents en Productique, 27 Sept., University of Nancy. Shan, X.H., Nee, A.Y.C. and Poo, A.N. (1990). An Integrated CAPP System for Parts Machined on Single Spindle Swiss-Type Automatics, Computers in Industry, 14, 4: 281-291. Shan, X.H., Nee, A.Y.C. and Poo, A.N. (1992). Integrated Application of Expert Systems and Neural Networks for Machining Operation Sequencing, ASME Annual Winter Meetings, 57: 117-126. Snoeys, R., Dekeyser, W. and Tricarico, C. (1988). Knowledge Based System for Wire EDM, Annals of CIRP, 37, 1: 197-202. Sohier, C. and Bourdet, P. (1993). Real Time Scheduling of a Production Cell Based on a Multi Agent System, Int. Conf. on Industrial Engineering and Production Management, 2-4 June, Mons: 931939. Sohlenius, G. and Kyellberg, T. (1986). Artificial Intelligence and its Potential Use in the Manufacturing System, Annals of CIRP, 35, 2: 425432. Sousa, P. and Ramos, C., (1997). A Dynamic Scheduling Holon for Manufacturing Orders, 2nd World Congress on Intelligent Manufacturing Processes & Systems, 10-13 Budapest: 542-547. Specht, D., Forkel, M. and Weber, H. (1993). Semantic Modelling of Technical Objects, Manufacturing Systems, 22, 3. Sugeno, M. (1985). Industrial Applications of Fuuy Control, Elsevier Science. Szelke, E. and Monostori, L. (1995). Reactive and Proactive Scheduling with Learning in Reactive

651

Operation Management, Int. Working Conf. on Managing Concurrent Manufacturing to Improve Industrial Performance, 11-15 Sept.: 456-471. Teshima, T., Shibasaka, T., Takuma, M. and Yamamuto, A. (1993). Estimation of Cutting Tool Life by Processing Image Data with Neural Nerwork, Annals O CIRP, 42, 1: 59-62. f Teti, R., (1995a). Fuzzy Logic Approach to Sensor Monitoring in Machining, WlLF '95,21-22 Sept., Naples: 189-199. Teti, R. and Buonadonna, P. (1995b). Tool Wear Monitoring in Turning Using Signal Frequency Analysis, I1 A l T f M Conf., 18-20 Sept., Padua: 353364. Teti, R., ed. (1995c). Proc. 2nd Meeting on Applications of A1 in Manufacturing Engineering, published by University of Naples/LAPT, CIRP, Aug. 1995, Enschede. Teti, R., ed., (1996a). Proc. 3rd Meeting on Applications of A1 in Manufacturing Engineering, published by University of Naples/LAPT, CIRP, Jan. 1996, Paris. Teti, R., ed., (1996b). Proc. 4th Meeting on Applications of A1 in Manufacturing Engineering, published by University of Naples/LAPT, CIRP, Aug. 1996, Como. Teti, R., Guida, C., Della Torre, G., Manzoni, A. and Parretta, F. (1997). Intelligent NDE of Composite Material Components, fPhInt. Conf. AIENG '97, 7-9 July, Capri. Thomas, D.A., Allanson, D.R., Moruzzi, J.L. and Rowe, W.B. (1995). ln-Process Identification of System Time Constant for the Adaptive Control of Grinding, J. of Engineering for Industry, 117, 5: 194-201. Ueda, K. and Vaario, J. (1997). The Biological Manufacturing System: Adaptation to Growing Complexity andh Dynamics in Manufacturing Environments, 29' CIRP Int. Sem. on Manufacturing Systems, I 1-13 May, Osaka. Valkenaers, P., Bonneville, F., Van Brussel, H. and Wyns, J. (1994). Results of the holonic system benchmark at KU Leuven, 4th Int. Conf. on Computer Integrated Manufacturing and Automation Technology, 10-12 Oct., Troy, NY: 128-133 Van Houten, F.J.A.M., Van't Erve, A.H., Boogert, R.M., Nauta, J.M. and Kals, H.J.J. (1990). PART Selection of Machining Methods and Tools, 2" 2' CIRP Int. Sem. on Manufacturing Systems, 20, 1, Enschede: 69-87. Villa, A., Quaglia, R., Chiara, R., Rutelli, G. and Levi, R. (1985). An Expert Control System for Tool Life Management in Flexible Manufacturing Cells, Annals of CIRP, 34, 1: 87-90. Wang, Z. and Dornfeld, D. (1992). In-Process Tool Wear Monitoring Using Neural Networks, JapanNsa Symp. on Flexible Automation, ASME, 1: 263-270. Wang, M., Zhu, J.Y. and Zhang, Y.Z. (1985). Fuzzy Pattern Recognition of the Metal Cutting States, Annals of CIRP, 34, 1 : 133-136. Warnecke, G. and Muller, M. (1994). Application of Artificial Neural Networks for the Identification of Cutting Processes, Production Engineering, II, 1: 65-68. Weck, M. and Bussenschutt, A. (1995). Proc. WCSMO7, Elsevier Science. Westkamper, E., Lange, D. and Schmidt, T. (1996). Modelling in Grinding Process with Regression Models and Artificial Neural Networks, Annals of German Society for Production Eng., 3, 1. Westkamper, E., Pirron, J. and Schmidt, T. (1997). Development of an Adaptive Simulation System, Annals of the German Society for Production Eng., 4, 1.

Wiendhal, H.-P. and Ahrens, V. (1995). Knowledge Based Support for Planning and Control in Distributed Production Systems, lFlP WG 5.7, Working Conference on Manufacturing to Improve Industrial Performance, Sept. 11-15, Seattle. Zhu, J.Y., Shumsheruddin, A.A. and Bollinger, J.C. (1982). Control of Machine Tools Using the Fuzzy Control Technique, Annals of CIRP, 31, 19: 347352. Zhu, J.Y. and Zhang, Y.Z. (1987). An Expert System for Part Classification, Annals of CIRP, 36, 1: 351-354.

652

You might also like