JAMRIS 2007 Vol 1 No 4

www.jamris.org

Publisher: Industrial Research Institute for Automation and Measurements PIAP

Journal of Automation, Mobile Robotics & Intelligent Systems

ISSN 1897-8649

VOLUME 1, No 4

December 2007

ISSN 1897-8649

VOLUME 1,

No 4

December 2007

www.jamris.org

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 1,

No 4

December 2007

JOURNAL of AUTOMATION, MOBILE ROBOTICS and INTELLIGENT SYSTEMS

1

JOURNAL of AUTOMATION, MOBILE ROBOTICS and INTELLIGENT SYSTEMS VOLUME 1, N° 4 A quarterly focusing on new achievements in the following fields: ● Fundamentals of automation and robotics ● Applied automatics ● Mobile robots control ● Distributed systems● Navigation ● Mechatronic systems in robotics ● Sensors and actuators ● Data transmission ● Biomechatronics ● Mobile computing

Special Issue on Knowledge-intensive support of decision making in engineering processes Guest Editors: I. Horváth, J. J. Broek and J. Duhovnik Based on the Tools and Methods of Competitive Engineering (TMCE 2006) International Symposium, Ljubljana, Slovenia

CONTENTS ARTICLES 59

5 Decision Making in Innovation Processes – a Concept toSupport Small and Medium Sized Enterprises Stefan Vorbach, Elke Perl

A Specialized Multi-agent Search Engine Model for the Extended Manufacturing Enterprise Ockmer L. Oosthuizen, Elizabeth M. Ehlers

16 Knowledge-based Support of Decision Making at the Example of Microtechnology Albert Albers, Norbert Burkardt, Tobias Deigendesch

DEPARTMENTS 3

EDITORIAL

21 Systematic Decision Making Process for Identifying the Contradictions to be Tackled by TRIZ to Accomplish Product Innovation Hajime Mizuyama, Kenichi Ishida

Knowledge-intensive support of decision making in engineering processes 70 INTERVIEW

30 Layout of Functional Modules and Routing for Preliminary Design of Automatic Teller Machines Katsumi Inoue, Tomoya Masuyama, Hayato Osaki, Tokachi Ito 41

73 IN THE SPOTLIGHT 74

Neural Network Based Selection of Optimal Tool - Path in Free Form Surface Machining Marjan Korosec, Janez Kopac 51 Evolutionary Prediction of Manufacturing Costs in Tool Manufacturing Mirko Ficko, Boštjan Vaupotic, ˇ Jože Balicˇ

2

with professor Lotfi Asker Zadeh

EVENTS

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 1,

No 4

December 2007

Editorial

Decision making plays an important role in solving design and engineering problems. It appears on multiple levels, for instance on governmental (industry and enterprise), organizational (department and team), and individual (designer, engineer, manager) levels. Complexity of design and engineering processes has grown rapidly in recent times. Engineering decision makers have to take into account complexity and dynamic nature of engineering processes, and have to cope with the difficulty of making decisions based on varying input and output.

Designers and engineers often need to frame decisions under uncertainties, i.e. based on incomplete qualitative or quantitative information. Depending on the measure of incompleteness, three distinct types of decision making are recognized, namely decision making under risk, uncertainty and ambiguity. They may occur on each of the above-mentioned levels.

Decision making interweaves with the entire process of solving engineering problems. It starts as early as during problem definition and accompanies the engineering process as long as solutions are soaked for the engineering problem at hand. Engineering decision making typically involves four major actions: aggregation of knowledge, information and factual data about the engineering problem, defining or selecting appropriate decision criteria, generating multiple alternative solutions, and selecting the most competitive alternative solutions.

Computer-based tools and methods allow us to investigate the engineering processes in a detailed manner and with sufficient precision, and to integrate in-process decision making into the main-stream engineering processes.

These tools assume and enforce order in decision making processes. Using computational tools increases the amount of information available for problem-solving and decision making, as well as the reliability and robustness of decisions.

This special issue concentrates on problem-solving and decision making related to concrete engineering tasks and applications. The papers were originally presented at the Sixth International Symposium on Tools and Methods of Competitive Engineering (TMCE 2006), which was held in Ljubljana, Slovenia, in April 2006. The papers have been revised substantially for this Special Issue and have gone through a two-stage peer review process. They report both on new comprehensive theories and proven development results, most of them tested through practical applications. The papers indicate that decision making related to solving complex problems is indeed a fundamental construct in engineering. They also present examples

on information generation for problem-solving and multicriteria decision making by computational techniques.

The first paper titled “Decision making in innovation processes - A concept to support small and medium sized enterprises”, written by Stefan Vorbach and Elke Perl, proposes a systematic approach for handling high risk and uncertainties in realization of new products. A structured innovation process model is presented and the limitations of this process model are analyzed by using an application example. In addition, an innovation tool box (in other words, a set of innovation methods) is proposed which considers the overall innovation expertise of the involved designers and decision makers (beginners, experienced and professionals) of the company, the activity fields of innovation management (marketing, quality and cost), as well as the degree of innovation. Various aspects of innovation, such as information, communication, organizational, cultural and sociopsychological are analyzed and their effects on the success of product innovation are pointed out.

The second paper titled “Knowledge-based Support of Decision Making at the Example of Microtechnology”, authored by Albert Albers, Norbert Burkardt and Tobias Deigendesch, sets the stage for further discussion by comparing various general and specific product development processes used in mechanical engineering, mechatronics, micro-electronics, mask-based micro-systems, and toolbased micro-technology. Likewise the products, which include multiple subsystems, the product development processes also show a complex structure. The authors claim that their Contact and Channel Model (C&CM) can be used in the above fields to systematize mechanical analysis and synthesis of products. C&CM combines an abstract functional description with a concrete description of technical components or systems and, by doing so, supports design decision making. The authors propose to use rule- and/or guideline-based methods for designing technology-driven micro-systems.

Hajime Mizuyama and Kenichi Ishida contributed a paper titled “Systematic decision making process for identifying the contradictions to be tackled by TRIZ to accomplish product innovation”. This paper proposes a method for consideration of customers’ requirements through technical innovation of the functions and structures of products in development. The proposed method is based on the concept of elementary conflicts. Actually, elementary conflicts that prevent achievement of the target quality characteristics are recognized and eliminated Editorial

3

Journal of Automation, Mobile Robotics & Intelligent Systems

by a systematic analysis of the relationships between the engineering solutions and the required quality characteristics. As practical means of analysis, function trees and mechanism trees are used. Both serial and parallel relationships are taken into consideration, and the criticality of the elemental conflicts is enumerated. The method has been applied in the innovation process of a die-casting machine by using TRIZ tools.

The fourth paper titled “Layout of functional modules and routing for preliminary design of automatic teller machines”, also submitted by Japanese researchers, namely

by Katsumi Inoue, Tomoya Masuyama, Hayato Osaki and Tokachi Ito, uses genetic algorithm to resolve lay-out conflicts. The design problem is formulated as a minimization problem, which is mathematically represented as a sum of weighted design parameters. For the spatial arrangement of components and for routing the Bottom-Left method and the Route-Design-Oriented Packing method, respectively, are used. The arrangement is guided by constraints and rules. The genetic algorithm is used to achieve simultaneous optimization of the module layout and the route design. This contribution is a good example of how a combination of computational methods can be used to solve complex lay-out problem to ultimately support human decision making.

The fifth paper titled “Neural network based selection of optimal tool-path in free form surface machining”, coauthored by Marjan Korosec and Janez Kopac, is another proof of the applicability of dedicated computation methods to solving specific (non-deterministic) engineering problems. The challenge of the optimal tool-path selection originates in the relatively large number of engineering parameters and the need to achieve a multiple parameters dependent optimum. A large number of variations of five basic milling tool-path strategies are presented to the employed probabilistic neural network, having two hidden layers. The authors tested not only the efficiency of the proposed approach, but also the quality of the surface that can be achieved with the predicted optimal tool-path. They argue that human process planners are typically good in finding an efficient tool-path, but the neural network based computational approach takes the surface quality into consideration too. The computational approach allows consideration of other objectives, such as minimal tool wear and shortest machining time.

Another example of involvement of computational methods to support complex engineering decision making is presented in the paper titled “Evolutionary prediction of manufacturing costs in tool manufacturing”, auˇ It thored by Mirko Ficko, Boštjan Vaupoticˇ and Jože Balic. is well-known that determination of the manufacturing costs is challenging due to the large number of aspects (e.g. materials, design, machining, tools, measuring, transportation and cooperation), and due to the inherent uncertainties of product development processes. The authors analyze the problem of total cost determination and cost prediction methods. They propose a cased-based approach, which imitates the reasoning of a human expert.

The basis of computational solution is genetic program4

Editorial

VOLUME 1,

No 4

December 2007

ming, which enables automated prediction based on a set of strongly similar source cases. The cases are described by so called case vectors, which are used as input for the genetic code. Although the results are less precise than those achievable by humans, the authors claim that they can still be considered satisfactory. Future work will focus on the improvement of predictions.

Finally, the paper titled “Specialized multi-agent search engine model for the extended manufacturing enterprise”, submitted by Ockmer L. Oosthuizen and Elizabeth M. Ehlers, first review the advances in the field of Internet search engines. The authors claim that the major problem is that context-restricted multi-agent search engines are not yet available. They argue that search engines can be made context sensitive by personalization through user modeling and by specialization for a given knowledge domain. They introduce a reasoning model and a multiagent architecture that serve as a functional framework for specialized search engines for virtual manufacturing enterprises. Special agents such as user role, query, partner, and result analysis agents have been defined for the proposed COPEMSA search engine. To facilitate the performance evaluation of the functional framework in virtual manufacturing enterprises, the authors propose various metrics.

There is an emerging consensus that a framework for design and engineering is decision-making, and that context-dependent decisions are the fundamental construct in these processes. Application of comprehensive reasoning, engineering and computational models, combining human intuition and heuristics with robust computational analysis and simulation techniques, and using formal risk management and error minimization methods will be the main elements of future best practices. We hope that with this special issue, we managed to provide good examples of computer oriented methods and tools that can be introduced in the industry to support engineering problemsolving. Though further work is needed, the presented methods also support fact-based decision making in various engineering processes. We also hope that by these papers, we demonstrated the best research practices for fellow researchers.

We are grateful for the efforts and willing collaboration of authors, and we appreciate their scientific contribution to this special issue. Finally, we are very grateful to reviewers who helped to improve the quality of papers and the value of this special issue.

Prof. Dr. Imre Horváth Faculty of Industrial Design Engineering Delft University of Technology, the Netherlands ir. Johan J. Broek Faculty of Industrial Design Engineering Delft University of Technology, the Netherlands Prof. Dr. Jože Duhovnik Faculty of Mechanical Engineering University of Ljubljana, Slovenia

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 1,

No 4

December 2007

Decision Making in Innovation Processes – a Concept to Support Small and Medium Sized Enterprises Stefan Vorbach, Elke Perl

Abstract: Innovation is a vital factor in today’s markets and can dramatically change a company’s competitive position.

Despite its importance, innovation can also be a high risk adventure for companies due to the high level of uncertainty it entails. This article describes how well structured procedures can help companies guide through the whole innovation process, from idea generation to market launch.

Furthermore, it illustrates how the decision making process regarding innovation management can be supported by appropriate software. First, currently existing innovation process models are investigated. How they support small and medium sized enterprises (SMEs) in their innovation activities is analysed and a modification Thom’s well known process model is presented. Furthermore, a description of useful methods and instruments and their use within the innovation process is provided in the form of the innovation tool box. The innovation tool box helps companies both select and implement appropriate methods for the context at hand. A discussion of core factors and basic conditions for successful innovation management in SMEs rounds out the paper. A software tool for innovation management is then presented. Such software systematically supports innovation management and decision making throughout the whole innovation process. Furthermore, by demanding specific description and guidance of innovation methods it can additionally help companies in executing their innovation projects. How the database within the software can also be used to help companies distribute relevant information effectively is also described.

Keywords: innovation, process model, decision making, innovation toolbox, innovation software system, small and medium sized enterprises

1. Introduction The management of innovation in companies is a complex, insecure process fraught with risk, and often entails dramatic changes within the company (Preissl, Solimene 2003). These may involve increased levels of competition, rapidly changing market environments, higher rates of technical obsolescence and shorter product life cycles (Griffin, 1997). Companies face with a multitude of challenges. On the one hand, dynamic markets are seen as the pulling force in the innovation process, demanding new and improved products. On the other hand, fast technological developments and rival inventions act as driving forces (Perl 2003).

To deal with this complexity, a thorough, well thought out structure encompassing the whole innovation process from idea generation up to final dissemination of

the product or service is vital for success. Proper structure leads to a reduction of time to market, a reduction of R&D and production cost, and an improvement of product quality (Eversheim 1997; Sanchez, Perez 2003). However, a detailed look at companies, especially at small and medium sized enterprises (SMEs) is highly revealing. In contrast to large organizations, structured and well planned processes for innovation are often lacking in SMEs (Karapidis 2005, p. 437; Gelbmann et al. 2003). This is especially true with respect to evaluation and decision making during different stages of the innovation processes, and its considerable drawback. As already Cooper & Kleinschmidt (1986) noted, “what the literature prescribes and what most firms do are miles apart”. SMEs in particular, are not used to supporting planning and decision making methods in innovation management, and they also suffer from a general lack of awareness with respect to the whole process. Furthermore, in contrast to big enterprises there is often no strict division of responsibilities, duties and work packages. Sometimes the head of the company is responsible for the whole innovation process on his own.

As a consequence, very little time may be available for innovation, thorough structures and clear responsibilities are often missing. Successful realization of new ideas obviously requires specific forms of organization.

1.1. Objectives The overall goal of the paper is to develop and illustrate an integrated concept which can be used to aid innovation management, as well as provide methodical support and software assistance.

We first present a structured process model which serves as a guide through the whole innovation process.

This process model has been developed with the help of practitioners and is highly suited to the needs and demands of SMEs, especially concerning the limited financial and human resources available. Moreover, to add specific support for decision making in all phases of the product innovation process, we pay particular attention to the integration of methods and tools. This is illustrated with respect to the selection of proper methods and tools in the innovation process. Thus, the process model itself, together with the methods provided within the model and the software tool, act as guidelines through the whole innovation process and may help SMEs successfully execute innovation management. Finally, an important objective of this paper is to support innovation management by developing and evaluating specific software. The software used serves to integrate three key areas: the innovation process, support methods and the communication process between all the relevant actors.

Articles

5

Journal of Automation, Mobile Robotics & Intelligent Systems

1.2. Methods and means This paper is based on three research projects, ‘Innovital’, ‘Innovators’ and ‘Innoware’, which altogether took up more than 7500 hours, and involved four research institutes and 15 companies from different branches, mostly of small or medium size. Additionally, in order to validate the usability and practicability of the project results, a further 15 companies were linked to the project research in the form of workshops.

Initial research focused on describing the status quo with respect to innovation processes. The field appears to be well developed. Approximately 40 theoretical process models, e.g. stage-gate-models, were analysed. However, the practical implementation of these models is still in its infancy, and requires adaptation for SMEs. We thus developed an ideal process model which employs a bottom up approach in order to better take the needs and demands of the companies into account.

Secondly, an intensive review of literature revealed that lots of innovation methods exist, but that clear and systematic description, as well as assistance with implementation are often totally lacking. Hence, the utilization of such methods remains rather rudimental. An important focus research was thus to further adapt the methods such as to meet the needs and demands of SMEs. This meant finding means for aiding the transfer of theoretical knowledge of innovation processes and methods into practical usage in order to secure broader acceptance. The development of practical parameters designed to make method implementation self-explanatory was a main source of research success. Our attempt to connect process structures with innovation methods is important and for the field of innovation management and remains highly neglected topic in the German speaking world.

Another important area of research described here focuses on the need to develop appropriate software. After a thorough investigation of existing software no suitable software for the execution of innovation management was found. We thus combined the results of process analysis and the development of guidelines and developed a totally new innovation software tool. A case study of 8 companies from different size and branches was used to assess the practicability of the new software tool.

2. Innovation Processes Thomas A. Edison recognized innovation as the process of producing ideas into practical use. Thus, innovation covers not only the invention itself, but also its market implementation (Tidd et. al. 2001, p. 37). Is it possible to manage this innovation process? New product development is not just a series of predictable steps that can be identified and planned in advance. For many development projects neither the exact nature of the product nor the necessary means o production are known with certainty at the start of the development project. This uncertainty, along with the required amalgamation of project resources, calls on project managers to engage in project planning (Tatikonda, Rosenthal 2000, p. 402). Despite the great uncertainty surrounding innovation and its apparently random nature, several methods are available which help identify underlying patterns of success. We need to find to find out the extent to which innovation is based on 6

Articles

VOLUME 1,

No 4

December 2007

routine and whether such routine can be learned and subsequently used to form patterns and clear structures. This does not necessarily mean that the innovation process will always be the same, but it may mean that with the help of suitable structures and process models, execution may require less detailed, conscious thought (Tidd et al. 2001, pp. 46). Research has demonstrated that not skipping steps increases the probability of success for any project (Cooper 1990a). Firms therefore have to find a balance between firmness and flexibility in product development (Tatikonda, Rosenthal 2000, p. 417).

Addtionally, innovation not only comprises the creation of new products to satisfy customers but also deals with the development of new services, processes and organizational structures within a company (Perl 2003, p. 35). However, many of the key success factors for new service product development are identical to those identified for manufacturing firms (de Brentani 1989; for differences between service and goods see Griffin 1997, pp. 452). Nevertheless, differences between products and services in the execution of innovation management make it apparent that an innovation process model should fit all types of innovation.

2.1. Existing process models During the last few decades innovation management has become an essential scientific discipline. A large number of scientists now concentrates on the key issue of how to manage innovation. An overview of these endeavours can be found in the literature (Tidd et al. 2001, p. 51).

Common to almost all the literature on innovation management is the understanding that innovation must be viewed as a holistic process (Tidd et al. 2001, pp. 38, 44).

This has obvious implication for the management of innovation processes.

In recent years, four, or perhaps five, generations of innovation models have been produced (Huizenga 2004, pp. 52; Tidd et al. 2001, p. 43). Initially innovation was seen as simple and linear, and models reflected this point of view by focusing on linear processes, with either a technology push or a demand pull fostering the innovation process (for example see Schwery, Raurich 2004). Later, interactions between different phases of the innovation process were recognized and feedback loops were employed (Huizenga 2004, p. 55). In the next phase there was an emphasis on linkages and alliances, including an upstream with key suppliers, and a downstream with demanding and active customers. The latest generation of process models focuses on systems integration and extensive networking, flexible and customized response, and continuous innovation.

More than 40 innovation process models were analyzed for the present projects. These included process models such as those from Utterback and Abernathy (Utterback, Abernathy 1975), Thom (Thom 1980), Cooper (Cooper 1990b), Utterback (Utterback 1994), Pleschak and Sabisch (Pleschak, Sabisch 1996, pp. 24), Brockhoff (Brockhoff 1999), Van de Ven et al. (Van de Ven et al.

1999, pp. 23) Tidd et al. (Tidd et al. 2001, pp. 19) and Vahs and Burmester (Vahs, Burmester 2002) (for an additional overview of process models see also Zotter 2003, pp. 49 and Afuah 2003, pp. 13). However, although there is such

Journal of Automation, Mobile Robotics & Intelligent Systems

a huge number of process models, their practical dissemination is still in its infancy. They are often not appropriate for the needs and demands of SMEs, since they are very often too complex. Furthermore, they are not flexible enough to fit all specific kinds of innovation. As many of the process models described in the literature were found to be inadequate, the innovation process model described below was developed.

2.2. Development of a new innovation process model appropriate for SMEs As already explained above, the following innovation process model is based on a study of more than 40 process models for innovation management.

The starting point of the innovation process model is the innovation process of Thom (1980). According to Thom (1980), the innovation process as a whole is divided into three overall phases; idea generation, idea acceptance and idea realization. However, especially with respect to idea generation and idea realization, we introduce several modifications to make the model suitable for SMEs. Moreover, in contrast to previous process models the results of each phase are here clearly defined.

The innovation process begins with some form of initial impulse (see Figure 1). This may take the form of demands, requirements or complaints from customers, hints from the marketing and distribution department, ideas from employees, new technologies, ideas from the research and development department to name but a few (following Yates/ Stone 1992, Forlani/Mullins 2000, see also Hilzenbecher 2005, pp. 67, Disselkamp 2005, pp. 40).

In contrast to many process models, emphasis here is placed on problem analysis and strategy definition, i.e. the first phase of the innovation process. This phase focuses on defining innovation strategy an dhow the company intends to reach its goals. In the second phase it is

VOLUME 1,

Idea generation Idea acceptance

December 2007

important to generate as many ideas as possible. This also entails the collection of existing ideas and solutions. As a consequence of this rather creative and very vaguely structured process, a pool of new ideas should result. As only some of these ideas can be investigated in detail, a rough selection of ideas has to be made in phase III. Out of this selection, a couple of concrete problem solutions and ideas should then remain.

Within the overall phase of idea acceptance it is necessary to investigate and analyze the ideas in detail before a basic decision about the project can be reached. So it is important to evaluate all technical and economic aspects of the innovation that may have an influence on the success of the product. A feasibility study with detailed realization plans should form the basis for the subsequent decision for an innovation project in phase V. In comparison to other process models, we focus on those decision making phases which particularly affect company competitiveness.

Once a decision for an innovation project has been made, implementation may begin. First of all it is important to generate a detailed realization plan for the product or the service, covering aspects such as size, materials, resources needed, costs etc. This is a very important phase for subsequent realization since essential aspects and criteria of the product are defined and fixed (for integrating user needs see Franke, von Hippel 2003 or Jeppesen 2005). A strong focus is placed on this phase, since it is here that SMEs often appear to neglect key items.

The product/service plan is followed by development and prototyping, resulting in a complete prototype by the end of this phase. Once all aspects and features of the prototype have been appraised, production starts (this phase is absent in service innovation). Product launch is the final phase of this process.

When looking at these individual phases it is very important to realize that this innovation process model

Innovation initiation

Idea realization

No 4

I. Problem analysis and strategy definition II. Idea generating and collecting

Results Innovation strategy Pool of ideas

III.Rough selection of ideas

Concrete problem solutions

IV.Feasibility studies

Realization plan

V. Decision for a realization plan

Release for realization

VI. Concept definition

Concrete product/service concept

VII. Development and prototyping VIII. Production IX. Distribution and product launch

Prototype Product Market acceptance

Figure 1. An innovation process model specific for SMEs. Articles

7

Journal of Automation, Mobile Robotics & Intelligent Systems

should not be seen as linear. We have merely presented an ideal workflow illustration of the innovation process.

Loops and steps backwards to former phases may be both necessary and desirable, especially when earlier phases of the innovation process have to be revised.

The innovation process model described above can be seen as very suitable for SMEs as it is not too complex and covers all the important factors for the management of the innovation. The usefulness of the process model has already been tested in several companies and branches, e.g. in primary industry, steel manufacturing companies, logistics software enterprises and in machine manufactures. It is suitable to both service and manufacturing companies.

2.3. Barriers for a structured innovation process model Although the necessity and the possibilities for innovation process models are obvious, it cannot be denied that face several restrictions in practice.

First of all, a process model cannot replace organizational management and control of the innovation process.

It serves merely as a guide, and offers an ideal-typical path for optimal innovation. It is essential that it not be seen fixed and unalterable. Companies often have to adapt as they innovate and follow process loops. They have to be conscious of the nonlinearity of the innovation process and must be prepared to retreat one or two steps in order to revise the plans.

Another limitation of process models for innovation management concerns the different organizational styles that are often needed. Within the idea generating phase, a rather unconstrained management style is needed in order to boost creativity. Further on, when the idea is being implemented, it is vital that the process is fairly rigorously organized. This is often called the ‘organizational dilemma’ (see Wilson 1966, pp. 195). However, Souder recommends that the most appropriate structure depends upon the level of innovation desired, and on the stability of the market and technical environment (Souder 1987). The complex interdependencies between the organizational context, business processes, and individual performers have also to be considered (Massey et al. 2002).

Last but not least there is often personal resistance to fixed or structured advancement. Employees feel imprisoned within the process and fear losing their personal area of responsibility and scope for development. It is the responsibility of top management to reduce such barriers and minimize such areas of potential conflict. In particular, flexibility and openness is required to overcome conflict arising as a result of employee resistance (Wengel, Wallmeier 1999, p. 77).

3. The support of methods in the innovation tool box To successfully execute an innovation, it is not only of great importance to know each step of the process, it is also vital to know how to perform each phase of the innovation process in an ideal way. A study of 120 product development projects found a strong positive correlation between project methods and the project success achieved. Those companies that successfully employed specific methods to support their development process 8

Articles

VOLUME 1,

No 4

December 2007

were more likely to succeed in the overall product innovation process [Tatikonda, 2000, p. 402]. Nevertheless, as other studies show, the use of specific tools and methods remains marginal (Farris 2003, p. 31; Gelbmann et al.

2003). Companies, especially SMEs, are largely unaware of innovation methods, nor do they dare to use them.

The following innovation tool box describes and illustrates how specific innovation methods can be used to support the innovation process.

3.1. The innovation tool box The development of the innovation tool box was preceded by a study of existing guidelines on innovation management appropriate for SMEs carried out in 2003 (see Gelbmann et al. 2003, p. 86).

This research revealed that no complete guide for the management of innovation processes was available, and even fewer tools were found that might be suitable for sustaining innovation processes in SMEs. We thus decided to take up the challenge and develop a scheme suitable for the SMEs that could cover the whole innovation process, from initial idea generation, on to product/service development and distribution (see Figure 2).

The innovation tool box is based on the process model explained above. The phases from idea generation, idea acceptance and on to idea realization are represented on the left hand side. Specific methods are selected for each phase of the process. These can be seen on the right hand side of the tool box. However, circumstances and core conditions in companies always differ considerably.

As a consequence, the methods offered within the innovation tool box are subdivided into several groups and depending on respective phase within the innovation process and on specific company characteristics. A more detailed description of the division of the methods will be provided in the next section.

3.2. Presentation of methods in the innovation tool box Acceptance of the innovation tool box by SMEs depends on it being as comprehensible as possible. Furthermore, time is a very scarce resource for these companies; hence explanation of the methods involved needs to be simple and compact. It is also important to explain the methods and in such a way that the companies can use them without further assistance from consultants. The following structure for method description was thus chosen (see Table 1).

First of all general facts of the methods are presented.

Secondly, the complexity of the method is explained using specific criteria (see Table 1). Criteria such as time, participants needed, degree of difficulty, to name but a few, were chosen and validated according to their operability with the companies included in the projects.

A further description of the procedure depicts how the method has to be used. Additionally, an example illustrates this in practice and helps participants understand the method. Last but not least, references for further information are provided.

The exact allocation of methods in terms of phase and specific company characteristics is undertaken as follows.

Journal of Automation, Mobile Robotics & Intelligent Systems

II. Idea generating and collecting

Professional

GAPAnalysis

ProductMarketAnalysis

Scenario Method

Portfolio Methods

Corecomp. analysis

Brainstorming (BS)

Morphol.

Matrix

635 Brainwriting

SILMethod

Inverse Brainstorming

Destruc.construc.

BS

III.Rough selection of ideas

Screening with k.o.-criteria

Product modification

Radical innovation

IV.Feasibility studies

IX. Distribution and product launch Marketing-Management

Product ,margin

QFD

FMEA

VIII. Production

Lead User Concept

Marketing Information

VII. Development and prototyping

Target Costing

Scoring method Pairwise comparison R&D Calculation

Simple evaluation method Balance of arguments FMEA

Target Costing

Conjoint Analysis

V. Decision for a realization plan

calculation

Function analysis, Market analysis Quality function deployment

VI. Concept definition

Idea realization

Advanced

December 2007

SWOTAnalysis

Product Clinic

Idea acceptance

Idea generation

Beginner I. Problem analysis and strategy definition

No 4

VOLUME 1,

Quality-Management

Cost-Management

Figure 2. The innovation tool box.

Criteria

Evaluation of the criteria

Comments Depends on the object itself and on the experience of the project leader with this method

Time

Participants needed

3-7

Team work recommended

Spatial requirements

1 Room

No further requirements

Extras needed

Writing material, flip chart, pin board, additional presentation material

Moderator

needed

Moderator is responsible for: Coordination of the method Tracing of the method through its whole cycle Documentation of the results participants’ experiences during sessions

Degree of difficulty

high

External assistance and consultancies can be recommended if no experience with this method is available

Table 1. The complexity of the method “Target Costing.

Articles

9

Journal of Automation, Mobile Robotics & Intelligent Systems

Division according to overall company experience with innovation methods For the first three phases of the innovation process (problem analysis, idea generation and idea selection), the methods proposed are classified in terms of participants’ previous experience with innovation techniques, and reflect the needs of beginner, advanced and professional innovator. A checklist with 12 short questions helps the companies allocate themselves to one of these groups.

For the phase covering rough selection of ideas with knock out criteria, only one general method is suggested as this process phase has to be done very thoroughly, no matter which innovation type the company represents.

Division according to the degree of the innovation After rough selection of the idea, a further classification according to the innovation type is no longer appropriate. As the innovation is now more clear and specified, methods can be chosen corresponding to the type and novelty of innovation. Therefore, within the overall phase of idea acceptance, the innovation tool box distinguishes between incremental and radical innovation and allocates the methods accordingly.

In the literature many definitions concerning classification into degree of novelty exist (see for example Veryzer 1998; pp. 306, Tidd et al. 2001; Afuah, 2003, p. 14 Bourque 2005, p. 72; Hauschildt, Salomo 2007). If the degree of innovation is defined in terms of the extent to which it impacts a firm’s capability, then the term radical innovation means that the technological knowledge required to exploit it is very different from existing knowledge. In contrast, incremental innovation is based on existing knowledge and on a current organizational pattern (Afuah 2003, p. 14). The emphasis in incremental innovation is cost and feature improvement in already existing products whereas radical innovation concerns the development of new businesses and products (Bourque 2005, p. 72).

However, the dimension of novelty is not only important for predicting an innovation’s rate of adoption and diffusion. It also influences an innovation’s developmental pattern (Veryzer 1998, Van de Ven 1999, p. 63). In this way the novelty of an innovation also has an impact on the selection of appropriate methods to support the innovation process. The innovation tool box thus provides scope in idea acceptance, by allowing for a division into radical innovation or modification and improvement of existing products and services. This applies to both the phase of feasibility studies and to that involving decisions on the start of the innovation project.

Division into marketing, quality and cost management For the final phase of innovation realization, in order to reduce complexity, methods are only divided according to the key area they affect. Furthermore, the methods cannot be allocated to only one phase in the innovation process. Of course they do focus on particular phases, but as they can also affect other phases within the realization, the innovation tool box indicates both the focal point of the method and illustrates the other phases affected.

To ensure that method selection remains relatively straightforward for the companies concerned, methods are divided into the areas marketing management, qual10

Articles

VOLUME 1,

No 4

December 2007

ity management and cost management. If a company only needs help and information in one of these sectors it can easily find an appropriate method. However, it must be noted that a company still has to cope with all three areas to successfully manage their innovation process.

Although project management methods are also of importance, they have here been excluded from the general tool box and placed in a somewhat outer layer of the innovation process. Hence, in the realization phase, only methods suitable for marketing management such as the lead user concept, or Failure Mode and Effect Analysis in quality management, or for example target costing for cost management, are presented within the toolbox.

To sum up, the innovation tool box provides a systematic overview of both the innovation process as a whole and of the respective methods that can be used to support each specific phase of the process. Furthermore, by allowing for various subdivisions, it enables companies to identify the right methods for their specific problems.

4. A software program to support Decision making in innovation management The use of software programs can be helpful in supporting innovation management and related decision making processes.

In recent years, the market for innovation management software has spread rapidly. However, an analysis of existing software solutions revealed that no complete solution for innovation management is available that covers all phases from problem analysis up to market launch (Vorbach, Perl 2005). Moreover, existing commercial software systems often cannot be used directly within companies, and they especially fail toe meet the specific needs and demands of SMEs. Consequently, customization with respect to SME requirements is clearly necessary. As part of the Innoware research project, the needs and demands of companies were analyzed and discussed, and then used as the basis for further software development.

Additionally, such software application has to target several goals in order to successfully meet the needs and demands of small and medium sized companies. First, it is essential that the software encourages a team focus, as innovation is mostly undertaken in teams from different departments. Second, the software should induce the maximum participation of all employees, regardless of their level in the company hierarchy. This is achieved by making the software s self-explanatory as possible. Finally, the software needs to be capable of adapting a process perspective.

4.1. The innovation workflow as a basic feature of the software program and knowledge management perspectives Workflow conception preceded the above mentioned intensive analysis of innovation processes using the process model presented in Figure 1 and formed the basis for further development of the software. Workflow, as a basis for future software solutions, consists of the following general phases (see also Staltner et al. 2006): • Problem identification and innovation initiation, • Idea generation, • Idea assessment and selection, • Idea realization.

Journal of Automation, Mobile Robotics & Intelligent Systems

Finding the most appropriate structure depends upon the given level of innovation and on the prevailing stability of the market and technology environment (Griffin, 1997, p. 442). Therefore, with respect to innovation software, the various steps of the innovation process are not all obligatory for companies. Each company can decide for itself whether to execute all steps of the process. While complete execution is recommended, users may skip various steps. The system informs them of their omissions.

This guarantees as much freedom as possible for the users of the software; but, some kind of guidelines should still be implemented by the company.

4.2. The support of knowledge management through specific databases The innovation software includes a database to facilitate the storage of knowledge about the specific innovation project. In this database, data about the innovation project currently being undertaken is stored. All information about the innovation project, such as market data, technical features, the results of tests and assessment, need to be saved.

It is also important that knowledge of all past ideas is stored.

This enables employees to check whether specific ideas and solutions have already been attempted, and, if they were rejected, the reasons for rejection can be reviewed. The database should thus help tacit knowledge on innovation ideas and projects become common knowledge and make it easily accessible (see Karapidis, 2005 on integrating ideas into company knowledge management).

In the subsequent warrant of apprehension, the ‘new idea sheet’, the data management within the software is illustrated. At the top of the sheet, there is a clear iden-

VOLUME 1,

No 4

December 2007

tification of the project with an identification number, an abbreviated designation, and a few key words and the date of receipt. The names of the idea generators and their organisational units are also situated at the beginning. Afterwards, a short description of the project follows, as well as links and attachments for further information. These are the most important data for submitting and presenting ideas. However, the new idea sheet also includes information on the innovation advancement and the progress of the project. Data about classification of the idea, evaluation results, development concepts to name but a few are illustrated. Thus, the new idea sheet provides all essential information that is linked to the innovation. Additionally, it is possible for all employees to comment the idea, and, what is even more important, to add solutions within the new idea sheet.

4.3. Software support in the use of methods for innovation management As mentioned above in chapter 3, the systematic and structured use of methods and tools to support the execution of innovation can contribute greatly to the probability of innovation success.

The methods presented in the toolbox in chapter 3 are thus fully integrated into the software. To facilitate acceptance, it is very important that method description and illustration be a comprehensible as possible. The software thus uses the scheme and selection criteria as described in section 3.2 above.

First of all, general facts relating to the methods are presented. Second, the complexity of the method is explained using specific criteria (see Figure 3 above). This

Figure 3. New Idea Sheet.

Articles

11

Journal of Automation, Mobile Robotics & Intelligent Systems

short overview of the complexity of the method is a very important part of the description because it gives the companies a very brief overview concerning core factors and requirements. A further description of the procedure depicts how the method should be used. This is illustrated by further practical example in order to promote greater understanding. Finally, references for further information are provided. Companies can also input their personal experience to complement method description. The methods database thus provides company specific knowledge concerning method usage and evaluation.

4.4. Limitations of software support for innovation management Despite all the positive effects generated by the usage of software systems, knowledge driven innovation software lacks support in several key areas. First of all, sharing tacit knowledge through a virtual medium is particularly suitable for a stable, incrementally changing environment. Thus, storage of knowledge is only as good as the willingness of the employees to document their knowledge (Voelpel 2005, p. 19).

Furthermore, in transferring product and development knowledge, supplementary communication, such as face-to-face communication, is required. While direct interaction among personnel is of great importance for successful communication, this is only aided by the software, but not guaranteed. In addition, such interaction is vital for the transfer and identification of tacit knowledge throughout the company, and is one of the most important drivers of innovation (Voelpel 2005, p.19).

Product development is thus often characterized as an exercise in information processing (Clark, Fujimoto 2001; Tatikonda, Rosenthal 2000). Therefore, in many cases lack of information and communication is reported as a key factor in failure. Department-focussed thinking and non-interdisciplinary project teams often lead to project failure and inhibit innovation. Communication with customers also has to be planned and structured in a careful manner. Only appropriate communication within and between companies and the market can ensure that future products meet customer requirements. The innovation software presented her can provide assistance in all such problem areas. However, it has to be admitted that many of these barriers are of a socio-psychologically nature and thus cannot be easily solved by software systems.

Last, but not least, it needs to be noted that, when taking the implementation of innovation software into consideration, the success of software in supporting company innovation processes is highly dependent on the prevailing company culture and organizational structure.

These are relatively rigid organizational boundaries and related problems cannot be solved by technical means alone, such as the installation of new software.

5. Implementation process of the innovation software An intensive testing phase involving all 8 project companies was undertaken in an attempt to verify software results. The companies represented a diverse industrial background, mining, engineering, electronics, plastics etc. Furthermore, the companies were of different si12

Articles

VOLUME 1,

No 4

December 2007

zes. Testing verified that the software worked well in all branches and that it was well suited to the large organizations as well as to SMEs. During the test phase, one outstanding characteristic became apparent. Companies need almost constant help and monitoring during the whole implementation process. Where this is not the case, many functions and features of the software remain unclear and unutilized.

This is particularly true with respect to the innovation process. In practice, several of the 8 companies, especially the SMEs, did not have a structured process model in place to aid innovation processes. Providing adequate support to companies restructuring their company processes towards a more systematic approach for innovation was of great importance here. In fact, it was a vital prerequisite for further implementation of the innovation software.

Moreover, assistance was also needed with the implementation of the methods provided in the software.

Although a thorough and self-explanatory description of the methods is provided in the software, the companies still needed guidance and to engage in learning by doing.

Companies frequently stated that it was hard for them to convince their employees of the usefulness of the methods. Higher acceptance was achieved when scientific consultants became more involved. Companies did, however, admit that the description of the methods and the data sheets provided to aid method execution were vital, especially for future method application.

This reveals the important fact that changes in innovation culture, in information management and in attitudes towards idea generation and idea handling are needed in order to ensure successful implementation of innovation software. Cultural aspects together with leadership styles are among the most important key aspects in innovation management since they often reflect and characterize the importance of new product development. Lack of support and acknowledgement by top management is especially damaging to the acceptance of innovation among employees since it quickly undermines the generation of employee commitment.

Finally, in the testing phase, it became obvious that the companies needed to establish appropriate infrastructure in advance in order to implement such a software tool. Without such preparation, implementation becomes so expensive that the companies are not willing to pay for the software system. Furthermore, companies are often highly reluctant to restructure their whole IT-infrastructure owing to the perceived risks that this entails.

In our case, following the testing and validation phase the software system was revised in accordance with company feedback and companies are now highly satisfied with the results. The producer of the software is now planning to translate the software into other languages. The innovation software described here won the international IBM Lotus award 2007 for the best industry solution.

6. Conclusion Innovation drives corporate success and is a strategic endeavour contributing to the creation of differential advantage. It is precisely for this reason that a well structured and transparent procedure is needed to guide the

Journal of Automation, Mobile Robotics & Intelligent Systems

whole innovation process and to ensure effective and efficient management of the innovation project.

The innovation process described above is just such a well structured process and can thus be of great support.

However, as shown in several studies, SMEs still often lack such a clearly defined process.

Furthermore, there is a clear knowledge gap in many companies with respect to the methods and tools needed in the innovation management process. The innovation tool box helps close this gap between the advanced knowledge available in the literature and the relatively primitive methods applied in practice. Within this toolbox, the methods are described in a very short and understandable way and are furthermore linked to the various phases of the innovation process. The innovation tool box and its methods were tested within 30 different small and medium sized companies and were found to work well and improve innovation success.

The support of software systems in innovation management can have considerable positive effects on the execution of company programs in innovation management.

The innovation software can help companies deal with the enormous amount of information within their innovation processes, assist them in making knowledge-based, structured decisions which are supported by appropriate tools and methods.

However, not all barriers and obstacles can be overcome with tools, methods and software systems. Innovation also depends on other core factors and basic conditions. Hence, a company has to ensure that these core factors are also taken into consideration and that basic conditions relating to the provision of appropriate information and communication within the company are successfully dealt with. Until this is done, no amount of technical adjustment will ever lead to innovation success.

Acknowledgments The Institute of Innovation and Environmental Management at the Karl Franzens University Graz has been involved in the research of innovation management since 1990. In recent years investigation into the innovation process and the development of guidelines to support this process has continued to be one of the key research fields of the institute.

The basis of this paper derives three projects called ‘Innovital’, ‘Innovator‘ and ‘Innoware’, comprising more than 7500 hours of work. The project partner of the first two projects were the Institute of Innovation and Environmental Management at the University of Graz, the Institute of Industrial Management and Innovation Research at the Technical University of Graz and the ARC Systems Research, located at the University of Mining, Metallurgy and Materials in Leoben. Furthermore, the project Innovital and the project Innovators entailed the cooperation of 2 and 3 companies respectively, thus ensuring practical relevance. Additionally, in order to verify the practicability of project results, a further 15 companies were integrated.

The software system is based on the research project ‘Innoware’, a project funded by the Austrian Federal Ministry of Economics and Labour. Research partners for the 22 month project are the Profactor Research Institute in

VOLUME 1,

No 4

December 2007

Upper Austria, the Institute of Innovation and Environmental Management at the Karl Franzens University in Styria and the ARC Systems Research, University of Mining, Metallurgy and Materials in Styria. A further ten industrial partners complement the project team to ensure practicability.

AUTHORS Stefan Vorbach - Institute of Innovation and Environmental Management, Karl-Franzens-University Graz, Universitätsstrasse 15, 8010 Graz, Austria, Tel.. +43 316 380 3235, Fax. +43 316 380 9585, e-mail: stefan.vorbach@uni-graz.at Elke Perl - Institute of Innovation and Environmental Management, Karl-Franzens-University Graz, Austria, e-mail: elke.perl@uni-graz.at

References [1] Afuah A., Innovation management: Strategies, Implementation and Profits, 2nd ed., 2003, New York.

[2] Bourque D.D., “Radical Innovation: How established companies must compete”, in: Erfolgsfaktor Innovation, ed. by Berndt R., 2005, Berlin Heidelberg, pp. 71-80.

[3] Brockhoff K., Forschung und Entwicklung: Planung und Kontrolle, 5th ed., 1999, München, Wien.

[4] Cooper R.G., The Key Factors in Success, American Marketing Association, 1990a, Chicago.

[5] Cooper R.G., “A new tool for managing new products”, Business Horizons, May-June, 1990b, pp.

44-56.

[6] Cooper R.G., Kleinschmidt E., “An investigation into new product process: steps, deficiencies and impact“, Journal of Product Innovation Management, no. 3, 1986, pp. 71-85.

[7] de Brentani U., “Success and Failure in New Industrial Services”, Journal of Product Innovation Management, vol. 4, no. 6, Dec. 1989, pp. 239258.

[8] Disselkamp M., Innovationsmanagement: Instrumente und Methoden zur Umsetzung im Unternehmen, 2005, Wiesbaden.

[9] Eversheim W., “Models and Methods for an Integrated Design of Products and Processes”, European Journal of Operational Research no. 100, 1997, pp. 251-252.

[10] Farris G.F., Hartz C.A., Krishnamurthy K., McIlvaine B., Postle S.R., Taylor R.P., Whitwell E.E., “Web-enabled innovation in new product development”, Research-Technology Management, November-December 2003, pp. 24-35.

[11] Forlani D., Mullins J.W., “Perceived Risks and Choices in Entrepreneurs’ New Venture Decisions”, Journal of Business Venturing, vol.15, 2000, pp.305-322.

[12] Franke N., von Hippel E., “Satisfying heterogeneous user needs via innovation toolkits: the case of Apache security software”, Research Policy, no. 32, 2003, pp. 1199-1215.

Articles

13

Journal of Automation, Mobile Robotics & Intelligent Systems

[13] Friesenbichler M., Leitner W., Ninaus M., Perl E., Ritsch K., Seebacher F., Vorbach S., Winkler R., Innovationsleitfaden: Ideen systematisch umsetzen, 2004, Graz.

[14] Gelbmann U., Leitner W., Perl E., Primus A., Ritsch K., Seebacher F., Tuppinger J., Vorbach S., Innovationsleitfaden: Der Weg zu neuen Produkten, 2003, Graz.

[15] Griffin A., “PDMA Research on New Product Development Practices: Updating Trends and Benchmarking Best Practices”, Journal of Product Innovation Management, no. 14, 1997, pp. 429-458.

[16] Hauschildt J., “Promotors and champions in innovations – development of a research paradigm”, in: The Dynamics of Innovation, ed. by Brockhoff, K., Chakrabarti, A.K., Hauschildt, J., 1999, Berlin.

[17] Hauschildt J., Salomo S., Innovationsmanagement, 4. ed., 2007, München.

[18] Hilzenbecher U., “Innovategy”, in: Erfolgsfaktor Innovation, ed. by Berndt R., Berlin Heidelberg, 2005, pp. 47-70.

[19] Huizenga E.I., Innovation Management in the ICT Sector: How Frontrunners stay ahead, 2004, Cheltenham.

[20] Jeppesen L.B., User Toolkits for Innovation: Consumers Support Each Other”, Journal of Product Innovation Management, no. 22, 2005, pp. 347362.

[21] Karapidis A., Kienle A., Schneider H., “Creativity, Learning and Knowledge Management in the Process of Service Development – Results from a survey of experts”, in: Tochtermann K., Maurer H. (eds.), Proceedings of I-Know05, 5th International Conference on Knowledge Management, Graz, June 2005, pp. 432-440.

[22] Massey A.P., Montoya-Weiss M.M., O`Driscoll T.M., “Performance-centered Design of Knowledge-Intensive Processes”, Journal of Management Information Systems, vol. 18, Spring 2002, no. 4, pp. 37-58.

[23] More R.A., “Barriers to innovation: intraorganizational dislocations”, Journal of Product Innovation Management, no. 3, 1985, pp. 205-208.

[24] Perl E., “Grundlagen des Innovations- und Technologiemanagements”, in: Innovations- und Technologiemanagement, ed. by Strebel H., Wien, 2003, pp. 15-48.

[25] Pleschak F., Sabisch H., Innovationsmanagement, 1996, Stuttgart.

[26] Preissl B., Solimene L., “The dynamics of clusters and innovation: beyond systems and networks“, 2003, Heidelberg, New York.

[27] Sanchez A.M., Perez M.P., “Flexibility in new product development: a survey of practices and its relationship with the product’s technological complexity”, Technovation, no. 23, 2003, pp.

139–145.

[28] Schwery A., Raurich V.F., “Supporting the technology-push of a discontinuous innovation in practice”, R&D Management, no. 34, vol. 5, 2004, pp. 539-552.

14

Articles

VOLUME 1,

No 4

December 2007

[29] Souder W., Managing New Product Innovations, 1987, Lexington.

[30] Staltner T., Huemer H., Vorbach S., Perl E., “Knowledge driven continuous software aided innovation process for SME’s”, in: Horváth, I., Duhovnik, J. (eds), Tools and methods of competitive engineering,, Proceedings of the TMCE 2006, vol. 2, Ljubljana, April 18–22, 2006, pp . 12011202.

[31] Storey J., “The Management of Innovation Problem”, in: International Journal of Innovation Management, vol. 4, no. 3, Sep. 2000, pp. 347-369.

[32] Tatikonda M.V., Rosenthal St.R., “Sucessfull execution of product development projects: Balancing firmness and flexibility in the innovation process”, Journal of Operations Management, vol.

18, 2000, pp. 401-425.

[33] Thom N., Grundlagen des betrieblichen Innovationsmanagements, 2nd ed., 1980, Königstein.

[34] Tidd J., Bessant J., Pavitt K., Managing innovation: integrating technological, market and organizational change, 2nd ed., 2001, Chichester.

[35] Utterback J.M., Abernathy W.J., “A Dynamic Model of Process and Product Innovation”, Omega. The International Journal of Management Science, vol. 3, no. 6, 1975, pp. 639-656.

[36] Utterback J.M., Mastering the dynamics of innovation, 1994, Boston.

[37] Vahs D., Burmester R., Innovationsmanagement: Von der Produktidee zur erfolgreichen Vermarktung, 2nd rev. ed., 2002, Stuttgart.

[38] Van de Ven A.H., Polley D.E., Garud R., Venkataraman S., The Innovation Journey, 1999, New York et al.

[39] Veryzer R.W., “Discontinuous Innovation and the New Product Development Process”, Journal of Product Innovation Management, no. 15, 1998, pp. 304-321.

[40] Voelpel S.C., Dous M., Davenport T.H., “Five steps to creating a global knowledge-sharing system: Siemens´ ShareNet”, Academy of Management Executive, vol. 19, no. 2, 2005, pp. 9-23.

[41] Vorbach S., Perl E., “Software based support for innovation processes”, in: Proceedings of I-KNOW ´05, 5th International Conference on Knowledge Management, ed. by Tochtermann K., Maurer H., 2005, Graz, pp. 220-228.

[42] Wengel J., Wallmeier W., “Worker Participation and Process Innovations”, in: Innovation in Production, the Adoption and Impacts of New Manufacturing Concepts in German Industry, ed. by Lay, G., Shapira, P., Wengel, J., Heidelberg et al., 1999, pp. 65-78.

[43] Wilson J.Q., “Innovation in Organisation: Notes towards a theory”, in: Approaches to Organisational Design, ed. by Thompson, J.D., Pittsburgh, 1996, pp. 193-218.

[44] Yates J.F., Stone E.R., “The risk construct”, in: Yates J.F. (ed.) Risk taking behaviour, 1992, West Sussex, England.

[45] Zien K.A., Buckler S.A., “From Experience: Dreams to Market: Crafting a Culture of Innova-

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 1,

No 4

December 2007

tion“, Journal of Product Innovation Management, no. 14, 1997, pp. 274-287.

[46] Zotter K.A., “Modelle des Innovations- und Technologiemanagements”, Innovations- und Technologiemanagement, ed. by Strebel H., 2003, Wien, pp. 49-93.

Articles

15

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 1,

No 4

December 2007

Knowledge-based Support of Decision Making at the Example of Microtechnology Albert Albers, Norbert Burkardt, Tobias Deigendesch

Abstract:

2. Design in Microsystem Technology

Designing products is a sequence of decisions. Especially early decisions have great impact on later product life cycle stages. The more complex products become the more information is required in order to support design when making decisions. The present contribution shows for the example of the skill-intensive technology of microsystems, what tools can be applied. First, design processes in microtechnology are presented, which depict the necessity of bringing knowledge from later product life cycle stages to the early ones. For this front loading process design rules as a methodological means of support are proposed. Further on that paper presents two tools that are based on design rules, which support the designer. The first one, wikis, open content management systems represent knowledge from later stages for design engineers. The same design rules on a formal level can be implemented into the second proposed tool, a knowledge-based engineering system.

Within single branches, e.g. microtechnology, specific process models were developed due to the reason that universal development processes were too general and thus insufficient for depicting special aspects – which have to be considered for reducing cost and development time while increasing product quality. Hence, some specific design process models are presented in order to give the reader an idea of the special aspects that have to be considered.

The potential functionality and the development process of microsystems are heavily dependent on production technology. As a matter of principle, two areas can be distinguished. There are lithographic microtechnologies, e.g. silicon micromachining or LIGA (German acronym for lithography, electroforming, moulding), which both are called mask-based processes, since substantial structuring steps are performed by exposure to radiation through a patterned mask. Silicon-based microsystems often are referred to as MEMS, microelectromechanical systems.

And there is tool-based microtechnology employing mechanical micromachining, i.e. miniaturized tools known from macroscopic technology, e.g. milling of moulds and subsequent moulding by thermoplastic injection moulding or powder injection moulding of ceramic or metallic feedstocks.

The following sections briefly describe how design processes for microtechnology are modelled.

Keywords: design process, microsystem technology, design rules, knowledge-based engineering, wiki

1. Introduction Today’s products usually consist of a multitude of different subsystems and components, which derive from different domains, e.g. microelectronics, microsystem technology, precision engineering or classic macroscopic mechanical engineering. Also the integration of nanotechnology as another domain is expected. These systems are multi-scaled, i.e. they merge systems of different levels of magnitudes and of different production technologies.

Designers have to be aware of specific manufacturing details, often decisive knowledge, which usually is reserved to specialists. Thus, designers have to be supported by methodological means that lead them to products being in accordance to given requirements. Early decisions in product development strongly influence subsequent stages of a product’s life cycle. Hence, transfer and representation of knowledge to the design stage from subsequent stages, e.g. production or system integration, has to be enabled by integrating means of knowledge management.

Further on, this decision support enables decision tracking later on, if appropriate tools of knowledge management are applied.

The present contribution discusses the reference model of a micro-specific design process in contrast to other processes and points out the application of design rules as a means of support. Wiki systems for early design stages and a computer-aided knowledge-based engineering system tools are proposed.

16

Articles

3. Mask-based Microsystem Technology For mask-based microsystem technology there are different models describing the design flow, which strongly depends on production technology. For silicon-based microsystems the technology was derived from microelectronics. Manufacturing of silicon chips is based on structuring by lithography and subsequent deposing or etching of layers. By combining layers with different electric properties (physical design), structural components are established, that enable electronic functions within a microchip, e.g. microprocessors, amplifiers, etc.

While handling and treatment of silicon wafers, on which the chips are developed, several processing steps are required. Some of these steps are not compatible, e.g. one production step would require higher temperature, than a before deposited layer could withstand.

In 1984 Gajski and Kuhn [9] published a model, which is based on a tripartite representation of designs (Y-chart). Design refinement can be done on each of the three axes, which represent functional, structural and physical design. The functional representation describes what the chip does - without comprising anything about

Journal of Automation, Mobile Robotics & Intelligent Systems

geometrical or structural details. The structural representation describes the circuit itself, i.e. components and connections, while the physical representation deals with cells, layout planning and mask geometries. This model was enhanced by Walker and Thomas [16]. Their model also consists of three perspectives (behavioural, structural and physical perspectives arranged in y-shape with a common vertex). The design flow itself is characterized by changes of the perspective while keeping the level of abstraction and changes of the level of abstraction while keeping the same perspective (refinement).

For mechanical microsystem being made of silicon, Hahn [10] adapted the Y-model, whereas the evidently occurring difference can be found in the levels of abstraction, which are system, component and structural level.

But when designing micromechanical systems not only the object of design itself has to be created, also production, i.e. the technological processing sequence, has to be developed in parallel, especially due to the fact that not only design but also the manufacturing process is application-specific and heavily influences the resulting shape.

Improving the Y-model, Brück and Schumer [7] employ a highly iterative such-called “circle-model”, which is adapted to the requirements of designing microstructures.

The model consists of four steps of layout design, process development, verification and process modification being arranged in a circle and especially considers the parallelism of developing mask layout and production process.

Wagener’s and Hahn’s “pretzel model” [15] clearly shows the parallelism of developing behavioural design and processing sequence. On the one hand, based on a behavioural model and supported by a component library, a verifiable 3D model is synthesized. This 3D model comprises relevant information on design and structure of the single layers and the materials these are consisting of. Based on the known materials of the layers, process sequences and according process parameters are defined. On the other hand, when developing a new process sequence, analysis of this processing leads to mask properties and hence to a 3D model. In 2006 Watty [17] proposed another model for development of microelectromechanical devices being derived from VDI 2206 V-model for mechatronic design [14].

VOLUME 1,

No 4

December 2007

tion are represented by three concentric rings, whereas the outermost ring is most abstract. The design flow itself runs counter clockwise from conceptual to detail design becoming more concrete. For system design, during the first quarter, conceptual design is performed, then in the second quarter the system is basically designed and finally detailed. The design activity is a superposition of a bottom-up design approach from structural level to system level and of a top-down design approach from conceptual design to detail design. When deciding on system level for concepts, on structural level, structures already have to be detailed due to the strong influence of technology.

Visualization of this activity results in a sickle-shaped curve. For the transition from functional description to embodiment, the junction of design and detailing has to be considered. Therefore, a “methodological stage of transition” is introduced. The designer approaches this transition stage with the results of conceptual design on system level. Main functions and subfunctions are extracted. Employing methodological means of support, e.g. effect catalogues, the designer finds effects, i.e. working principles that fulfil the subfunctions. Conceptual design derives from the combination of these partial solutions and consists of functional items and basic shapes without any quantified dimensions or specified materials. The system itself is subdivided into components. On structural level, details for the functional items are designed with respect to technological conditions and restrictions. The latter are provided externally by a knowledge representation, for which design rules are used. These design rules represent knowledge from subsequent life cycle stages, e.g. process preparation in terms of realizable structural details, e.g. the minimum edge radius. Thus, while synthesizing, the transition stage is required to adhere to invariant structural details.

4. Tool-based Microsystem Technology Considering tool-based microtechnology, there are technological conditions and restrictions, e.g. achievable flow lengths, minimum milling cutter diameter or minimum wall thickness. This results in a strong orientation in what is producible and therefore in a technology-driven design flow – in contrast to macroscopic product design, which is driven by market requirements. In order to achieve a design compatible to production, specific knowledge from process preparation (e.g. mould manufacture) and production (e.g. injection moulding) is required. The special aspects of designing tool-based microsystems are visualized by a model introduced by Marz [12], which is called “sickle-model” (cp. Figure 1) due to its sickle-shape transition from the design stage to the detail stage. The model represents the design stages conceptual, basic and detail design on system, component and structural level, i.e. on different levels of abstraction. The levels of abstrac-

Figure 1. Design flow for tool-based microtechnology (sickle model [12]).

As an example, design of a micro gear is discussed: When beginning to concept the micro gear on system level, structural details like the tooth shape have to be considered. Assuming a powder injection moulding process, for production a mould insert is required. The cavity has to be scaled up by a certain percentage (shrinkage) and has to be milled. Employing one of the smallest off-shelf end mill cutters (100μm diameter), the minimum edge radius Articles

17

Journal of Automation, Mobile Robotics & Intelligent Systems

within the mould is 50μm. This strongly affects the tooth shape and results in a shorter involute and therefore in a reduced contact length. At the same time the tooth width is restricted by the cutting depth of 200μm, which influences the transmittable torque and hence the conceptual design of the gear system. [1] [4] [5] [12]

5. Tools and Methods for Knowledge-base Design Designing technical systems, regardless of size, is a complex process, which is influenced by many parameters. Especially strongly technology-driven design, like that of microsystems, requires the transmission of knowledge and experience, which are gained within other stages, e.g. process preparation or production, in order to be able to design a functioning product. Generally approved means of supporting design decisions are design principles, guidelines and rules. Another method for supporting an inexperienced designer is the methodological documentation and knowledge representation, for which the authors propose a wiki system.

Knowledge-based Design in Microsystem Technology Usage of design rules is a widely applied method for providing knowledge for designers. The terms design rule, design guideline, design advice or embodiment rule often are used in a synonym way, while the design rule term also is used for rules of varying levels of abstraction. Therefore, herein design rules are defined as instructions deriving from technological conditions and restrictions that have to be regarded stringently in order to achieve a realizable design. These technological conditions and restrictions derive from all methods and processes of process preparation, production and material science. They include all influences and effects being adjacent or subsequent to the design stage.

Due to the technology-driven design process and the hence higher importance of knowing about details from stages subsequent to design (production preparation, production) when designing microsystems, knowledgebased design of those systems is described.

Design of microsystems is a technology-driven process. Shape and material strongly are dependent on production. Especially within microelectronics, design rules are an established method for supporting designers in order to achieve products that are producible. The design methodology according to Mead and Conway [13] separates functional and physical design. This separation is enabled by design rules, which describe minimum geometrical values for certain widths, thicknesses, distances or superpositions on and in-between processes layers.

Due to this separation, enterprises with costly fabrication lines (foundries) can provide their production facilities to several designers, like universities, small companies, etc.

Often foundries and designer communicate by design rule kits. These kits include a set of design rules, which are relevant for the producer’s fabrication line. Computer-aided design tools for microelectronics support design rules and enable online design rule checks for direct feedback during design.

Since microelectronic circuits are characterized by simple and repeatedly occurring two-dimensional patterns, 18

Articles

VOLUME 1,

No 4

December 2007

design rule checking can be implemented. For mask-based microsystem design this approach cannot be applied due to the enabled third dimension, a greater shape variety and more applicable materials. Design rules within knowledge-based design are an established means of support within design in microelectronics and micromechanics.

There are several references having investigated such design rule systems, e.g. Hahn [10], Leßmöllmann [11], Buchberger [8], Albers [3] and Marz [12].

Application of design rules is a powerful instrument for supporting decision making in design. After derivation of manufacturing-specific knowledge, design decisions can be founded on a stable grounding. Implementation of these rules in computer-aided engineering environments (such called knowledge-based engineering systems, KBE) can support designers during modelling, while less formal derivatives of design rules may already influence design decisions during earlier stages.

Derivation of Design Rules During conceptual and embodiment design engineers have to decided on shape, material, manufacturing possibilities and many more aspects. Design rules as very concret and precise representative of the domain of design principles, guidelines and rules help to decide by providing information from subsequent product life cycle stages concerning manufacturing technology for example. The derivation of design rules is exemplarily described by those for tool-based microtechnology. First of all, the potential influence of existing domains needs to be detected. In a following step, all relevant machine and process parameters are extracted from process preparation, production or quality assurance. Regarding process preparation, exemplary conditions and restrictions are type and size of milling cutters, tool tolerances, machine tolerances or the realizable surface roughness. In a third step, these parameters have to be interpreted in a way relevant for design, i.e. external knowledge is transformed into a methodological knowledge that can be used by designers. Then rules are formulated, classified for the ease of access and filed in a database. Finally, they have to be provided by an information portal (cp. wiki in Figure 2) or can be integrated into a knowledge-based engineering system.

Wiki Support Wikis are software tools for computer supported cooperative work [18]. They are content management systems enabling users to integrated informational aspects and files. Not only for supporting groups in communication, coordination and cooperation, a wiki especially can be used for documenting design information knowledge.

Applying this tool for documentation, decisions cannot only be facilitated, but also the entire design process and design decisions can be reconstructed from a later point of view. Late design stages are characterised by a relative high level of structure, many participants, defined means of support and defined workflows. These late stages usually are supported by highly integrated software, e.g. product data management (PDM) or enterprise resource planning (ERP) tools. On the other side, early design stages, e.g. conceptual design, have a relative low structure,

Journal of Automation, Mobile Robotics & Intelligent Systems

is characterised by individuality and small design teams and is not supported by specified means of software. Especially for those wikis have the necessary flexibility when integrating new content.

VOLUME 1,

No 4

December 2007

(UG/Knowledge Fusion) as well as a programmable user interface. In a rule database design rules are filed, based on which an online design rule check can be performed.

The data comprises explanatory text, which is displayed in

Figure 2. Design rule representation in a wiki system.

Wikis allow easy and instant collaboration over the internet, an instant “online-documentation” that is webaccessible via browser while no client application is required. Other advantages are versioning an page locking mechanisms, always up-to-date-versions, online discussion and improvement. It is very easy to revert unwanted changes. Pages easily are created and edited. Text formatting is simple and does not require HTML skills. Editing designers can focus on content instead of formatting due to easy syntax. Some wikis come with an (optional) WhatYou-See-Is-What-You-Get (WYSIWYG) editing functionality. Edited work is immediately visible. Finally low install and maintenance costs are a big advantage.

In product development, wikis can be used for all documentation processes. This makes them being very interesting for design. Based on a good documentation, the entire design process including all decisions can be reviewed and reconstructed. Further on, the same platform enables production engineers to share their knowledge regarding relevant features and properties of production technology. This knowledge representation can be realized in an unstructured way or by using design rules. All information can be accesses by using implemented search funtions. Figure 2 shows an exemplary design rule depicted in a wiki system for supporting design of tool-based microsystems. [6]

case of rule infringement, and the rule itself being formulated in IF-THEN-ELSE conditions. The IF-part is formally described as a mathematical equation. If the IF-condition is fulfilled, the THEN-part is carried out, otherwise the ELSE-part offers alternative actions, e.g. an automatic correction. [2][12]

Knowledge-based Engineering Realization of a knowledge-based design environment for primary-shaped microparts was shown by Albers [2]. Unigraphics (UG) was used as a compuer-aided design (CAD) system, which offers all possibilities of parametric design and includes a knowledge-based module

Figure 3 shows a micro planetary gear. As the feature being relevant for function, the module, is within micron range (m=169μm), the gear itself is classified as a microsystem, whereas its outer dimensions are within millimetre range (approx. 7mm). Design followed – obviously successfully – the presented design flow for tool-

Figure 3. Micro planetary gear on a Cent coin.

Articles

19

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 1,

based microtechnology (sickle model) in combination the knowledge-based design as a methodological means of support.

[9]

6. Summary and Conclusion The paper summarized models of development processes for the specific application of microtechnology to motivate the necessity of knowledge-based support tools for decision making in design. Increasingly becoming more complex, products can only be designed by (globally distributed) engineering teams, when supporting design decisions as early as possible and on a sustained basis that can be reconstructed from a later point of view. The paper points out wikis for decision documentation and knowledge representation by design rules. These rules then support decisions on geometry or material by just providing information or even by being integrated into CAD-systems.

[10]

[11]

[12]

[13] [14]

AUTHORS Albert Albers* - Universität Karlsruhe, Institute of Product Development, Kaiserstrasse 10, 76131 Karlsruhe, Germany, Phone: +49 721 608 23 71, Fax: +49 721 608 60 51, e-mail: albers@ipek.uka.de Norbert Burkardt, Tobias Deigendesch – Universität Karlsruhe, Institute of Product Development, Kaiserstrasse 10, 76131 Karlsruhe, Germany * Corresponding author

[15]

References

[18]

[1] A. Albers et al., “Micro-specific Design Flow for Tool-based Microtechnologies”, in: Procs. High Aspect Ratio Micro Structure Technology Workshop, HARMST 2005, Gyeongju, Korea, 2005.

[2] A. Albers et al., “Knowledge-based design environment for primary shaped micro parts”, Microsystem Technologies, vol. 11, 2005, no. 4-5, pp. 254-260 [3] A. Albers et al., “Design Methodology in Micro Technology”, in: Procs. Int. Conf. on Engineering Design ICED’03, Stockholm, Sweden, 2003.

[4] A. Albers, J. Marz, “Restrictions of Production Engineering on Micro-specific Product Development”, Microsystem Technologies, vol. 10, 2004, pp.205-210.

[5] A. Albers, J. Marz, “Design Environment and Design Flow”, in: Advanced Micro and Nanosystems vol. 3, ed. by Löhe, D., Hausselt, J., Weinheim, Germany: Wiley VCH, 2005, pp.3-28.

[6] A. Albers et al., “Wikis as a Cooperation and Communication Platform within Product Development”, in: Int. Conf. on Engineering Design ICED’07, 28-31 August 2007, Paris [7] R. Brück, C. Schumer, “INTERLIDO – Web-basierte Werkzeuge für den Mikrostrukturentwurf”, Workshop Multimedia und Mikrotechnik, vol. 1 Lüdinghausen, Germany, 1998, pp.82-102.

[8] P. Buchberger et al., “MIDAS: Ein wissensbasiertes System zur Unterstützung des fertigungs20

Articles

[16]

[17]

No 4

December 2007

gerechten Entwurfs von LIGA-Mikrostrukturen”, in: Wissenschaftliche Berichte no. 5843, Forschungszentrum Karlsruhe, 1997.

D. Gajski, R. Kuhn, “New VLSI-Tools”, Computer, vol. 16, 1983, no. 12, pp. 11-14.

K. Hahn, Methoden und Werkzeuge zur fertigungsnahen Entwurfsverifikation in der Mikrotechnik, Diss., Fortschritt-Berichte VDI 20, no. 286, Düsseldorf, Germany: VDI-Verlag, 1999.

C. Leßmöllmann, Fertigungsgerechte Gestaltung von Mikrostrukturen für die LIGA-Technik, Diss., Karlsruhe, 1992.

J. Marz, Mikrospezifischer Produktentwicklungsprozess (μPEP) für werkzeuggebundene Mikrotechniken, Diss., IPEK Forschungsberichte, no. 17, Karlsruhe, Germany, 2005 C. Mead, L. Conway, Introduction to VLSI Systems, 2nd ed., Reading, USA: Addison-Wesley, 1980.

VDI guideline 2206, Design methodology for mechatronic systems, Berlin, Germany: Beuth, 2004.

A. Wagener, K. Hahn, Eine Entwurfsmethodik für die Mikrosystemtechnik und Post-CMOS, Austrochip 2003, Linz, Austria, 2003.

R. A. Walker, D. E. Thomas, Model of Design Representation and Synthesis, Procs. 22nd ACM/IEEE Conf. on Design Automation, Las Vegas, USA, 1985, pp.453-459.

R. Watty, Methodik zur Produktentwicklung in der Mikrosystemtechnik, Diss., Institut für Konstruktionstechnik und Technisches Design, no. 533, 2006.

T. Gross, M. Koch, Computer-Supported Cooperative Work, München, Germany: Oldenbourg, 2006.

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 1,

No 4

December 2007

Systematic Decision Making Process for Identifying the Contradictions to be Tackled by TRIZ to Accomplish Product Innovation Hajime Mizuyama, Kenichi Ishida

Abstract This paper presents a systematic decision making process for accomplishing product innovation in accordance with the target quality, i.e. the target values and relative weights of the relevant quality characteristics of the product to be developed, with the help of TRIZ (the Russian abbreviation of the theory of inventive problem solving).

Since TRIZ methodology deems innovation as resolving a contradiction, the proposed approach first reveals the contradictions that block the target quality from being reached, based on the engineering solutions that the current base product employs and the phenomena that take place while the base product is performing its function. Then, the approach structures the contradictions and distinguishes the causal conflicts from the resultant ones. It also calculates the criticality of each causal conflict according to the relationships with the quality characteristics. These steps make it possible to properly highlight the focus of innovation within the whole function and mechanism structure of the base product. The proposed method is described using a diecasting machine as an illustrative example, and the example confirms that some innovative conceptual design solutions can be successfully derived through the proposed decision making process.

Keywords: conceptual design, conflict resolution, DEMATEL, product innovation, QFD, TRIZ

1. Introduction It is a critical issue for a manufacturing firm to develop an innovative product that not only meets customers’ needs but also distinguishes itself clearly from existing competing products, especially in the recent mature market environment. Even when developing a new derivative product from an existing base product, it is desirable that the resultant product should be deemed innovative as well as responding to customers’ needs. That is, now it is an important issue how to accomplish product innovation in a target-quality-oriented way. Upon this background, it has been pointed out that combining QFD (Quality Function Deployment) and TRIZ (the Russian abbreviation of what can be translated as the theory of inventive problem solving) will make a powerful approach for the conceptual design of such an innovative product. QFD will set an attractive target quality to the product, and TRIZ will help solve technical problems if any and thus enable to reach the given target quality.

In order to make such an approach actually work, the target quality to be achieved must be translated into technical problems in a form that can be handled by TRIZ. TRIZ methodology deems innovation as resolving a contradic-

tion, and usually expresses the technical problem to tackle as a contradiction. Thus, this paper presents a systematic decision making process for capturing and structuring the contradictions that should be resolved to achieve the target quality, and thus enables product innovation to be accomplished in a target-quality-oriented way with the help of TRIZ. The proposed approach calls it an engineering solution to perform each elemental function by the corresponding elemental mechanism, where elemental functions and elemental mechanisms are the smallest constituents of the function and mechanism structure of a product. Then, the approach reveals the contradictions based on the engineering solutions that the current base product employs and the phenomena that take place while the base product is performing its function. Any product realises its function by intentionally inducing some phenomena with its engineering solutions. Unfortunately, these beneficial phenomena often accompany some harmful phenomena as their side effects, and they cause the problems. This paper names the obtained contradictions the elemental conflicts. Then, the approach structures the complicated causal relationships among the elemental conflicts and distinguishes the causal conflicts from the resultant ones.

It also calculates the criticality of each causal conflict according to the relationships with the quality characteristics. These steps make it possible to properly highlight the focus of innovation within the whole function and mechanism structure even when the quality characteristics to be enhanced are affected by many interrelated contradictions located in various parts of the base product. The proposed approach also provides a procedure for choosing an appropriate TRIZ tool and posing the technical problem to the chosen tool.

In the remainder of this paper, after a literature review section, the problem to be dealt with and the proposed decision making process for the problem are described using a die-casting machine as an illustrative example, and how the proposed approach works is demonstrated with the example. Finally, conclusions follow.

2. Related work QFD is a well-known and widely-used methodology for prioritizing the customers’ requirements on a product, and translating them into its design specifications [1][2].

Hence, it can be used to set appropriate target values for the key quality characteristics of the innovative product to be developed. Since this target quality should clearly differentiate the resultant product from competing ones, it is usually difficult to reach with the current function and mechanism structure of the base product. Thus, in order Articles

21

Journal of Automation, Mobile Robotics & Intelligent Systems

to attain the target quality, it must be determined what changes should be made to which parts of the base product. However, QFD itself does not provide any means to make the decision.

Whereas, when a technical problem is given in an appropriate form, TRIZ will provide several tools to support solving the problem [3][4][5][6][7]. It has been successfully applied to real-life design problems [8][9][10][11].

Thus, TRIZ can be utilised to identify what sort of innovation should be introduced into the function and mechanism structure of the base product so as to attain the specified target quality. Accordingly, it has been pointed out that combining QFD and TRIZ will make an effective approach for systematic product innovation [12][13].

To make such an approach actually work, the target quality set by QFD must be translated into technical problems in a suitable form for TRIZ, that is, a contradiction.

However, this critical part of the innovation process still depends largely on the capabilities of human engineers, and only a few approaches have been proposed for supporting it. As a pioneer work in this field, Yamashina et al. [14] developed a systematic approach for bridging the gap between QFD and TRIZ. This approach first determines which part or sub-mechanism of the base product should be given changes according to the relationships between the submechanism and the quality characteristics to be enhanced, and then defines contradictions to resolve for the chosen sub-mechanism based on its house of quality matrix. Hua et al. [15] adjusted this approach to a specific TRIZ software. Wang et al. [16] proposed a similar approach, which defines contradictions based only on the house of quality matrix of the whole product and hence can be deemed as a simplified version of Yamashina et al. [14].

What is common to these approaches is that they do not distinctly consider what technical problems prevent the quality characteristics from being improved as intended and only capture some resultant contradictions appeared on the house of quality matrix, for example, as trade-off relationships among quality characteristics. However, it is often the case, in practice, that any trade-off relationship may be caused by not just one but several different technical problems located in various parts of the base product, a same technical problem can affect many quality characteristics, and the technical problems are themselves interrelated. In such circumstances, it seems that the conventional approaches cannot always identify the root cause of the problems and may result in a tedious iteration process that reveals and resolves those problems one by one. This is obviously inefficient and may take a long time to converge. Therefore, more desirable will be a decision making process that can first exhaustively capture what technical problems exist in which parts of the base product and then properly narrow down the focus of innovation. However, such a decision making process has not yet been established, and hence will be newly developed in this paper.

3. Target-quality-oriented product innovation problem 3.1 Assumptions on target quality This paper deals with the problem to establish a conceptual design of a new innovative product from an exist22

Articles

VOLUME 1,

No 4

December 2007

ing base product. It is assumed that the quality characteristics {y1, y2, …} which should be considered for fulfilling the customers’ requirements on the product have been identified, and the target values {t1, t2, …} and the relative weights {u1, u2, …} for the quality characteristics have been set. The relative weights are assumed to have been normalised as below:

∑u

i

=1

(1)

i

where the higher the relative weight ui, the more important to enhance the quality characteristic yi to satisfy the customers’ requirements. For example, QFD can be utilised to determine the target quality. However, the proposed approach does not assume, and is independent of, the use of QFD. To make the below discussion easier to understand, we hereafter take up a die-casting machine, shown in Figure 1, for a certain aluminium product as an illustrative example. We assume that the below five variables have been identified in this example as the quality characteristics to be considered to satisfy customers’ requirements: y1: fraction of defectives due to voids (%), y2: fraction of defectives in the other defect modes (%), y3: raw material yield (%), y4: cycle time (seconds), y5: die life (shots), and that ambitious target values and appropriate relative weights have been already assigned to those characteristics so that the resultant product should be deemed innovative.

Figure 1. Example die-casting machine.

3.2 Assumptions on function and mechanism structure It is also assumed that the function and mechanism structure of the existing base product has been deployed as a corresponding pair of a function tree and a mechanism tree. This deployment is not a new practice, and has been introduced and utilised, for example, in axiomatic design [17].

The current die-casting machine we consider consists of a molten metal reservoir, a molten metal feeder, a molten metal injector, an ejector, a moulding mechanism, and a die lubricator (which is omitted in Figure 1). The molten metal reservoir further comprises a pot which reserves molten metal, a heater which raises its temperature, and a lid which maintains the temperature. The molten metal feeder is composed of a ladle which ladles molten metal,

Journal of Automation, Mobile Robotics & Intelligent Systems

and a manipulator which manipulates the ladle. The subsystems of the molten metal injector include a sleeve which reserves molten metal and keeps its temperature, an accumulator which generates a force, a plunger rod which conveys the force, and the plunger which injects molten metal into a die and eliminates voids. The subsystems of the moulding mechanism contain the die which moulds molten metal, a die-cooler which cools down the die, a die heater – which heats the die, a die closer – which shuts the die, and air vents, which exhaust gas from the die. Further, the die lubricator is composed of a spray, which sprays lubricant and cools down the die, and a tank, which reserves lubricant.

Thus, the function and mechanism structure of the base product can be captured by the function tree shown in Figure 2 and the mechanism tree shown in Figure 3.

Hereafter, we refer to each sub-function corresponding to a leaf node of the function tree and each sub-mechanism represented by a leaf node of the mechanism tree as

VOLUME 1,

No 4

December 2007

an elemental function and an elemental mechanism respectively. In the die-casting machine example, dividing those elemental functions and mechanisms further down to smaller constituents was judged to be meaningless for devising innovation.

Further, we term it an engineering solution to utilise a specific elemental mechanism for performing each elemental function. Then, a product can be deemed as a set of engineering solutions {a1, a2, …}. For example, the current die-casting machine can be captured as the set of the below engineering solutions: a1: reserve molten metal by a pot, a2: raise molten metal temperature by a heater, a3: maintain molten metal temperature by a lid, a4: ladle molten metal by a ladle, a5: manipulate the ladle by a manipulator, a6: reserve molten metal by a sleeve, a7: generate a force by an accumulator, a8: convey the force by a plunger rod,

Figure 2. Function tree of example die-casting machine.

Figure 3. Mechanism tree of example die-casting machine.

Articles

23

Journal of Automation, Mobile Robotics & Intelligent Systems

a9: inject molten metal and eliminate voids by a plunger, a10: eject a casting by an ejector, a11: mould molten metal by a die, a12: cool down the die by a die cooler, a13: heat the die by a die heater, a14: shut the die by a die closer, a15: exhaust gas by air vents, a16: spray lubricant and cool down the die by a spray, a17: reserve lubricant by a tank.

3.3 Product innovation using TRIZ The target quality, i.e. the set of the target values assigned to the quality characteristics, must be able to clearly differentiate the resultant product from competing ones in the mature market environment, and hence will be difficult to attain with the function and mechanism structure of the base product. This paper does not compromise the target quality, but instead aims at deriving a new innovative conceptual design solution by changing the function and mechanism structure through an appropriate technical innovation. What is required to accomplish this product innovation will be to determine what changes should be made to which parts of the base product to reach the target quality.

This paper uses a TRIZ-based approach to this problem. The TRIZ tools that we can use to derive an innovation concept for an engineering solution include the 40 principles of innovation, the principle of separation, the effect database, etc. The 40 principles of innovation are a collection of typical principles for resolving a technical contradiction, which is a situation that improving an attribute degrades another attribute. The principle of separation resolves a physical contradiction, which is a situation that mutually conflicting target values are given to a certain single attribute, by separation in space, time, etc. The effects database shows utilizable physical, chemical and geometrical effects that can substitute for a current engineering solution having a contradiction, when the function to be realised by the solution is specified.

Thus, what TRIZ tools try to accomplish is not to reach a given target quality but to resolve contradictions. Therefore, in order to attain the target quality by the help of TRIZ, we need first to study the contradictions in the function and mechanism structure of the base product that block the target quality from being reached, and then to pose an appropriate contradiction to a suitable TRIZ tool in a proper form. The next section proposes a structured decision making process for this purpose.

4. Proposed decision making process for product innovation 4.1 Capturing elemental conflicts The outline of the proposed approach is shown by a diagrammatic presentation in Figure 4. The first step reveals the technical problems preventing the quality characteristics from being enhanced as prescribed, each in the form of a contradiction, on the engineering solutions of the base product. We start this step by enumerating major phenomena {pi1, pi2, …} that affect each quality characteristic yi, and clarifying the relationships between the phenomena and the engineering solutions {a1, a2, …}.

24

Articles

VOLUME 1,

No 4

December 2007

Then, we grasp the technical problems concerning each engineering solution in the form of a contradiction; the engineering solution gives a desirable effect to a phenomenon and at the same time an undesirable effect to another phenomenon.

Figure 4. Diagrammatic presentation of proposed approach.

Hereafter, we call the contradictions the elemental conflicts, and denote them by {c1, c2, …}. When the former phenomenon of an elemental conflict concerns the quality characteristic yi and the latter phenomenon influences yj, then the elemental conflict will cause a trade-off relationship between the characteristics yi and yj and as a result will make the innovative target quality difficult to attain. Thus, the elemental conflict must be resolved so as to reach the target quality.

In the die-casting machine example, the major phenomena concerning the quality characteristics are the followings: p11: generation of voids from remaining air in the die, p12: generation of voids from die lubricant, p21: soldering between the die and molten metal, p22: temperature drop and incomplete filling of molten metal, p31: molten metal consumption in the runner of the die, p32: molten metal leakage from the parting surface of the die, p41: time consumption by spraying lubricant, p42: time consumption by injecting molten metal, p51: die fatigue due to temperature stress, p52: die fatigue due to the closing force.

We can capture the elemental conflicts of the base product according to the relationships between those phenomena and each engineering solution.

Each elemental conflict has a structure that an engineering solution gives a desirable effect to a phenomenon and at the same time an undesirable effect to another phenomenon. We refer to the former phenomenon as the upstream phenomenon and the latter one as the downstream phenomenon of the conflict. Then, the elemental conflict can be expressed by a model: an upstream phenomenon

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 1,

-> an engineering solution -> a downstream phenomenon.

For example, the engineering solution a9 (inject molten metal and eliminate voids by a plunger) gives a desirable effect to the phenomenon p11 (generation of voids from remaining air in the die) and at the same time an undesirable effect to the phenomenon p32 (molten metal leakage from the parting surface of the die). Hence, the current die-casting machine has the elemental conflict: c1: p11 -> a9 -> p32 .

Similarly, the below nine elemental conflicts have been captured in the example die-casting machine: c1: generation of voids from remaining air in the die (p11) -> inject molten metal and eliminate voids by a plunger (a9) -> molten metal leakage from the parting surface of the die (p32), c2: generation of voids from die lubricant (p12) -> inject molten metal and eliminate voids by a plunger (a9) -> molten metal leakage from the parting surface of the die (p32), c3: soldering between the die and molten metal (p21) -> spray lubricant and cool down the die by a spray (a16) -> generation of voids from die lubricant (p12), c4: soldering between the die and molten metal (p21) -> spray lubricant and cool down the die by a spray (a16) -> time consumption by spraying lubricant (p41), c5: soldering between the die and molten metal (p21) -> cool down the die by a die cooler (a12) -> die fatigue due to temperature stress (p51), c6: temperature drop and incomplete filling of molten metal (p22) -> mould molten metal by a die (a11) -> molten metal consumption in the runner of the die (p31), c7: temperature drop and incomplete filling of molten metal (p22) -> inject molten metal and eliminate voids by a plunger (a9) -> generation of voids from remaining air in the die (p11), c8: time consumption by injecting molten metal (p42) -> inject molten metal and eliminate voids by a plunger (a9) -> generation of voids from remaining air in the die (p11), c9: molten metal leakage from the parting surface of the die (p32) -> shut the die by a die closer (a14) -> die fatigue due to the closing force (p52).

The elemental conflicts c1, c2, …, c5 and c9 are straightforward to understand. The conflict c6 occurs because of enlarging the runner cross section of the die to slow down the temperature drop. Whereas, the conflicts c7 and c8 arise due to increasing the speed and the pressure of molten metal injection.

These elemental conflicts are the technical problems that prevent the target quality from being reached. By revealing which engineering solution causes each elemental conflict, we have the set of the engineering solutions having technical problems. We denote the set by AC. In the die-casting machine example, this set is given by: AC = {a9, a11, a12, a14, a16}

(2)

No 4

December 2007

4.2 Structuring elemental conflicts When there are many elemental conflicts, we should pay attention to the causal conflicts but the resultant ones that are brought about by the causal ones. Hence, our product innovation approach next clarifies the relationships among the elemental conflicts and highlights the causal ones to be considered according to the relationships.

If eliminating an elemental conflict will make it unnecessary for another elemental conflict to be resolved, we say that these conflicts are in an influential relationship.

There can be the following two types of influential relationships between elemental conflicts: Serial influential relationship: It is the influential relationship that occurs between elemental conflicts cm and cn when the downstream phenomenon of cm is identical to the upstream phenomenon of cn and resolving cm will make the engineering solution itself or its condition that causes cn unnecessary. In this case, cn needs not to be dealt with any more if cm is resolved.

Parallel influential relationship: It is the influential relationship that occurs between elemental conflicts cm and cn when they have a same upstream phenomenon and resolving cm will make the engineering solution itself or its condition that causes cn unnecessary. In this case as well, cn needs not to be dealt with if cm is resolved.

Hereafter, we represent these influential relationships by: cm -> cn, and call cm the elemental conflict on the cause-side and cn the one on the effect-side.

In the die-casting machine example, the downstream phenomenon of c3 is p12 and which is also the upstream phenomenon of c2, and resolving c3 will make it unnecessary to consider c2. That is, if the conflict c3 is resolved and the phenomenon p12 is no longer problematic, the engineering solution a9 will not need to deal with the voids generated by p12. Thus, there is a serial influential relationship c3 -> c2 . Whereas, the elemental conflicts c6 and c7 share the same upstream phenomenon p22 and the solution a9, which causes c7, is the primary engineering solution to this phenomenon p22. If a9 does not cause the conflict, the engineering solution a11, which causes c6, will become unnecessary. Hence, there is a parallel influential relationship c7 -> c6. Similarly, the following eight influential relationships can be obtained in this example: c1 -> c9, c2 -> c9, c3 -> c2, c5 -> c3, c5 -> c4, c7 -> c1, c7 -> c6, c8 -> c1.

As we can see in this example, when there are many elemental conflicts, the influential relationships among them are likely to be complicated. Since the influential relationships defined above satisfy the transitive law, not only the direct influential relationships but also indirect ones must be grasped and unified into the aggregate relationships. They can be captured by structuring the influential relationships among the elemental conflicts through DEMATEL (Decision Making Trial and Evaluation Laboratory) [18][19]. DEMATEL is a matrix-based approach for quantifying the overall relationships among multiple items and has been used for structural modelling in various fields [20][21].

Articles

25

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 1,

We first set the direct influence rmn = 0 for each pair of elemental conflicts cm and cn that has no direct influential relationship between them. Since no elemental conflict has a direct influential relationship with itself, rmm = 0 holds for every cm. If cn is directly influenced only by cm, we set rmn = 1. In the die-casting machine example, we have: r32 = r53 = r54 = r76 = 1

(3)

No 4

December 2007

When, and only when, the elemental conflict cm is a causal conflict which is not influenced by any other elemental conflicts, the mth column of the aggregate influence matrix will become a zero vector. Thus, we can identify the causal elemental conflicts by extracting every column of a zero vector from the aggregate influence matrix. In our die-casting machine example, as shown in Figure 5, the causal conflicts to be really resolved have been revealed as {c5, c7, c8}.

Further, if cn is directly influenced by more than two elemental conflicts, we define the value of each corresponding rmn according to the relative degree of its influence so that the below condition should be satisfied:

∑r

m n

=1

(4)

m

In our example, c1 is directly influenced by c7 and c8, and c9 is affected by c1 and c2. Hence, we set: r71 = r81 = r19 = r29 = 0.5

(5)

We now have the direct influence matrix R = (rmn), which then will be used to derive the aggregate influence ~ matrix R = ( ~rmn ) as below:

~ ∞ R = ∑ R k = R (I − R ) −1

(6)

k =1

where I is an identity matrix. Each element ~ rmn of the ~ obtained matrix R expresses the aggregate influence from cm to cn.

In the die-casting machine example, the direct influence matrix is given by:

⎛ 0 ⎜ ⎜ 0 ⎜ 0 ⎜ ⎜ 0 R = ⎜⎜ 0 ⎜ 0 ⎜ ⎜ 0.5 ⎜ 0.5 ⎜⎜ ⎝ 0

0 0 0 0 0 0 0 0.5 ⎞ ⎟ 0 0 0 0 0 0 0 0.5 ⎟ 1 0 0 0 0 0 0 0 ⎟ ⎟ 0 0 0 0 0 0 0 0 ⎟ 0 1 1 0 0 0 0 0 ⎟⎟ 0 0 0 0 0 0 0 0 ⎟ ⎟ 0 0 0 0 1 0 0 0 ⎟ 0 0 0 0 0 0 0 0 ⎟ ⎟ 0 0 0 0 0 0 0 0 ⎟⎠

(7)

Accordingly, the aggregate influence matrix is obtained as:

§ 0 ¨ ¨ 0 ¨ 0 ¨ ¨ 0 ~ ¨ R=¨ 0 ¨ 0 ¨ ¨ 0 .5 ¨ 0 .5 ¨¨ © 0

0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

0 .5 · ¸ 0 .5 ¸ 0 .5 ¸ ¸ 0 ¸ ¸ 0 .5 ¸ 0 ¸ ¸ 0.25 ¸ 0.25 ¸ ¸ 0 ¸¹

Figure 5. Relationships among elemental conflicts.

By checking which engineering solution brings about each causal elemental conflict, we can further narrow down the set of the engineering solutions to be considered from the above AC. We denote the obtained subset of ~ AC by AC . In the die-casting machine example, this set is given by:

~ AC = { a 9 , a12 }

(9)

That is, the engineering solutions to be really considered were successfully narrowed from the above five down to only two engineering solutions.

4.3 Criticality analysis of elemental conflicts When more than two engineering solutions have been found to have a causal conflict as in the case of the diecasting machine, the proposed approach next prioritises the engineering solutions.

We first quantify how important it is to resolve each elemental conflict so as to satisfy customers’ requirements. For this purpose, we define the phenomenon weight wP(pij) of each phenomenon pij concerning the quality characteristic yi by distributing the relative weight ui of the quality characteristic to every phenomenon pij according to the magnitude of its effect on the characteristic. Then, we define the raw conflict weight wC(cm) of each elemental conflict as below: wC(cm) = wP(pij) + wP(pkl)

(8)

(10)

where pij and pkl represent the upstream phenomenon and the downstream one of the elemental conflict cm respectively. Accordingly, we can obtain the aggregate con~ (c ) of the same elemental conflict c by flict weight w C m m the below equation:

~ (c ) = w (c ) + ~ w ∑ rmn wC (c n ) C m C m n

26

Articles

(11)

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 1,

~ (c ) represents This aggregate conflict weight w C m that the greater its value, the higher the number and/or the larger the relative weights of the quality characteristics to be improved by resolving the elemental conflict cm.

Next, we prioritise the engineering solutions that we should focus on as the target of innovation, according to the aggregate conflict weights calculated above. When ak is included in the set of the relevant engineering solu~ tions AC , we define its solution weight wA(ak) as below: w A (a k ) =

∑ w~

es ( cm ) = a k

C

(c m )

(12)

where es(cm) expresses the engineering solution (∈{ a1 , a 2 , K}) that causes the elemental conflict cm.

This is the sum of the aggregate weights of the elemental conflicts located in the engineering solution. We can prioritise the engineering solutions to be considered in the decreasing order of this value. For convenience, we can deem that the weight of any engineering solution that ~ is not included in the set AC is zero.

In the die-casting machine example, the process of this criticality analysis has been undergone as shown in Figure 6, and the relevant engineering solutions were prioritised as: wA(a9) > wA(a12)

(13)

No 4

December 2007

4.4 How to pose technical problems to TRIZ When the relevant engineering solutions are determined and prioritised through the above criticality analysis, our product innovation approach will pose a technical problem concerning the chosen engineering solution in one of the following forms: • formulate an elemental conflict of the chosen engineering solution as a technical contradiction, and apply one of the suggested principles of innovation, • formulate an elemental conflict of the chosen engineering solution as a physical contradiction, and apply the principle of separation, and • describe the elemental function of the selected engineering solution as a function to be realised, and obtain an applicable one from the physical, chemical and geometric effects in the effects database.

Thus, the procedure of posing technical problems to TRIZ and deriving innovation concepts to be introduced to the function and mechanism structure of the base product can be summarised as below, where A0 is the set of the candidate engineering solutions to be focused on as the target of innovation: ~ Step 1: Set A0 = AC .

Step 2: If holds, then select the most satisfactory innovation concept among those which have been devised, and end the procedure.

Step 3: Select the engineering solution ai having the highest solution weight from the set A0, and update the set as:

Figure 6. Process of criticality analysis.

Articles

27

Journal of Automation, Mobile Robotics & Intelligent Systems

A0 = A0 − {ai }

VOLUME 1,

(14)

Step 4: Derive innovation concepts by attempting innovation through TRIZ focusing on the chosen engineering solution ai. If a satisfactory concept is obtained, then end the procedure.

Step 5: Examine possible combinations among the innovation concepts obtained so far. If a satisfactory concept is obtained through combination, then end the procedure. Otherwise, go back to Step 2.

4.5 Example application results This subsection presents some results of applying the proposed approach to the die-casting machine example.

As shown earlier in Figure 4, the role of the proposed approach itself is to define appropriate technical problems for TRIZ and applying TRIZ to the problems is outside the scope of the approach. However, in this subsection, how TRIZ handles the problems derived by the proposed approach is also described briefly for illustrative purpose.

For the sake of intellectual property preservation, the innovation concepts that are not regarded as a part of the common knowledge cannot be disclosed here and hence those described below are only a part of the obtained innovation concepts. Despite this, we can still see in the below how the proposed approach guides the process of deriving innovation concepts.

According to the above criticality analysis, we first select the engineering solution a9 (inject molten metal and eliminate voids by a plunger) as the target of innovation.

It has two causal elemental conflicts c7 and c8. Since the effect database is not to resolve these conflicts directly, here we give some results of applying the 40 principles of innovation and the principle of separation.

Suppose that we took up the elemental conflict c7 concerning the chosen engineering solution a9. Then, we can formulate it into a technical contradiction. When we choose durability of moving object as the attribute to improve and harmful side effects as the attribute to be degraded, TRIZ will recommend the principle of skipping (rushing through). By applying this principle, we can obtain an innovation concept that radically shortens the route to the gate of the die and injects molten metal at a high speed without causing a significant temperature drop. This innovation concept will shorten the cycle time without increasing the fraction of defectives due to voids.

Instead of a technical contradiction, we can formulate the same elemental conflict c7 into a physical contradiction: “molten metal injection should be at a high speed and a high pressure to prevent incomplete filling, and should be at a low speed and a low pressure not to generate voids” and utilise the principle of separation. However, separation in time has been already adopted by many die-casting machines through changing the plunger speed in the middle of the injecting process. Hence, here we apply separation in space to this technical problem. Accordingly, we can obtain an innovation concept that raises the pressure of the molten metal only in some areas in the die after completing injection.

At this step, we can also take up the other elemental conflict c8 concerning the engineering solution a9. When formulating it into a technical contradiction by choosing 28

Articles

No 4

December 2007

speed as the attribute to improve and accuracy of manufacturing as the attribute to be degraded, TRIZ will suggest the principle of preliminary action and the principle of self-service and self-organization. Applying the principle of preliminary action will lead to an innovation concept that exhausts the gas from the die prior to molten metal injection through a vacuum pump, and utilizing the principle of self-service and self-organization will result in an innovation concept that eliminates the gas through a reaction with the injected molten metal itself, instead of exhausting it, by fulfilling the die with a reactive gas before starting injection. These innovation concepts will significantly decrease the fraction of defectives due to voids.

We can further combine the above innovation concept obtained by the principle of skipping (rushing through) with the innovation concept achieved by the principle of self-service and self-organization into the function and mechanism structure of the product to be developed.

As demonstrated above, the proposed approach enabled to highlight the target of innovation systematically within the whole function and mechanism structure of the current base product, and the above die-casting machine example suggested that the proposed approach will be actually able to guide us to methodically derive some innovative conceptual design solutions. When dealing with a product having a small number of engineering solutions and elemental conflicts, a simple drawing like Figure 5 can also guide the decision making process instead of DEMATEL analysis. Hence, the full version of the proposed approach will be more effective for a large-scale product.

However, it is also true that, the larger the scale of the product, the more time-consuming to rigidly follow the proposed decision making process. In order to use the proposed approach efficiently for a large-scale product, a conventional approach, such as Yamashina et al. [14], can be used as a pre-processor to narrow down in advance the engineering solutions and elemental conflicts to be considered.

5. Conclusions This paper proposed a systematic decision making process for identifying the contradictions to be tackled by TRIZ to accomplish product innovation in a target-quality-oriented way. An illustrative example showed that it is capable of systematically guiding the conceptual design process of an innovative product. The approach utilises the engineers’ knowledge on the base product, for example, the knowledge on the phenomena taking place while the product is performing its function. Thus, when the engineer using this approach is not so familiar with the base product and hence some of the phenomena were remained unknown, the available innovation concepts may be limited. However, it is also true that the knowledge on the phenomena only is often not sufficient for effective product innovation. Therefore, the proposed approach and a deep understanding of the base product should go hand in hand to make the most of the approach.

How to handle a large-scale product efficiently is an important future research topic. Further, we need to note that the proposed approach is not the one and only way to define contradictions, which prevent the given target quality from being reached, in the base product. Thus, it is

Journal of Automation, Mobile Robotics & Intelligent Systems

also an interesting future research topic to establish other approaches for capturing contradictions. A same contradiction may be described in a different form, and it may result in a different innovation concept. It is not astonishment since the conceptual design problem treated in this paper is not a problem with a single right answer. That is, the main objective of a systematic innovation approach, as the one proposed here, is not to derive one right answer, but to enhance the efficiency of the innovation process.

AUTHORS Hajime Mizuyama* and Kenichi Ishida – Department of Mechanical Engineering and Science, Kyoto University, Kyoto 606-8501, Japan. Tel.: +81-75-753-5237, Fax: +8175-753-5239.

E-mail: mizuyama@mbox.kudpc.kyoto-u.ac.jp * Corresponding author

References [1] Hauser J. R. and Clausing D., “The House of Quality”, Harvard Business Review, no. 66, 1988, pp.

63-73.

[2] Chan L. K. and Wu M. L., “Quality Function Deployment: A Literature Review”, European Journal of Operational Research, vol. 143, 2002, pp.

463-497.

[3] Altshuller G., 40 Principles: TRIZ Keys to Technical Innovation, Technical Innovation Center, Worcester, MA, 1998.

[4] Altshuller G., Innovation Algorithm, Technical Innovation Center, Worcester, MA, 1999.

[5] Kim Y. S. and Cochran D. S., “Reviewing TRIZ from the Perspective of Axiomatic Design”, Journal of Engineering Design, no. 11, 2000, pp. 79-94.

[6] Savransky S. D., Engineering of Creativity: Introduction to TRIZ Methodology of Inventive Problem Solving, CRC Pr I Llc, Boca Raton, 2000.

[7] Orloff M. A., Inventive Thinking through TRIZ: a Practical Guide, Springer Verlag, New York, 2003.

[8] Bariani P. F., Berti G. A. and Lucchetta G., “A Combined DFMA and TRIZ Approach to the Simplification of Product Structure”, Proceedings of the I MECH E. Part B. Journal of Engineering Manufacture, no. 218 B8, 2004, pp. 1023-1027.

[9] Chang H. T. and Chen J. L., “The Conflict-Problem-Solving CAD Software Integrating TRIZ into Eco-Innovation”, Advances in Engineering Software, no. 35, 2004, pp. 553-566.

[10] Mao Y. J. and Tseng C. H., “An Innovative Piston Retractor for Bicycle Hydraulic Disc Braking Systems”, Proceedings of the I MECH E. Part D. Journal of Automobile Engineering, no. 218 D3, 2004, pp. 295-303.

[11] Tsai C. C., Chang C. Y. and Tseng C. H., “Optimal Design of Metal Seated Ball Valve Mechanism”, Structural and Multidisciplinary Optimization, no.

26, 2004, pp. 249-255.

[12] Terninko J., “The QFD, TRIZ and Taguchi Connection: Customer-Driven Robust Innovation”.

VOLUME 1,

[13]

[14]

[15]

[16]

[17] [18] [19]

[20]

[21]

No 4

December 2007

In: Proceedings of the 9th Symposium on Quality Function Deployment, 1997, pp. 441-445.

Hua Z., Yang J., Coulibaly S. and Zhang B., “Integrating TRIZ with Problem-Solving Tools: A Literature Review from 1995 to 2006”, International Journal of Business Innovation and Research, no.

1, 2006, pp. 111-128.

Yamashina H., Ito T. and Kawada H., “Innovative Product Development Process by Integrating QFD and TRIZ”, International Journal of Production Research, no. 40, 2002, pp. 1031-1050.

Hua Z., Huang F. and Wang W., “A Method of Product Improvement by Integrating FA with TRIZ Software Tools”, International Journal of Product Development, no. 4, 2007, pp. 122-135.

Wang, H., Chen, G., Lin, Z. and Wang, H., “Algorithm of Integrating QFD and TRIZ for the Innovative Design Process”, International Journal of Computer Applications in Technology, no. 23, 2005, pp. 41-52.

Suh N. P., The Principles of Design, Oxford University Press, New York, 1990.

Fontela E. and Gabus A., “DEMATEL - Progress Achieved”, Futures, no. 6, 1974, pp. 361-363.

Fontela E. and Gabus A., “World Crisis - Decision Making Trial and Evaluation Laboratory”, Futuribles, no. 14, 1978, pp. 211-221.

Tamura H. and Akazawa K., “Structural Modeling and Systems Analysis of Uneasy Factors for Realizing Safe, Secure and Reliable Society”, Journal of Telecommunications and Information Technology, no. 3, 2005, pp. 64-72.

Tamura H. and Akazawa K., “Stochastic DEMATEL for Structural Modeling of a Complex Problematique for Realizing Safe, Secure and Reliable Society”, Journal of Telecommunications and Information Technology, no. 4, 2005, pp. 139-146.

Articles

29

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 1,

No 4

December 2007

Layout of Functional Modules and Routing for Preliminary Design of Automatic Teller Machines Katsumi Inoue, Tomoya Masuyama, Hayato Osaki, Tokachi Ito

Abstract: In this study we address the preliminary design for the module layout and bill conveyance routes of automatic teller machines (ATMs). We determine a two-dimensional layout for the modules as below that are approximately rectangular if the ATM is viewed from the side. ATMs require the compact placement of modules within the chassis and conveyance routes that smoothly circulate bills. However, the intersection and overlapping of routes by which the bills are conveyed in opposite directions are not allowed. Applying the bottom-left method and route-design-oriented packing method to the layout of the modules and the direction-oriented maze routing expediting branching and interflow of routes to the bill conveyance route, the application orders are optimized simultaneously using genetic algorithms (GAs). Results show that suitable designs for the ATM including the case when modules are selected as well as placed are achievable using the above simultaneous optimization.

The design intention is expressible by changing the weights associated with chassis dimensions, route lengths and the number of route bends, which compose the objective function. The proposed method is useful for efficiently advancing the preliminary design of ATMs. Finally, if island models pursuing individual targets are used along with a GA, the design becomes even more efficient.

Keywords: module layout, routing, simultaneous optimization, genetic algorithms, design intention, Island Model

1. Introduction A high demand exists for optimal design through partslocation decision making. This is particularly true for the layout of parts with particular shapes and functions and sequential routing in machinery for chemical plants and ductwork, manufacturing machine location, and conveyer routing. Numerous promising solutions exist for such design problems despite the fact that the problems themselves are numerically limited. For this reason, it is impossible to evaluate each solution. Subsequently, it is necessary to obtain an optimal solution. Regarding simultaneous decisionmaking processes for the location and routing of parts, Sessomboon and coworkers [1]-[3] examined the problem of machine layout and conveyance routes for automated guided vehicles involved in manufacturing processes. That study concluded that the entire process, namely, the workplace layout designed by the bottom-left (BL) method and the route obtained using graphical method and genetic algorithms (GAs), should be optimized by simulated annealing (SA). Shirai and Matsumoto [4] dealt with workplace block and aisle location issues using the Packing Method for lo30

Articles

cation and the Maze Routing Method for determining the aisles’ orientation; they carried out optimization by SA as did Sessomboon and coworkers.

The authors [5] studied the above-mentioned optimization issue targeting the preliminary design of automatic teller machines (ATMs). ATMs require the compact placement of modules while avoiding their overlap and a route for bills to circulate smoothly. Moreover, neither the intersection of routes nor the overlapping of routes that convey the bills in opposite directions is allowed, because the direction of bill conveyance in ATMs is fixed. Therefore, the authors [6] decided to use a Maze Routing Method that has high flexibility regarding the shape of the route and can be sed to determine the shortest primary path. We improved this method so that the branching and merging can be efficiently performed for ATM functions. We called our method direction-oriented maze routing (DOMR). Meanwhile, in the case of setting multiple routes using the maze-routing method, the route order is generally determined on the basis of the route priorities, which are obtained from information on nesting methods and the amount of traffic, because the solution depends on the search order. In contrast, the authors optimize the route-setting order using GAs and thereby suggest an optimal design method for module layout using the BL method while determining the optimal route using DOMR. The proposed method is generally effective, but some improvements remain regarding the compact placement of modules when there are no restrictions on module shapes.

The arrangement of modules using the BL method is restricted by the module shapes. In this paper, therefore, we first suggest a layout method that combines the secondary allocation issue and the packing method; subsequently, we clarify its effectiveness by applying the method to the preliminary design of ATMs. Next, we evaluate the validity of the design method issues pertaining to its flexibility when modules are to be selected as well as placed. In addition, we demonstrate that designing the arrangement according to specific intentions is possible by changing the weights of the objective function. Lastly, we consider a method of efficient processing by introducing decentralized processing in the GAs using an island model.

2. Module layout and route-design issues in preliminary atm design 2.1. Modules and Routes to Realize ATM Functions The overlapping of modules in terms of depth is not apparent if an ATM is viewed from the side. For this reason, ATM design can be conceptualized as a two-dimensional layout problem. Figure 1 shows the placement of the modules in

Journal of Automation, Mobile Robotics & Intelligent Systems

a grid space (x, y). We refer to the individual grid areas as cells. All modules are rectangular; the location of module i is expressed in centroid coordinates as Pgi(xi, yi). To simplify the design task, the route is set as being parallel to the axis.

Our method does not deal with shapes with diagonals. In addition, the route has a direction for bill conveyance; only one route exists in each cell. Table 1 shows the names of the modules and their dimensions in cell units. Figure 2 illustrates the bill conveyance routes that connect the modules containing the bill-out and bill-in apertures.

VOLUME 1,

No 4

December 2007

Every ATM requires the ows of bills to realize the four functions shown in Figure 3. For example, in the case of a deposit, after validating the bill in the bill-verifying unit B, which has come from the bill entrance unit A, different processes might ensue: i) a bill might be returned to the bill entrance unit; ii) a defective bill might be stored in reject box R1 or R2; or iii) a bill might be stored in cashbox K1-K3 for subsequent reuse. The bills are conveyed along different routes according to the situation. A total of 15 routes are necessary to implement all the necessary functions; these routes are illustrated in Figure 2. The module codes and route numbers indicated in Figure 3 are those shown in Table 1 and Figure 2.

Figure 1. Module and route modeling.

Table 1. Modules of ATMs. Code

Module

Size

A

Bill entrance unit

5*3

B

Bill verifying unit

3*5

CA

Temporary stacker A

3*3

CB

Temporary stacker B

3*3

D

Loading cassette

5*9

K1

Cashbox 1

3*9

K2

Cashbox 2

3*9

K3

Cashbox 3

3*9

R1

Reject box 1

5*3

R2

Reject box 2

5*3

Figure 3. Modules and routes necessary for ATM functions ((1) deposit, (2) withdrawal, (3) load, (4) check).

Figure 2. Module layout and routing.

2.2 Formulation of the Design Problem This design method aims to place m rectangular modules that differ in dimensions and to set n routes. The following information is required.

Articles

31

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 1,

1. Dimensions, i.e., vertical and horizontal lengths, of modules, 2. Locations of bill-in and bill-out apertures for each module, 3. Modules at the starting point and ending point of each route.

On the basis of this information, the design problem is formulated as follows. min. f0 = w1L+w2B+w3C+w4S

(1)

In this equation, L, B, C and S are penalties, as below: L: Total route length by cell unit; B: Number of route bends; C: Number of crossings with other routes; S: Size of the smallest rectangle that contains the modules.

As mentioned above, those penalties are expressed by the number of cells. Weights wj correspond to the respective performances; the values assigned to them reflect the designer’s intention. For instance, if the weight of S increases, the design places more emphasis on compactness. If the weights of L and B are increased, the route length and number of bends are expected to decrease; such a design accelerates ATM processing. The bill conveyance routes do not intersect. Therefore, C must be zero.

However, it is difficult to express this condition as positive in the objective function. Therefore, the objective function is expressed as a linear combination including C; and intersections are avoided by assigning a large weight to C. In this design, the number of modules m is 10 and the number of routes n is 15.

3. Method of module layout and routing 3.1. Layout Design of Modules Various methods such as (a) the quadratic assignment program (QAP) [7], (b) computerized relationship layout planning (CORELAP) [8], and (c) flexible bay structure (FBS) [9] have been proposed for the layout design of modules. Method (a) deals with the problem of distributing N modules among N candidate sites. The modules, however, are defined as points; they are not considered as shapes. In method (b), the modules are considered as shapes, but because other modules are sequentially adjacent to one module, the concavity and convexity of the outer shapes become critical. In method (c), candidate sites at which modules are placed first determined in rows. Then the modules are loaded every row. Accordingly, this method results in a comparatively orderly module layout.

If an ATM is viewed from the side, it be regarded as a two-dimensional layout. Methods (b) and (c) thereby become applicable. Size and shape differences among the modules are considerable, and the number of modules to be arranged is comparatively small. Consequently, the following two layout methods were chosen for this study.

(1) BL Method [10] In the original method, modules are inserted from the upper right of the layout area; they are first moved to the bottom and then to the left until the modules can no longer be moved. Figure 4 shows an example of layout 32

Articles

No 4

December 2007

design using this method. Here, module 3 is inserted at first, next, module 2, 4 and 1 are inserted in order. In this study, the upper and left boundaries are set. Module A, the bill entrance unit, is fixed in the upper left of the area; thus, the BL method is applied under the upper-left condition. One empty cell is reserved around individual modules to secure a route-setting area. The BL method determines the layout according to the order of module placement. Therefore, the order of placing the modules in the layout area is the variable in this design problem.

Figure 4. Algorithm of the BL method.

(2) RDOP Method The modules’ shapes sometimes restrict layout design using the BL method. Therefore, we suggest a layout design method that combines the secondary allocation issue and the packing method. We call this route-design-oriented packing (RDOP) method. The flow of this method is depicted in Figure 5. It is similar to the BL method in that the upper and left boundaries of the layout area are fixed. The modules are placed in the following order.

Step 1: To secure the route-setting area, one empty cell is placed to the right and below the modules containing the bill-in and bill-out apertures, as shown by the shaded areas in the Figure 5. Equal-sized rectangular areas, which can accommodate all the modules including the extra spaces mentioned above, are allocated in the grid.

The number of rectangular areas is identical in both the vertical and horizontal directions and is the minimum number required to accommodate all the modules. The number of modules is 10 in this design problem. Therefore, the number of rectangular areas is 4*4=16, as shown in Figure 5(a).

Step 2: Figure 5(b) shows that modules are moved to the upper left of the rectangular areas based on the module layout information.

Step 3: The border lines of the design area are moved simultaneously up and to the left, until they abut other modules.

This process is portrayed in Figure 5(c).

Using this method, it is possible to produce a compact layout without being restricted by the modules’ shapes.

The design variables for this method are the module arrangement information that is initially given for the rectangular areas.

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 1,

No 4

December 2007

Figure 5. Algorithm of the route-design-oriented packing method.

3.2. Route Design by Direction-Oriented Maze Routing Method The maze-routing method has been used in such cases as wiring design in VLSI [11]; this method provides the shortest route from the starting point to the end point in the design area while avoiding obstacles. For instance, to avoid obstacles in Lee’s algorithm [12], weights are applied to the labels. Existing routes are considered as obstacles. Therefore, this weighting is effective for the avoidance of route crossings. However, route crossings are actually unavoidable in this problem because the routing for bill conveyance in ATMs is complicated.

For ATM design, the bill-conveyance direction is set for every route. Although the intersection and overlapping of conveyance routes with route in the opposite direction are not permitted, branching and merging are allowed in route design, under the condition that the routes proceed in the same direction. Branching and merging should be promoted because merging has the advantage of shortening the total length of routes. Consequently, a new maze-routing method is suggested using labeling that encourages merging while prohibiting intersection and overlapping with routes in the opposite direction. This method is called direction-oriented maze routing (DOMR) [6].

Figure 6 shows the process of labeling and route searching using the DOMR method. In Figure 6(a), the white rectangular blocks represent modules. Solid lines and arrowheads respectively indicate existing routes and their directions; the broken line is the route that is chosen from starting point S to terminal point T. The obstacles described above are denoted as X in Figure 6(b). The cell that is under consideration is S’, and the newly labeled cells are S’’; the initial state is S’=S, and label LS is 1. Label numbers are given to S’’, which are located abovw, below, left and right of S’ by the following procedure. Here, four more cells in the same directions are also considered at each stage [4].

Rule 1: The stage is terminated without labeling if cell S’’ is X. Rule 2: If the label value of S’’ already exists and it is smaller than the label value to be given, then the stage is terminated without updating the label.

Rule 3: If the label value does not exist for S’’ or if it does exist but is larger than the label value to be given, then the label value is updated.

Rule 4: If S’’ is not on an existing route, S’’ is moved to the cell next to where it was initially located as in LS’’ = LS+20.

(Here, the added value to the label, +20, is to be adjusted according to the area size and the route). In the example shown in Figure 6(b), S’ is the cell (3, 4) and the labeling of S’’ is performed in the regions (1, 4)-(2, 4) and (4, 4)(7, 4).

Rule 5: The following sub rules are applied if S’’ is on an existing route that is in the same direction as the cell to be labeled or S’’ is on the point of inflection.

1) If the new label value to be given to S’’ is less than that of the cell next to S’’ and on the S’ side, then the stage is terminated without updating S’’.

2) In other cases, LS’’ = LS’+1 is applied. S’’ is moved by one more cell if it is on the forward-direction route.

An example of this case when S’ is at (3, 5) and S’’ is at (3, 6)-(3, 9) in Figure 6(c).

3) The stage is terminated if S’’ is on the point of inflection. For example, when S’ is at (3, 4) and S’’ is at (3, 5) in Figure 6(b).

Rule 6: If S’’ is on an existing route and is neither in the forward direction nor at a point of inflection, the turn is terminated by labeling LS’’ = LS’+120. (This additional value of +120 is also to be adjusted according to the area size and the route.) This case is represented by the situation when S’ is at (3, 5) and S’’ is at (4, 5) in Figure 6(c).

Articles

33

Journal of Automation, Mobile Robotics & Intelligent Systems

Rules 4-6 enable branching and merging. Using these rules, the minimum label is given to the cell that is on an existing route in the same direction. Therefore, the newly determined route merges with on existing route as soon as possible and proceeds along this existing route for as long as possible.

Following these rules, the operation is repeated with the newly labeled S’’ as S’ until all cells have been completely updated; the result is shown in Figure 6(c). After having finished labeling, as in Lee’s algorithm, the algorithm searches up/down and left/right from the starting point T. A new starting point T’ is created if a smaller label number exists than the current value. The label with the smaller number of cell movements is adopted if the same label numbers exist. The process continues until T’ reaches S. This process is portrayed in Figure 6(d). The symbol denotes T’. The route from T to S passing through T’, if the point T’ reaches S.

VOLUME 1,

Articles

December 2007

4. Simultaneous optimization of design based on ga 4.1. Simultaneous Optimization of Module Layout and Route Design The design of the module layout and route is affected by the order of module placement, as mentioned above.

Therefore, it is necessary to optimize this order. For the optimization method, we use a GA [13].

In the BL method, to express 10 modules and 15 routes, the integers 0-24 are used as strings. Figure 7 shows the string composition. The genetic loci 0-9 represent the module numbers shown in Table 1. As shown in the figure, the order of information extracted from the genetic locus determines the insertion order for the modules. Loci 10-24 are the route numbers shown in Figure 2 or Figure 3 with 10 added to them. Therefore, deducting 10 from these values in the genetic locus gives the information is

Figure 6. Creation of a route from S to T by proposed direction — oriented maze routing.

34

No 4

Journal of Automation, Mobile Robotics & Intelligent Systems

useful for the route-setting order. In the RDOP method, the module numbers are determined by using the numbers 10 to 15, which refer to “empty” modules, as well as actual module numbers 0-9. These numbers correspond to the 16 rectangular areas that accommodate the modules. While extracting information from strings 0-15 in the genetic locus, the corresponding modules are placed in accordance with the numbers given to the rectangular areas in advance. The following string represents the module arrangement shown in Figure 5(a).

0-3-10-8-2-11-12-9-5-13-1-4-14-6-7-15 The route-setting procedure is identical to that of the BL method.

VOLUME 1,

No 4

December 2007

tions is 30,000. Figure 9(a) shows the optimized solution based on the BL method; the module layout and routes are obtained without overlapping between modules to modules, or between modules and routes. In addition, there are no intersections or overlapping routes in opposite directions. The relative locations of modules B, CA, CB, D, K1, K2, K3, and R2 are similar to the module arrangement for an actual machine. The obtained layout of modules is compact, and the route length L is 78. The objective function calculated using Equation (1) is 2454.

The solution in the case of using the RDOP method is shown in Figure 9(b). As presented in the previous section, a concern exists regarding the possibility of a declining rate of superior-individual creation because of the use of longer strings than those used for the BL method. Therefore, we set the population size to 200, the mutation evolution rate to 0.2, the number of generation to 100,000, which increases the calculation load. In this case, the weights in Equation (1) were set to w1=3, w2=20, w3=100, and w4=1 because the RDOP method provides efficiently minimizes the layout space. All the terms of the objective function are improved compared with those shown in Figure 9(a). The improvement of the area and route length appears to be attributable to the effect of the empty cells placed around the modules when setting the routes. Furthermore, the number of bends is greatly decreased.

Figure 7. String used for simultaneous optimization and decoding of present design variables.

The flowchart of the design optimization method is shown in Figure 8. The steady-state GA [14] based on a continuous generational model is used in this study. In the case that an individual created by a crossover or mutation is superior to the worst individual in the population, the worst individual is replaced by the created one. A high probability of creating individuals with a fatal gene that includes the same modules and routes exists if a simple crossover is used. Therefore, the ordered crossover is used here.

Figure 8. Flowchart of design optimization method.

4.2. Solution Obtained by Simultaneous Optimization Design of ATM is calculated by setting the weights to w1=3, w2=20, w3=100, and w4=4, and the population size to 100; the mutation rate is 0.3, and the number of genera-

Figure 9. Module layout and routes optimized by the (a) BL method and (b) RDOP method.

Articles

35

Journal of Automation, Mobile Robotics & Intelligent Systems

5. Layout design including module selection In actual design, it is often the case that great flexibility exists in selecting the modules to be used. It is easy to select from among distinctively different options. For instance, whether to select high-cost and high-reliability modules or low-cost and low-reliability modules depends solely on the design goal as to whether low cost or high reliability is sought as a design priority. Alternatively, it is often possible to select modules based on product specifications. Here, we apply the design method presented in the previous section to a problem that does not appear to be affected by module selection.

Table 2 shows the modules that can be chosen for the design. The sizes of the modules shown in the table are almost identical to those of the modules used in the layout design in the previous section. On the other hand, the shapes and locations of the bill-in and bill-out apertures differ. The information on the module selection is expressed in two bits in the form of a string that is newly created for the 10 modules that are placed. This coding method differs from that used for the string described earlier. Therefore, we apply a one-point crossover to the module selection string.

VOLUME 1,

No 4

December 2007

improvement is found in the value of the objective function. The reason for this lack of improvement has not been clarified, but a contributing cause might be the fact that the RDOP method is less affected by the modules’ shapes than the BL method. The superiority to the BL method is the same as that of the design that does not include module selection.

Table 2. Selectable modules for ATM design.

The optimal layout is calculated using the weights w1=3, w2=20, w3=100, and w4=4, population size of 500, a mutation evolution rate of 0.3, and the 200,000 generations. Figure 10(a) shows the optimal design in the case of using the BL method. The numbers shown in the figure represent the module numbers. By selecting appropriate modules, a better solution is obtained than that shown in Figure 9(a). The decrease in the number of bends is particularly noteworthy.

Figure 10(b) illustrates the optimal solution in the case of using the RDOP method. Although different modules are selected from those shown in Figure 9(b), no marked 36

Articles

Figure 10. Module layout and routes optimized by the (a) BL method and (b) RDOP method with module selection.

The distribution of solutions for each method is shown in Figure 11. Error bars in the figure indicate standard deviations. Despite the greater dispersion, the RDOP method is superior to the BL method. Moreover, its superiority for design with module selection is clear.

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 1,

No 4

December 2007

and (c) result in solutions that maximize the machine’s speed.

Table 3. Weights of objective function for each design intention.

Design

Intention

w1

w2

w3

w4

1

10

100

4

(a)

minimize space

(b)

minimizeroute length

10

10

100

1

(c)

minimize number of bends

1

40

150

1

In each case, applying the BL method and the RDOP method, designs including the selection of parts were generated three times. The optimum solutions that were obtained in each case are shown in Figure 12. All components of the objective function, for which the weight factors were increased, are reduced, thereby realizing the design intentions. In design (c), which used the BL method, the design contained a route intersection. The cause is inferred to be that the search for a solution did not converge, but another possible cause is the coexistence of the difficult objectives of minimizing route bends and avoiding route intersections. To illustrate a similar result, recall the previous section, in which a good solution was obtained when the RDOP method was used.

7. Introduction to distributed process of GA and its effect

Figure 11. Penalties incurred by each method.

Good results can be achieved by applying the design methods along with the selection of modules with various characteristics. In addition, the method described in this section might be applicable to the research and development of modules with the characteristics required for product design with high performance.

6. Realization of the design intention As described in section 2.2, design intentions can be expressed by tuning the value of each weight wj in the objective function, Equation (1). This section examines this matter. We discuss three design intentions; (a) minimization of module layout space, (b) minimization of routing length, and (c) minimization of number of route bends.

The values of weights for these design intentions are shown in Table 3. Design intention (a) result in a solution that realizes the miniaturization of the machine, and (b)

An island GA is sometimes used as a method of increasing GA efficiency. Here, relatively small populations evolve independently by repeating the genetic operations in each island, then they are mixed through mutual immigration. This idea was adapted to the present design task using two different island models. One is a conventional island model and the other is, as shown in Figure 13, an island model based on the idea of realizing the design intention by changing the weights, which was considered in the previous section. These island models are called island model 1 and island model 2. Island model 2 is introduced based on the idea that an independent design with highperformance can be implemented a comprehensive highperformance design. Using this method, it is expected that the search for a solution has greater efficiency than that of a conventional island model (island model 1). In addition, because increasing the performance of individuals can provide more optional solutions, it is expected to diversify the populations. For the proposed island models, the weight distribution (w1=1, w2=10, w3=100, and w4=1) is used for a multipurpose design intention, as shown in Table 3.

Island models 1 and 2 have 500 individuals each.

The top 100 individuals are received as immigrants from another island; the other individuals are rejected. After 30,000 generations, the initial immigration is carried out; subsequent immigrations are performed every 10,000 generations. The simulation was ended after the 100,000th generation.

The objective functions for both island methods starting from the same initial group of individuals are shown in Articles

37

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 1,

No 4

December 2007

Figure 12. Module layout and routes reecting different design intentions.

Figure 13. Island model concept.

Figure 14. Both models converge to excellent solutions. The designs obtained for island model 2 and a non island model are shown in Figure 15. For each design intention, objective functions obtained by applying the island models are summarized in Figure 16. Values given are averages over three trials for each condition, and are expressed as a ratios to the results achieved without using island models.

38

Articles

Figure 14. Comparison of reduction of the objective function using island models.

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 1,

No 4

December 2007

For the selection of modules, it is inferred that the application of island models is more effective because the range of the solution search is greater. For both the selection and non selection of modules, superior solutions were obtained using island model 1 for design intentions (a) and (c). For this reason, minimizing the area and minimizing the number of route bends are antithetical items.

It seems to be ineffective to receive immigrants that have achieved performance gains for the realization of these intentions. On the other hand, for design intentions (b) and (d), good solutions were obtained using island model 2; results were particularly outstanding for (d). It is considered that receiving immigrants that achieved individual performance gains was effective for obtaining multipurpose optimum solutions.

8. CONCLUSIONS

Figure 15. Comparison of module layout and routes obtained using island model 2 and a non island model.

ATMs require the compact placement of modules while avoiding their overlap and routes for bills to circulate smoothly. Moreover, the intersection of routes is not allowed; the overlapping of routes that convey the bills in opposite directions is also not allowed because the billconveyance direction in the ATM is fixed.

On the basis of these conditions, we present and discuss methods of optimizing the layout of modules with various functions and the routes connecting them to obtain appropriate solutions simultaneously during the preliminary design. The main results of this study are the following.

1) A method was suggested for solving this design problem by combining the BL method or the RDOP method for module layout design with DOMR for route design and by simultaneously optimizing the process using a GA. 2) The RDOP method demonstrates better performance than the BL method.

3) The suggested method was applied to design problems with different intentions including the case when modules are selected. Good design solutions were obtained using the method.

4) While independently advancing the design to enhance the performance of individuals, their achievement was used as a basis for multipurpose design. On the basis of the concepts presented above, unique island models were proposed and their effectiveness was confirmed.

ACKNOWLEDGMENT The authors sincerely appreciate Dr. Kunio Fukatsu from Toshiba Social Automation Systems Co., Ltd., for his useful advice regarding this work.

AUTHORS

Figure 16. Objective functions obtained using two island models.

Katsumi Inoue – Department of Mechanical Systems and Design, Tohoku University, Japan.

E-mail: inoue@elm.mech.tohoku.ac.jp Tomoya Masuyama* – Department of Mechanical Engineering, Tsuruoka National College of Technology, Japan.

E-mail: masu@tsuruoka-nct.ac.jp Hayato Osaki – YAMAHA Motor Co., Ltd. Japan.

E-mail: oosakihayato@yamaha-motor.co.jp Articles

39

Journal of Automation, Mobile Robotics & Intelligent Systems

Tokachi Ito – TOYOTA Motor Co., Ltd. Japan.

E-mail: ito@elm.mech.tohoku.ac.jp * Corresponding author

References [1] Sessomboon W., Nakano H., Irohara T. and Yoshimoto, K., “A Technique to Solve Dynamic Layout Problem with Fixed and Rectangular Departments”; Transactions of the Japan Society of Mechanical Engineers, Series C, vol. 63-615, 1997, pp. 4050-4056. (in Japanese) [2] Sessomboon W., Murofushi T., Irohara T. and Yoshimoto K., “A Layout Technique for Unequal Area and Shape Departments Considering Main Aisle”; Transactions of the Japan Society of Mechanical Engineers, Series C, vol. 63-615, 1997, pp. 4057-4064. (in Japanese) [3] Sessomboon W., Sugiyama M., Irohara T. and Yoshimoto K., “A Design Technique of AGF Flow Path by Considering Machine Layout”; Transactions of the Japan Society of Mechanical Engineers, Series C, vol. 64-617, 1998, pp. 370-376. (in Japanese) [4] Shirai Y. and Matsumoto N., “Simultaneous Optimization of Floor and Aisle Planning for Facility Layout Problems”; Transactions of the Japan Society of Mechanical Engineers, Series C, vol. 65632, 1999, pp. 1593-1600. (in Japanese) [5] Masuyama T., Kobayashi T., Fukatsu K., Yamanaka M. and Inoue K., “Layout Design of Functional Modules and Connecting Routes (Tuning of Weight in GA-based Design of Automated Teller Machine)”; Journal of Japan Society for Design Engineering, vol. 36, no. 4, 2001, pp. 157-163. (in Japanese) [6] Inoue K., Osaki H., Masuyama T. and Fukatsu K., “Layout Design of Functional Modules and Routing for Their Connection Considering the Route Divergence and Confluence (Layout Design of Modules and Routing for Automatic Teller Machine)”; Transactions of the Japan Society of Mechanical Engineers, Series C, vol. 68-655, 2002, pp. 331-338. (in Japanese) [7] Koopmans T.C. and Beckmann M.J., “Assignment Programs and Location of Economic Activities”; Econometrica, 1957, vol. 25, 1957, pp. 53.

[8] Lee R.C. and Moore J.M., “Computerized Relationship Layout Planning”; Journal of Industrial Engineering, vol. 18, no. 3, 1967, pp. 194-200.

[9] Tate D.M. and Smith A.E., “Unequal-area Facility Layout by Genetic Search”; IIE Transactions, vol.

27, no. 4, 1995, pp. 465-472.

[10] Stefan J., “On Genetic Algorithm for the Packing of Polygons”; European Journal of Operational Research, vol. 88, no. 1, 1996, pp. 165-181.

[11] Kimura M., Shiraishi Y., Himeno F., Hanaki, S. and Kusaoke M., “A Grid-Free Maze Router”; The Institute of Electronics, Information and Communication Engineers Transactions, Series A, vol. J74-A-8, 1991, pp. 1294-1301. (in Japanese) 40

Articles

VOLUME 1,

No 4

December 2007

[12] Lee C.Y., “An Algorithm for Path Connections and Its Application”; IRE Transactions Electronic Computers, vol. EC-10, 1961, pp. 346-365.

[13] Goldberg D.E., “Genetic Algorithms in Search, Optimization and Machine Learning”, 1989, Addison Wesley.

[14] Inoue K., Fueki H., Ohmachi T. and Kato M., “An application of Genetic Algorithms to Design of Stiffened Plates”. In: Proc. International Conference on Engineering Design, International Society for the Science of Engineering Design, 1995, pp. 1369-1377.

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 1,

No 4

December 2007

Neural Network Based Selection of Optimal Tool - Path in Free Form Surface Machining Marjan Korosec, Janez Kopac

Abstract: The purpose of the presented paper is to show how with the help of artificial Neural Network (NN) the prediction of milling tool-path strategies could be performed in order to determine which milling tool - path strategies or their sequences will yield the best results (i.e. the most appropriate ones) of free form surface machining, in accordance with a selected technological aim. Usually, the machining task could be completed successfully using different tool-path strategies or their sequences. They can all perform the machining task according to the demands but always only one of the all possible applied strategies is optimal in terms of the desired technological goal (surface quality in most cases). In the presented paper, the best possible surface quality of a machined surface was taken as the primary technological aim. Configuration of the applied Neural Network is presented and the whole procedure of determining the optimal tool-path sequence is shown through an example of a light switch mould. Verification of the machined surface quality, in relation to the average mean roughness Ra is also being performed and compared with the NN predicted results.

Keywords: (NN) neural network, CAD/CAM system, CAPP, Intelligent CAM (ICAM), milling strategy

1. Problem formulation Many efforts have been made in order to simplify and make NC programming procedures easier. Nowadays, the trend in CAM systems development is to make different CAM systems capable of recognizing particular features which compose a 3D model of the part and then generate the most important machining procedures and parameters [1] according to geometric shape recognition.

Some researchers employ Neural Networks and Genetic Algorithms (GA) at this stage but they all face the problem of recognizing very complex free form surfaces, which are far away from being only the basic geometric shapes, such as a cylinder, a cube, a cone etc. So the problem arises how to present a complex surface configuration of free form model to a Neural Network. NN should be capable of predicting the right or optimal machining strategy in order to achieve a high surface quality. So NN must be somehow acquainted with the complex surface configuration of machined workpieces [2]. One possible solution of this problem is shown in this paper. For machining 3D complex surfaces, it is often not enough to use basic tool-path milling strategies only [3]. Specific combinations, which even change during travelling of the cutting tool across the surface, should often be used. Combining milling strategies, there are no simple relations between machining parameters, because they are changing in time

and depend on a particular sequence of milling strategies.

Their mutual relations are mostly non-linear. The question also arises as to which milling tool-path strategy will be the most adequate to satisfy the demands according to selected technological aims [4]. Generally, it is possible to make optimization according to these main technological aims: • best possible surface quality, • minimum tool wear, • shortest achieved machining time and, • minimum machining costs.

Since this problem was initiated by the tool shop industry, which produces tools for car lights equipment, our technological aim was to achieve the best possible surface quality of machined workpieces. Different milling strategies can be applied to machine the same complex surface on a workpiece, but surface quality after each different applied combination of milling strategies will differ a lot.

It has been realized that by changing the feed rate and cutting speed only, it is very hard to achieve the best possible surface quality in 3D complex surfaces.

2. Current state-of-the-art Many researchers and developers of CAM systems try to incorporate some intelligence in their applications in order to improve technological knowledge. Some of them are trying to introduce NN, GA and expert systems in their solutions in order to be able to predict crucial machining parameters. A modified Backpropagation NN is proposed for on-line modelling of the milling system and a modified NN is proposed for the real-time optimal control of the milling system [4]. Also a self-organized Kohonen NN is used for path finding and for feed rate adjustment [5,6].

New approaches tend toward integrating CAD, CAPP (computer added process planning) and CAM system [7].

By integrating those three systems, feature-based technology becomes an important tool. The goal of this integration is to replace a typical procedure of manual process planning (Figure 1a.) with CAPP, which represents a basis for automatic generation of NC machining programme (Figure 1b.) [8,9] In such integrated system, the so-called feature based design or design by features should be used instead of conventional CAD methods. Design feature understands the properties of the region to be machined, such as the geometric shape, the dimensions, the dimensional and geometrical tolerances, etc. However, such features, being primarily design-oriented, have to be converted into manufacturing features in order to build the interface between the feature-based design and automated process planning. According to the definition, manufacturing feaArticles

41

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 1,

No 4

December 2007

Figure 1a. Manual process planning.

Figure 1b. Extracting of manufacturing features. tures are surfaces or volumes, which are produced by one or a series of material-removal operations [10]. Without properly defined manufacturing features it is not possible to perform automated NC code generation.

Activities in the area of CAPP and intelligent CAM can be divided into four areas: • -feature recognition, • -extracting manufacturing features from a featurebased design model, • -operation selection as part of CAPP and • -operation sequencing as a part of CAPP. Four modules are included in obtaining manufacturing features, presented in Fig. 1b. Feature recognition has been one of the major research issues in the area of automated CAD/CAPP interface. So far, it has also been also the basis for applying intelligent CAM (ICAM) systems.

Some main approaches in this area include the Cover Set Graph (CSG)-based approach, the graph-based approach and the neural-network-based approach. Most suggested methods for feature recognition apply a solid model as their input, which represents only the purely geometric aspects of the design information. In 1992, a Super Relation Graph (SRG) system, using artificial Neural Networks, was developed for the purpose of feature recognition.

The objective of this system is to recognize and extract prismatic features from 3-D CAD databases [10]. This system recognizes some volumetric primitive features and classifies them into: holes, pockets, blind-slots, blind-steps, thru-slots, and steps. Using the techniques of artificial neural networks and computational geometry, the SRG identifies only features based on two types of relationships between faces: super concavity and face to face. Basically, the SRG is a matrix representation of the relationship between the faces. Later, the SRG based system was improved by adding cover set model (CSM) 42

Articles

approach. The CSM is built on the SRG system and determines the essential and non-essential features [11].

In the field of operation sequencing, two methods, i.e. knowledge-based evaluation and fuzzy quantitative evaluation, are widely applied. It means that attributes of a feature are quantitatively fuzzified into a number of measures, which can build up a numeric data array for modelling important features. After that, the fuzzy evaluation function created with neural network can be used to automatically execute feature prioritization. It is clearly that NN based approach is advancing very fast in all CAPP segments, mainly because of the ability of incremental learning and the capability of modelling non-linear inter-relationships. The other advantage of NN based approach is providing a more precise technique and representing the complex inter-connections between the fuzzified feature parameters and the manufacturability of the features [12].

All these approaches are needed to provide the necessary link between CAD and CAM systems and integrate them into the CAPP system. The described systems add technological data to the features from some general data bases, but those added machining data are not optimized for every single feature within the CAD model. Actually, there is no need for feature optimization because of their simple topological and geometrical nature (cylinders, spheres and prismatic features only). But in the case of a complex surface recognition, machining parameters have to be optimized, most effectively by the NN approach [13, 14, 15]. The common limitation of all the above-mentioned methods is that they can recognize and define only volumetric features, based on solid models [16]. These methods can still not recognize features in surface models. When talking about complex free form surfaces, it is not possible to simply divide them into some elementary prismatic or cylindrical features because of their irregular shapes. The

Journal of Automation, Mobile Robotics & Intelligent Systems

other problem with free form surfaces is also their non-linear technological and topological properties relationship, which are impossible to be captured with the above-mentioned methods [17,18]. Considering the fact that nowadays mould design industry uses circa 70% of surface modellers, it is obvious that there is a need for a more general method of manufacturing feature recognition.

The method presented in this paper is different from the above-mentioned approaches especially in the following attributes: it is applicable in solid, as well as in surface 3D free form models, the concept of NN training takes account of all geometrical and topological non-linear relationships, it can easily represent free form surface to the NN, it does not have to be supported from the feature based design concept (therefore it can be used in any CAM system, or in the frame of a widely used CAPP system), and the method is adaptive according to the selected technological goal.

3. Presentation of free form surfaces to the neural network As mentioned before, many machine-technological parameters depend on the workpiece surface configuration.

But for a successful tool-path optimization with the use of NN, the main problem remains: how to present the surface configuration to the Neural Network. So the workpiece surface configuration must first be recognized by NN. 3.1 Selection of representative 3D models and their corresponding milling-path strategies According to the programming of machining with CAM system (Hypermill-Open Mind), we created five representative 3D models shown in Figure 2. They are the most frequently used tool path strategies and surface forms in our tool-shop company. These selected milling strategies proved to be the best for selected 3D models according to the checked surface quality (mean roughness Ra). For machining material, we used 54 HRc steel. The selected tool path strategies for models presented in Figure 2 are: • Combination of Profile finishing and Z finishing (slope mode option) representing model NN1. First, flat surfaces are machined in the “profile finishing” mode - surfaces which have the slope angle smaller than the boundary set angle - and then the rest of the surface is machined in the “Z finishing” mode.

• Profile finish, or 3D finish, representing NN2 model.

• Profile finish (scallop height mode), representing NN3 model.

• Z level finish, representing NN4 model.

• Profile finish, (equidistant machining, in feed is constant on the whole surface area), representing NN5 model.

VOLUME 1,

No 4

December 2007

Shapes of the shown 3D models were machined with the proposed milling strategies and in this way the best surface quality results were achieved.

3.2 Multiplication of basic 3D models In order to get enough training data for NN learning data-base and in order to make the application more universal, a C++ executable programme named Saturnus. exe was written. Among other tasks, it also rotates every basic model in increments by 10° (degrees), starting from -50° to +50°. After the first ten rotations, the basic model was turned upside down by 180° degrees and rotated again by 10 degrees increments. The starting, final and incremental angles are arbitrarily chosen by the user as an input in our written programme (the input file organization will be presented later in this paper). In this way, each basic model produced 20 additional submodels. So each basic model together with the rotated sub-models provided 24 different models, thus 24 NN model vectors. With five basic models, we provided 120 NN model vectors.

3.3 Projection of points in training models This task was also automatically performed in our C++ executable program. At first, the programme clipped models in the smallest possible rectangular shape, preserving the a/b relations of rectangular sides constant for each model. After that, models were transferred into the so-called “model space”, where they were lifted over the rectangular ground plane and strewed with a raster of points, on the upper part of the model. The strewed points must be settled in appropriate raster, which is arbitrary.

In our case, it was 1 mm in the X and Y directions. The more points there are, the more precise the interpolation, which is later performed in the same programme. The whole set (raster) of points was directly projected on each of the five basic models. The number of points and their raster must be equal for every model. After the points were projected, the models were removed and only the distributed points remained in the picture. This was done because later, in the appropriate data transfer format, only points were presented, which was actually of our interest. Namely, we used coordinates of those points in order to get information about 3D surface configuration. As a software tool for all described manipulations with models, Mechanical desktop V4.0 was used.

The purpose of the point’s projection across the model was to gain up the surface configuration of the models in the Z direction, which has proved to have the biggest influence on the important technological parameters (feed rate and cutting speed).

Figure 2. Five models representing basic milling tool-path strategies.

Articles

43

Journal of Automation, Mobile Robotics & Intelligent Systems

3.4 Preparing VDA files for input into NN The strewed points were translated with the VDA file format, and then those VDA files are used as an input in our executable programme Saturnus.exe, which performed interpolation between points and their reorganization , and produced ASCII data files, organized in a way to be convenient for entering into NN as input data.

Therefore, our programme performed two very important tasks: it made an interpolation between strewed points so the necessary amounts of data points were reduced (but the Z height configuration of surface was still retained), and it clipped, strewed and rotated the models.

With the interpolation, the starting amount of points was reduced to a rectangle having 15 points in the Y direction and 15 points in the X direction for each model.

3.5 Structure of model vectors as an input to the NN The programme Saturnus.exe produced ASCII files organized as model vectors suitable for direct entering into NN. Each model vector consists of an input and output part (variables).

Concatenation of both vectors gives the original model vector. This can be written as:

mv = P ⊕ Q = (m1 , m2 , ..., m M , m M +1 , ..., m L ) (1) and in a matrix form:

VOLUME 1,

3.6 Structure of output part of model vectors Looking at Table 1, it is noticeable that the discrete output variable has eight variables of probability. Five of them are used to designate the milling path strategy. Three output variables are left in case of expanding the number of milling path strategies from five to eight. The meaning of the variables of probability is as follows: (out 1) 10000000….profile finish + Z finish (slope mode option) (out 2) 01000000….3D finish, respectively profile finish (out 3) 00100000….profile finish (scallop height mode) (out 4) 00010000….Z level finish (out 5) 00001000….profile finish (equidistant machining, constant infeed)

Table 1. Organization of model vectors in the NN training data base.

Articles

December 2007

where the shadowed part belongs to the output part of the model vector.

In the presented case, the input part of the model vector consists of 225 (15x15) interpolated coordinate points, and the output part consists of eight probability variables, for one milling strategy each. So there are 120 (5x24) model vectors, each of them consisting of 225 input variables, and one discrete output variable. This presents a learning base for our NN. Training model vectors and their organization are shown in Table 1.

Figure 3. Structure of model vectors in a matrix form.

44

No 4

Journal of Automation, Mobile Robotics & Intelligent Systems

These are five most often used milling strategies for finish machining. Those strategies will not be described in this paper because this is not the intention of this paper. However, their details can be found in almost every CAM system.

4. Neural network setup and algorithm used for probability prediction of milling toolpath strategies The NN algorithm used for solving the problem of surface recognition is different from the traditional artificial Neural network, in the sense that it is derived from the probabilistic approach and it uses a new self-organizing system algorithm. It is based on a self-organizing system, called the neural network-like system, presented by Grabec [19,20]. It is similar to the method of the nearest neighbour, Learning Vector Quatization Network [21], and also to the probabilistic neural network, proposed by Specht [22].

All of the above-mentioned methods have the same foundation and similar rules for describing various phenomena. On the other hand, they are different compared to each other, similarly to differences among various paradigms used in backpropagation artificial NN, or among various types of artificial NN. Most NN that can learn to generalize effectively from noisy data are similar or identical to statistical methods. For example, probabilistic neural nets are identical to kernel discriminant analysis. On the other hand, Kohonen’s self organizing maps have no close relativities in the existing statistical literature, but self-organization of neurons, proposed by Grabec, is very similar to Kohonen’s self-organization process and is based on statistical principles [21,23]. Also, feedforward nets are a subset of the class of non-linear regression and discriminant models.

Neural network can learn from cases. It predicts probability in %, which determines milling strategies or their combination that is to yield the best machining results, according to surface roughness. It means that the machining time of predicted strategies is not necessarily the shortest time because our technological aim was to yield the best possible quality of machined surface.

4.1 Definition of probability assumption However, it is very unlikely that a perfect match exists in reality. Thus, a second probability-based assumption is needed. It states that if the input parts of model vectors P and C are “near”, there is a high probability that the output part of C is similar to the output part of P. Conversely, if the input parts of P and C are “far”, there is only a low probability as shown in Figure 4. [24].

VOLUME 1,

No 4

December 2007

The words “near” and “far” from assumption are then converted into numbers. Thus, two vectors are near if the vector norm of their difference is a small value. Usually, an Euclidean norm is used in such cases. Equation (5) shows the Euclidean norm for the difference (distance) of the vectors P and C [25].

d PC = ¦ (x Pi − x Ci )

2

(5)

i

where: dPC is the distance between the input parts of model vectors P and C xpi is input part of model vector P xci is input part of model vector C When the distances between the model vectors are defined, a Gaussian probability function can be selected.

Moreover, if the probability function and the distance are known, the similarity can also be calculated. Thus, the similarity between P and C is represented by [26, 27]:

− d PC 2

S PC = e

α

(6)

where: SPC is the similarity between model vector P and C a is the penalty coefficient, replacing the standard deviation value Equation (6) is a slightly modified Gaussian function because the standard deviation cannot be calculated. Therefore, the standard deviation is replaced by a constant value, which is called “the penalty coefficient”, and has a significant influence on the shape of the probability function. The penalty coefficient is selected a priori by the user.

For each model vector in M, its distance from P and their similarity can be calculated. To simplify the final calculation of the output part of vector P, the similarity coefficient must be normalized; that is, their sum must equal 1 [26]:

s px =

s px ; → ¦ s pi = 1.0 ¦ s pi i

(7)

where: s pi normalized similarity coefficient of model vector P s px normalized similarity coefficient of model vector X Once the similarity coefficients are normalized, the final result is obtained by a combination of the output parts of all model vectors :

P0 = ¦ s px ⋅ X 0

(8)

x

where: P0 is the final calculation of the output part of model vector P The index X in Equation (8) runs over all model vectors in the model. It should be stressed that the most important thing the user has to do is choosing a penalty coefficient that minimizes the mean square errors from the output variables. The programme uses a method, which prevents “over-training”. Namely, it first deletes a case and then uses the remaining cases for training. This “trained” NN is then used to validate this deleted case. This operation is reiterated until all cases have been processed [22, 26].

Figure 4. Probability assumption.

Articles

45

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 1,

4.2 Neural network construction In this particular NN, the training phase is very quick and corresponds to the presentation of the model vectors (loading the database) to the network (Manual from NeuralWare, Neural, 1991). The prediction phase corresponds to the calculation of values of processing elements and to the calculation of unknown output values of prediction vector (in case of prediction) or output values of model vectors (in case of filtration of verification to determine the penalty coefficient value). The weights on connections equal either one or zero. The expression for weight adaptation can be written as: (9) wij = wijδ

kj

where

wij equals 1.0, and ij is defined as: ­1; i = j

δij = °®

°¯0; i ≠ j

NN has two hidden layers (layer B and layer C). The number of neurons in layer B equals the product of the number of all model vectors N and the number of input variables M (N.M), while the number of neurons in layer C equals 2 times the number of model vectors. Graphical presentation is shown in Figure 5.

No 4

December 2007

The presented procedure corresponds to the associative recognition of some unknown properties of the phenomenon on the basis of incomplete observation or experience, obtained by previous complete observation.

5. Training of neural network The form of model vectors prepared with executable program Saturnus.exe is shown in Table 1. This NN does not learn in a conventional manner but it actually learns simultaneously during the prediction phase. In conventional NNs, a lot of time is spent for training (determination of weights) the NN. But once it has been trained, it predicts quickly. Here it is just the opposite: it predicts a little bit slower, but it learns very quickly. Partially, it happens also because the presented NN uses only two weight values: 0 and 1.

Before running the test set, all vectors had to be normalized and penalty coefficient a must be chosen. The training test was also applied in order to determine the right value of the penalty coefficient. The actual value of the penalty coefficient depends on the density of model vectors. Automatic determination of the penalty coefficient is based on minimizing the RMS verification error.

Penalty coefficient is strongly correlated with the learning error in the back propagation neural network (BPNN) and is a very important parameter.

5.1 Results of a training test with known data

Figure 5. Construction of neural network.

Notations in Figure 5 have the following meanings: • p prediction vector, • m model vector, • i indicates the neuron, belonging to the input variable, • o indicates the neuron, belonging to the output variable.

• N number of model vectors, • M number of input variables of the phenomenon, The processors (neurons) are connected by unidirectional communication channels, which carry numeric data: this can be seen in Figure 5. It should be mentioned that connections (weights) are not changed; they have values either 0 or 1 [22]. All four layers have linear transfer functions and the final output of the neuron on layer D is :

(

)

po = Y D = f X D , X D =

XD XD

(10)

where: X D and X D mean neuron value before and after weight adjustment.

46

Articles

Figure 6. Predicting probabilities of milling strategies in a training set for the first 3D model.

First, data points are normalized. Since the variable values are not falling within particular limits, the statistical normalization was used. Figure 6 shows probability values for the first 3D model (NN1 shown in Figure 2) using data base of all 120 model vectors. It is clearly seen that NN proposes using the first milling path strategy (profile finish + Z finish in slope mode option, as marked in Figure 2) for machining the model since the first 24 model vectors (which represent the first strategy) achieved the highest probability (0.8 to 0.85 in the verification curve in Figure 6) and as such, will give the best surface roughness results. This prediction is quite correct considering the fact that in the learning set, the first 3D model (i.e. the first 24 model vectors) is machined with profile finish + Z finish strategy as the most convenient strategy. In Figure 6, two curves are presented. The verification curve actually excludes the model vectors for which the milling path strategy is predicted. The filtration tool also includes the model vectors for which the

Journal of Automation, Mobile Robotics & Intelligent Systems

prediction is being made. If the filtration and verification curves are getting along fairly well it means that the data noise is small, and vice versa [24,26]. The second proposed milling strategy according to the predicted probability in Figure 6 is equidistant machining with constant infeed over the surface area (verification probability 0.4 to 0.45), and so on. In later experiments with models, which NN hasn’t “seen” yet, only their VDA files are needed. The procedure of preparing the VDA files for NN is done automatically, before running the NN.

VOLUME 1,

No 4

December 2007

ble 1) and therefore represent a really serious proof for NN model.

6. Testing Neural Network model The proposed NN model is tested on a two experimental 3D models, which have never before been “seen”

by NN. The milling path strategies will be predicted with a view to the best possible surface quality. Both 3D models represent the upper part of a mould for plastic injection, and are taken out of the tool-shop practice. The projection of point set was made on a 3D model as shown in Figure 7a and 7b. Points were projected from a rectangular net, lifted over the 3D model as described in chapter 3. In this particular case the spacing between points arrayed on a rectangular raster was 1 mm in the X and 1 mm in the Y direction, because of very steep walls. The spacing between points is picked out arbitrary and depends on a model surface configuration. The results of projecting the points are shown in Figure 7. The points were projected and strewed as described in chapter 3.

Figure 7a. 3D free form model and point model of a lightswitch body.

Figure 7b. 3D free form model and point model of a water tank body.

When DWG files of point set were translated into a VDA files, a record of 3832 strewed points (x, y, z coordinates) in a case of light switch body and 18.235 points in a case of water tank came out. Those points describe the 3D surface configuration, which is then imported into NN. 6.1 Results of the test set When the training model has proved to work well, the milling path strategies for both models from figure 7. are predicted.. It has to be emphasized one more time that these models have never been “seen” before by the NN (they are not included in the training model from Ta-

Table 2. Predicted probabilities of milling path strategies for model 1.

Figure 8. NN results for predicted probabilities of tool path strategies for light switch model and water tank model.

The following finish milling tool-path strategies were used: out 1……profile finishing + Z finishing (slope mode option) out 2……3D finishing out 3……profile finish (scallop height mode) out 4……Z level finish out 5……profile finish (equidistant machining, constant infeed) Looking at Table 2, and observing the prediction for the light switch model, it is very obviously that NN gave the highest probability to strategy number 5 (Mv 1 gave 0.45, and total sum 5.46), that is equidistant machining with constant infeed. The second and the third predicted strategy probability are almost the same (total sum 3.56 and 3.47). Figure 8 shows a graph of predicted probabilities of milling strategies. When observing the prediction for for water tank (model 2), the highest probability was given to the profile finishing + Z finishing with slope mode option (this is strategy “out 1” in Figure 8). The second and third probabilites were 3D finishing (“out 2”) and profile finishing with scallop height mode (“out 3”), which were actually used by NC programmers.

Articles

47

Journal of Automation, Mobile Robotics & Intelligent Systems

In order to get sufficiently satisfactory predicted results, the basic 3D model of a light-switch was rotated around each axis (step of rotation in degrees and axes are arbitrary selected, before running our executable program Saturnus.exe ) and strewed with points again. In this way, we got 15 model vectors, and hence the prediction is more reliable. It is important to notice that in most predicted model vectors the strategy with the highest probability is the same, in our example the milling path strategy number 5 (see Table 2). This also implies that the predicted probability might be correct.

If the highest predicted probability were divided among different milling path strategies in each model vector the result would probably be doubtful, and the step in the X and Y directions should be changed or the learning model should be redefined. In the presented example this was not the case.

6.2 Checking machined surface quality for NN predicted tool-path strategies and their assessment Usually, the most frequently used strategy for finish machining of proposed experimental parts used by NC programmers was a conventional 3D finish (strategy in output 3 in Table 2). Sometimes some NC programmers also used a combination of 3D finish with Z level finish machining (strategy in output 1). The part was machined with all 3 milling path strategies, and the centreline average roughness was compared. Finish machining was performed using a ø 2 mm ball mill two flutes cutter, with machine parameters: n = 18.000 min-1, vf = 1900 mm/min, aa = 0.04 mm, pf = 0.2 mm, employing 3+2 axes simultaneously. The parameters were picked out as appropriate for a high-speed-cutting process and a machining workpiece of 54 HRc. A new tool was used for every milling strategy to exclude the influence of tool wear on surface roughness. The results are shown in Figure 9 and Figure 10.

Ra (0,001 mm)

Figure 9. Surface roughness achieved using three milling tool-path strategies for model 1.

Figure 10. Surface roughness achieved using four milling tool-path strategies for model 2.

48

Articles

VOLUME 1,

No 4

December 2007

In the presented case, NN predicted the probabilities in order of precedence, according to the achieved centreline average roughness Ra on a machined surface. The best Ra at model 1 (light switch) was achieved with the milling strategy “out 5”. In this strategy almost 50% percent of machined surface has the Ra value of about 0.45 mm, 30% achieved Ra≅0.7 mm, and 25% of machined surface achieved the value of Ra≅0.85 mm. The next two strategies according to the surface quality were milling strategy “out 1” and “out 3”. As shown in Figure 9, the machining results are well in agreement with the predicted results from NN in Figure 8. Despite selecting a rather simple machined surface, it is obviously that NC programmers mostly selected milling strategies “out 1” and “out 3”, which satisfied the requirements in the tool shop industry, but yielded worse results with a view to surface quality (see Figure 8) than strategy “out 5”, proposed by the developed NN. In such a cases, NN can be of great help for NC programmers, operators and technologist in tool-shops, especially in the sphere of fine machining of 3D complex and functional surfaces, where the surface quality plays a major role.

Inspecting the model 2 (water tank) for average surface roughness Ra, it is noticed that the smallest Ra is achieved by using profile + Z level finishing (“out 1”) tool-path strategy. By using this strategy, almost 50% of surface has the Ra value of about 43 mm. The second best strategy regarding the achieved Ra is 3D finishing (“out 2”). Comparing results with predicted probabilities in Figure 8, it is clear that actually tool-path strategy “out 1” and “out 2” yielded the highest probability, which is in agreement with surface roughness results from figure 10.

7. Conclusion A method for optimal choosing and optimizing the milling tool – path strategies based on the use of NN has been presented. The presented method could be used in solid or surface models and can be applied in all modern CAM systems. The surface quality was set as the primary technological aim, and it was focused on it, considering the tool-shop industry. Of course, one may wish to set up a different technological aim, such as achieving the smallest tool wear, or shortening the machining time etc.

When changing the technological aim, the learning model should be reorganized according to the new technological aim and NN should be trained again. The more stirring and confabulated machined surface, the more complex and interlaced are the machining parameters, more difficult is to combine the right order of precedence for milling strategies, or to select the most suitable strategies to achieve the best possible machining surface results. In the case of large complex surfaces, reliability is improved when surfaces are divided into technologically and geometrically reasonable subsurfaces before applying the method. It must be also stressed that representative models for NN training phase depends on the type of milling path strategies which will be used inside particular CAM system. So representative models must be choosen in accordance with the capabilities of CAM system which will be used for machining.

After predicting milling path strategies, it is also possible to make a new NN learning model for predicting maximum possible feedrate and rotational speed of spindle.

Journal of Automation, Mobile Robotics & Intelligent Systems

Input variables in a new NN are: probability of predicted milling path strategies, hardness of workpiece material and stock allowance left during machining. Output variables in a new NN are: maximum possible feedrate and rotational speed of spindle [28]. Predicted feedrate and rotational speed could be then used as the upper limit value for milling path optimization in applications such as OPTIPATH or OPTIMILL (Vericut v. 5.0, CG Tech Ltd).

For solving those problems, NN can serve as an ideal tool for helping NC programmer make the right decision, or at least serving as an orientation tool. The advantage of NN based approach presented in the paper is its ability to learn and recognize all possible complex non-linear topological and geometrical relationships, which cannot be recognized by other graph based or similar techniques.

In this way, it is also possible to save time because many additional post machining operations are reduced to a minimum or even zero amount of time.

AUTHORS Marjan* Korosec, Janez Kopac – University of Ljubljana, Faculty of Mechanical Engineering Askerceva 6, 1000 Ljubljana; Slovenia.

E-mails: Marjan.Korosec@lecad.uni-lj.si, Janez.Kopac@fs.uni-lj.si. *Corresponding author

References [1] A.C. Lin, S.Y. Lin and. S.B. Cheng, “Extraction of manufacturing features from a feature-based design model”, Int. J. Prod. Res., vol. 35, 1997, no. 12, pp. 3249-3288.

[2] G.A. Stark, K.S. Moon, “Modeling surface texture in the peripheral milling process using Neural network”, Journal of Manufacturing science and Engineering, ASME, ISSN 1087-1357, May 1999.

[3] S.H. Suh, Y.S. Shin, “Neural network modeling for tool path planning of the rough cut in complex pocket milling”, Journal of Manufacturing Systems, ISSN 0278-6125, 1996, pp. 295-304.

[4] J. Balic, A. Nestler, G. Schulz, “Prediction and optimization of cutting conditions using neural networks and genetic algorithm”, Journal of Mechanical Engineering, Association of Mechanical Engineers and Technicians of Slovenia, ISSN 0039-2480, 1999, pp. 192-203.

[5] Y.M. Liu, C.J. Wang, “Neural network based adaptive control and optimisation in the milling process”, International Journal of Advanced Manufacturing Technology, ISSN 0268-3768, vol. 14, no. 11, 1999, pp. 791-795, [6] M. Korosec , “Optimization of free form surface machining, using neural networks”, Doctor thesis, 2003, University of Maribor, Faculty of technical engineering.

[7] Sankha Deb, Kalyan Ghosh; S. Paul, “A neural network based methodology for machining operations selection in Computer-Aided Process Planning for rotationally symmetrical parts”,

VOLUME 1,

No 4

December 2007

Journal of Intelligent Manufacturing, vol. 17, no.

5, October 2006 , pp. 557-569(13).

[8] M. Brezocnik, I. Pahole, J. Balic, “Feature recognition from boundary model of a part” (intelligent CAD-CAP interface), in: Proc. International Conference Design to Manufacture in Modern Industry, Bled, Slovenia, 29th-30th May 1995, pp. 395-404.

[9] J. Dong and S. Vijayan, “Feature extraction with the consideration of manufacturing processes”, Int. J. Prod. Res., vol. 35, no. 8, 1997, pp. 21352155.

[10] T.N. Wong and Wong K.N., “Feature-based design by volumetric machining features”, Int. J. Prod. Res., vol. 36, no. 10, 1998, pp. 2839-2862.

[11] K. A. Aldakhilallah and R. Ramesh, “Recognition of minimal feature covers of prismatic objects: A prelude to automated process planning”, Int. J. Prod. Res., vol. 35, no. 3, 1997, pp. 635-650.

[12] C. Bishop, Neural networks for pattern recognition, Oxford Press, 1995.

[13] I. Grabec, “Optimization of kernel-type density estimator by the principle ofmaximal self-consistency”, Neural Parallel & Scientific Computations, no. 1, 1993, pp. 83-92.

[14] S. G. Wang, Y. L. Hsu , “One-pass milling machining parameter optimization to achieve mirror surface roughness”. In: Proceedings of the I MECH E Part B Journal of Engineering Manufacture, vol.

219, no. 1, 2005, pp. 177-181(5) .

[15] J. Wang, “Multiple-objective optimisation of machining operations based on neural networks”, The Int. J. of Advanced manufacturing technology, Springer London, vol. 8, no. 4, July 1993, pp. 235-243.

[16] C.K. Mok and F.S.Y. Wong, “Automatic feature recognition for plastic injection moulded part design”, The International Journal of Advanced Manufacturing Technology, Springer: London, vol.34, no. 5-6, September 2007, pp. 1058-1070.

[17] G. Jung Hyun Han, Inho Han, Eunseok Lee, Juneho Yi, “Manufacturing feature recognition toward integration with process planning systems”, Man and Cybernetics. Part B, IEEE Transactions on, vol.

31, issue 3, June 2001, pp. 373 – 380.

[18] Helen L. Locket, Marin D. Guenov, “Graph based feature recognition for injection moulding based on a mid-surface approach”, Computer –Aided Design, vol. 37, issue 2, 2005, pp. 251-262.

[19] I. Grabec, “Self-Organization of Neurons Described by the Maximum Entropy Principle”, Biol.

Cybern., no. 63, 1990a, pp. 403-409.

[20] I. Grabec, “Modeling of Natural Phenomena by a Self-Organizing System”, Proc. of the ECPD NEURO COMPUTING, vol. 1, no. 1, 1990b, pp. 142-150.

[21] D. F. Specht , Probabilistic Neural Networks for Classification, Mapping or Associative Memory, ICNN-88, Conference Proc., 1988, pp.525-532.

[22] C. Principe, R. Euliano, Neural and adaptive systems, John Wiley&Sons, 2000.

[23] T. Kohonen et al., “Statistical Pattern Recognition with Neural Networks: Benchmark Studies”. In: Articles

49

Journal of Automation, Mobile Robotics & Intelligent Systems

Proceedings of the 2nd Annual IEEE International Conference on Neural Networks, 1988, vol. 1.

[24] J. Guh et al, “Predicting equilibrated postdialysis BUN an artificial neural network in high-efficiency hemodialysis”, Am. J. Kidney Dis , no. 31 (4), April 1998, pp. 638-46.

[25] S. Prabhu, “Automatic extraction of manufacturable features from CADD models using syntactic pattern recognition techniques”, International Journal of Production Research, vol. 37, issue 6, 1999, pp. 1259-1281.

[26] Neural Ware, Inc., Neural Computing Manual, Wiley, 1991.

[27] T.C. Li, Y.S. Tarng, M.C. Chen, “A self-organising Neural network for chatter identification in milling”, International Journal of Computer Applications in Technology, ISSN 0952-8091, vol.9, no.56, 1996, pp. 239-248.

[28] M. Korosec, J. Balic, J. Kopac, “Neural network based manufacturability evaluation of free form machining”, I. Journal of Machine Tools & Manufacture, no.45, 2005, pp. 13-20.

50

Articles

VOLUME 1,

No 4

December 2007

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 1,

No 4

December 2007

Evolutionary Prediction of Manufacturing Costs in Tool Manufacturing Mirko Ficko, Boštjan Vaupotic, ˇ Jože Balicˇ

Abstract: One of the most important factors in the offer for tool manufacture is the total manufacturing cost. Although the total manufacturing costs can be rather precisely determined by the cost analysis, this approach is not well applicable in tool-making due to cost and, particularly, time demand. Therefore, the authors propose a new approach to prediction of total manufacturing costs, which is based on case based-reasoning method and imitates the human expert. The system first abstracts from CAD-models the geometrical features, and then it calculates the similarities between the source cases and target case. The most similar cases are used for preparation of prediction by genetic programming. The genetic programming method provides the model connecting the individual geometrical features with the costs searched for. Regarding to the connections between geometrical features and tool cost of source cases the formula for calculation of tool cost of target case is being made. The experimental results show that the quality of predictions made by the intelligent system is comparable to the quality assured by the experienced expert.

Keywords: prediction of tool manufacturing costs, case based reasoning, genetic programming

1. Introduction Quick response to business opportunity has been considered as one of the important factors to ensure company competitiveness. Therefore, new products must be more quickly developed, manufactured and introduced to the market. This fact is obvious for example in automotive industry, where the time to develop a car has been reduced from 60 months 10 years ago to 18 months today (Ding, Y. et al., 2004). In such competitive conditions, where new products appear on the market within shorter time intervals the development time is shorter and shorter, the branch of industry busy with tool manufacture assumes a vital role. The capacity of the tool-making shop to respond quickly to the demand in today’s competitive environment is a key factor of competitiveness.

The buyers of tools have a worked out idea about finished product, whereas the tool-makers are responsible for the tool design, preparation of the manufacturing technology and final manufacture of the tool. The first activity of tool making shop connected to new order is preparation of the offer. In most cases the offer itself is quite simple. Most often it comprises only business information, above all, the price. This information is of key importance for the economic success of the order both for the buyer and seller of tools. However in the multi-project environment, characteristics of the tool-making industry it is dif-

ficult to predict the total costs at execution of the order. It would be very useful to have a tool for preparation of total manufacturing costs. The paper tries to find a solution for this problem. Differently from other approaches, we have developed a system which imitates the natural intelligent system – expert for solving this problem.

This paper comprises five sections. The introductory section presents the problems of the tool-making industry occurring in preparation of the order. The second section discusses the cost prediction activity focusing on costs prediction in the tool-making companies. The third section presents the model of the cost prediction by intelligent methods. The subsections explain the individual components and working principles of that system. The fourth section deals with the use of the presented system on a problem and with the test results of the system. In the last section the results are discussed and the guidelines for future research indicated.

2. Present situation of cost determination 2.1 Tool manufacturing costs The manufacturing costs are divided into the costs of materials, work, cooperation, design, manufacture, tests, measurements and transport. While some of these costs have the nature of fixed costs incurred in the manufacture of different tools of approximately same size, other costs can occur depending on the size and shape of finished product. Usually, fixed costs are simple to determine, whereas variable costs mostly depend very much on the tool features. Out of the total costs the costs of work represent approximately one half of all costs, therefore, for prediction of total costs it is of utmost importance to predict the required number of man-hours as accurately as possible. The work costs are all cost related to mechanical and manual work.

2.2 From demand to order Usually sheet metal stamping tools are made individually upon the order of costumer. Shortly after order the tool-makers must obtain the answer to the following questions within the shortest possible time: 1. Are we in a position to make the tool for the product concerned?

2. Do we have the means for the manufacture of that tool?

3. How much time do we need to be able to make the tool?

4. How much the tool will cost?

The answers to the first two questions are rather trivial; if the company does not know the answers to these two Articles

51

Journal of Automation, Mobile Robotics & Intelligent Systems

questions it is probably better for it not to undertake the order at all. As the matter of fact the answers to these two questions are the result of cooperation between designers and technologist and depend on the state of skills and resources in the company (Fulder, T. et al., 2001). The answer to the third question is very important particularly in the tool-making activities since adhering to the delivery time is one of most important factors of success. However, it is often difficult to answer that question, since the answer depends on a variety of interconnected factors (Pahole, I. et al., 2003). In the process of agreeing the tool-makers usually do not have the chance to determine the tool manufacturing time since they are specified by the clients.

The tool-making shop must answer like this: “The delivery time can/cannot be met.” The answer to the fourth question, too, is very important, since only if it is precise, on the one hand the preparation of a competitive offer is possible and, on the other hand, undertaking jobs, bringing loss, is avoided.

2.3 Problem of total cost determination The tool manufacture is a complex process including a variety of personnel, machines and technologies. Therefore, specifying the manufacturing costs poses a serious problem. In addition, this activity is very time-limited. The tool manufacturing costs can be rather precisely analytically determined, but analyses require additional time and cause additional costs. The tool-makers can afford none of these. In answer to the demand the offer must be prepared as fast as possible, possibly within a few hours, but not later than in a few days.

Thus, the cost prediction is connected with quite a few difficulties. Most of the difficulties appear due to the dependence between quantity of information and the capability of cost prediction. The principal difficulties can be summarized as follows: o It is hard to obtain a high-quality cost prediction from the design drawings, although in this stage of the product development it is the product target price which is the most important. o For a sufficiently accurate cost price considerably more technological information is required. o Cost prediction is connected with costs. Therefore, the cost prediction is not always economical (Wierda, L. S., 1990). It may happen that the prize of cost prediction is higher than the benefit of it. o The information about costs is dynamical. Due to internal changes in the company and external changes in the economic environment the costs change. o The cost prediction is limited to one company and one product only. A general rule, applicable everywhere cannot be contrived. o Due to characteristics of the problem itself the cost prediction accuracy is often not satisfactory.

If the findings about the problems appearing in determining the total manufacturing costs in tool manufacture are summarized it is found out that two basic problems are in question: • lack of technological information about fulfilment of the order • lack of time for preparation of cost analysis of the orders.

52

Articles

VOLUME 1,

No 4

December 2007

Lack of technological information is caused by the nature of make-to-order production, which is common in tool manufacturing: in the beginning, only the CAD model of the final product and practically no technological information is available and no sooner as on the end, when the tool has been finished and is ready for the manufacture of the final products, all technological information resulting from the tool manufacture process, is available. Therefore in the stage of securing the order not the determination but prediction of total costs of tool manufacture is at issue.

Most frequently, in tool manufacture, the experts with long-standing experience deal with prediction of cost of tool manufacture. It can be claimed that the problem of prediction of the total manufacturing costs has not been satisfactorily solved. Prediction relies too much on subjective influences of the expert. It is evident that the described problem needs a better solution. A system is needed in the offering stage to be able to determine the tool manufacture costs directly from the CAD-model of the finished product fast and without the necessary expert knowledge.

2.4 Cost prediction methods As we mentioned before there has been developed many approximate cost prediction methods. They can be divided (Duverlie, P. and Castelain, J. M., 1999): o Intuitive methods; based exclusively on the expert’s capabilities o Analogue methods; costs are evaluated on the basis of similarity with other products o Parametric methods; costs are evaluated on the basis of the product characteristics which are in the form of parameters (influencing factors) o Analytical methods; costs are evaluated on the basis of the sum of the individual planned costs.

Figure 1: Area of use of cost prediction methods (adopted from Duverlie, P. and Castelain, J. M., 1999).

None of the above methods is appropriate in all stages of the development cycle. They differ in the requirements and area of use. Figure 1 shows the areas of the use of the cost prediction methods per product development cycle.

It can be seen that intuitive methods are useful in early development stages. Such prediction method does not require special preparation and is not demanding with respect to time and cost, but it is unreliable and needs a qualified expert.

Journal of Automation, Mobile Robotics & Intelligent Systems

3. Model of intelligent system for prediction of tool manufacturing costs 3.1 Taking example by intuitive methods The cost prediction methods, enumerated in the previous section, do not regard the type of product. However, all the methods enumerated are not adequate in toolmaking, but only those meeting the specific requirements of the tool-making industry for very fast and precise predictions. In the business environment of the tool-makers only the analogue and parametric methods are applicable and in no way the analytical ones. Although many methods of prediction of the tool manufacturing costs have been developed, the intuitive cost prediction is most frequently used for the reasons stated in the introduction. That task is performed by experts with proper technical-economic knowledge. In this case the experts are experienced individuals having gathered in their work enough knowledge to be able to perform this task. When gathering experience they resorted to acquiring the knowledge by deliberate exercise in order to improve the effect, which according to Ericsson, K. A. and Charness, N. (1994) is the best method of recruiting the experts. However, in spite of optimal learning the future experts must do 10000 hours of deliberate exercise (Charness, N. and Schultetus, R.S. 1999) which in practice, implies 10 years of working experience in this area. Gradually, the expert develops the capability of rather good cost prediction. Such cost prediction is used since it is not demanding with respect to time and cost. However, this approach is today obsolete and the problem requires a better solution.

In all methods developed so far, besides intuitive prediction, difficulties are met, which have not yet been satisfactorily solved. Associations between geometrical information and tool manufacturing costs practically cannot be covered by deterministic methods. Therefore for the determination of dependence between the geometric features and the manufacturing costs the evolutionary methods have been used. By using these methods we have tried to avoid difficulties arising in describing the complex system by deterministic rules. We have conceived an intelligent system using the principle of operation of the analogue and parametric methods. The so-called intelligent system is similar to the natural intelligent system, i.e., expert. Like the expert the system has the memory structured in the form of relation data base. While the expert uses his intelligence for reasoning, the artificial system uses genetic programming method.

3.2 Case-based reasoning The case based reasoning concept has been in use since mid eighties of the past century. It is based on the findings in psychology and adopts one of the solving processes used by experts. Researchers in artificial intelligence have found out that this concept ensures working out of intelligent systems which are useful and non-exacting at a time (Kolodner, J. K. et al., 1985). As a matter of fact, this is one of the most universal manners of problem solving, used frequently by the human in his work. It uses the recognizing way to modelling and explaining the human approach to solving the problems in the areas where experience has a very important role (Strube, G. and Janetzko, D., 1990).

VOLUME 1,

No 4

December 2007

In case-based reasoning it is assumed that interconnections between the descriptions of problems can be found. The knowledge about the area is saved in the form of cases similarly as the knowledge owned by the expert and not in the form of deterministic knowledge. The case is defined as the record of the problem and of its situation.

It can be presented in the form of vector or as a complex composed object (Althoff, K. D. and B. Bartsch-Spörl, 1996). The target case is the description of the problem whose solution is searched for, whereas the source case is the description of the problem with known solution.

3.3 Description of model The model is built on the basis of the improved model of the global cost prediction and the case-based reasoning concept. For preparing the prediction it uses the following steps (Figure 2): o Collecting the geometrical and technological information in the computer data base. o Abstracting the geometrical features from the target case (CAD-model of product). o Selecting the most similar cases (source cases) from the data base. o Working out the formula for cost prediction. o Use of formula – preparing the prediction.

Source cases are necessary for the use of case-based reasoning. Therefore, geometrical and technological information must be collected. It is saved in the data base as logically connected geometrical and technological information about the individual cases. Selection of the source cases, most similar to the observed case, facilitates searching for the dependence and preparation of the formula and ensures higher precision of the prediction. In the next step, the parametric dependence is prepared by system for genetic programming. In the last step the resulting parametric dependence is used like in the case of ordinary parametric method for prediction of costs.

As soon as the system gets a new case, i.e., the problem description in the form of CAD-model, it must translate it into the form which is suitable for artificial system.

We must be aware that by today’s artificial intelligence it is impossible to treat the entire complex product model as perceived by the human. However, even the experts do not have in memory the complete information about the product but only the most important parts and summaries. The system first abstracts the geometrical features from the CAD-model. Most frequently, this means that the system isolates the physical properties, the quantity description of the product and the geometrical features from the CAD-model. The output of abstraction of the CAD-model is a record of the problem in vector form. The individual features are comprised parametrically as components of that vector. In the next step, the similarity of the target case against other cases saved in the data base is calculated. The similarity is calculated as the distance between the final points of vectors in the vector space. The greater that distance the smaller is the similarity between the two products. In the further step, those most similar cases, which are then the input into the reasoning subsystem, are chosen. Those isolated cases are the source cases for reasoning about the solution of the target case.

Articles

53

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 1,

Figure 2. Case based reasoning cycle in predicting total costs.

For reasoning about the solution on the basis of similar cases the reasoning subsystem uses the artificial intelligence method – genetic programming. We decided on genetic programming because of its ability of robust forming of formula. Evolutionary methods are the most general approach to solving such problems. The genetic programming method forms the solution in accordance with evolutionary principles. Here the source case components are the programme terminals. For evaluation of the solution the system needs the value of costs – the solutions of the most similar cases, therefore, in this step it transfers them from the data base. In our case, in the stage of adoption of solution the system uses the approach similar to parametric method. The result is the formula containing parameterized geometrical features of the finished product as variables.

3.4 Abstraction of CAD-model For cost prediction much information, contained in the CAD-model, is excessive. This is the information having no influence on the manufacturing costs or having insignificant influence. It must be isolated not to hinder establishing of similarities and reasoning about solution.

By abstracting the precise numerical description in the form of CAD-model is reduced to the only one vector. That vector is called the case vector: * v p = {g1 , g 2 , g 3 ,...g i ,...g n }

(1)

The vector components from g1 to gn are parametrically comprised individual mostly geometrical features; however, the vector can contain also the known technological features and auxiliary data. Thus g1 can be the thickness of sheet metal, g2 the number of surfaces etc. When selecting the vector components, utmost attention is required, since it is desirable to describe the product with smallest possible number of components, i.e., as adequately as possible. It is in the nature of evolutionary methods that they work faster, if they have fewer terminals. In our case the terminals are the vector components.

54

Articles

No 4

December 2007

3.5 Selection of the most similar source cases Selection of the most similar cases is intended to increase the quality of the formula obtained by the reasoning system. The formula applicable only for similar cases will be much easier to obtain than the universal formula.

Usually, it will contain fewer terminals and will be more precise. When speaking about the most similar cases the cases are meant which are not equally similar all of them but they are ranked on the top of the scale of similar cases.

For forming the formula by genetic programming method more cases are urgently needed.

As the system of cost prediction imitates the natural intelligent system – i.e. the human – the similarity, having the same meaning as in the every day conversation, is introduced. Similarity is calculated on the basis of case vectors. The target case vector is compared with all vectors of source cases. Similarity is defined as the distance between the two final points of vectors. The smaller the distance, the more the two products are similar. vcp designates the target vector and vpi the vector of the source case i. Similarity Pi between the vectors vcp and vpi, or between the abstracted target and source case is equal to absolute value of difference between two vectors: * * Pi = vcp − v pi

(2)

However, the similarity thus calculated is not a good enough criterion of similarity since the vector components have different value extents. Therefore when calculating the similarity P, all components must be normalized. When normalizing the components, the importance of the individual components or geometrical features can be considered. Therefore each component is multiplied by the normalization multiplier dj, which can increase or decrease the influence of the individual component on the value of similarity. gcj designates geometric feature c of case j. Multiplier dj is: 1 ­ ½ ; g cj ≠ 0° °r j ⋅ d j = ® g cj ¾ ° 0; g cj = 0 ° ¿ ¯

(3)

The multiplier of influence of component r j can assume the values on the interval from 0 to 1.

Similarity between the vectors of products is equal to: Pin =

(

1 n ⋅ ¦ d j ⋅ g cj − g pij n j =1

)

2

(4)

The similarity determined in this way has a value between 0 and 1. Here, lower value of Pin means greater similarity.

The number of selected cases depends on the number of similar cases. It’s not adequate to select a case which is not similar at all and on other hand it’s not adequate to make a prediction on the basis of a small number of cases.

Therefore the first condition to make a good prediction is to have enough similar cases.

3.6 Reasoning with genetic programming In the reasoning part of our system the genetic programming methods is used. In this environment this method of evolutionary computation proves to be excellent. Together with preparation of input data on the basis

Journal of Automation, Mobile Robotics & Intelligent Systems

of determination of similarity this method has proved to be efficient.

The idea of evolutionary computation was presented in 1974 by Rechenberg in his work Evolutionary strategies.

His work was then pursued by other researchers. Thus in 1975 John Holland developed genetic algorithms, and some 15 years later John Koza J. R. (1992) developed still the genetic programming. In these methods the evolution is used as an optimization process in which the organisms become increasingly adapted to environment in which they live (Kovacic, ˇ J., 2003). Two main ˇ ˇ M. and Balic, characteristics of evolutionary methods are: they do not search for the solution in the ways determined in advance (deterministic) and they simultaneously treat a variety of simple objects (Brezocnik, M. et al., 2003). Structural solution is left to the evolutionary process. Because of the probabilistic nature of the evolutionary computation methods there is no guarantee that each evolution arrives at a satisfactory result.

In any evolutionary method we have to do with structures subject to adaptation. In conventional genetic algorithms and genetic programming a population of points is subject to adaptation in search space. In genetic programming hierarchically structured computer programmes are subject to adaptation (Koza, J. R., 1992). The set of possible solutions in genetic programming is the set of all possible combinations of functions which can be composed in recursive way from the set of functions and from the set of terminals.

Solving of the problem starts with creation of a random population of solutions. In our case the solution is the formula for calculation of costs. This initial collection of problem solutions, which is usually created at random, is left to evolution. Each individual organism represents solution of the problem. Then the organisms are evaluated and greater probability of taking part in operations of selection and changes is assigned to those organisms which better solve a certain problem. By genetic operations of crossover, mutation and reproduction better and better solutions are then gradually approached from generation to generation. Reproduction is the basic way of continuation of a species of living organisms. Mutation is the component of evolution bringing novelties. Competition and selection are two processes always repeating where several organisms have available limited quantities of resources. Selection assures the survival of more successful members of the population and their passage in unchanged form into the next population. Changes influence one or several organisms and create the descendants from them. Selection results in a new generation which, again, is evaluated. The procedure is repeated until the establishment criterion of the process has been fulfilled.

This can be the greatest prescribed number of generations or the sufficient quality of solutions.

Because of the nature of genetic programming, preparation of a high-quality formula requires a high number of vectors of source cases, which actually means much source cases. In practice this condition is hard to meet.

Only rarely a great number of very similar cases are available. Further, the case vector contains too many components. Many components mean many variables in formula and, of course, many terminals in the tree-like structure

VOLUME 1,

No 4

December 2007

of the organisms. Together with the number of terminals also the computation exactingness increases. Preparation of the formula by genetic programming, containing many terminals and operators, is not rational with the computation power available today.

For these reasons the number of components of the case vectors must be reduced. Another abstraction is effected, but now the case vectors are abstracted in order to reduce the number of components. For reducing the number of components the following approaches are used: o Components only slightly influencing the costs are isolated o Doubled components, i.e., components containing identical information are isolated o Computation operations between two or more components are carried out by uniting the information into one component.

Vectors of case vip, having the extent of size n are transformed into converted vectors of m scope where m<n applies:

* * vipk → v' ipk

(5)

v’ipk is the transformed source vector of case k.

For the reasoning subsystem for the genetic programming method the input data are prepared in the form of a list of converted source vectors of cases with added values of costs and/or solutions. Now the input data for the reasoning subsystem have been prepared, the latter has yet to be set.

For the reasoning subsystem the application for determining multi-parametric function on the basis of known cases, written in programme language AutoLISP, has been used.

Procedure of solving by genetic programming is presented in the following steps: o Determination of set of terminals; terminals are the components of the transformed case vectors and the real numbers created at random. o Determination of set of basic functions; these are in particular the basic mathematical functions. o Insertion of cases for calculation of adaptation; the cases are the lists of converted source case vectors o Determination of parameters of evolution; the evolution parameters are the number of organisms in the population, the maximum depth in crossover, the maximum depth in creation o Determination of criterion for stopping the evolution; for stopping of evolution the number of evaluated evolutions has been selected.

The output of reasoning subsystem is the functional dependence between components of the converted vector of the target case and costs. tc designates the solution of the target case, i.e. the solution of problem:

(

* tc = f (vc ) = f g c1 , g c 2 , g c 3 ,...g cj ,...g cm

)

(6)

After having the formula in hand, the components of the transformed target vector are entered and thus the costs are calculated. It must be emphasized that this functional dependence applies only to this target case, thus the function obtained is usable only once.

3.7 Automation of predicting In order to make global prediction of total manufacturing costs efficiently the method must be automated to Articles

55

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 1,

certain extent (Wierda, L. S., 1990), otherwise the method cannot produce up-to-date results due to changing environment. If the method is not automated, there are no reasons to use it, although it is time and cost non-demanding.

Automation must cover the acquisition and storing of data. The geometrical and technological data must be stored in suitably structured form. It is proper to use relation and object computer data bases. Also the module for finding the dependence between geometrical and technological features must be automated for efficient work.

Consequently, such automated system must contain at least the data base for storing geometrical and technological data and the module for finding the dependence between geometrical and technological features.

If we have to do with the manufacture of a great number of different products also the modules for finding similarities between the products and the modules for selecting the most influencing geometrical features – and/or CAD model abstracting are desirable.

No 4

December 2007

Afterwards the similarity between the target and the source cases is calculated. Similarities were calculated on the basis of normalized vectors of cases. We selected five most similar cases as shown in Figure 3.

4. Example and results The input information into our model is the CAD-model of the final product, which is also the target case. The other input information into the model are the cases of tools already made. These CAD-models with costs are the source cases.

From CAD model it is necessary first to abstract the data on the basis of which the case vectors will be determined and similarity between the source cases and the target case calculated. For the first test we used the most general approach and from the target and source cases we abstracted the most recognizable geometrical features and not the features the features most influencing the variable searched for i. e. the total manufacturing costs.

From CAD-models the following geometrical features have been identified: o Number of geometrical features made by cutting (secondary features) – R o Number of geometrical features made by bending (secondary features) – U o Extent of bends – SU o Number of faces of CAD-model – F o Thickness of main geometrical feature – D o Surface area of main geometrical feature – P o Volume of main geometrical feature – V o Total outside length of cutting of main geometrical feature – LRZ o Total inside length of cutting main geometrical feature – LRN o Total length of bending lines – LU o Number of triangles in STL-format – T o Greatest distance between two points of CAD-model – DI o Greatest distance between two points of CAD-model in direction of largest plane of CAD-model – DH o Greatest distance between two points of CAD-model in direction rectangular to largest plane of CADmodel – DV o Ratio between DV and DH – K. After abstracting the features of all source cases and target case the case vectors for each case were obtained: * vi = {Ri ,U i , SU i , Fi ,Vi , Pi , Di , LRZi , LRN i , LU i , Ti , DI i , DH i , DVi , K i }

56

Articles

(7)

Figure 3. Source cases used in reasoning process by genetic programming.

Due to limitations of number of source cases and especially limitations of computation power we decreased number of vector components. In the next step we transformed the cases in such a way that we obtained the vectors of source cases which are of the form suitable for reasoning by means of genetic programming. The following transformation of the case vectors was effected: * * v pi → v' pi ½° ­° R pi ,U pi , SU pi ,VI pi , F pi ,V pi , Ppi , * v pi = ® ¾ °¯ D pi , LRZ pi , LRN pi , LU pi ,T pi , DI pi , H pi ,V pi °¿ * v' pi = SU pi , Ppi , DI pi , K pi ,t pi

{

(8)

}

The following actions were taken: o components SU, P, DI were transferred o components R, U, VI, F, V, D, LRZ, LRN, LU, T, H, V were removed o components DH in DV were transformed into K K=

DV DH

(9)

Transformed source vectors and target vector, presented in Table 1, were obtained.

For reasoning the genetic programming system proˇ ˇ (M. Kovacic, ˇ ˇ 2003), was used. After posed by Miha Kovacic transformation of the selected vectors the genetic programming system was prepared in following steps: o Determination of set of terminals; in our case the terminals are SU, P, DI and K. o Determination of set of primitive functions; the basic mathematical operations, i.e., addition, subtraction, multiplication and division were selected. o Insertion of cases for calculation of adaptation; in the form of a list of transformed vectors. o Determination of evolution parameters: — Number of organisms in population is 2000 — Maximum depth in creation is 8

Journal of Automation, Mobile Robotics & Intelligent Systems

— Maximum depth in crossover is 15 — Probability of crossover on cells and organs is 0.7 — Probability of crossover on organs is 0.2. o Determination of criterion for stopping of evolution; the greatest number of generations is 200.

After inserting of data and adjusting of all parameters of genetic programming the evolution was run a few times. The results obtained were in the form of functional dependence between the components of the transformed vector of product and costs: (10) t = f (SU , P , DI , K ) s

Where ts are the tool manufacturing costs.

Table 1. Example of data to be entered into reasoning system.

SU

P

DI

K

total manufacturing costs

t arget case

2,0

2259,0

59,1

0,3

searched

4,0

2909,0

59,0

1,1

305,0

2,0

4235,0

67,6

0,4

298,2

2,0

4632,0

82,1

1,2

324,3

1,0

2461,0

59,2

0,2

300,0

2,0

2155,0

47,4

0,3

365,3

source cases

case

We several times ran the evolution and each time the genetic programming system worked out a formula. The formulas were more or less complicated; the comparison of quality of prediction of our system and expert is shown in Table 2.

VOLUME 1,

No 4

December 2007

been made by an artificial system not having the capacity of intuition.

After simplification the worked out formulas are casually very complicated. Entering the parameters and calculation of the variable searched for take place in an automated manner in the computer system, so that the complication of the formula does not make predicting difficulty and/or the formula need not be simplified at all.

5. Conclusion This paper presents quite uncommon approach to cost prediction. We have decided on building intelligent system due to awareness that the problems treated cannot be adequately solved by deterministic approaches. Testing of the system has brought interesting insights and many future challenges. Already the hitherto results show that they are of good quality compared to those made by expert. It is hard to expect the system will make very precise predictions, since even experienced human experts cannot do that. It must be borne in mind that the tool manufacturing takes place in changing environment where also rules of chaos apply. The objective of our model is not to surpass the expert but to support him and maybe to replace him in the future. It can be established that the system is capable to work out a good prediction.

Our further research will be oriented towards making a system capable to abstract and to convert intelligently the data into a form suitable for processing by genetic programming. It is expected that with the increase of the computer power also the capacity and usability of the system will increase. In the future the system can be adapted for predicting the manufacturing costs of other types of forming tools.

Table 2. Comparison of error of system and expert.

Prediction

Error [%]

Expert

3,60

System (Run 1)

3,63

System (Run 2)

1,59

System (Run 3)

9,89

System (Run 4)

4,45

System (Run 5)

0,52

System (Run 6)

7,42

System (Run 7)

1,17

System (Run 8)

2,33

System (Run 9)

6,36

System (Run 10)

10,56

In Table 2 it can be seen that the quality of prediction of the expert and of our system are somehow comparable. The average error committed by our system is 4.79 %.

Although the error is higher than that of the expert, the results can be considered to be satisfactory. Especially, if it is borne in mind the analysis of the influence of vector components on total manufacturing costs was not made.

Experience shows that a qualified expert commits up to 15 % of error. From this point of view the predictions can be considered as good taking into account that they have

AUTHORS ˇ Jože Balicˇ – Laboratory Mirko Ficko*, Boštjan Vaupotic, for Intelligent Manufacturing Systems, Faculty of Mechanical Engineering, University of Maribor; Smetanova ulica 17, SI-2000 Maribor, Slovenia. Tel.:+386 2 220 1595.

Fax.:+386 2 220 7990.

E-mail: mirko.ficko@uni-mb.si * Corresponding author

REFERENCES [1] K. D. Althoff, B. Bartsch-Spörl, “Decision Support for Case-Based Applications”, Wirtschaftsinformatik, no. 38, 1996, pp. 8-16.

[2] M. Brezocnik, J. Balic, Z. Brezocnik, “Emergence of intelligence in next-generation manufacturing systems”, Robotics and Computer Integrated Manufacturing, vol. 19, no. 1-2, 2003, pp 55-63.

[3] N. Charness, R.S. Schultetus, Knowledge and Expertise, Handbook of Applied Cognition, John Wiley & Sons: New York, 1999.

[4] Y. Ding, H. Lan, J. Hong, V. Wu, “An integrated manufacturing system for rapid tooling based on rapid prototyping”, Robotics and Computer-Integrated Manufacturing, vol. 20, 2004, pp. 281-288.

[5] P. Duverlie, J. M. Castelain, Cost Estimation During Design Step, “Parametric Method versus Articles

57

Journal of Automation, Mobile Robotics & Intelligent Systems

[6]

[7]

[8]

[9]

[10]

[11]

[12]

[13]

[14]

[15]

58

Articles

Case Based Reasoning Method”, International Journal of Advanced Manufacturing Technology, vol. 15, no. 12, 1999, pp. 895-906.

K. A. Ericsson, N. Charness, “Expert performance: its structure and acquisition”, American Psychologist, 1994, vol. 49, pp. 725–47.

T. Fulder, A. Polajnar, K. Pandža, “Sposobnost – orodjarna. Ideje – motor gospodarske rasti orodjarn”. In: Proceedings of Orodjarstvo, 2001, Portorož, Slovenija.

J. Holland, Adaptation in Natural and Artificial Systems, University of Michigan Press, Ann Arbor, 1975.

J. K. Kolodner, R. Simpson, K. Sycara, “A Proces Model of Case-Based Reasoning”. In: Problem Solving, Proceedings IJCAI-85, MorganKaufmann, Los Angeles, 1985, pp. 284–290. ˇ “Evolutionary programming ˇ ˇ J. Balic, M. Kovacic, of a CNC cutting machine”, International journal of advanced manufacturing technology, vol. 22, no. 1-2, 2003, pp. 118-124.

J. R. Koza, Genetic Programming: On the Programming of Computers by Means of Natural Selection, MIT Press, Cambridge, 1992. ˇ “OptiI. Pahole, M. Ficko, I. Drstvenšek, J. Balic, misation of the flow of orders in a flexible manufacturing systems”, Strojniški vestnik, vol. 49, no. 10, 2003, pp 499-508.

I. Rechenberg, Evolutionsstrategie: Optimierung technischer Systeme nach Prinzipien der biologischen Evolution, Frommann-Holzboog Verlag, Stuttgart, 1974.

G. Strube, D. Janetzko, „Episodisches Wissen und fallbasiertes Schließen: Aufgaben für die Wissendiagnostik und die Wissenpsychologie“, Sweicherische Zeitschrift für die Psychologie, vol.

49, no. 4, 1990, pp. 211–221.

L. S. Wierda, Cost information tools for designers: a survey of problems and possibilities with an emphasis on mass produced sheet metal parts, Delft University Press, Delft, 1990.

VOLUME 1,

No 4

December 2007

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 1,

No 4

December 2007

A Specialized Multi-agent Search Engine Model for the Extended Manufacturing Enterprise Ockmer L. Oosthuizen, Elizabeth M. Ehlers

Abstract: The Internet and Internet search engines have revolutionized the way in which we search for, and integrate information into our daily lives. One of the problems with general Internet search engines is the diversity of contexts queries can assume. Therefore, the construction of context-restricted search engines designed for use within a specialized domain has been proposed (S. Lawrence, 2000).

The objective of this article is to suggest a model for a collaborative, personalized meta-search agent system for the virtual manufacturing enterprise in the context of resource location and integration across scope of the extended manufacturing enterprise.

The key differences between the proposed model and general search engines are (1) the ability to personalize the search task according to an individual user need, (2) the utilization of similar search sessions through partner collaboration at an information retrieval level, (3) leveraging on existing search services available on the virtual enterprise extranet and (4) autonomous behaviour through the use of intelligent agents.

The perceived benefits of a such search agent system for the virtual enterprise are: (1) Enabling a consistent and uniform view of business entities through improved access to information, (2) Simplified access to operational data, (3) Effective access and retrieval of organizational model elements and (4) Improving data location and sharing between trading partners providing a platform for tighter process integration between trading partners. These benefits are critical for effective decision making in a collaborative engineering process followed by the differing entities in a virtual manufacturing organization to ultimately enable final product realization.

An evaluation strategy for the COPEMSA system presented is then briefly discussed, metrics for evaluation of focused crawlers and justification for development and implementation of the search system proposed. Finally, this article concludes with remarks about the model presented and future research to be undertaken.

Keywords: virtual manufacturing enterprises, agent based systems, personalised web search engines

1. Introduction The growth of the Internet over recent years has not only enabled easy access to volumes of publicly accessible information globally, but has also substantially contributed to the arsenal of useful, effective and most importantly open technologies available to developers.

One of the most useful applications deployed on the Internet and, more specifically, the World Wide Web

(WWW) are so called web search engines. The search engine’s primary goal is to effectively locate and retrieve information that is the closest match to a specific user’s input query. Advances in Web searching technology have made Web search engines an indispensable tool for the effective use of the WWW. This article investigates the viability and usefulness of the application of Web search technology to virtual manufacturing enterprises. The ultimate goal of such a proposed search engine is the support of process integration of content between different organizations in the same enterprise and that of trading partners (S. Khoshafian, 2002).

The existence of a collaborative search infrastructure could greatly assist in the integration of existing resources in the extended manufacturing enterprise to improve operational activities and service levels by ensuring that information is easily accessible and readily available to all entities in the organization. The ultimate goal of the organization-wide use of such an infrastructure is the support of effective decision making in the (possibly distributed) engineering process for product or service realization across the virtual manufacturing enterprise.

In the following sections, a brief introduction to the extended manufacturing enterprise and specialized search engines is given. The collaborative, personalized meta-search agent (COPEMSA) system architecture is then introduced.

The COPEMSA search system is realized through the use of a multi-agent architecture. Each agent in the search system represents a specialization of the overall search task. We define a user agent, search agent, resource crawler, partner agent and results analysis agent.

The user agent learns about the user in order to introduce personal preference into user queries. The search agent attempts to match these personalized queries it receives from the user agent to multiple resources identified by the Resource Crawler. The Resource Crawler continually navigates the extranet and indexes resource locations and descriptions. The community agent enables the search system to communicate and leverage on the search experiences of other members of the extended manufacturing enterprise. Finally, the results analysis agent organizes and ranks the returned results according to user and partner profiles.

A section on evaluation strategy and evaluation metrics is then presented. The goal of this section is to elaborate on the strategy to be followed to test the proposed search system and also the possible performance metrics that could be used for testing and evaluation purposes.

Finally, the article closes with conclusions and further research opportunities.

Articles

59

Journal of Automation, Mobile Robotics & Intelligent Systems

2. The Extended Manufacturing Enterprise L. Song and R. Nagi (1997) define a virtual enterprise (VE) as an organization constructed by partners from different companies who collaborate with each other to design and manufacture high quality, customized products.

The authors also note that VE’s are usually product-orientated; team-collaboration styled and are fast and flexible to respond to changing market conditions. L.M. Camarinha-Matos (1999) define the virtual enterprise paradigm as a temporary alliance of enterprises that collaboratively share skills and resources to competitively respond to business opportunities.

Both the definitions given above, although slightly different, define the following three generic characteristics: • The virtual enterprise is composed of different entities allied to achieve some goal.

• The virtual enterprise relies on collaboration and communication of information and skills between these entities to achieve its goals.

• The virtual enterprise is highly flexible and adaptive to remain competitive in changing market conditions.

The first characteristic is related to the scope of the virtual enterprise. The question that must be asked here is if these entities are merely different units in the same organization or totally different and independent companies. The second characteristic defines some sort of collaboration between differing entities to achieve goals.

The main question of interest here is at what level this collaboration and co-operation is done. Finally, the third characteristic requires that virtual enterprises are adaptable and flexible to changing business needs. The question here is what business model and interactions are needed to achieve this flexibility and adaptability.

These three characteristics and the questions associated with them are briefly discussed in the following subsections.

2.1. The scope and core features of virtual enterprises S. Khoshafian (2002) summarizes the scope of virtual enterprises into the following three categories: (1) Process integration between operational units in the same organization. (2) Process integration between organizations in the same enterprise. (3) Process integration between trading partners.

One of the characterizing features of the virtual enterprise is the collaboration between entities to achieve goals and to react to changing conditions. It is obvious that this collaboration between entities must be facilitated on various levels throughout the virtual enterprise.

S. Khoshafian (2002) defines four aspects or levels of particular importance to virtual enterprises, namely: (1) A consistent view of business entities. (2) Consistent operations on business entities. (3) Uniform organizational model. (4) Processes with consistent roles and activities VE wide.

2.2. The virtual enterpise business model As have been previously mentioned, flexibility and adaptability are two enabling features of the virtual enterprise to respond to changing market conditions. These features are, in part, achieved through the inherent net60

Articles

No 4

VOLUME 1,

December 2007

work-structure of the virtual enterprise to enable flexibility and adaptability.

As illustrated in Figure 1, trading partners (TPs) provide different value and support products and/or services to the value chain of the virtual enterprise. This effort ultimately contributes to the realization of a real product/ service to the end consumer.

As all resources of trading partners are not necessarily centrally located, it is natural to assume that these trading partners are linked by some form of electronic communications network to enable information exchange and inter-organizational communication (Rautenstrauch, T., 2002). This extended information network between trading partners is commonly referred to as an extranet.

Upstream Supplier

Service/Product

TP

Upstream Supplier

Service/Product

TP

Marketing TP

R&D

Upstream Supplier

Service/Product

TP

Finance

Logistics

IT

TP

TP

Vision, Strategy Network Management

Real Product /Service

End Consumer

Sales & Distribution

Production Procurement

TP

TP

Figure 1. Virtual Enterprise Organizational Model [Adapted from (Rautenstrauch, T., 2002)].

2.3. Information exchange between entities in the virtual enterprise The final consideration of interest in the context of this section is the type of information exchange between entities in a virtual enterprise.

Virtual enterprises can exchange a variety of information including expertise, product models, business process information, quality-related information, commercial information etc. The only requirement is that some common reference model exists so that all the entities involved in the exchange can participate in the exchange in a pre-defined consistent manner.

Standards like the standard for the exchange of product model data (STEP) and the standard for electronic data interchange for administration, commerce and transport (EDIFACT) are both examples of common reference models for product and commercial information (ISO103031:1994, 1994) and ( ISO9735-3:2002, 2002) .

The unifying idea between all these standards is to provide a common format for representation and exchange of operational data between collaborating entities.

3. Specialized search engines The ability to search large volumes of information in a relatively short period of time has been one of the greatest enabling factors of the widespread use of the World

Journal of Automation, Mobile Robotics & Intelligent Systems

Wide Web in recent years. In this section we briefly discuss the basic ideas behind general web search engines and introduce the concept of a meta-search engine. We then turn our attention to a critical evaluation of current search services and why there is a need for specialization and personalization of the search experience as well as the benefits of context and search space restrictions for search systems. The section closes with a discussion on the benefits of such a context restricted specialized search system for the virtual manufacturing enterprise.

3.1. General web search engines The Web can be seen as a large collection of highly-topical, geographically dispersed, interlinked information. To enable the effective location and retrieval of information from this very large corpus of information, so called Web search engines were developed. The search process can be summarized in five steps (A. Broder, 2002): • The search engine user has a certain task he/she wishes to complete and identifies an associated information need.

• The information needed is expressed in terms of a query to be submitted to the search engine.

• The search engine interprets the query.

• It selects those documents that match the query from a corpus of web documents/content.

• Finally, the search engine displays the results of the query to the requesting user (Further refinement of the query based on the results is, of course, also possible at this point).

As a highly successful and effective search engine Google® and (more specifically) Google’s desktop search deserves special mention in the context of this paper.

The desktop search product is intended as a local search service for the desktop computer with (typically) a single user. This is achieved through building an index of email, files and web history stored on the user’s machine and updating it as new information becomes available (Google, 2007). As correctly noted by Chirita et al. the desktop search product includes no metadata whatsoever in their system, but just a regular text-based index (Chirita et al., 2005). The collection of metadata on user search sessions and augmenting results based on user context derived from the collected metadata could lead to great improvements in the search experience and perceived usefulness of the search system.

One of the main problems facing many search engines is the issue of equal coverage of the corpus of documents.

A related problem from the user perspective is the complexity associated with the use and management of multiple, usually differing search systems. The following section introduces the concept of meta-searching as a technique to address these problems.

3.2. Meta-search engines Web search engines typically do not offer equal coverage of the World Wide Web (S. Lawrence and C.L. Giles, 1999). In addition to this there are a multitude of search engines available, each offering a different strategy and a different interface for searching the web. To receive the broadest coverage and get more relevant results for a given query, users could use multiple search services.

No 4

VOLUME 1,

December 2007

This can be problematic however, as the overhead involved in collating and merging multiple results can become quite a monolithic task for the user. Moreover, if a large amount of search engines are available, it can also be difficult for a user to keep track of which search services are better for which queries (Glover E. et al., 1999), (S. Lawrence and C.L. Giles, 1998), (E. Selberg and O. Etzioni, 1997).

In order to leverage the coverage of multiple search engines and to provide a single interface to a host of search engines, the concept of a meta-search engine was born. A meta-search engine is a computer program that uses multiple search engines (usually in parallel) to process user queries. The process is briefly summarized in Figure 2 below.

The obvious benefits are that the meta-search engine’s coverage is improved through the use of different search engines and that the user is presented with a single interface through which he/she can search the web. Another, less obvious, benefit is that a meta-search engine can be programmed to “know” a lot more about a search engine than a casual human searcher. Many search engines have special features and optimizations that meta-search engines could exploit and take advantage of, thereby improving the results returned from the various search engines (Glover E. et al., 1999), (S. Lawrence and C.L. Giles, 1998), (E. Selberg and O. Etzioni, 1997).

Some search engines like ProFusion (Gauch et al., 1996) try to select the most appropriate search engine(s) for a user query using a learning approach. In essence, the approach keeps track of how well search engines perform for submitted queries of a given context. For each category, n training queries are submitted to the known number of search engines and each engine’s performance ranked by some function, indicating how relevant the top x returned documents were to the query submitted. The average scores of the n training queries is calculated to give an overall score for a given engine in a given context. This approach “scores” search engines based on the perceived relevance of documents retrieved for a specific category. These scores are then used to select the most appropriate engine for a query in a given category.

Search Services Search Engine 1 Engine specific queries

User query

User

Meta-search engine

Collated and re-ranked results

Search Engine 2 Search Engine 3 Search Engine 4

Distributed Corpus

. . .

Multiple results for user query

Figure 2. The Meta-Searching Process [Adapted from (Glover E. et al., 1999), (S. Lawrence and C.L. Giles, 1998), (E. Selberg and O. Etzioni, 1997)].

Articles

61

Journal of Automation, Mobile Robotics & Intelligent Systems

3.3. The need for personalisation and specialisation Search services are typically used by users with differing goals and informational needs. Currently, the most popular method for the representation of user queries is with a textual string containing keywords. One of the core challenges general search engines are faced with is to determine the context in which keywords contained in a user query was meant. Unfortunately, because of the usually diverse user base and the high volume of requests usually received by such a search engine, personalization is extremely difficult or impossible. The goal of the personalization of search is the construction of a search engine that “knows” their users in terms of the context of their queries, previous requests made to the search engine as well as personal interests and searching habits of the specific user.

Making search engines “context-sensitive” can be achieved in two primary ways: • Personalization through user modelling.

• Specialization of the search engine to a specific domain or task thereby limiting the context and scope of potential user queries.

These two differing approaches are very briefly discussed below.

3.3.1. PersonaliZation through User Modelling The idea is to supply the search engine with information pertaining to the user in terms of the user’s search habits, frequently used keywords and reaction to presented results. Users can be modelled by either requiring them to complete a user profile before he/she commences to use the search engine or the system can dynamically build a user profile through observation of the user’s interaction with the search engine.

3.3.2. Search Engine SpecialiZation Context sensitivity can also be achieved by restricting the search engine’s scope, effectively focusing specialization of the engine to a particular domain or context.

This has the benefit of simplifying the interpretation and processing of keyword-based user queries as the number of contexts a specific keyword could have been meant in is automatically limited by the restricted domain.

Another benefit of specialization is a potentially smaller corpus that the search engine has to cover. This could enable the engine to do a more comprehensive analysis on the corpus documents without a significant time trade off.

Furthermore, as noted by Poblete and Baeza-Yates (2006), the application of text-mining and link analysis techniques to the analysis of the content distribution and link structure of a website could lead to improvements in relevant content provision as well as interconnectivity between similar contents. The application of models for improving the distribution of content and link structure like the one suggested would benefit search engines by simplifying content location and restricting the number of links to be followed for retrieval.

3.4. Specialized search for the virtual manufacturing enterprise In the context of the virtual manufacturing enterprise, specialized and personalized search engine technology 62

Articles

VOLUME 1,

No 4

December 2007

could improve communication, collaboration and interoperation between entities in the enterprise.

As was noted in the previous sections, one of the enabling requirements for the virtual enterprise is the existence of a digital computer network between entities. This implies that users located at a certain entity have access to information and services located on the networks of other entities (i.e. departments, organizations or trading partners).

It is also not uncommon for companies deploy applications and content on their intranets using Web technologies. As an example, consider a manufacturing enterprise that publishes and maintains a comprehensive parts list and ordering system on their Intranet. This parts inventory and ordering system should then also be available to any trading partners the enterprise might have through the network link that exists between the two organizations.

The value proposition posed by search engine technology is the ability to locate and retrieve information about products/services/processes published on the organizational extranet by partner entities. Additionally, the improved access to information helps stimulate trust and improve communication between entities in the virtual enterprise, both which are critical components for effective decision making in a collaborative engineering process followed by different entities for final product realization.

The ability to easily search and retrieve information on partner extranets helps to enable the four core virtual enterprise features (see Section 2.2).

• Search capability enables easy access to information held by partners, thereby promoting a consistent view of business entities.

• Searching exposed operational databases on a partner’s intranet can help ensure consistent operations on business entities and provide a convenient method for extracting audit trails.

• The search system can help promote a uniform organizational model by enabling access to documents pertaining to the corporate structure of a trading partner.

• Finally, a search system deployed on the virtual enterprise extranet can help ensure that processes are understood to have consistent roles and activities by partner entities by granting access to policy and procedural documentation stored on the trading partner’s intranet.

To support the abovementioned, the search system can be specialized according to the user’s organizational role, thereby automatically limiting the context of submitted queries. It would also be advantageous if the search system were to operate autonomously, freeing up the user’s time to pursue other activities.

In the following section, we present a search system model to achieve context-sensitive query formulation and refinement as well as autonomous operation.

4. THE Collaborative, PersonaliZed META-SEARCH AGENT (COPEMSA) SYSTEM ARCHITECTURE The remainder of this article focuses on the development of a collaborative personalized meta-search agent

Journal of Automation, Mobile Robotics & Intelligent Systems

system for the virtual manufacturing enterprise. In section 4.1 the design goals of the system are outlined as well as the benefits and limitations of the chosen approach.

Section 4.2 presents the COPEMSA system model and outlines key technical details of the model.

4.1. COPEMSA System Design Goals The potentially vast amount of information available on the organizational extranet could make it one of the largest sources of information available to the virtual organization. Unfortunately, because of the part structured, part semi-structured nature of the information on the extranet; it can be a difficult and extremely time consuming task to locate and retrieve relevant and useful information. As more information and services become available on extranets between partners, it is of paramount importance that methods are developed to aid in the easy recovery of this information.

As was mentioned in section 3.3, users typically differ in their information needs and have differing interests and characteristics. Personalization of partner search could lead to more accurate query formulation and thereby increase the relevance of returned results. The idea of a community of people, perhaps located at different organizations in the virtual enterprise, with similar interests could also aid in improving search results and exchange of search experiences.

One of the key issues to be addressed by any personalized information recovery system for the virtual organization is the issue of coverage and is discussed in the following subsection.

4.1.1. Issue of coverage Users would obviously want the maximum coverage of the extended enterprise possible when their search queries are processed. The only currently feasible method for achieving this is to attempt to leverage on search services already available on partner organizations extranets, thus leveraging search coverage. Meta-searching is therefore also a key concept in the design of the system.

Intelligent agents act autonomously on behalf of their users. This is desirable for the search system because of potential time savings. A multi-agent system could autonomously continue searching and evaluating extranet content, thereby automating the information retrieval task to a certain extent. Web content mining techniques could also be successfully applied to analyze retrieved pages from the partner’s extranet for potentially useful information. Web mining techniques could also be applied to the results retrieved from meta-searching the extranet.

Users could then focus more of their time on evaluation of retrieved results instead of exhaustively searching for information.

Based on the discussion above, the critical evaluation of current search services and the need for specialization and personalization given in section 3, the scope of development for the collaborative personalized metasearch agent system can be summarized in the following 5 points: • Personalization of search through user modelling.

• Collaborative filtering of search engine results and partner-sensitive searching.

VOLUME 1,

No 4

December 2007

• Meta-searching the extranet for improved coverage.

• Using multiple intelligent agents for autonomous and continuous extranet searching and result evaluation.

• Use of Web content mining techniques for analysis of multiple search engine results and content published on the organizational extranet and/or the World Wide Web. A further design issue is the notion of where the search mechanism should be located. There are two possible approaches to this, either on the client side or server side and is discussed in section 4.1.2.

4.1.2. Issue of Location of Search Mechanism Conventional search engines typically follow the server side model, i.e. user queries are processed remotely by a server. Large scale personalization of a server based search service could potentially be extremely expensive in terms of processing. Keeping track of a large user base’s previous queries, selected documents and other personalization information could be an extremely difficult and processing intensive task, slowing down search engine query processing speed as a side-effect.

A client-based approach could therefore be more feasible for a personalized search agent system. This is discussed in more detail below.

There are a number of significant benefits in a clientside approach to personalized searching. Such a system could more effectively monitor the behaviour of a user, thereby building a more reliable user profile. A clientbased system could then effectively modify user queries to help retrieve documents that are relevant to a given context. Another benefit, in the context of a collaborative multi-agent system, is that the processing of retrieved content could be distributed across multiple system users in the same organization or partners, thereby reducing the load that would normally be on a server or group of servers in a server based approach. This information could then be shared among members of the virtual enterprise in an attempt to minimize reprocessing of results.

One of the limitations of a client-side approach is that such a service would not have local access to a large scale index of the content available on the entire extranet, thereby limiting their functionality (S. Lawrence, 2000).

As mentioned before, the use of meta-search techniques could improve this limitation significantly.

4.2. The collaborative personalized meta-search agent (copemsa) system model In this section, a model is given for the COPEMSA system. The system’s key components are then briefly discussed.

The model given in Figure 3 is a client-based multiagent system for meta-searching the extended manufacturing enterprise. The system also provides for collaboration between multiple users.

The system consists of five key components: role agent, query agent, partner agent, result analysis agent and a directed extranet crawler.

The above core components are discussed in more detail below.

Articles

63

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 1,

Collaborative personalized meta-search agent system

User

Role Agent

Role Profile

Content mining unit

Result analysis agent

Trading Partner 1

Query agent

Trading Partner 2

Partner profile

Parts Inventory

Trading Partner 3

Partner agent

Directed Extranet Crawler

December 2007

Partner search services

Collaborative Meta-Search Unit

Topics-links database

No 4

Trading Partner 4

Manufacturing Resources

. . .

Operational Data Organisational Extranet

Figure 3. The COPEMSA System Model [Adapted from (O. L. Oosthuizen, 2004)].

4.2.1. Role agent The role agent is responsible for information collection and modelling of its user(s). It stores information related to the individual user’s role in the organization in a role profile database. It is important to note that this database can contain the role profiles for multiple roles, either for the same user or multiple users as a search system like this could potentially be used by multiple users of a given organization.

The role agent receives search query strings from the user and then uses the user profile database to rewrite the query to include context specific terms. The modified query is then passed on to a query agent for further processing.

The profile agent achieves this by maintaining a keyword hierarchy for each profile as illustrated in Figure 4.

The structure of the keyword hierarchy tree is based on that of the Open Directory Project (http://rdf.dmoz.org/).

The role agent monitors the queries submitted to the search system as well as the user/role’s reaction to the returned results. The agent then modifies the keyword hierarchy based on its observations. The internal nodes represent topics with further contextual specialization as the tree is traversed downwards. The leaf nodes represent specific keywords associated with a specific context. Additional information stored at leaf nodes about keywords are their usage frequency and their perceived relevance to the specific role being profiled.

ROOT

AR TS

NODE_NM: Watercolor NODE_ID: 0001 USAGE_NUM: 223 RELAVANCE: Range -1 to 1.

HIS

...

TO RY

S IN BU

ME DIE VA LA RT

...

BO OK S

...

S ES

MA NA GE ME NT

...

Figure 4. ODP Node Annotation Approach [Adapted from (O.L. Oosthuizen, 2004)].

64

Articles

The process described above can be seen as consisting of three crucial steps, namely: • Building a user profile for each user represented as an annotated ODP-tree in the system and using this profile in query augmentations.

• Storing the role-profile in a database for each user using an individualized ODP tree storing descriptive words/phrases about contexts defined in the general ODP tree. The descriptive words stored in the individual ODP tree are words that define a given context for a specific user. This implies that, although all individual trees have a similar structure, the words stored at the leaf nodes may differ for each user of the system.

• Using user feedback to refine the agent’s perception of its user. Implicit and explicit feedback can both be used to refine the role agent’s perceptions about its user’s interests. This means that the role agent interacts with the results analysis agent to determine what keywords/key phrases were present in a given result. The ODP-tree node annotation scheme relies on keywords/key phrases to represent user interests. By combining user ratings (either explicitly or implicitly) with an analysis of the most frequently used keywords/key phrases in a document may be an indicator of how good (or bad) a certain keyword/key phrase is for describing a context in the ODP tree structure. Through the feedback process, the user agent can modify the relevance indicators stored at each leaf node in his/her individual ODP tree. This constitutes the refining of the agent’s belief about its user in the model presented in this section.

A concern that must be addressed in the functioning of the role agent as described above is the possibility of the agent modifying queries and thereby filtering documents of interest to the user due to an underdeveloped role-profile stored in the role profile database for that specific user. This situation is avoided by pre-training the user agent with initial interests and keywords frequently utilised by the user. This ensures that the agent is able to

Journal of Automation, Mobile Robotics & Intelligent Systems

rewrite queries to obtain documents that are more likely to be perceived useful to the user from initial use.

4.2.2. Collaborative meta-search unit In the context of the virtual enterprise the ideas of a community or alliance between organizations (and implicitly also between individual roles or members of the organization) and shared resources across the enterprise between different entities play a critical part in the operation of the organization.

Collaboration among these roles is necessary to ensure the efficient operation of the virtual enterprise as a whole. The promotion of access to shared resources not only helps to streamline the enterprise’s processes, but also promotes improved relationships between entities by enabling transparency of data and operations.

The collaborative meta-search unit in the COPEMSA system consists of two agents to incorporate collaboration between search users and integration of shared resources.

The query agent processes user queries and the partner agent is responsible for interfacing with other search system users. These will be briefly summarized below.

Query agent The query agent is a specialized meta-search engine. It is responsible for the processing of modified queries. This involves rewriting the queries in a search-service specific form (if applicable) and then posting them to a host of different organizational search services (if any are available). Organizational search services could include services like searchable parts inventory, manufacturing resources published on a partner’s intranet or a repository of operational data such as transaction logs etc to name a few.

One of the main challenges facing the query agent is the rewriting of the user/role query in a search service specific way. This is accomplished in the COMPEMSA system through 1) the use of the role profile maintained by the role agent and 2) the use of an interface definition for search services.

Queries are represented internally in the system through an XML mark-up scheme. The actual query string is typically coupled with information such as the query type, node ID etc. The query is then augmented by using the role profile maintained by the role agent. This includes the contexts the keywords in the query have been associated with and any additional keywords associated with the identified contexts.

Next, the query agent uses an interface definition to guide the transformation of the annotated query for specific inter-organizational search services. These interface definitions contain information about what kind of queries the search service can process. For example, does the service support search phrases, forced exclusion of keywords, forced inclusion of keywords, etc. In the context of a virtual organization, the interface definitions for search services deployed on the extranet can quite easily be agreed upon and shared between trading partners.

Using the annotated query and interface definition, the query agent generates a number of different queries and submits them to the identified search services, effectively meta-searching the organizational extranet for relevant results.

VOLUME 1,

No 4

December 2007

After the various search services have responded to the submitted queries, the query agent is also responsible for receiving and collating multiple search service results (i.e. removing duplicates, ranking of results etc.).

The results received from this process are then stored in a topics-links database that lists the topics a user or role is interested in and the locations identified by the query agent to potentially be within that topic.

PARTNER agent The partner agent initiates contact with other multiagent systems of other users located either in the same organization or another entity outside of the organization.

The agent acts as an information broker between the role and query agents and the virtual enterprise’s search community. The partner agent supplies additional communitysensitive information to the other agents on demand and facilitates the exchange of user sessions and analyzed results between members of the virtual enterprise’s search community.

In the COPEMSA system, this is done through the use of a centralized partner server. The primary focus of the partner server is to provide multiple instances (potentially extended enterprise wide) of the search system with additional keywords and contexts. The server is responsible for storing and providing access to information submitted by the various partners to aid each other in their searching activities.

The partner agent periodically collects all the local ODP-trees located with other users of the search system for personalizing user queries from the user agent and submits these to the partner server. This action forms the basis for building a partner ODP tree that consists of common words for describing a particular context. The structure of the partner tree is similar to that of the individual ODP trees maintained by the user agents of the various members of the search community. In order to factor out the words most commonly used to describe a given topic/ context by the community, the partner agents submit each individual user’s local ODP-tree to the server for processing. On the partner server, the ODP-trees submitted by the agents of different community members are analysed and the most common words associated by community members with certain topics are included in the partner ODPtree as common words.

User agents may also submit a topic/context to the server for lookup in the ODP-Tree. The topic is matched in the partner ODP-tree with an internal node of the same name, and any common words (if any) is returned to the requesting user agent. These can then be used by the search partner’s user agents for query augmentation.

The use of a central server has the advantage of simplifying agent discovery and communication. Instead of taxing every system’s partner agent with the task of automatic discovery of other users using the same search system, the agents could simply communicate through a central server thereby ensuring higher participation and information submission by partners. The additional contexts received from partners is then passed to the query agent to assist in query formulation or used to modify resource rankings directly in the topics-links database.

Articles

65

Journal of Automation, Mobile Robotics & Intelligent Systems

4.2.3. Content mining unit The type of content published on the extranet that exists between entities in the virtual organization could possibly be of a highly diverse in terms of context and structured, unstructured or semi-structured in terms of organization. The content mining unit consists of two components, the results analysis agent and a directed extranet crawler to enable the search system to effectively deal with these two challenges.

The results analysis agent continually monitors the topics-links database for new content to analyze. The agent performs a post-retrieval analysis on the content to enable improved ranking and presentation of results.

The directed extranet crawler is responsible for retrieval of actual content (e.g. data from operational databases, content published on the extranet as web pages etc.) from the organizational extranet, as directed by the results analysis agent.

These two agents will be briefly elaborated in the sections below.

Results analysis agent The results analysis agent uses web content mining techniques to initially analyze results retrieved from the query agent to determine the context and topic(s) of the page. The process the results analysis agent follows to achieve this consists of three phases: identification, analysis and ranking.

In the identification phase the agent identifies potential pages from the topics-links database and then uses a directed extranet crawler to retrieve information from that page or resource directly. The benefit of direct retrieval and analysis is the customization of the ranking of results for various roles based on the role profile maintained by the system. Post-retrieval analysis of results can also assist in the presentation of the results in a role-specific way.

In the analysis phase, the agent mines the content of the identified resources to further its knowledge about the resource. The approach taken by the agent to analyze the identified resources is a process consisting of the following steps: • Document representation and feature extraction.

A standard “bag-of-words” (D. Mladenic,1999) approach could be utilized where an individual document, di is then represented as a vector of features d where each feature is associated with a word frequency (i.e. the number of times the specific word appears in the document). One of the major issues with the use of word-based features from resources is the problem of the resultant document vectors being of very high dimensionality. This may have a serious impact on the computational effort required for processing large numbers of vectors in a relatively small space of time. A common approach is to remove words that occur in a stop list from the feature vector. The stop list typically contains words that are common to the language most retrieved results will be written in. Other, more advanced techniques like latent semantic indexing (LSI) with singular value decomposition (SVD) have also been applied to the dimensionality reduction problem (S. Chakrabarti,2003).

66

Articles

VOLUME 1,

No 4

December 2007

• Neighboring resources. Careful analysis of anchortext in hyperlinked resources could prove invaluable to additional resource discovery by the results analysis agent. An issue that must be addressed is the problem of resources with a deep and dense link structure. These resources could have many circular references to each other and could force the agent to analyze the same pages repeatedly. The agent could address this by only considering results that are a preset distance from the original page, thereby ensuring that identical resources are not exhaustively retrieved and analyzed.

• Document classification and Clustering. In this step the agent attempts to automatically discover general patterns present across the multiple results and group the results according to similarities.

A common approach is term frequency inverse document frequency (TFIDF) document classification. The TFIDF scheme represents a key classification method and variants of it are used by many content mining and meta-search systems (D. Mladenic, 1999), (D. Dreilinger and A.E. Howe, 1997). Document clustering techniques can be used to group similar documents into clusters, with each cluster representing a topic or subtopic. This idea of automatically grouping similar documents is of key importance to the results analysis agent as it would then be able to automatically create a taxonomy of results grouped by topic. The clustered results could then be presented to the user in a much more structured way and/or the generated taxonomy could be used as the basis for ranking and filtering operations on the result set or even further query specialization for submission to the query agent (E. Han, et al, 1998).

In the context of the virtual manufacturing enterprise, the structure of the information shared between trading partners is usually well defined (see section 2.3). The context of the information received is also typically restricted to the activities of the manufacturing enterprise. Leveraging on these two assumptions in the analysis phase could greatly increase the quality of the returned results in terms of relevance to the user query.

In the final ranking phase, the agent modifies the topics-links database with the newly discovered knowledge gained (from the clustering and classification operations discussed above) about the resource(s). After the clustering and classification process, a taxonomy of resources classified into different clusters or topics will have been generated. The agent then associates the newly discovered knowledge with the various topics defined in the database. This information is then used in the following ways by the search system: • Future query augmentations by the query agent.

• Results presentation through clustering by topic.

• Further Role-profile specialization.

The results analysis agent could naturally be customized to suit the needs of different manufacturing enterprises and the different informational needs of the user base.

Directed EXTRANET CRAWLER The directed extranet crawler is a small scale web spider capable of retrieving text-oriented content and link-

Journal of Automation, Mobile Robotics & Intelligent Systems

following. It has the key goal of retrieving pages and data from the organizational extranet for analysis by the results analysis agent.

A general web crawler (also known as a web spider) can be defined as a software program that automatically traverses the web by downloading documents using the standard Hypertext Transfer Protocol (HTTP) and following hypertext links inside these documents to other documents. General web crawlers usually traverse the web on a large scale, i.e. they download a wide variety of web pages, process the information contained inside these pages and follow hypertext links inside the processed page to a linked page where the process repeats itself.

A special class of web crawlers called focused crawlers exists where the goal is not to download pages and follow links on the massive scale general web spiders do, but rather to only gather documents and information on a specific topic or from a specific resource. Focused crawlers typically use less network bandwidth and download fewer items when compared to general spiders as only items on a certain topic or from a certain resource is retrieved.

The directed extranet crawler in the COPEMSA system is a focused spider that collects information from specific sources on the organizational extranet related to some topic. The spider is controlled by the results analysis agent and requests pages as described for analysis.

As noted by (S. Mukherjea, 2000), the main bottleneck for a web crawlers is the time spent downloading documents. In order to speed up the results analysis process, the directed spider could fetch all the documents associated with a specific query in parallel and store the HTML source of the documents or extracted database information in a local cache for fast retrieval by the results analysis agent. Additionally, by only considering topically related information the download of irrelevant pages is avoided.

5. Copemsa system evaluation strategy and metrics With the discussion of the COPEMSA system given above, attention can now be given to the questions of perceived efficiency of COPEMSA compared to other web crawling software, and evaluation metrics that can be used for the evaluation and performance testing of web crawling software and finally if the development and implementation of COPEMSA is justified in terms of the given metrics.

5.1. COPEMSA EFFICIENCY One of the development goals of the COPEMSA system was the ability to meta-search the organizational extranet for improved coverage (see Section 4.1). COPEMSA also utilizes a focused crawler for retrieval of topical pages related to user queries, if needed. This guarantees that COPEMSA is at least as efficient as the underlying search services it uses. Furthermore, because post-retrieval analysis is done on results by the results-analysis agent based on a personal profile, COPEMSA is potentially better at satisfying the informational need driving a specific user’s queries than more generalized crawlers.

VOLUME 1,

No 4

December 2007

5.2. Evaluation metrics Two classic metrics for the evaluation of the efficiency of a retrieval system exist in the information retrieval literature, namely precision and recall. Precision refers to the proportion of relevant documents retrieved from all documents in the corpus. Recall refers to the proportion of relevant documents retrieved from all relevant documents available in the corpus.

In their paper, P. Srinivasan et al. (2002) define a general framework for evaluation topical crawlers. They utilize four measures: precision and recall for target pages that have been identified by the specific topic and precision and recall for relevance assessments based on the lexical similarity between crawled pages and predefined topic descriptions (P. Srinivasan et al., 2002).

The directed web crawler component of the COPEMSA system presented in this article can be evaluated using a similar approach and generalized metrics as presented

by P. Srinivasan et al. The approach can also be extended to evaluate the precision and recall for the search system as a whole.

A third metric, user perceived relevance, can also be integrated into the performance metrics for the COPEMSA system to gauge the users perception of the results returned by the search system. Through interaction with the user agent, users rate and score keywords/key phrases as described in section 4.2.1. The annotated ODP-tree maintained by the user agent can then be used as the basis for determining the perceived relevance of search terms through the usage number and/or relevance modifiers associated with each term.

Further analysis can additionally be done using the usage number and relevance modifiers using clustering techniques to generate clusters of keywords/key phrases against which the system can be gauged.

5.3. Development justification and advancement relative to the current state of the art Based on the COPEMSA system development goals stated in section 4.1, the main justification for development of the COPEMSA system in terms of the two metrics stated in the previous section are improved precision through the use of meta-searching techniques by the query agent and improved recall through post-retrieval analysis and reranking of results through the use of web-content mining techniques by the results-analysis agent.

Models like the one proposed in this paper represent only one of many steps toward the ultimate goal of enabling users to construct individual views of the web according to their own personal criteria. More specifically the model adds value in the following areas: • Content location and personalization of searching.

Even with the use of general search engines, finding useful information on a large, interconnected network is still a tedious time-consuming effort. Search systems that autonomously locate documents on behalf of users and then proceed to analyze and categorize the located documents based on personal preferences represent a significant enhancement to current generalized search solutions.

• Collaborative rating of content. The inclusion of a partner agent in the search system model could Articles

67

Journal of Automation, Mobile Robotics & Intelligent Systems

prove invaluable for determining the (perceived) quality of content deployed on the intranet and the design of ranking algorithms that take these ratings into account when ranking results. Content with a perceived low quality can then be gradually phased out of the result sets presented to users.

• The issue of coverage. The query agent presented in the proposed model leverage on the combined efforts of various search services provided on a corporate extranet. The meta-search ability of the query agent contributes significant value to the search system due to improved coverage of contents as well as the ability to interface with each individual search service in an optimized manner.

6. The next generation internet (NGI) and Internet2 Internet2 is made up of a consortium of leading U.S. universities working in partnership with industry and the U.S. government’s next generation internet (NGI) initiative for the development of a faster, more reliable Internet for research and education. Included in this initiative is the development of enhanced, high-performance networking services and the advanced applications that are enabled by those services (Katz et al., 2001). One of the more interesting features of Internet2 is its planned inherent support for middleware. This layer of software provides core services and the idea of core services becoming part of the networking infrastructure is of prime importance in the context of this paper. It can be argued that these core services provided by the networking infrastructure should include resource location and retrieval services. The integration of models similar to COPEMSA into the set of core services offered by such a networking infrastructure could lead to improved resource location and personalised results presentation to the final user.

7. Conclusions and future research In this article we discussed the scope, core features and business model behind the virtual manufacturing enterprise. We also discussed the ideas behind specialized search engines, meta-searching and personalization of the search experience.

The main focus of the article was the introduction of the COPEMSA search system architecture for the virtual manufacturing enterprise. COPEMSA is a multi-agent system designed for the autonomous location and retrieval of information on an organizational extranet in a context sensitive manner.

The key benefits of a common search system integrating resources between different organizations in the same enterprise and the resources of trading partners is higher access to information on an enterprise level and increased transparency of operations between trading partners, stimulating trust.

Further research will be focused on the improvement and refinement of the search model specifically for the extended manufacturing enterprise domain. Specifically, practical implementation considerations for the model proposed in this paper for a concrete manufacturing enterprise will receive special attention. The key aspect to success of this system in a manufacturing enterprise is 68

Articles

VOLUME 1,

No 4

December 2007

the adoption of the system not only by departments internal to the enterprise but also trading partners. Through leveraging on the inherent trust that exists between trading partners, resources external to the enterprise could be integrated into the results retrieved by the system, thereby delivering results that are more accurate and useful for various users’ enterprise wide.

AUTHORS Ockmer L. Oosthuizen* – Academy for Information Technology, University of Johannesburg, Auckland Park Campus, South-Africa.

E-mail: oloosthuizen@uj.ac.za.

Elizabeth M. Ehlers – Academy for Information Technology, University of Johannesburg Auckland Park Campus, South-Africa.

E-mail: emehlers@uj.ac.za. *Corresponding author

References [1] A. Broder, “A Taxonomy of Web Search”. SIGIR Forum, vol. 36, 2002, no.2, pages 3 – 10.

[2] Camarinha-Matos L.M., “Introduction to Virtual Enterprises”, Uninova, MonteCaparica, 1999.

Available online at: http://www.uninova.pt/ ~escn/ttt_portugal.html [3] Chakrabarti S., Mining the Web: Discovering Knowledge from Hypertext Data, Morgan Kaufmann Publishers, 2003.

[4] Chirita P.-A., Costache S., Nejdl W., and Paiu R., “Semantically Enhanced Searching and Ranking on the Desktop”. In: Proceedings of the International Semantic Web Conference Workshop on The Semantic Desktop – Next Generation Personal Information Management and Collaboration Infrastructure, ISWC, Galway, Ireland, November 2005.

[5] Dreilinger D. and Howe A.E. Experiences with Selecting Search Engines Using Metasearch. ACM Transactions on Information Systems, vol. 15, 1997, no.3, pages 195–222.

[6] Gauch S., Wang G., and Gomez M,. “ProFusion: Intelligent Fusion from Multiple, Distributed Search Engines”, Journal of Universal Computer Science, vol. 2, no. 9, 1996, pages 637–649.

[7] Glover E., Lawrence S., Gordon M.D., Birmingham W., and Giles C.L., “Recommending Web Documents Based on User Preferences”. In: SIGIR 99 Workshop on Recommender Systems, Berkeley, CA. 1999.

[8] Google, Google Desktop Features, Available at http://desktop.google.com/features.html, Accessed 18/10/2007.

[9] Han E., Boley D., Gini M., Gross R., Hastings K., Karypis G., Kumar V., Mobasher B., and Moore J., “Webace: a web agent for document categorization and exploration”. In: Proceedings of the Second International Conference on Autonomous Agents, ACM Press, 1998, pages 408–415.

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 1,

No 4

December 2007

[10] Khoshafian S., “Web Services and Virtual Enterprises”, Tect, 2002, Available online at: http:// www.webservicesarchitect.com/content/articles/khoshafian01.asp [11] ISO 10303-1:1994, Industrial automation systems and integration Product data representation and exchange - Overview and Fundamental Principles, International Standard, ISO TC184/SC4, 1994.

[12] ISO 9735-3:2002, Electronic data interchange for administration, commerce and transport (EDIFACT) -- Application level syntax rules, International Standard, ISO TC154, 2002.

[13] Kratz M., Ackerman M., Hanss T. and Corbato S., “NGI and Internet2: Accelerating the Creation of Tomorrow’s Internet”. In: V. Patel et al., editor, MEDINFO 2001, Amsterdam, 2001. IMIA, IOS Press.

[14] Lawrence S. and Giles C.L., “Inquirus, the NECI meta Search Engine”. In: Proceedings of the Seventh International World Wide Web Conference, Brisbane, Australia. Elsevier Science. 1998, pp.

95–105.

[15] Lawrence S. and Giles C.L., “Accessibility of information on the web”, Nature, vol. 400, 1999, pp.

107-109.

[16] Lawrence S., “Context in Web Search”, IEEE Data Engineering Bulletin, vol. 23, issue 3, 2000, pp.

25-32.

[17] Mladenic D., “Text-learning and related Intelligent Agents: a Survey”, IEEE Intelligent Systems, vol.14, 1999, no. 4, pp. 44–54.

[18] Mukherjea S., “WTMS: A system for collecting and analysing topic-specific web information”, In: Proceedings of the 9th International World Wide Web Conference, Amsterdam, Netherlands, 15th-19th May, 2000.

[19] Oosthuizen O.L., “A Multi-Agent Collaborative Personalised Web Mining System Model”, MSc Dissertation, University of Johannesburg, 2004.

[20] Poblete B., Baeza-Yates R., “A content and structure Website mining model”. In: Proceedings of the 15th International Conference on the World Wide Web, Edinburgh, Scotland, 23-26 May, 2006 957-958, 2006. Review date: 4 Oct 2006. Review published with ACM Computing Reviews.

[21] Rautenstrauch T., The Virtual Corporation: A Strategic option for Small and Medium Enterprises (SME’S), Association for Small Business & Entrepreneurship Annual Conference, 2002.

[22] E. Selberg and O. Etzioni., “The MetaCrawler architecture for resource aggregation on the Web”, IEEE Expert, January-February, 1997, pages 11–14.

[23] Srinivasan P., Pant G., and Menczer F., “A general evaluation framework for topical crawlers”, IEEE Trans. on Knowledge and Data Engineering, Submitted, 2002.

[24] Song, L. and Nagi R., “Design and Implementation of a Virtual Information System for Agile Manufacturing,” IIE Transactions on Design and Manufacturing, Special issue on Agile Manufacturing, vol. 29, 1997, no. 10, pp. 839-857.

Articles

69

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 1,

No 4

December 2007

We All Have a Duty of Leaving Our Ideas Behind A conversation with creator of fuzzy logic Lotfi Zadeh

Creator of fuzzy logic, fuzzy mathematics and fuzzy sets theory Lotfi Asker Zadeh was born in Baku (Soviet Union) in 1921. His father was an Azerbaijani journalist, mother – Russian physician. In the early 1930’s family moved to Tehran. In 1942 Lotfi was graduated from the University of Tehran in electrical engineering. Since 1944 he lives in USA, where he took a Master’s degree from Massachusetts Institute of Technology (MIT) in 1946 and a Ph.D. from Columbia (New York) in 1949.

Since 1959 Zadeh has taught at Berkeley University. He has been officially on retirement since 1991, but his schedule is still tight.

The list of Professor Zadeh’s achievements and awards is extraordinarily long – including medals, fellowships, honorary memberships and doctorates, editorships, and chairmanships from all over the world. He was awarded the esteemed Honda Prize Japan in 1991, and in 1995 the IEEE Medal of Honor For pioneering development of fuzzy logic and its many diverse applications. Zadeh is also listed in “Who’s Who in the World”. In 1993 Azerbaijan bestowed him an honorary Professorship from the Azerbaijan State Oil Academy. Since 2005 he is also foreign member of Polish Academy of Sciences, which awarded him Nicolaus Copernicus Medal.

Anna Ladan: Professor, in the year 1973 you proposed new logic, called fuzzy logic. Why fuzzy logic is, described as cheaper and easier than traditional methods of logic?

Lotfi Zadeh: First of all, there are many misconceptions about fuzzy logic. When people hear fuzzy logic they think that logic is fuzzy. The logic is not fuzzy. The logic is precise, but the objects that logic deals with are fuzzy. It is precise logic of imprecision. In classical logic where you have two values: true and false, in fuzzy logic you have many intermediate values. Truth is not so simply as “true – false” rule, it can have degree. The same is true in multiple valued logic, but fuzzy logic is much more general than Łukasiewicz’s logic, because fuzzy logic is much more than logical system. Most of the application of fuzzy logic today has nothing to do with logic, for example this recorder may use fuzzy logic. However, there is no logic in it. What is it used are the concept of linguistics variable and fuzzy “if… then…” rules. Linguistics variable is variable whose values are words. For example, age can be one year, two years, three years – you can be young, middle aged, old – and these are linguistics variables. Fuzzy logic tends to be 70

Interview

more like human logic, and because of that fact it is easier to understand. You feel comfortable with fuzzy logic because this is like the logic that you use all the time yourself.

Classical two-valued logic is not like human logic. It is very deep, but difficult to understand even for mathematicians. In my first paper on fuzziness, which was published in 1965, I used the term „fuzzy sets” that had nothing to do with logic. My first paper in which I began use word “logic” I wrote in period 1974-74. Today fuzzy logic has broad sense, it is used in many products, particularly in consumer products: television cameras, microwave ovens, washing machines, and many others. The word “fuzzy” is usually used in English in a pejorative sense – if something is fuzzy, it is not good. People have some prejudices, because when they use the word “fuzzy” they think that fuzzy logic is something that is not good. As you see, in English-speaking countries the name “fuzzy logic” creates some problems.

In non-English-speaking countries it is not a problem. It happened that in Japan they became interested in it and they had starting use this term. Then from Japan it has spread to other countries.

Fuzzy logic was very popular in the Soviet Union before it collapsed. Also in members of Comecon – Poland, Romania, Czech Republic fuzzy logic was well known and acclaimed. In all of these countries fuzzy logic was popular, and it was used in many, many applications. In 1980 there was a big conference in Moscow – all Union conference on fuzzy sets, in 1990 – in Beijing. After the Soviet Union collapsed, works on fuzzy logic in neighbourhood countries went down too. There’s not having the kind of government support before the collapse. In Poland today there are many people doing very good work You have mentioned Japan. I heard Japanese scientists are very interested in fuzzy logic. Is that true?

Yes, it is. In Japan they are very advanced and they have many engineers working on fuzzy logic products. Not only in Japan but in Asian countries; also China, Vietnam, Hong Kong, South Korea, Singapore. What I see that in those countries the governments are pushing development of this branch. This is not done in Europe. Siemens has fuzzy logic products.

Do you really think that the main reason of that are governments’ actions?

In general in these countries government pushes all science.

Governments are much more active than they are in Europe.

Journal of Automation, Mobile Robotics & Intelligent Systems

You left Europe many years ago and you have lived in USA to these days.

I am the citizen of the United States. I was born in Baku, but I was not Soviet citizen, I was an Iranian citizen. In 1944 I came to the States as an immigrant, not as a student. Iran is a wonderful country but I could not do scientific work. Becoming rich was possible, but I did not want to be rich and spending my life playing cards. That is why I came to the States. I started my graduate studies at MIT, where in 1946 I received my master degree. I could to stay there to continue my doctoral studies, but my parents moved to New York this time. I wanted to be near them; therefore I started study at Columbia University.

When did you start thinking about fuzzy logic?

My first paper about it was published in 1965. The work was done in 1964, of course, but I felt earlier that something has to be done, because classical mathematics is little too black and white. Human concept is not black and white.

You are honest to a certain degree; you are beautiful, young, and tall - all to certain degree. Classical logic consists on two values only, true and false. It does not fit well to reality; it is two-dimensional. At this point there are 50 thousands people in the INSPEC’s database with “fuzzy” in title. In another mathematical sciences database there are 14 thousands titles. As you see, today many people do on fuzzy logic. However initially people were very critical.

Sometimes I give a lecture. For all hour I talk about all kinds of the bad things that people say about fuzzy logic.

Luckily I have very thick skin.

What kinds of application of fuzzy logic are most interesting for you? Not only today, but also in the future.

It is really hard to say, because there are really many applications. At this point, so far as my work is concerned, I am very much interested in application to natural language understanding. I feel this is crucial. This is a central problem, because there are many other problems whose solutions depends on natural language. But things change, so my presentation at London is concerned with something what I called FL+ – “extended fuzzy logic”. It is a new direction.

What direction? Could you explain this?

For example you want to climb the mountain. You drive by car as far as you can, but there will be the point, from which you cannot drive a car. You need a mule. Later you have just to climb. This is what I do now. I am climbing. Let me give you an example problem: automation of driving a car in city. You can automate in city with very light traffic, so this impossible in a city like London, New York or Warsaw. If you put together all the computers from the world you can solve the problem. Problem with traffic cannot be solved by using traditional theories. In my London presentation I wanted to start with the problem. Let’s take a situation: I called a taxi and I say to a driver: take me to this address the shortest way. And to another driver I say: take me to this address the fastest way. Two different problems. The first is a solution, but the second is not. You can take me but you cannot prove that is a good solution.

In the case of the second problem the shortest way you can prove that is the best solution or not. I consider this

VOLUME 1,

No 4

December 2007

problem now. Some problems have scientific solutions, and some problems have not. Only human can do it, but one cannot automate. Problem of automation a car in city traffic does not have a scientific solution.

Another problem which we have difficulties is the problem with summarization. Today we have programs, which can summarize only very simple things. The question is what can I use? Classical logic does not work. Fuzzy logic does not work. I need extended fuzzy logic. Let me tell you how systems became available now. In the USA at this point you can subscribe GPS system, today in combination with databases. If you are looking for sushi, it tells you that nearest sushi restaurant is here. GPS knows where you are.

The nearest does not mean the best. Do not you think the next step should be the best place?

Yes, indeed, GPS shows only the nearest place (laugh). Do you use Google? Google is a search engine. Questionnaire system is a system that answers questions. Google cannot do that. For example, if you ask the question: what is my telephone number? How old was Clinton when he graduated Yale University? It is not able to do it. It cannot answer the questions. Existing search engines do not have deductive capability, they cannot reason. Salesman was demonstrating questionnaire system to the costumer.

He said, “Ask any question that you wish and the system will answer it”. Client asked a question and the system answered. Second question, the system answered. Then client said to the salesman: “You know, I think I can ask your system a question on which it will not ask”. Smiling salesman, “Try, but it will answer”. Client said, “What is my father doing now?” The system was working for a while and printed, “At this moment your father is fishing in Maine”.

“It’s wrong”, said the client, “My father passed away few years ago.” The system knows only that what you put on it.

It does not know that your father died, you have put this information into the system. System printed second sheet of paper, “You do not understand. The person who passed away ten years ago was the husband of your mother, but your father is fishing in Maine”. Therefore, we find out that this client was son of the lover. Now is dangerous to ask questions. The system knows more than you do.

When we are talking about family: your roots are very intriguing. Your father came from Iran and your mother was Russian. Goethe, whose mother was Italian, always stressed that German and Italian elements compounded on his originality and genius. What is your approach to such theories?

Firstly, I am not a man of two cultures! My culture is the Russian culture. My mother was Russian and so I am. Of course, I have been lived in United States for over sixty years, but I speak Russian, I think in Russian, my accent is still Russian, I am totally interested what is going on in Moscow. I am reading Russian writers and listen to Russian music. Shostakovich, Prokofiev – these are my heroes.

Then, if you do not like talking about the past, tell me what would you want to be remembered for most?

I do not like to forecast, because I live today and I could be dead tomorrow. I don’t care about fame or something Interview

71

Journal of Automation, Mobile Robotics & Intelligent Systems

like that. What I try to do at this point is to take some of my ideas and write them up. Some something will be left.

If you have ideas and you do not write them up, they will disappear and no one memorizes them.

What advice would you give to young people as they enter to XXI century?

Basically they are very good feeling with computer science field. Those people who are keen on with computers find good jobs, especially in Poland. It is not for everybody; you have to have some capabilities. There are some people who are unsuited for career in computer sciences. Within computer sciences is computational intelligence – it also very good field. In United States robotics was very popular, but then were expectations, which did not materialise, so robotics starting going down. It became a bad word. So, it used in industry mostly of automation of production. In Japan I visited Matushito factory, where they make VCR’s.

In the whole factory were two-two people. Everything is doing by machines. Automated assembly, it’s not a robot.

On the other hand, robotics became popular in Japan, especially if we talk about humanoid robotics. Japan has no competition in this field. Many companies – Honda, Sony - all have humanoids robots. It is prestige. Maybe not useful, but when Honda humanoid robots dancing, people are thinking, “if they can do this, they can also produce good cars”. From advertising point of view it’s very useful, so they use it mostly for that purpose. Japan is number one in robotics today.

London, 23rd July 2007.

72

Interview

VOLUME 1,

No 4

December 2007

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 1,

No 4

December 2007

In the Spotlight

Omron´s ZXF vision sensors The built-in touch screen offers interactive menus which guide the user through the entire set-up of individual tasks in three simple steps: select the inspection tools, choose the inspection regions and set the proposed inspection parameters with a push of the AUTO button.

Multiple inspections per image can be configured. No PC is required for set-up, maintenance and operation.

They are available with a field of view from 10 mm to 150 mm. All cameras are available as colour or monochrome versions. more at: www.never-fail.info and www.omron.com

Nanogenerator Scientists from Georgia Institute of Technology constructed a prototype of tiny generator, which obtains energy from acoustic vibrations or blood pulsation in the veins. Having only 2 mm2 surface and working like a perpetuum mobile – it can obtain energy from any motion, it may be used as power supplier for nanodevices.

Source: Georgia Tech Photo (Gary Meek) – official site

more at: www.gatech.edu/news-room/release.php?id=1326

Spying Dragonfly

AI Palm iLIMB, produced by Scottish company Touch Bionics, works almost like natural human hand, moreover it can be look humanly because is covered by a semitransparent “cosmesis”, which is computer modeled to look like human skin. First users are Iraq war veterans. Their opinions are enthusiastic. iLIMB costs £8,500. more at: www.touchbionics.com/professionals.php?section=4

Scientists from Technische Universiteit Delft constructed quiet, light, fast and nimble dragonfly-like robot.

DelFly II weighs 16 kg and can fly trough 15 minutes with maximum speed up to 50 km/h; it also can fly backwards or make starting and landing horizontally. Small camera and wireless communication system enable video recording in the real time. Taking into consideration fact that the drive is noiseless, efficient, and relatively resistant, and also the robot can be remote controlled, DelFly seems to be ideal spying device. Scientists are working currently on smaller DelFly Micro, and are announcing DelFly Nano. more at www.robonet.pl

In the spotlight

73

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 1,

No 4

December 2007

Events winter 2007-2008

74

3-7 December

˘ Portugal.

IROBOT’07 – 2nd Intelligent Robotics Workshop, Guimaraes, http://epia2007.appia.pt

4-6 December

Industrial Wireless & Ethernet Conference, Sydney, Australia. http://www.idc-online.com/newsletters/IWEconf.html

10-12 December

ICCMB International Conference on Cellular & Molecular Bioengineering, Singapore http://www.ntu.edu.sg/iccmb/home.htm

20–22 December

IEEE Conference on Electron Devices and Solid-State Circuits Tainan, Taiwan. http://www2.eecs.stut.edu.tw/~edssc2007/

5–7 January

9th IBIMA conference on Information Management in Modern Organizations, Marrakesh, Morocco http://www.ibima.org

28-31 January

BIOSTEC 2008 International Joint Conference on Biomedical Engineering Systems and Technologies, Funchal, Madeira, Portugal. http://www.biostec.org

28–31 January

SICPRO’08 – 7th International Conference System Identification and Control Problems http://www.sicpro.org/sicpro08/code/e08_01.htm

28–31 January

IMI’s 3rd Annual Engineered Fine Particle Applications Conference, Orlando, Florida, USA. http://www.imiconf.com

1-2 February

ICDM’08 – International Conference on Frontiers in Design & Manufacturing Engineering, Coimbtore, Tamil Nadu, India. http://www.karunya.edu

1-2 February

ISCO 2008 – Intelligent Systems and Control, Coimbtore, Tamil Nadu, India. http:// www.kce.ac.in

8 February

National Seminar on Information, Communication & Intelligent Systems, Cochin, Kerala, India. http://www.mec.ac.in/news/attached/41150_sem.doc

20-21 February

7th International Heinz Nixdorf Symposium - Self-optimizing Mechatronic Systems: Design the Future, Pardeborn, Germany. http://wwwhni.uni-paderborn.de/en/symposium2008/

Events