A data-mining approach for product conceptualization in a web-basedarchitecture

Wei Yan a, Chun-Hsien Chen b,*, Youfang Huang a, Weijian Mi a

a Logistics Engineering School, Shanghai Maritime University, 1550 Pudong Dadao, Shanghai 200135, PR Chinab School of Mechanical and Aerospace Engineering, Nanyang Technological University, North Spine (N3), Level 2, 50 Nanyang Avenue, Singapore 639798, Singapore

Computers in Industry 60 (2009) 21–34

A R T I C L E I N F O

Article history:

Received 24 April 2006

Received in revised form 22 August 2008

Accepted 4 September 2008

Available online 5 November 2008

Keywords:

Axiomatic product conceptualization

Web-based data-mining

Laddering technique

Design knowledge hierarchy

Restricted Coulomb energy neural network

A B S T R A C T

Rapid advancing information technology (IT) has improved the efficiency and effectiveness of product

conceptualization and increased the importance of its role in new product development (NPD). However,

there are two major omissions in existing work: firstly, a unified framework in the process of product

conceptualization has not been well addressed; and secondly, it is imperative to attain an effective data-

mining approach to support the product conceptualization process. Based on this understanding, the

proposed approach aims at postulating an axiomatic product conceptualization system (APCS) to meet

the demand of product concept development. The proposed APCS comprises three cohesively interacting

modules, namely, knowledge elicitation module using laddering technique; knowledge representation

module using design knowledge hierarchy (DKH); and knowledge synthesis module using restricted

Coulomb energy (RCE) neural network. Accordingly, this system offers a method of making design

decisions via a web-based data-mining product conceptualization approach. A case study on wood golf

club design has been used for system illustration.

� 2008 Elsevier B.V. All rights reserved.

Contents lists available at ScienceDirect

Computers in Industry

journal homepage: www.e lsev ier .com/ locate /compind

1. Introduction

Few companies, especially small- and medium-sized enter-prises, now possess sufficient expertise or proficiency to develop acomplete product [1]. Nevertheless, companies can gain morecontrol in their competitive arena by cooperating with othercompanies. Furthermore, enterprises are recognizing that theymust devote more effort to product conceptualization rather thanlater stages of the new product development (NPD) life cycle,because of its disproportionate impact on the final product.Accordingly, it is crucial to improve design consistency yet managedesign conflict amongst design participants. To this end, Pahnget al. [2] integrated designer-specified mathematical models formulti-disciplinary and multi-objective design problems. Alterna-tively, Lu et al. [3] analyzed the relationship between designprocess and design conflict, and thereafter developed a frameworkin terms of technical and social factors.

Rapid advancing information technology (IT) has increased thepossibilities for product conceptualization and the importance ofits role in NPD. This is further enhanced if the companies can useITs to form an alliance for the purposes of in-depth technical and

* Corresponding author. Tel.: +65 6790 4888; fax: +65 6791 1859.

E-mail address: mchchen@ntu.edu.sg (C.-H. Chen).

0166-3615/$ – see front matter � 2008 Elsevier B.V. All rights reserved.

doi:10.1016/j.compind.2008.09.003

business process integration [4]. Technologies for product designhave been frequently explored and include: a standard for theexchange of product model data (STEP) translation [5]; databasemanagement systems (DBMS) (e.g. [6]); real-time 3D CAD systems(e.g. [7]); and virtual reality modelling language (VRML) displays[8]. On the other hand, to enhance the ability of productconceptualization, rather than individual capability alone, researchwork has focused on communication and coordination amongstdistributed resources, e.g., knowledge-based system (KBS) [9],design management system (DMS) [10], and conceptual designtool [11].

Thus, all these methodologies have emphasized embodimentdesign, rather than product conceptualization, so that exploitationof early design creativity and efficiency has not been fully explored.Accordingly, there still exist a number of critical issues in productconceptualization. As such, product concept development mustperform a number of complex functions with respect to designmethodology, concurrency, teamwork, knowledge managementand design representation [12]. In so doing, the IT realization ofconceptualization systems is likely to be a major problem. In thisregard, the data-mining technology presents a logical alternative.

In recent years, data mining has been increasingly advocated inacademia and industries. Its applications are widespread in suchdisciplines as marketing [13], engineering [14], biology [15], andweb analysis (i.e., web mining) [16]. Specifically for product

W. Yan et al. / Computers in Industry 60 (2009) 21–3422

development, a number of research efforts were attempted inproduct data management (PDM), the scope of which has evolvedfrom the internal efficiency of a company into the incorporation ofboth internal and external issues [17]. In the past few years, thefundamental paradigm shift in data mining for product develop-ment has emphasized on improving outward-facing activities in anorganization, such as electronics data interchange (EDI) [18],customer relationship management (CRM) [19], enterpriseresource planning (ERP) [20], virtual enterprise (VE) [21], supplychain planning (SCP) [22], and Internet-based commerce (IBC)[23].

However, some issues have not been well addressed in theprevious work, such as lacking of quantitative analysis methods,low knowledge transparency, extensibility and predictability,and scarcity of effective customer management. To deal withthese problems, some researchers recognized the importance ofknowledge-based data-mining approaches, such as featurespace theory [24], knowledge refinement [25] and rule-basedclassification [26]. Furthermore, in NPD, the product develop-ment team should incorporate customer concerns into productconcepts. This may bring a significant benefit to the companybecause of higher customer satisfaction to the product, forexample, CRM-based methodology [13] and web-mining model[27].

Nevertheless, there are two major omissions in existing work:first, a unified framework in the process of product conceptua-lization, which integrates customer requirements with designknowledge management in the early stages of product develop-ment, has not been well addressed; and second, it is imperative toattain an effective data-mining approach to support the productconceptualizing process. Based on this understanding, anaxiomatic product conceptualization system (APCS) has beendeveloped to meet the demand of product concept development.The proposed APCS comprises three cohesively interactingmodules: namely, knowledge elicitation module using ladderingtechnique; knowledge representation module using designknowledge hierarchy (DKH); and knowledge synthesis moduleusing restricted Coulomb energy (RCE) neural network. Accord-ingly, this system offers a method of making design decisions viaa web-based data-mining product conceptualization approach. Acase study on wood golf club design has been used to illustrateand validate the system. The details of the validation arediscussed.

2. Axiomatic design for product conceptualization

2.1. Background

Axiomatic design has gained wide attention in recent years. Todate, many researchers have attempted applying axiomatic designtheory to such disciplines as product design [28], software systemdesign [29], mechanical system design [30], manufacturing systemdesign [31], design for environment [32], and design forergonomics [33]. Moreover, owing to axiomatic design is a generaltheoretical framework, rather than a specific modelling methodol-ogy, it has frequently been integrated with other technologies inproduct design. Amongst them, Chen [34] proposed a robustconcept exploration method (RCEM) for enhancing productivity inproduct design, in which robust design technique as well asconcurrent engineering technology were combined with Suh’sdesign axioms [35]. For the purpose of improving system qualityand problem solving, Engelhart [36] suggested a design analysisapproach that integrates axiomatic design theory with severalquality control tools and designed experiments. In addition, Seliger[37] treated axiomatic design as one of the methods for product

innovation with regard to inductive and deductive innovation. Inadvent of rapidly developing Internet technologies, Huang [38]postulated a web-based design infrastructure, so-called systematictheory of axiomatic design review (STAR), which defines the designreview as a mapping process between the design objects and thereview criteria.

In essence, the key concepts of axiomatic design [35] involve:(1) the existence of four domains, viz., the customer, functional,physical and process domain; and (2) the characteristic vectorswithin the domains can be decomposed into hierarchies throughzigzagging. Based on these notions, research efforts have beenmade on synthesizing or mapping between consecutive domains.In this respect, most work has been contributed to the mappingbetween functional requirements (FRs) and design parameters(DPs) [39], or that between design parameters (DPs) and processvariables (PVs) [40]. As stated by Suh [35], a product or system caneffectively be classified according to the number of FRs. Accord-ingly, the FRs play a crucial role in axiomatic design process. Acomplete axiomatic design process provides a traceable pathbeginning with the original customer attributes (CAs) througheach domain of a sequential procedure. Hence, the CAs elicitationbecomes the starting point of employing axiomatic design theoryfor product conceptualization, viz. the mapping between CAs andFRs. In this regard, Tesng and Jiao [41] developed a tool-kit tosupport the organization of FRs based on CAs. The approachfocuses on recognizing functional requirement patterns from pastdesign efforts via axiomatic design and taking into account productmigration, technological trends and customer voices. Similarly,Suh and Do [42] attempted to synthesize the elicited CAs with FRsin designing a software system. However, it still remainsproblematic in product conceptualization, especially customervoices involvement and incorporation, such as

� h

ow to genuinely solicit CAs form the voice of customers (VoCs); � h ow to subsequently represent the CAs in a well-organized

manner, such as a hierarchical structure;

� h ow to effectively synthesize the relationship between CAs and

FRs; and

� h ow to consequently flag competitive opportunities and

technical targets through the synthesis between customer andfunctional domain.

The mapping between CAs and FRs is one of the essentialpremises for developing successful product concepts. In this stage, adesign team explores a combination of customer needs, corporateobjectives, product ideas and related technological capabilities, andconcludes the process with a set of product definition, i.e., FRs.Usually, FRs can be represented as a list of product specifications ortarget values, which are often a mixture of quantitative values andqualitative descriptions of a product. The imperatives in dealingwith CAs and FRs via axiomatic design include:

1. A

cquisition of CAs and FRs using a single and effective techniquethat has a broad coverage of design knowledge from suchknowledge carriers as customers and designers.

2. R

epresentation of CAs and FRs under a unified yet simpleframework, e.g., a hierarchical structure.

3. S

ynthesis between CAs and FRs for product conceptualization.

2.2. System architecture

In general, the APCS starts from the customer domain, which isstemmed from the voice of customers (VoCs), followed by thefunctional domain. In order to solicit broad knowledge (or

Fig. 1. Infrastructure of axiomatic product conceptualization system.

W. Yan et al. / Computers in Industry 60 (2009) 21–34 23

requirements) from both customer and functional domain,requirements elicitation becomes an imperative to further (1)represent the customer and functional domain as CAs and FRs,respectively; and (2) identify the relationship between customerand functional domain, i.e., the mapping between CAs and FRs. Forthese purposes, as shown in Fig. 1, the APCS consists of threecorrelated modules, namely, CAs/FRs acquisition module, CAs/FRsrepresentation module, and CAs/FRs synthesis module. These threemodules are described as follows.

(i) T

he CAs/FRs acquisition module

Under the circumstances of rapidly increasing competi-tiveness in today’s business environment, product develop-ment has become more and more complicated. In addition, anumber of complex knowledge carrier’s behaviors, such asperceptions, motivations, attitudes, and personality, influencethe way in which customers or designers organize andinterpret a company and its products. In this respect, aneffective technique to acquire the customer’s and designer’sknowledge by domain experts is highly desirable in forming aproduct concept. Therefore, it is necessary to employ a simpleyet effective technique for dealing with the bespoke complexknowledge acquisition problem during product conceptuali-

zation. The technique should be able to elicit broad customer’sand designer’s knowledge, i.e., CAs and FRs respectively, undera unified framework, and be conducted in a controllableprocess by domain experts. Furthermore, the knowledgeacquisition technique should possess nice compatibility withthe subsequent design knowledge representation scheme,such as a hierarchical structure, in both theoretical orientationand format. Based on these understandings, a well-establishedknowledge or requirements acquisition technique, the ladder-ing technique [43], presents a logical alternative for designknowledge acquisition in product conceptualization and is,therefore, employed and specified for CAs/FRs acquisition inthis study.

(ii) T

he CAs/FRs representation module

After the CAs/FRs are acquired from both customer andfunctional domain, it is necessary using a simplified way torepresent the acquired customer’s or designer’s knowledge bydomain experts in different abstraction levels within a unifiedknowledge representation scheme. For this purpose, a so-called design knowledge hierarchy (DKH) is developed bydomain experts in this module. Under the framework of DKH, acustomer attributes hierarchy (CAH) and a functional require-ments hierarchy (FRH) are stemmed on the basis of the elicited

W. Yan et al. / Computers in Industry 60 (2009) 21–3424

customer’s and designer’s knowledge, respectively. Subse-quently, the comprehensive design knowledge acquired fromdiverse knowledge carriers (e.g., customers and designers) canbe used as the inputs for CAs/FRs mapping. The potentialadvantages of employing the DKH involve: (1) both CAH andFRH can be generated and represented under a unifiedframework; (2) the DKH represents both the decompositionrelationship and the inheritance relationship within a productplatform; (3) by exploring the DKH, design alternatives can beeasily and intuitively generated from different combinationsof alternative values; and (4) it is suitable to quantify thedesign properties via such information as customer ratings inthe heuristic design evaluation.

(iii) T

he CAs/FRs synthesis module

Design knowledge reasoning and decision-making play acore role in narrowing down the design solution (or designdecision) space based on design criteria and constraints. Thisis due to each design space (or product concept) depends onlyon a finite set of entities, whereas infinite possible designsolutions exist. Thus, while represented by a hierarchicalstructure alone, product concept becomes quite qualitativeand uncertain, because (1) the nodes of the hierarchy containsemantic or nominal values; and (2) there are few uncoupledrelations between the properties of the product platformduring product concept development. As such, a typicalmeasurement from design knowledge carriers (e.g., customerimportance ratings) is required to offer the numerical valuesto the properties of the hierarchy and to handle the productconceptualization more normatively and quantitatively. Inthis respect, a restricted Coulomb energy (RCE) neuralnetwork [44] is applied for CAs/FRs synthesis based onaxiomatic design. That is, the mapping between customer andfunctional domain is identified using the RCE network so as tofurther determine technical targets and formulate productconcepts.

3. Data-mining technology for axiomatic design

Referring to Fig. 1, for the purpose of product conceptualization,such knowledge carriers as customers, designers and domainexperts could be concurrently involved in the APCS. This aims toelicit CAs/FRs using such tool as the laddering technique that ishandled by domain experts. As stated by Pyle [45], data mining,which was originated from classical statistics, is associated with aprocess of analyzing data from various sources and solicitinghidden patterns or trends within them by means of association,sequence-based analysis, clustering or classification. Specificallyfor this study, such information as customer ratings is used as datasource for CAs/FRs mapping through RCE network classification. Inthis respect, an integrated approach that comprises ladderingtechnique, DKH and RCE neural network is investigated via data-mining technology for the purpose of CAs/FRs acquisition,representation and synthesis.

3.1. CAs/FRs elicitation using laddering technique

Laddering is a structured questioning methodology derivedfrom Kelly’s repertory grid technique [46]. It was initiallydeveloped by Hinkle [47] for classifying the relation betweenthe constructs and organizing them into hierarchical relations.Similar to other ‘contrived’ knowledge elicitation techniques suchas sorting techniques [48], it was originated in psychology. It hasalso been applied and improved with increasing frequency in thefield of knowledge and requirements acquisition for productconceptualization in recent years [49–51]. Compared with other

knowledge acquisition techniques, there are a number ofadvantages of using the laddering technique for CAs/FRs acquisi-tion as follows.

� T

he laddering technique assumes that customers or designersknow their objectives and are able to group products intodifferent categories. It provides a novel way for transformingcognitive psychology factors into useful inputs for design issues. � T he laddering technique is able to generate more rules and

relevant clauses, has a wider coverage of domain, and requiresless ‘effort’ in terms of mean total time for elicitation and coding.Moreover, it possesses more focused ‘control’ in terms of thesignal/noise ratio over the direction of elicitation for processautomation.

� T he laddering technique requires less effort to transform output

into alternative formats of ‘part-of’ or ‘is-a’ hierarchies, andimposes more categorical hierarchies or discrete classes. As such,it can be employed to bridge the knowledge carrier space (e.g.,customers or designers) and the design space (e.g., a hierarchicalstructure for representing CAs and FRs).

Laddering resembles a form of structured interview in whichthe interviewer (or elicitor) uses a limited set of standard questionsto elicit respondent requirements. It is based on the assumptionthat respondent requirements are organized as a poly-hierarchy,i.e., a multi-dimensional or multi-faceted set of hierarchies.Laddering provides a structure for the elicitation of informationusing a ‘facet’, which is a convenient way to describe individualhierarchy and decomposition requirements. The procedure of theladdering technique, originally presented by Rugg and McGeorge[43], is summarized as follows.

Step 1: S

electing/faceting a seed item. An interviewer first selects aseed item, which is a point within the domain in question,from any level within the hierarchy. It is recommendedthat several sessions be conducted, each time for a ‘facet’ or‘dimension’.

Step 2: P

reparing/phrasing the probes. The interviewer uses probingquestions to move around the structure embedding theseed item. Some of the frequently used probes or phrasingsinclude ‘is-a’, ‘has-goal’ and ‘part-of’. Note that the natureof the probes (phrasings) can be rephrased according to thetypes of elicited hierarchy. In general, a wide range of linksbetween the nodes of the hierarchy is advisable while thebasic question format (as opposed to phrasing) remains thesame.

Step 3: D

irecting/levelling the semantics. Although laddering pro-ceeds simply and recursively, different prompts arerecommended to alter the direction once laddering isnot possible to go any further in a particular direction.Moreover, so-called ‘bottoming out’ (or ‘topping out’) isreached when it is not possible to proceed any furtherdownwards (or upwards) in that line of questioning.

Step 4: D

ecomposing/classing the explanations. As the depth ofexplanations can be treated as an indication of require-ment complexity, also known as elucidatory depth,explanations are then decomposed recursively until termssuch as classes, attributes and entities have bottomed out.Basically, the maximum elucidatory depth is domaindependent. Flexibility that allows terms to be classifiedinto relevant domain features, such as observable andunobservable ones and directly and indirectly measurableones is provided.

Step 5: R

ecording/coding the sessions. Several coding methods areavailable for laddering; including paper record, graphic

W. Yan et al. / Computers in Industry 60 (2009) 21–34 25

representation and pseudo-production rule. Appropriatelabelling that displays the names of classes and attributesis advisable. In general, abbreviations of terms should beavoided.

Step 6: A

nalyzing/post-processing the results. This enables theelicitors to gain insights into the results of laddering.Quantitative analysis can be employed to post-process theresults obtained.

After applying laddering technique, CAs/FRs respectivelyacquired from customers and designers can be representedcomprehensively by a unified scheme, viz. design knowledgehierarchy (DKH), which is described in Section 3.2.

3.2. CAs/FRs representation using design knowledge hierarchy

Product concept development is a complicated process. Itrequires a comprehensive design knowledge representationscheme to describe (1) the multi-disciplinary knowledge fromdiverse viewpoints; and (2) the multi-dimensional goals fordecision-making. A hierarchical structure might be helpful inorganizing the multi-disciplinary knowledge, viz. the diagram-matic, symbolic, cognitive, psychological, semantic and mathe-matical knowledge, in a logical manner. It possesses the followingcharacteristics.

� U

sing the hierarchical structure, the diagrammatic representa-tion of design knowledge is able to provide visible and readilyaccessible illustrations compared to other forms of representa-tion, such as text-based representation. � T he hierarchical structure is an effective means to provide a

high-level overview of the entire product architecture. In sodoing, the categorization at different abstraction level ofproduct concepts, and their relationships and interactions canbe easily formulated.

� T he logical and spatial-visual properties of the hierarchical

structure are effective in providing direct association with thesemantic domain and facilitating the reasoning process.

Fig. 2. Representa

� A

hierarchical structure possesses the properties of descriptivesuch as a tree-like architecture, and normative such as thenumerical values resulted from some computations. It isrelatively easy to represent an entity by decomposing, recom-posing and validating the hierarchical structure.

Accordingly, the proposed design knowledge hierarchy (DKH)possesses a generic hierarchical structure to represent the designknowledge for product conceptualization. The functional taxon-omy used in the DKH (Fig. 2) is defined in Table 1. The DKH has twointegral domains, namely the functional and relational compo-nents.

(i) T

tion

he functional domain

Basically, the DKH is a four-level top-down knowledge-carrier-oriented architecture for the representation of productconcepts. Based on the multi-level taxonomy described inTable 1, each property, which stems from the high-levelprototype, can be decomposed into several sub-properties, andeach sub-property contains several alternatives. Typicalinstances can be selected from different combinations ofalternatives using the DKH to form a specific category. As aresult, the DKH can be employed to represent a knowledge-carrier-perceived category in different abstraction levels fromthe generic (i.e., prototype) to specific (i.e., instances) levels.

9H : Hl 2H;Hl! Tl 8Cl ¼ fTlig;H

l ¼ fPlkg (1)

where 9 is the existential quantifier, C is the category of the

DKH, H is the prototype of the category, T is the instances of the

category, i is the sequential number of instances, P is the

elements or properties of the DKH, l is the level number of the

DKH, and k is the sequential number of the elements within a

specific level.

(ii) T he relational domain

The entities of the DKH interrelate and interact with eachother in both directions via such relations as abstraction,instantiation, inheritance and polymorphism, and expressed in

of DKH.

Table 1Functional terminology used in DKH.

Terminology Definition

Category A group in which an entity is classified or categorized

Prototype A central tendency sharing properties with most or all of the instances of a category

Property The knowledge registered by an entity in the ‘property’ level of the hierarchy is inherited from the respective parent entities in the ‘prototype’ level

Sub-property The knowledge registered by the entities in the ‘sub-property’ level of the hierarchy is derived from the respective parent entities in the ‘property’ level

Alternative The knowledge registered by the entities in ‘alternative’ level of the hierarchy is derived from the respective parent entities in the ‘sub-property’ level

Instance A typical example of a category derived by combining the alternatives

Criterion The rule or theorem, which is used as the basis for the formation of the hierarchy and the solicitation of relations between entities of that hierarchy

Constraint The prerequisite or condition that restrains the resulting outcomes from the hierarchy

Table 2Relational terminology used in DKH.

Terminology Definition

Abstraction The ability to distinguish an entity from all other kinds of entities using the essential characteristics

Instantiation The ability to lead an entity to different forms, which appear to proceed the same functionality

Inheritance The ability of an instance to inherit from its category while sharing the entities or properties of a prototype

Polymorphism The ability to specialize a prototype into more concrete forms

Semantic Relation The ability to describe relations in semantics including ‘is-a’ or ‘part-of’ relation

W. Yan et al. / Computers in Industry 60 (2009) 21–3426

semantic relations. The relational taxonomy used in the DKH isdefined in Table 2. Note that these relations are restricted bythe criteria and constraints (Fig. 2) as follows.1. The abstraction-instantiation pair is a dialectic relation,

existing respectively between the prototype and category, aswell as between the higher level elements and the lowerlevel elements of the DKH.

2. The inheritance-polymorphism pair is another dialecticrelation, existing respectively between the category andinstance, as well as between the instance and differentcombinations of alternatives.

3. The semantic relation occurs in the DKH. For instances, therelation from a property to the prototype, as well as thatfrom a sub-property to a relevant property, is a ‘part-of’relation; the relation from an alternative to relevant sub-property, as well as that from the category to instance, is an

‘is-a’ relation.

Fig. 3. Relationship betw

A hierarchical structure extended from the DKH denotes theproduct concept from the design knowledge carrier’s point of view,viz. the knowledge configuration and re-configuration amongstdesign knowledge carriers. In this respect, the structure of tworelated hierarchies, namely the functional requirements hierarchy(FRH) and customer attributes hierarchy (CAH), is proposed(Fig. 3). These structures inherit the characteristics as well ascriteria and constraints from the DKH.

The FRH, which accesses the designer’s knowledge, i.e., FRs,registered in the DKH, comes with a four-level top-down designer-directed architecture for the decomposition of a specific productconcept. In this multi-level taxonomy, each design specification,which stems from the high-level product platform, can bedecomposed into several sub-specifications, and each sub-speci-fication contains several alternative values (Fig. 3). Typical designalternatives can be selected from different combinations ofalternative values using the FRH to form a specific product family.

een CAH and FRH.

W. Yan et al. / Computers in Industry 60 (2009) 21–34 27

As such, this hierarchical structure can be employed to representthe product concept in different abstraction levels from the generic(e.g., the designer-perceived product platform) to the specific (e.g.,the design alternatives) conception. The results from the FRH andthe selection of typical design alternatives may differ from oneanother due to different designers or domain experts involved. Bythe same token, the customer-perceived product platform, i.e., CAs,can be effectively formulated by the CAH as shown in Fig. 3. Similarto the FRH, the hierarchy of the CAH also starts from the root on top(i.e., the customer-perceived product platform) and then spreadsdownward to other nodes (or properties). In this multi-leveltaxonomy, each imposed construct, which stems from the high-level product platform, can be decomposed into several super-ordinate constructs, and each superordinate construct containsseveral verbatim constructs.

3.3. CAs/FRs synthesis using restricted Coulomb energy (RCE) neural

network

The mapping between CAs and FRs for axiomatic design can beestablished on the basis of customer’s and designer’s knowledge,which is acquired form the CAH and FRH respectively. Forexample, the imposed constructs elicited from the CAH can betreated as one dimension of inputs for CAs/FRs mapping.Concurrently, the design specifications solicited from the FRHcan be regarded as the other dimension of inputs. As such, thecorrelation between such as imposed constructs and designspecifications is further evaluated to generate technical targets.Consequently, product concepts can be formulated based on theFRH, that is, diverse combination of targeted alternative values ofdesign specifications according to such multicultural customerfactors as gender and age. Specifically in this study, the restrictedCoulomb energy (RCE) neural network [44] is applied for thepurpose of design knowledge organization, viz. evaluation of thebe-spoke CAs/FRs mapping.

However, the integration of product concept development andcustomer requirements management has not been effectively dealtwith, e.g., how to translate customer attributes (CAs) intofunctional requirements (FRs) is still problematic. Neural networkhas been proven to be one of the most effective artificialintelligence (AI) techniques for engineering applications, of whichNPD is an important domain [52]. Furthermore, researchers have

Fig. 4. Architecture of

paid more attentions to customer issues in product conceptdevelopment, such as customer segmentation [53] and marketinganalysis [49]. In the proposed approach, the RCE network, designedas an adaptive pattern classification engine, is a supervised neuralnetwork. Compared with other neural networks, its advantagescan be described as follows [54]:

� G

th

uaranteed convergence of thresholds in hidden-units.

� U sually 3–4 epochs for fast training. � P re-determining hidden-unit number not required. � C lass handling even with disjoint regions. � C lassification results from supervised strategy being controlled

more accurately.

� D ynamic new class learning despite of re-training of entire

sample set.

(i) A

e R

lgorithm of the RCE neural network

The RCE network generally consists of three layers, i.e., theinput layer, the hidden layer and the output layer. The inputlayer made up of sensory units from input samples of featurevector is fully interconnected with the hidden layer, whichprovides a set of functions of the input patterns. The hiddenlayer is partially interconnected with the output layer, whichsupplies the response of the network to the activation patterns,e.g., output patterns of weak, moderate and strong level.Moreover, each hidden-unit projects its output to one and onlyone output unit. The architecture of the RCE network is shownin Fig. 4.

Mathematically, the activation function of jth hidden-unit isgiven by [44]

z jðxÞ ¼ f ðjÞ ¼ f ½r j � Dðm j; xÞ� (2)

where mj is a parameter vector called centre, rj is a threshold

and D is the pre-defined distance between mj and x, e.g.,

Euclidean distance. The jth hidden-unit in the RCE network

defines the unit’s influence region with location at mj and size

of rj. Here, f is the threshold activation function given by

f ðjÞ ¼ 1 ðfireÞ if j�00 ðinactiveÞ otherwise

�(3)

The RCE network training involves two mechanisms: unitcommitment and modification of hidden-unit threshold.

CE network.

TabClas

Poss

1

2

3

4

5

6

7

8

whe

W. Yan et al. / Computers in Industry 60 (2009) 21–3428

Initially, a random sample pattern x1 is selected from thetraining set. The allocated hidden-unit centre m1, whichprojects to its output unit z1 representing the class of x1, isset equal to x1. Meanwhile, its threshold r1 is set equal to a pre-defined maximum size of influence region rmax. Next, a secondrandom example x2 is fed into the current network. One of twosituations occurs as follow:1. If x2 causes the output unit to fire, and x2 belongs to the class

that is represented by this unit, training continues with anew input. On the other hand, the correct output may firemore than one unit representing that the input pattern liesinside the overlap region of various classes. This proceeds onby reducing the thresholds of all active hidden-unitsassociated with classes other than the correct one, untilthey become inactive. Note that a pre-defined minimum sizeof influence region rmin is set for the hidden-unit to be shrunkat most.

2. If x2 happens to belong to the same class as x1 but does notcause the output to fire, a new hidden-unit is allocated withcentre at m2 = x2 and threshold rmax. The output z2 of this unitis connected to the output units. Now, if the input patterncauses no output units to fire, a new hidden-unit centred atthe current input pattern is allocated with a thresholdr = min(rmax, max(dmin, rmin)), where dmin is the distance fromthis new centre to the nearest centre of a hidden-unitrepresenting any class different from that of the currentinput pattern. Note that it may cause one or more outputunits that represent the wrong class to fire. Thereafter, theshrinking of the region of influence mechanism aforemen-tioned will ultimately rectify the situation.

le 3sifi

ibl

re

The training process continues until no new units areallocated and the size of the influence region of all hidden-unitsconverges. After training, the classification proceeds using thetrained network.

(ii) S

pecifying the RCE network for design knowledge organization

The customer importance rating to imposed constructs ofthe CAH (an M � N matrix with M-dimensional input featuresand N-dimensional respondents for each design specification ofthe FRH) is used as an input to an RCE network for designknowledge organization for axiomatic design. The RCE neuralnetwork possesses the capability of representing arbitraryfunctions. It converges quickly due to the small number ofhidden-units and training epochs used to respond to anyparticular input vector, which results in rapid training. Thereliability of the RCE network output, such as weak, moderateor strong level for each design specification regarding suchmulticultural factors as gender or age, may be affected by thefeatures and the volume of data used for training. This isbecause each design specification is represented by an M � N

training matrix from customer importance ratings for imposed

cation strategy regarding output activation of the RCE network.

e situation State 0 (weak level) State 1 (moder

0 0

1 0

0 1

0 0

1 1

1 0

0 1

1 1

0 stands for inactive and 1 stands for fire.

constructs. In addition, the training set from the respondents’ratings should be sufficient and effective, because not onlyexternal respondents such as the surveyed company’s and itscompetitors’ customers are involved, but also internal respon-dents such as designers and domain experts are included. Themore comprehensive database the network is trained, the morereliable the results will be produced.

As the RCE neural network is a supervised network, the networktraining for output patterns in relation to input features proceedsunder certain pre-determined schemes or categories of levels. Eachoutput unit performs a simple OR function on the input arrivingfrom the hidden-units connected. The classification strategyregarding output activation and related decision-making aresummarized in Table 3.

While ‘Identified’ situation is detected, a weak level between apair of input (customer importance ratings for imposed constructmatrix) and output (the level for design specifications relating tomulticultural factors) is denoted as State 0. Similarly, a moderatecorrelation and a strong correlation can be denoted as States 1 and2, respectively. On the other hand, there exist ‘Unidentified’ and‘Uncertain’ situations corresponding to different output activationshown in Table 3.

4. A case study on wood golf club design

4.1. Application of a web-based data-mining system

The web-enabled APCS is established for organizations to meetthe demand of an efficient and flexible strategy for axiomaticproduct conceptualization in the early product conceptualizingstage. In the prototype system, the geographically distributeddesigners and customers can use the web site as a shared platform.The data-mining applications of the web-enabled APCS are shownin Fig. 5. The fundamental components required to set up the webenvironment include active server pages (ASP), hypertext mark-uplanguage (HTML) pages, ActiveX data objects (ADO), and database(DB) connections. In more detail:

� T

at

he system has a web-enabled, client/server architecture, withASP as a server-side scripting environment for Microsoft Internetinformation server (IIS).

� T he web server processes the ASP, invokes ActiveX server, and

sends the resulting plain HTML to requesting browsers. TheActiveX server supplies built-in objects, such as application andsession for processing hypertext transport protocol (HTTP)requests and responses, as well as for creating and managingweb applications.

� A ccess to data or knowledge bases is given by the ADO, which is

built on an object linking and embedding database (OLE DB).Moreover, the developed OLE DB depends on the component

e level) State 2 (strong level) Decision-making

0 Unidentified

0 Identified

0 Identified

1 Identified

0 Uncertain

1 Uncertain

1 Uncertain

1 Uncertain

Fig. 5. GUIs of the APCS.

W. Yan et al. / Computers in Industry 60 (2009) 21–34 29

object model (COM) to access the structured query languagebased relational database management system, i.e., SQL-basedRDBMS, through entity-relations (E-R).

The application objects encapsulate the specific functionalitiesof the system for a web-based data-mining approach.

� T

TaPr

Im

M

Co

D

M

D

Pe

his is also implemented using ActiveX and VBScript objects.ActiveX components are used to provide special functionsto such as downloadable visual basic (VB) objects foractivities.

� D esign knowledge management processes (e.g., design knowl-

edge acquisition, representation and synthesis, customer ratingsolicitation, and computer-aided-design, i.e., CAD drawingdisplay) are implemented as an in-process DLL (dynamic linklibrary).

Results obtained from a case study on a wood golf club designwill be provided and discussed in the next sub-section, Section 4.2.

ble 4operties of the CAH.

posed construct Superordinate

construct

Example of verbatim

construct

arket/business Customer segmentation Likes the design

Design alternative Provide more market

information

st/price Pricing strategy Priced reasonably

Cost estimation Cost as low as possible

esign principle Standard issue Follow design standards

Part standardization Use standard parts

anuf./assembly/

repair

Manufacturing factor Is easy to manufacture

Assembly measurement Is easy of assembly or repair

urability Product material Use new material to

reduce weight

Durability consideration Can be used for a long time

Mechanics constraint Possess desirable mechanism

Usability perspective Is easy to use

rsonal preference Design style Provide fashionable styles

Outside appearance Possess good appearance

Tailored specification Provide adequate dimensions

Product fitting Design to suit specific

customers

4.2. Results and discussion

This case study involved the design of a wood golf club, whichwas based on the assumption that the respondents possess someknowledge about golfing. By applying the laddering technique,twelve (12) customers and two (2) designers, who weregeographically distributed, involved in a web-based elicitationfor establishing the CAH and FRH, respectively. The respondentswere asked to contribute their constructs upon probes by a web-based surveying via a domain expert during laddering process.Fig. 6 presents the graphical user interfaces (GUIs) of knowledgeacquisition and representation. The CAH and FRH are respec-tively generated via the ActiveX DLL. The design knowledgehierarchy (DKH), which is handled by a domain expert, possessesa generic hierarchical structure to represent the design knowl-edge for product conceptualization on the basis of its functionaland relational domains described in Section 3.2 (Fig. 2). Theproperties in the four-level CAH and FRH are accordinglyorganized based on the DKH as shown in Tables 4 and 5. In

able 5roperties of the FRH.

esign specification Sub-specification Alternative value

ead material (face/body) General Stainless Steel/Forged

Specific Ti Alloy/Forged,

Ti Alloy/Casting,

Maraging Steel/Forged

ength Short 112 cm

Long 115 cm

ead angle (loft/lie) Standard 128/578, 118/578Specific 108/568, 108/558, 98/558

otal weight Low 250 g, 260 g

Medium 270 g, 280 g

High 290 g, 300 g

ead volume Low 260 cm3

Medium 270 cm3, 280 cm3

High 290 cm3, 300 cm3

haft material General Carbon

Specific Light Carbon, Fiberglass

lex Reflex R, SR

Stiff S, X, RS

stimated price Low $500, $600

Medium $700, $800

High $900, $1,000

TP

D

H

L

H

T

H

S

F

E

Fig. 6. GUIs of knowledge acquisition and representation.

W. Yan et al. / Computers in Industry 60 (2009) 21–3430

this respect, the properties of CAH and FRH can be generatedusing the OLE DB.

Subsequently, after the CAH and FRH have been established,customer importance ratings (from ‘1 - least important’ to ‘10 -most important’) on imposed constructs (Table 4) with respect toeach design specification (Table 5) were elicited together withcustomer information via a web-based solicitation (Fig. 7). In thiscase, eighty (80) respondents were chosen to contribute theircustomer ratings for the RCE network classification according todiverse multicultural customer groups, such as age, gender andskill. The respondents were divided into two groups of differentgenders, i.e., male and female. Each group consisted of fortyrespondents from younger golfers (�35) and forty from oldergolfers (>35). Furthermore, half of them belong to low-skill playersand the others belong to better-skill players.

In this work, the RCE network was employed after therespondents completed customer ratings on imposed constructsfor each design specification. The graded imposed constructs

were then organized into a feature vector and used as inputs tothe RCE network for training and classification via ActiveX DLL.For instance, if six (6) imposed constructs are graded by forty(40) respondents, an input matrix of 6 � 40 dimensions willform the input features. The network’s output is the patternfor a specific design specification regarding those input featuresof imposed constructs, such as low level (State 0), moderatelevel (State 1) and high level (State 2), as explained inSection 3.3.

In more detail, the classification scheme of the RCE neuralnetwork is dependent on the multicultural factors elicited from theoutput patterns for the design specifications, which are organizedin the form of input matrices and gathered from groups ofrespondents having different gender, age and skill (Fig. 8). Asmentioned previously, two diversities exist in the output patterns:one is selected from any of the three patterns (State 0, 1 or 2); theother is ‘Uncertain’ or ‘Unidentified’ due to the output activationshown in Table 3. The results obtained are organized according to

Fig. 7. GUIs of customer importance ratings.

W. Yan et al. / Computers in Industry 60 (2009) 21–34 31

the three multicultural factors, i.e., gender, age and skill, inTables 6–8, respectively.

It appears that the outputs under each gender, age and skillgrouping exhibited certain patterns: either Patterns 0 and 1 or

Table 6Statistical results based on gender.

Des Spec Male Female

Output pattern Output pattern

0 1 2 UC UI 0 1 2 UC UI

I 9 28 1 2 0 24 13 0 2 1

II 0 6 31 2 1 0 5 34 0 1

III 22 15 0 1 2 12 19 6 3 0

IV 0 21 16 3 0 0 11 27 1 1

V 30 8 0 0 2 15 23 2 0 0

VI 19 17 2 1 1 25 13 0 2 0

VII 0 9 29 2 0 0 7 32 1 0

VIII 18 11 7 2 2 8 20 9 1 2

Table 7Statistical results based on age.

Des Spec �35 >35

Output pattern Output pattern

0 1 2 UC UI 0 1 2 UC UI

I 1 14 22 1 2 14 23 2 0 1

II 36 3 0 1 0 32 8 0 0 0

III 23 13 0 3 1 15 20 2 1 2

IV 18 19 0 1 2 31 8 0 1 0

V 16 21 2 0 1 7 28 3 2 0

VI 17 20 1 1 1 28 10 0 1 1

VII 10 27 1 2 0 19 16 2 2 1

VIII 20 12 5 0 3 0 3 34 1 2

Patterns 1 and 2 were instantiated for most design specifications.‘Uncertain pattern’ and ‘Unidentified’, though not significantlyprominent, were also observed in the distribution due to theoutput activation of the RCE network classification. Some form ofsimilarities (commonality of distribution) can be observedbetween two different gender, age and skill groups. For example,male and female golfers emphasized Design Specification ‘Flex’,because the majority of output patterns was linked to State 2 (highlevel). Hence, differences (adverse correlation) can still be spottedas different groups possessed different distribution patterns forsome design specifications.

� It

TaSt

D

I

II

III

IV

V

VI

VI

VI

was found from Table 6 that, in different gender groups, maleplayers concerned more with Design Specification ‘Flex’, whilefemale golfers considered more about Design Specifications‘Length’ as well as ‘Flex’.

� It was detected from Table 7 that respondents over 35-year-old

focused on Design Specification ‘Estimated Price’ much more

ble 8atistical results based on skill.

es Spec Beginner Better amateur

Output pattern Output pattern

0 1 2 UC UI 0 1 2 UC UI

20 14 2 2 2 0 6 33 1 0

0 31 8 0 1 2 15 19 1 3

0 19 17 3 1 0 12 26 1 1

2 23 14 1 0 0 10 29 0 1

13 24 1 1 1 0 16 21 2 1

31 7 0 0 2 0 17 20 1 2

I 0 8 32 0 0 0 22 17 1 0

II 1 10 28 1 0 37 2 0 0 1

Fig. 8. GUIs of knowledge synthesis.

W. Yan et al. / Computers in Industry 60 (2009) 21–3432

than respondents below 35-year-old did, as 34 out of 40responses from those who are over 35-year-old were linked toState 2 (high level) as compared to only 5 out of 40 from thosewho are below 35-year-old. As for Design Specifications ‘HeadMaterial’, ‘Total Weight’, ‘Shaft Material’ and ‘Flex’, it is obviousthat respondents below 35-year-old paid more attentions tothem than those who are over 35-year-old did.

� It

Table 9Response to design specification from diverse customer groups.

Imposed

construct

Gender Age Skill

Male Female �35 >35 Beginner Better amateur

I M L H M L H

II H H L L M H

III L M L M M H

IV M H M L M H

V L M M M M H

VI L L M L L H

VII H H M L H M

VIII L M L H H L

was observed from Table 8 that beginner golfers concentratedmore on Design Specifications ‘Flex’ and ‘Estimated Price’whereas better amateur golfers paid more attentions to theother design specifications.

The final response of the RCE network decisions, involvingmulticultural customer groups and design specifications inparticular, is listed in Table 9, together with a web-based CADdrawing display as shown in Fig. 8. It has an effect on thesubsequent process, viz. the product concept generation viatargeted design alternatives selection. In this work, the voices ofmulticultural customer groups with respect to each designspecification were treated as low (L), medium (M) or high (H)degree (as shown in Table 9), when either one of the followingconditions is satisfied: (1) more than half the number of thenetwork outputs is fired in State 0, State 1 or State 2,

respectively; or (2) the maximum number of fired outputs isdistributed in Sate 0, State 1 or State 2, respectively. Table 10presents the final results of the preferred product conceptgeneration in relation to different multicultural customer groups(forty each out of eighty respondents). It was observed that acustomer preferred product concept, which corresponded toeach of the multicultural customer segmentation, was elicited byselecting the targeted alternative values according to thedetected degrees shown in Table 9.

Table 10Results from the RCE network decision-making.

Design specification Gender Age Skill

Male Female �35 >35 Beginner Better amateur

Face material Ti Alloy Stainless Steel Maraging Steel Ti Alloy Stainless Steel Ti Alloy

Body material Forged Forged Forged Forged Forged Casting

Length (cm) 115 112 115 115 115 115

Loft angle (8) 11 12 11 10 118 98Lie angle (8) 56 57 57 56 578 558Total weight (g) 290 260 270 290 280 300

Head volume (cm3) 270 300 280 270 290 260

Shaft material Carbon Carbon Fiberglass Carbon Carbon L Carbon

Flex S R SR S RS X

Estimated price ($) 900 700 800 600 500 1,000

Product concept #1 #2 #3 #4 #5 #6

Notes: Ti denotes Titanium, L denotes light, S denotes stiffness, R denotes reflex, X denotes extraordinary stiffness, SR and RS is respectively the level between R and S and

between S and X.

W. Yan et al. / Computers in Industry 60 (2009) 21–34 33

5. Concluding remarks

This approach has revealed the potential of improvingconventional axiomatic design theory in terms of effective productconcept development. For this purpose, a prototype axiomaticproduct conceptualization system (APCS) has been established.Compared with previous axiomatic-design-related approaches, itpossesses the following strengths:

� T

he customer’s and designer’s knowledge, which is used as theinputs to CAs/FRs mapping, can be genuinely elicited using apsychology-originated technique, i.e., laddering. The ladderingtechnique can systematically acquire design knowledge fromsuch knowledge carriers as customers and designers. � A so-called design knowledge hierarchy (DKH) has been

developed as a logical and novel knowledge representationscheme associated with laddering technique. Based on the DKH,both the customer attributes hierarchy (CAH) and functionalrequirements hierarchy (FRH) were established for representingCAs and FRs, respectively.

� A novel classification strategy based on the RCE network has

been proposed to analyze multicultural customer factors (e.g.,gender and age), i.e., identification of output patterns withrespect to diverse multicultural customer groups, and evaluatethe relationship between CAs and FRs.

� T he product concepts are specifically generated according to

diverse multicultural customer groups. In other words, a specificproduct concept can be obtained from the combination oftargeted alternative values, which are dependent on classifica-tion results from the RCE network.

A case study on wood golf club design was used to illustrate theperformance of the proposed approach. From the case study, aweb-based data-mining approach has demonstrated its effective-ness in CAs/FRs acquisition, representation and organization at theearly stage of NPD. It is envisaged that with the genuineness ofdesign knowledge elicited and the effectiveness of multiculturalcustomer factors identified, more reasonable product concepts canbe gleaned. As a result, organizations can gain a competitive edgein NPD.

Acknowledgements

The research was supported by Shanghai Education CommitteeResearch Project (Project No. 07SG52), Shanghai Science &Technology Committee NSF Project (Project No. 08ZR1409200),and National High-Tech R & D Plan (Project No. 2007AA04Z105).

References

[1] W. Yan, M.C. Pritchard, C.H. Chen, L.P. Khoo, A study of bidding-oriented colla-borative product conceptualization, International Journal of Computer IntegratedManufacturing 19 (2) (2006) 161–177.

[2] F. Pahng, N. Senin, D. Wallace, Distribution modelling and evaluation of productdesign problems, Computer-Aided Design 30 (6) (1998) 411–423.

[3] S.C.Y. Lu, J. Cai, W. Burkett, F. Udwadia, A methodology for collaborative designprocess and conflict analysis, Annals of the CIRP 49 (1) (2000) 69–73.

[4] F. Giannini, M. Monti, D. Biondi, F. Bonfatti, P.D. Monari, A modelling tool for themanagement of product data in a co-design environment, Computer-AidedDesign 34 (2002) 1063–1073.

[5] C.Y. Kim, N. Kim, Y. Kim, S.H. Kaug, P. O’Grady, Distributed concurrent engineeringInternet-based interactive 3-D dynamic browsing and makeup of STEP data,Concurrent Engineering: Research and Applications 6 (1) (1998) 53–69.

[6] Y.P. Zhang, C.C. Zhang, H.P.B. Wang, An Internet based STEP data exchangeframework for virtual enterprises, Computers in Industry 41 (2000) 51–63.

[7] T.J. Nam, D. Wright, The development and evaluation of Syco3D: a real-timecollaborative 3D CAD system, Design Studies 22 (2001) 557–582.

[8] H.Y. Kan, V.G. Duffy, C.J. Su, An Internet virtual reality collaborative environmentfor effective product design, Computers in Industry 45 (2001) 197–213.

[9] P.A. Rodgers, A.P. Huxor, N.H.M. Caldwell, Design support using distributed Web-based AI tools, Research in Engineering Design 11 (1999) 31–44.

[10] R.I. Whitfield, A.H.B. Duffy, G. Coates, W. Hills, Distributed design coordination,Research in Engineering Design 13 (2002) 243–252.

[11] N. Shyamsundar, R. Gadh, Collaborative virtual prototyping of product assembliesover the Internet, Computers in Industry 34 (2002) 755–768.

[12] S.Y.T. Lang, J. Dickinson, R.O. Buchal, Cognitive factors in distributed design,Computers in Industry 48 (2002) 89–98.

[13] M.J. Shaw, C. Subramaniam, G.W. Tan, M.E. Welge, Knowledge management anddata mining for marketing, Decision Support Systems 31 (2) (2001) 205–222.

[14] A. Hameri, J. Nihtila, Product data management: exploratory study on state-of-art-in-a-kind industry, Computers in Industry 35 (1998) 195–206.

[15] M. Maddouri, M. Elloumi, A data mining approach based on machine learningtechniques to classify biological sequences, Knowledge-Based Systems 15 (4)(2002) 217–233.

[16] J.H. Lee, W.K. Shiu, An adaptive website system to improve efficiency with webmining techniques, Advanced Engineering Informatics 18 (3) (2004) 129–142.

[17] K.K. Leong, K.M. Yu, W.B. Lee, A security model for distributed product datamanagement system, Computers in Industry 50 (2) (2003) 179–193.

[18] S. Lee, I. Han, EDI controls design support system using relational databasesystem, Decision Support Systems 29 (2) (2000) 169–193.

[19] E. Colombo, C. Francalanci, Selecting CRM packages based on architectural,functional, and cost requirements: empirical validation of a hierarchical rankingmodel, Requirements Engineering 9 (2004) 186–203.

[20] C.W. Holsapple, M.P. Sena, ERP plans and decision-support benefits, DecisionSupport Systems 38 (2005) 575–590.

[21] M. Martinez, P. Fouletier, K. Park, J. Favrel, Virtual enterprise—organization,evaluation and control, International Journal of Production Economics 74(2001) 225–238.

[22] C.W.R. Lin, H.Y.S. Chen, A fuzzy alliance selection framework for supply chainpartnering under limited evaluation resources, Computers in Industry 55 (2004)159–179.

[23] S.S. Zhang, W.M. Shen, H. Ghenniwa, A review of Internet-based product infor-mation sharing and visualization, Computers in Industry 54 (2004) 1–15.

[24] H.X. Li, L.D. Xu, Feature space theory—a mathematical foundation for data mining,Knowledge-Based Systems 14 (5–6) (2001) 253–257.

[25] S.C. Park, S. Piramuthu, M.J. Shaw, Dynamic rule refinement in knowledge-baseddata mining systems, Decision Support Systems 31 (2) (2001) 205–222.

[26] Y.C. Hu, G.H. Tzeng, Elicitation of classification rules by fuzzy data mining,Engineering Applications of Artificial Intelligence 16 (7–8) (2003) 709–716.

W. Yan et al. / Computers in Industry 60 (2009) 21–3434

[27] Y. Li, N. Zhong, Web mining model and its applications for information gathering,Knowledge-Based Systems 17 (5–6) (2004) 207–217.

[28] C.G. Park, D.F. Baldwin, N.P. Suh, Axiomatic design of a microcellular filamentextrusion system, Research in Engineering Design 8 (3) (1996) 166–177.

[29] S.J. Kim, N.P. Suh, S.G. Kim, Design of software systems based on axiomatic design,International Journal of Robotics & Computer-integrated Manufacturing 8 (4)(1991) 243–255.

[30] S. Bae, J.M. Lee, C.N. Chu, Axiomatic design of automotive suspension systems,Annals of the CIRP 51 (1) (2002) 115–118.

[31] N.P. Suh, Design of systems, Annals of the CIRP 46 (1) (1997) 75–80.[32] D.R. Wallace, N.P. Suh, Information-based design for environmental problem

solving, Annals of the CIRP 42 (1) (1993) 175–179.[33] M.G. Helander, L. Lin, Axiomatic design in ergonomics and an extension of the

information axiom, Journal of Engineering Design 13 (4) (2002) 321–339.[34] W. Chen, A robust concept exploration method for enhancing productivity in

concurrent systems design, Concurrent Engineering: Research and Applications 5(3) (1997) 203–217.

[35] N.P. Suh, The Principles of Design, Oxford Series on Advanced Manufacturing,Oxford, 1990.

[36] F. Engelhart, Improving systems by combining axiomatic design, quality controltools and designed experiments, Research in Engineering Design 12 (2000) 204–219.

[37] G. Seliger, Product innovation—industrial approach, Annals of the CIRP 50 (2)(2001) 425–443.

[38] G.Q. Huang, Web-based support for collaborative product design review, Com-puters in Industry 48 (2002) 71–88.

[39] P. Ge, S.C.Y. Lu, N.P. Suh, An axiomatic approach for ‘‘Target Cascading’’ ofparametric design of enhancing systems, Annals of the CIRP 51 (1) (2002)111–114.

[40] N.P. Suh, Axiomatic design theory for systems, Research in Engineering Design 10(1998) 189–209.

[41] M.M. Tesng, J.X. Jiao, Computer-aided requirements management for productdefinition: a methodology and implementation, Concurrent Engineering:Research and Applications 6 (2) (1998) 145–160.

[42] N.P. Suh, S.H. Do, Axiomatic design of software systems, Annals of the CIRP 49 (1)(2000) 95–100.

[43] G. Rugg, P. McGeorge, Laddering, Expert Systems 12 (4) (1995) 279–291.[44] D.L. Reilly, The RCE Neural Network, Industrial Electronics Handbook, CRC Press,

New York, 1983.[45] D. Pyle, Making the case, DB2 Magazine 8 (4) (2003) 12–15.[46] G.A. Kelly, The Psychology of Personal Constructs, W.W. Norton, New York, 1955.[47] D. Hinkle, The change of personal constructs from the viewpoint of a theory of

construct implications, Unpublished Ph.D. thesis, Ohio State University, OH(1965).

[48] W. Yan, C.-H. Chen, L.P. Khoo, An integrated approach to the elicitation ofcustomer requirements for engineering design using picture sorts and fuzzyevaluation, Artificial Intelligence for Engineering Design, Analysis and Manufac-turing 16 (2002) 59–71.

[49] W. Yan, C.-H. Chen, L.P. Khoo, A radial basis function neural network multiculturalfactors evaluation engine for product concept development, Expert Systems 18(5) (2001) 219–232.

[50] L.P. Khoo, C.-H. Chen, W. Yan, An investigation on a prototype customer-orientedinformation system for product concept development, Computers in Industry 49(2) (2002) 157–174.

[51] C.-H. Chen, L.P. Khoo, W. Yan, PDCS—a product definition and customisationsystem for product concept development, Expert Systems with Applications 28(3) (2005) 591–602.

[52] C.-H. Chen, L.P. Khoo, W. Yan, A strategy for acquiring customer requirementpatterns using laddering technique and ART2 neural network, Advanced Engi-neering Informatics 16 (2002) 229–240.

[53] C.-H. Chen, L.P. Khoo, W. Yan, Customer-oriented product concept formationbased on laddering technique and Kohonen association, Concurrent Engineering:Research and Applications 10 (4) (2002) 299–310.

[54] M.H. Hassoun, Fundamentals of Artificial Neural Networks, MIT Press, MA, 1995.

Wei Yan is an associate professor in the Logistics

Engineering School at Shanghai Maritime University,

China. His research work is oriented to customer-

oriented product concept development, knowledge

management and artificial intelligence, supply chain

management and logistics engineering. He has a BEng

degree in mechanical engineering from Shanghai Jiao

Tong University, China, an MEng dgree in mechanical

engineering from National University of Singapore, and

a PhD degree in mechanical and production engineer-

ing from Nanyang Technological University, Singapore.

He has several years of engineering experience in

industry prior to his graduate study. He is also an editorial board member of the

journal Transactions of Chinese Construction Engineering Association.

Chun-Hsien Chen is an Associate Professor in the

School of Mechanical & Aerospace Engineering at

Nanyang Technological University, Singapore. He

received his BS degree in Industrial Design from

National Cheng Kung University, Taiwan, MS and

Ph.D. degrees in Industrial Engineering from the

University of Missouri-Columbia, USA. He has several

years of product design & development experience in

industry. His research interests are in collaborative/

consumer-oriented product development, knowl-

edge management for design and manufacturing,

and artificial intelligence in product/engineering

design. He is an editorial board member of the

journals Advanced Engineering Informatics and Recent

Patents in Engineering.

Youfang Huang is a professor, the director of Logistics

Research Center, the Vice President at Shanghai

Maritime University, China. His research work is

focused on logistics management and engineering.

He has a BEng degree in mechanical engineering from

Shanghai Maritime University, an MEng dgree and a

PhD degree in mechanical engineering from Tongji

University in Shanghai, China. He is also an interna-

tional consultant of UNCTAD/WTO, the chairman of

Logistics Association in Northeast Asia, and the deputy

president of Chinese Logistics Association.

Weijian Mi is a professor and the Dean of Logistics

Engineering School at Shanghai Maritime University,

China. He has a BEng degree in mechanical engineering

from Shanghai Maritime University, an MEng dgree and

a PhD degree in mechanical engineering from Tongji

University in Shanghai, China. He is also a committee

member of the Chinese Logistics Association and

Chinese Construction Engineering Association. His

research and teaching interests are involved in large-

scaled container equipment and port information

system design and implementation.