Systematic causal knowledge acquisition using FCM Constructor for product design decision support

Abstract

Despite its usefulness, design knowledge is not often captured or documented, and is therefore lost or damaged after a product design is completed. As a way to address this issue, two major formalisms can be used for modeling, representing, and reasoning about causal design knowledge: fuzzy cognitive map (FCM) and Bayesian belief network (BBN). Although FCM has been used extensively in knowledge engineering, few methodologies exist for systematically constructing it. In this paper, we present a methodology and application—FCM Constructor—to systematically acquire design knowledge from domain experts, and to construct a corresponding BBN. To show the system’s usability, we use three realistic product design cases to compare BBNs that are directly generated by domain experts, with BBNs that are generated using the FCM Constructor. We find that the BBN constructed through the FCM Constructor is similar, based on reasoning results, to the BBN constructed directly by specifying conditional probability tables of BBNs.

Introduction

The mechanical industry sector continues to require higher levels of productivity and quality performance in order to maintain a competitive edge in the global economy. Most mechanical products are non-monolithic and assembly. Understanding and retaining a product’s design rationale is important to be able to verify and trace the ever-evolving product design. The rationale is even more important in large and complex mechanical systems. Despite its usefulness, however, the design rationale is seldom documented, and therefore knowledge is lost or damaged after the product design is completed. Without such knowledge, the effects of proposed, new system design changes are not accurately assessed. Furthermore, problems in various product life-cycle activities may arise because expertise is no longer available or the knowledge has been forgotten. This situation contributes to long delays in recognizing potential failures in a product design. When a potential failure is not promptly identified in the product development cycle, it causes greatly increased costs in terms of warranty and maintenance.

Many systems propose methods of capturing and storing design knowledge. Examples of such systems include DEKLARE, a methodology that supports engineering redesign, and PROSus, which captures the rationale behind designs (Arana et al., 2000, Blessing, 1996). Examples of systems specifically aimed at capturing the rationale behind decisions include Issue-Based Information Systems (IBIS) developed by Rittel (circa 1970s), which formed the basis of many subsequent systems (Conklin and Yakemovic, 1991, Kunz and Rittel, 1970). There are many reasons why knowledge is reused from such systems, including enabling others to understand the original design process and the rationale behind the decisions made, or searching for past designs when working on a similar product or problem.

Causal design knowledge, which utilizes causal reasoning, is particularly useful because it seeks to establish the relationship between causes and effects in the design process. By modeling these relationships, the causes of certain events are diagnosed and their effects are predicted. Causal reasoning is useful in decision making for two reasons: first, it is natural and easy to understand because this ability is inherent in human beings; second, it is convincing as it explains why a particular conclusion was reached. However, causality utterances are often used in situations with uncertainty. Fuzzy cognitive map (FCM) and Bayesian belief networks (BBN) are two major formalisms for modeling, representing, and reasoning about causal design knowledge (Gopnik and Glymour, 2002, Gopnik et al., 2004, Liu, 2001). The two models combined are graphical, and use nodes for representing domain variables and directed links between nodes for representing cause-and-effect relationships between variables.

To obtain cause-and-effect relationships between variables, one significant process is a causal design knowledge elicitation from a domain expert. The process should be simple and intuitive so that it is done by the domain expert with little or no help from a knowledge engineer. Although FCM has been used extensively in knowledge engineering, few tools exist for its systematic construction. Various techniques can be used in each construction process; however, these techniques are either not systematically documented or too specific to a problem at hand. Moreover, many of these techniques are not formalized; hence, automation has not been realized.

In this paper, we present a systematic process to generate an FCM from multiple experts. The knowledge acquisition system, called FCM Constructor, builds a FCM by direct knowledge acquisition. The generated FCM is converted to a BBN for the representation and reasoning about design knowledge. To show the usability of this approach, we compare BBNs that are directly generated by domain experts, to those BBNs generated from the FCM Constructor. To do this, we use three realistic product design cases: an automobile rim design, a fuel nozzle for an aerospace engine, and a product design decision making. The three cases show that the BBN constructed from a FCM is similar to the BBN constructed directly by specifying conditional probability tables (CPTs). They are 87% similar in the best case and 55% in the worst case. The similarity is measured based on their reasoning results; specifically, it is measured based on the number and the order of common variables among the top five variables with highest probability values, when evidence is found on one or more other variables.

The remainder of the paper is as follows: in Section 2, we discuss important literature related to the fundamentals and applications of FCM and BBN. Section 3 shows a comparison between FCM and BBN. In Section 4, we present a systematic method of constructing FCM using the FCM Constructor. In Section 5, we present a systematic conversion of FCM into BBN. Section 6 presents a case study to validate a new network mitigation method (called FCM–BBN). In Section 7, we conclude the results of the comparison and indicate future research issues.