The Application of Self Organizing Maps in Conceptual Design
Author: Matthews, Peter C
Supervisor: Ball, Nigel; Blessing, Luciënne; Wallace, Ken
Institution: Cambridge University Engineering Department
In engineering design, there is a need for designers to have a good understanding of the design domain. This will typically take the form of some model of the domain. However, such models do not necessarily exist for the early stages of the design process. Designers need to rely on coarse models to guide them through this stage. These models tend to be the result of several years of experience of the domain, and as a result are not necessarily explicit. This research addresses the issue of extracting coarse design domain models using previous design examples. The approach adopted is to rerepresent the design space using Self Organizing Maps (SOM). This rerepresentation is analysed using novel techniques developed in this dissertation to extract a set of relationships. Due to the coarseness of these relationships, they are referred to as heuristics. This is to emphasise that the relationship is not guaranteed to hold, however, it is likely to increase the probability of a good or successful conceptual or preliminary design. For comparison purposes, the same heuristics extraction method is also performed using Principal Components Analysis (PCA) instead of the SOM approach as the core data analysis algorithm. Three case studies illustrate and verify the heuristics extraction method. These investigate the design of autonomously guided vehicles, gas turbine combustors, and aircraft wings. These represent non-linear design domains, that are not necessarily fully understood. Of these case studies, the rst two generated heuristics that were successfully veri ed by domain experts while the third (aircraft wings) highlighted conditions where the method does not initially produce a coarse model that is useful for designers. This research contributes to the machine learning, data mining and mechanical design domains. The machine learning and data mining domains are extended by developing a means of providing an explicit understanding of the inner representation that a SOM generates. The research provides a novel method for analysing a given mechanical design domain, and helping to extract the implicit knowledge that resides within previous designs.