Adaptive data mining in a variant design support system

In variant design, getting a product to market quickly is critically important to maximize sales profits. Often, designers waste time and effort "re-inventing the wheel" by redesigning parts or components simply because they do not know the items exist, or because finding a similar item is difficult. Therefore, much research has been done to develop classification systems that facilitate the search for similar parts and components and promote reuse of these existing designs. One approach to categorizing and classifying proven end items is to develop a generic bill of material (GBOM), a single entity that encapsulates all variants of a product family. The GBOM can be searched for similar products or components, and, since one GBOM represents many individual bills of material (BOMs), the search space for these similar items is greatly reduced. In Romanowski and Nagi [1,2,3] GBOMs were formed by first clustering individual BOMs into families and then unifying the cluster members into a single GBOM entity. Thus, each GBOM is representative of several similar products. But because products are generally not static - new designs are created, others become obsolete, still others undergo revisions - the clusters and resulting GBOMs must be updated periodically to keep abreast of these changes. Certain questions arise: how often do we update? How much change is needed before the clusters need to be recalculated? Are changes in some components more important than others? In this research, we discuss the types of changes that take place in a bill of materials database and answer the when question by using a cusum method to signal density changes in the underlying clusters affecting the GBOM structure. We present a methodology for determining when and how to adapt the support system to evolving data, and give an example of our approach as applied to an industrial dataset.