Neural network based shaped neighborhoods: a design retrieval system

Abstract

The life cycle cost of a single detail part in a large assembly such as an airframe includes design, tooling, manufacture, maintenance and other associated computing overhead throughout the 40 year life of the part. These expenses are a significant factor in the final cost of the assembled product. As such, opportunities exist in preliminary design to identify and exploit existing "similar parts" that are already configured in the release system. The notion of a similar part properties includes geometric parameters as well as other non-geometric attributes such as material and finish.This dissertation is based on an 8-dimensional hyperspace where the axes of the space represent an appropriate set of the detail part properties discussed above. Thus, any given detail part can be mapped directly into this space. In this domain, the concept of similar part discovery reduces to identifying an appropriate neighborhood around a given target location. This dissertation develops "shaped neighborhoods" in hyperspace using a neural network based vector approach. The method provides the required flexibility to define highly specific neighborhoods which in turn yield highly refined search results. Put another way, these 8-dimensional shaped neighborhoods, called "cookie cutters", provide the means for defining individual property variability as a simultaneous function of all other property variability.Finally, an improved NN training algorithm specifically suited to the network topology used to represent the shaped neighborhoods is presented. In this algorithm, Kohonen self-organizing feature maps (SOFM) are used to define a topology preserving mapping between a high dimensional NN input space and a two dimensional output map space. NN training error is associated with respective SOFM map nodes. Thus, the Kohonen map nodes represent vector prototypes over the input domain and the values associated with each map node are representative of training error for a generalized region in the input domain. This dissertation exploits this relationship by defining a weight update expression based on this error leading to an accelerated network convergence algorithm.