Decomposition of interacting features using a Kohonen self-organizing feature map neural network

Abstract

Recognition of interacting features has been a difficult task in many existing feature-recognition systems. The unique topological patterns of isolated features change drastically when they interact. Hence many surface-based methods encounter problems in accommodating these changes in their generic feature definitions. Recently, much effort has been concentrated on the volumetric approach. However, many of these systems suffer from a problem of combinatorial explosion as the interaction between features becomes more complex. This paper presents a simple and robust system, in which the interacting features are decomposed into simple primitive features prior to recognition. The system starts by searching a B-rep solid model, using a layering technique, for volumes corresponding to interacting features. The interacting features considered in the paper are of the type that has a uniform thickness and a common bottom face, referred to here as the “to-be-decomposed” type. The volume of an interacting feature is then represented in a simple 2D framework as the resultant area. The vertices of the resultant area are clustered using a Kohonen self-organizing feature map (SOFM) neural network to generate maximal rectangular regions (MRRs). A decomposition process utilizing boolean operations intersects the resultant area with the MRRs to generate regions that represent primitive features, referred to as primitive regions. These primitive regions are then subtracted from the resultant area. Any remaining region is further decomposed into primitive regions, using a second stage of the SOFM and decomposition process. The feature patterns in these primitive regions are used as input in a multilayer feedforward neural network to recognize the features. Self-organization, competitive learning and the clustering of data are some of the SOFM’s attributes, exploited in this work to deal with interacting features.