Recognizing D shape features using a neural network and heuristics

This work presents a Feature Recognition system developed using a previously trained Artificial Neural Network. The part description is taken from a B-rep solid modeller's data base. This description refers only to topological information about the faces in the part in the form of an Attributed Adjacency Graph. A set of heuristics is used for breaking down this compound feature graph into subgraphs, that correspond to simple features. Special representation patterns are then constructed for each of these subgraphs. These patterns are presented to a Neural Network which classifies them into feature classes: pockets, slots, passages, protrusions, steps, blind slots, corner pockets, and holes. The scope of instances/ variations of these features that can be recognised is very wide. A commercially available neural network modelling tool was used for training. The user interface to the neural network recogniser has been written in Pascal. The program can handle parts with up to 200 planar or curved faces. The performance of the recogniser in terms of speed is far better than that of any other rule-based system due to the Neural Network approach employed. The basic limitation is that of the heuristics used to break down compound features into simple ones which are fed to the ANN, but this is still a step ahead compared to other approaches.