Encoding mechanical design features for recognition via neural nets

Within mechanical computer-aided design (CAD), pattern-recognition techniques are fundamental to feature recognition. The use of neural net software as the pattern-recognition element depends upon encoding schemes which extract critical information from candidate geometric subsets. The trained memory can then determine if a particular candidate geometric subset corresponds to a feature of interest. Successful experiments with particular encoding schemes over a restricted class of features will be presented. Neural nets were chosen with the long-term view toward a feature-recognition architecture where the end-user could customize the domain of features that can be recognized. The training of the neural net memory would be achieved through a graceful graphics interface. Extensive programming and knowledge bases would be avoided. This envisioned architecture will be presented to provide a context for the encoding schemes.