Recursive Neural Networks for the Automatic Analysis of STEP Files

Intelligent interpretation of three-dimensional data represents a complex problem for artificial intelligence models. This is a relatively young, relatively small field of research, in which a key consideration is the format of the input data. This thesis proposes an entirely novel approach, in which natural language processing techniques are applied directly to CAD (computer aided design) models in the STEP file format. The research presented here represents the early development of a unique and flexible method for intelligent extraction of complex information directly from CAD models, without requiring any transformation away from the input data. The primary contribution of this thesis is the development of a recursive encoder network capable of encoding the hierarchical data structures present in a STEP file into a single-vector representation. To investigate the potential of this approach, encoder-decoder models are trained to perform machining feature recognition, to estimate the values of geometric parameters, and to perform comparisons between pairs of CAD models. Further analysis is performed into the output representations of the encoder network, demonstrating the successful encoding of information relating to shape and dimensions.