Approaching STEP file analysis as a language processing task: A robust and scale-invariant solution for machining feature recognition

Abstract

Machining feature recognition is a key task in the intelligent analysis of 3D CAD models as it represents a bridge between a part design and the manufacturing processes required for manufacture and can, therefore, increase automation in the manufacturing process. As 3D model files do not naturally conform to the fixed size necessary as the input to most varieties of neural network, most existing solutions for machining feature recognition rely on either transforming CAD models into a fixed shape representation, accepting some loss of information in the process, or employ rigid rules-based feature extraction techniques prior to applying any learning-based algorithm, resulting in solutions which may display high performance for specific applications but which lack in the flexibility provided by a purely learning-based approach. In this paper, we present a novel machining feature recognition model, which is capable of interpreting the data present in a STEP (standard for the exchange of product data) file using purely learning-based algorithms, with no need for human input. Our model builds on the basic framework for feature extraction from STEP file data proposed in Miles et al. (2022), with the design of a decoder network capable of using extracted features to perform the complex task of machining feature recognition. Model performance is evaluated based on accuracy at the task of identifying 24 classes of machining feature in CAD models containing between two and ten intersecting features. Results demonstrate that our solution achieves comparable performance with existing solutions when given data similar to that used during training and significantly increased robustness when compared to existing solutions when presented with CAD models which vary from those seen during training and contain small features.