Toward non-default partitioning for compound feature identification in engineering design

Geometrical operations, such as extraction, partitioning and reconstruction, are defined in ISO standards on Geometrical Product Specifications and Verification (GPS) in order to obtain ideal and non-ideal features on mechanical parts. Default partitioning enables to decompose the workpiece into independent surface portions regarding kinematic invariance classes. For both specification and verification purposes, non-default partitioning is utilized to create compound features and assist functional tolerancing in the design process. Therefore, it is essential to formalize non-default partitioning and exploit it for supporting further operations within the design activities. In this paper, after a state-of-the-art survey of partitioning and segmentation methods for both default and non-default partitioning, a non-default partitioning process is proposed from both rule-based (explicit knowledge) and data-driven (implicit knowledge) perspectives. The rule-based process addresses non-default partitioning by using Technologically and Topologically Related Surfaces (TTRS) concept while the data-driven method benefits from the recent developments brought by a convolutional neural network (CNN) on point sets. A pre-labeled dataset of mechanical parts is established in the paper for training the network. Experiments and results on CAD models are presented to illustrate the effectiveness of the proposed non-default partitioning method.