PointPPE: A precise recognition method for complex machining features based on point cloud analysis network with polynomial positional encoding

Abstract

Machining feature recognition is a pivotal step of computer-aided manufacturing, providing the analytical foundation for subsequent machining processes. However, the insufficient utilization of point cloud positional information and redundant information in hierarchical network learning hinder the precise recognition capability of complex features. To address these problems, this work introduces an improved machining feature recognition method, termed PointPPE. Given the precision parts’ feature complexity and similarity, the polynomial position encoding module is designed to learn geometric structures efficiently to encode point cloud position information. A channel attention context fusion module is developed to enhance local feature analysis through channel feature weights assignment and contextual information integration. The results demonstrate that PointPPE exhibits precise recognition capability on constructed precision mold part point cloud datasets, with an instance mean Intersection over Union (IoU) of 90.57%, and shows great generalization on the ShapeNetPart dataset, with class and instance mean IoUs reaching 83.9% and 86.0%, respectively, manifesting superior prospects for complex feature recognition in advanced manufacturing.