Research on Point Cloud Classification Method of Machining Features Based on Improved PointNet++

• 1、School of intelligent Engineering and Automation，Beijing University of Posts and Telecommunications，BeiJing，10087

Abstract：Traditional machining feature recognition methods have problems such as weak generalization ability, and although deep learning has made breakthrough progress, it still faces the technical bottleneck of multi-scale feature processing. In this study, a point cloud dataset containing 36 typical machining features is constructed, and an improved PointNet++ network model is proposed. The dataset is constructed by parametric modeling, eigenface extraction and Poisson disk sampling; the network architecture integrating self-attention and multi-attention mechanisms is designed to enhance the ability of local geometric relationship modeling; a multi-granularity feature fusion strategy is introduced, and residual feature fusion is used to solve the problem of gradient vanishing and dimension explosion. The experimental results show that the accuracy of the improved network in the machining feature classification task is improved up to 5.63%, effectively captures multi-scale features, and exhibits strong recognition ability and generalization performance for machining features of different sizes and shapes, which provides new technical support for the field of intelligent manufacturing.

Keywords： Point cloud classification PointNet++ Machining features Attention mechanism Residual Connection