Abstract: To address the limitations of traditional point cloud segmentation methods that they rely excessively on manually designed geometric feature and statistical rules, as well as manual feature design depends on prior knowledge and suffers from insufficient capabilities of feature characterization while extracting machining features of parts, a segmentation approach based on improved dynamic graph convolutional neural networks (DGCNN) is proposed in the paper. Machining features of parts typically exhibit strong locality, and the DGCNN′s powerful capabilities of extracting local features enable it to identify detailed information about these machining features and distinguish their boundaries during segmentation. On this basis, to establish an effective correlation between global and local features as well as to improve the overall feature characterization capability of the network, a shuffle attention module has been introduced into DGCNN, which can fully leverage the correlation between spatial attention and channel attention, and combine their advantages to learn and utilize the point cloud features more efficiently. To verify the effectiveness of the improved network, a point cloud dataset containing 24 types of common parts is constructed. Experimental results show that the overall segmentation accuracy of the proposed method reaches 99.79%, which is 2.22%, 1.16%, and 0.38%higher than that of PointNet, PointNet++, and the original DGCNN respectively, significantly improving the accuracy of part point cloud segmentation.