关键词: 加工特征识别, UV网格化, 图注意力, 特征融合, 几何属性邻接图

Abstract: Aiming at the difficulty of machining feature recognition in 3D CAD models, a part machining feature recognition method based on feature fusion and an improved graph attention network is proposed. The Boundary Representation (B-Rep) model is used to represent the 3D CAD model. The parametric surfaces and curves of the B-Rep model are discretized through UV meshing, and the 2D/1D Convolutional Neural Network (CNN) is used to extract the geometric features of faces and edges. Meanwhile, the Multi-Layer Perceptron (MLP) is adopted to extract the attribute features of faces and edges. The obtained geometric features and attribute features are imported into the face adjacency graph to form the geometric-attribute adjacency graph. Edge features and node features are concatenated through a target-oriented aggregation strategy to form fused features; edge features are introduced into the attention coefficient calculation to dynamically assign neighborhood weights, which enhances the ability to parse complex features and topologies. At the same time, residual connections are designed to optimize gradient propagation. Experiments are conducted on the MFCAD and MFCAD++ datasets respectively. On the MFCAD dataset, the accuracy and mean Intersection over Union (mIoU) are 99.97 % and 99.96 % respectively; on the MFCAD++ dataset, the accuracy and mIoU are 97.03 % and 94.70 % respectively.

Key words: machining feature recognition, UV meshing, graph attention, multi-features fusion, geometric attribute adjacency graph