Feature Recognition Method Based on Multi-Feature Fusion and Graph Attention

For the task of automatic recognition of machining features in 3D CAD models, existing methods have problems of information loss and insufficient generalization ability, so this paper proposes a recognition method based on an improved graph neural network.. First, the method constructs Boundary Representation (B-Rep) models into a geometric attribute adjacency graph that integrates geometric, topological, and attribute information. Precise geometric features are extracted by Convolutional Neural Network (CNN) after UV meshing, while rich attribute features are encoded by Multi-Layer Perceptron (MLP). On this basis, an improved graph neural network architecture is designed, with the core being the introduction of Graph External Attention Network (GEANet) into the graph neural network. Through a learnable external shared memory unit, this mechanism enables the model to not only learn local features within a single part but also capture and utilize common structural patterns across parts (graphs). This deepens the understanding of complex and intersecting features and effectively improves the model's generalization ability. Experiments are conducted on the MFCAD and MFCAD++ datasets. The accuracy and mIoU on the MFCAD dataset are 99.73% and 99.48% respectively, while those on the MFCAD++ dataset are 98.78% and 97.87% respectively. The network model complexity is only 0.5M, demonstrating high recognition accuracy as well as excellent computational efficiency and resource utilization.