Machining feature recognition is a key enabling technology in intelligent manufacturing that extracts manufacturing semantics from the boundary representation (B-Rep) of 3D CAD models to bridge design and process planning. Recent advances in deep learning have accelerated data-driven feature recognition methods. Among these, graph neural networks (GNNs) have gained significant attention due to their natural compatibility with the non-Euclidean, hierarchical topological structure of B-Rep data, enabling efficient and lossless encoding of geometric and topological attributes. However, existing GNN-based methods primarily leverage the topological structure and geometric attributes of B-Rep models, often neglecting the inherent geometric relationships present in the B-Rep data structure. To address this gap, we propose a dual-graph fusion network (Topo-Geom DualGNN) that integrates a topological attribute adjacency graph and a geometric relationship graph. Our approach employs a GatedGCN-based graph encoder and an FiLM-based cross-stream fusion mechanism to jointly encode topological and geometric information from the B-Rep model. Evaluations on open-source synthetic datasets, including MFInstSeg and MFRCAD, demonstrate that the proposed method achieves competitive comprehensive recognition performance and exhibits promising capability in recognizing machining features in complex parts.
Wang et al. (Thu,) studied this question.