An edge-modulated dual-graph neural network for interacting machining feature recognition

Machining feature recognition represents a pivotal function in the integration of computer-aided design (CAD), computer-aided process planning (CAPP), and computer-aided manufacturing (CAM) systems. Existing methods have exhibited limited recognition accuracy when dealing with interacting features from CAD models, which makes it difficult to meet the demands of real-world engineering scenarios. To address the issue, this paper presents a novel edge-modulated dual-graph neural network (EMD-GNN) for performing interacting feature recognition. EMD-GNN adopts a new dual-graph representation to integrate geometric details from triangular meshes with topological relations from B-Rep faces to facilitate interacting feature representation. Based on it, a dual-graph neural network is designed to deduce interacting features, where a residual mesh graph encoder is used to capture fine-grained geometric details, and an edge-modulated B-Rep face encoder integrates edge-type and dihedral angle information to enhance topology reasoning. To support model training and evaluation, a large-scale synthetic dataset, GeoMFD (Geometric Mixed Feature Dataset), is constructed with multi-feature CAD models and multi-level annotations. The experimental results on GeoMFD, MFCAD++, and MFInstSeg demonstrate that EMD-GNN consistently outperforms existing methods in both recognition accuracy and generalization capability.