A face-level machining operation strategy prediction method for prismatic parts based on transformer and B-Rep learning

Driven by increasing product customisation and shortened product life cycles, modern manufacturing places higher demands on the efficiency and consistency of machining operation strategy selection. Traditional strategy selection relies heavily on expert knowledge and manual intervention, resulting in limited efficiency and the risk of human error. B-Rep-based learning methods can better preserve the geometric and topological information of CAD models, but existing graph neural network-based approaches mainly rely on layer-wise message passing within local neighbourhoods and remain limited in modelling dependencies between topologically distant regions. To address this issue, this paper proposes BRepMSNet, which integrates the B-Rep data structure with the Transformer architecture for direct face-level machining operation strategy prediction from 3D CAD models. BRepMSNet takes directed coedges as local topology centres, gathers face, edge, and coedge information through predefined topological walks, and organises local B-Rep neighbourhoods into coedge-centred topology tokens. Multi-head kernelized linear self-attention is then employed to establish global contextual relationships among different local topological neighbourhoods, jointly modelling local topological structures and long-range contextual information. Experimental results show that BRepMSNet achieves an accuracy of 98.02% on the machining operation strategy prediction task and also demonstrates competitive performance on multiple public datasets.