A novel machining operation strategy prediction method based on self-supervised B-rep learning

Manufacturing operation strategy, as a crucial element in part manufacturing, directly impacts processing efficiency and manufacturing quality. Current machine learning-based manufacturing operation decision methods inevitably lose geometric and topological information when converting 3D CAD models into intermediate formats such as point clouds, voxels, or meshes, while also facing the challenge of insufficient training data. To address these issues, this study proposes an innovative hybrid learning approach combining self-supervised pre-training with supervised fine-tuning that learns B-Rep representations from large-scale unlabeled CAD data to achieve manufacturing operation strategy prediction on small-sample datasets. The method consists of two key modules: first, a pre-training model based on an encoder-decoder architecture, where the encoder learns latent representations of CAD model face topological entities, and the decoder reconstructs the signed distance function of model faces; second, a supervised learning phase that utilizes pre-trained facial latent space representations to complete model segmentation tasks based on manufacturing strategies and modeling information using labeled data. Compared to traditional rule-based 3D shape representation sampling methods, directly using B-Rep data representation significantly improves accuracy; compared to existing supervised learning methods based on B-Rep input, the hybrid approach proposed in this study demonstrates superior performance on small-scale training sets.