Geometric feature extraction from complex-surface point clouds for adaptive CNC machining

Abstract

Measured point clouds of complex surfaces often lack pointwise machining semantics, which restricts CAM adaptation near abrupt curvature changes and feature edges. This paper presents a serialized point-cloud Transformer for geometric feature extraction and adaptive machining mapping. After denoising and topology construction, the proposed MachFormer combines graph-based conditional positional encoding with serialized window attention to predict sharp edges, surface types, principal curvatures, and surface normals. The predicted features are mapped to machining strategies, curvature-adaptive iso-scallop stepovers, and B-spline feature curves. To avoid manual labeling, we construct a benchmark with analytical labels and closed-form pointwise ground truth. On this benchmark, the proposed method achieves an edge F1- score of 0.587, a surface-type mIoU of 0.531, and a principal-curvature RMSE of 0.118. On a 325,000-point automotive component, it reduces the tool-path length by 5.9% and supports tolerance-adaptive curve fitting, demonstrating practical feasibility.