Burr Segmentation Method for Oil Injection Holes of High-Speed Train Wheelsets Based on Improved PointNet++

Accurate segmentation of burrs in the oil injection holes of high-speed train wheelsets is essential for automated quality inspection. To address the small size, irregular morphology, and indistinct local boundary features of these burrs, we propose an improved PointNet++-based point cloud segmentation method for oil injection-hole burrs. First, we established an experimental platform for acquiring point clouds of oil injection-hole burrs and used a line laser sensor to acquire three-dimensional point clouds of the oil injection-hole region. We then preprocessed the acquired point clouds to construct a dataset containing burrs with diverse morphologies. Subsequently, to address the indistinct local structural features of burrs, we incorporated a Kernel Point Convolution (KPConv) module into the shallow feature extraction stage of PointNet++ to strengthen the network’s ability to extract local geometric features of burrs. On this basis, we designed a Kernel-Guided Geometric Point Attention (KGPA) module. By jointly incorporating point features, relative coordinates, local density, and distances to kernel points, KGPA assigns weights to neighboring features, thereby enabling the network to focus more on burr boundaries and small protruding regions. Experimental results show that the proposed network achieved an mIoU of 94.25%, an OA of 98.38%, a recall of 98.65%, and a precision of 98.50%, representing improvements of 2.28, 1.62, 2.25, and 2.34 percentage points, respectively, over PointNet++. The results demonstrate that the proposed method can accurately segment oil injection-hole burrs from point clouds, providing technical support for subsequent burr localization, size evaluation, and determination of removal regions.