WireGC-Former: Surface defect segmentation method of steel wire ropes based on 3D point clouds

Three-dimensional point clouds can accurately represent the spatial geometric information of industrial product surfaces, offering significant value in small-target defect segmentation tasks. To address the scarcity of high-quality datasets in existing research on industrial defect point cloud segmentation, as well as the challenges posed by the small scale, sparse distribution, and complex morphology of wire rope surface defects, this paper constructs a wire rope defect point cloud dataset, WireRope3D, covering three typical types of defects: wear, wire protrusion, and broken wire. Furthermore, we propose a point cloud segmentation method named WireGC-Former (Wire Rope Graph Convolution Transformer), which integrates a graph convolutional network with a Transformer. The method enhances the model’s ability to perceive local geometric structures and defect boundaries through an edge feature extraction module; improves the representation of local spatial relationships by incorporating a spatial attention module that fuses absolute coordinate information with relative positional encoding; integrates multi‑level local and global features via a feature fusion module; and models long‑range dependencies with a feature attention module, thereby achieving precise segmentation of small‑target defects on wire rope surfaces. Experimental results show that WireGC-Former achieves a mean Intersection over Union (mIoU) of 80.32% on the WireRope3D dataset, demonstrating the effectiveness of the proposed method for small‑target defect segmentation on wire rope surfaces. The presented method can provide a reliable data foundation and technical support for refined detection and subsequent quantitative analysis of wire rope surface defects, and serve as a reference for point cloud semantic segmentation research in complex industrial scenarios.