Graph Neural Networks for patterned welds detection on point clouds

Remote laser welding is an increasingly common process in automotive assembly, particularly in body-in-white and battery manufacturing. While 3D vision cameras enable quick-turnaround inspections of the welds, traditional machine vision algorithms to detect welds, namely template-matching raster-search implementations, often struggle with the variability in weld appearance and quality. Consequently, these implementations may fail to detect partial welds, distorted welds, or welds with any significant deviation from the templates used within the algorithms. Also, the raster scan sometimes detects false positives, welds that are, in fact, not present. To address these challenges, this study introduces a weld surface detection system that uses Graph Neural Networks (GNNs) to detect the presence and locations of welds in 3D point cloud data, implicitly incorporating the variability inherent in the training data samples. The proposed deep-learning algorithm comprises two GNN models chained together: one for segmenting welds by the direction and identity, and another for locating the center of welds. This approach also operates directly on point clouds, offering a computationally efficient alternative to the typical 2D normal map projection or 3D voxelization pre-processing operations on point clouds. In the experimental results, the proposed algorithm identified all welds present, including those with shape deviations, demonstrating a relative strength compared to a template-matching baseline currently used in production.