A Graph Neural Network and Keypoint‐Guided Fitting Based Method for 3D Weld Seam Trajectory Extraction

To address inaccurate keypoint recognition and noise‐sensitive trajectory fitting in automated 3D weld seam extraction, a learning‐to‐fitting framework combining 3DSeamNet and KGC‐LO‐RANSAC is proposed. 3DSeamNet jointly predicts weld regions, centerline offsets, local tangential directions, and keypoint confidence from workpiece point clouds, providing structured geometric priors for subsequent trajectory fitting. KGC‐LO‐RANSAC then reconstructs the weld centerline using offset‐corrected centerline candidates, high‐confidence keypoint guidance, and spatial‐tangential consistency constraints. The fitted trajectory is further mapped to the original point cloud to generate welding path points. Experiments on flat‐weld and fillet‐weld workpieces demonstrate stable and accurate trajectory extraction under different geometric conditions. The root‐mean‐square error remains below 0.98 mm under different plate thicknesses, while the total processing time from point‐cloud acquisition to trajectory output is limited to 0.68 s. These results demonstrate the feasibility of the proposed framework for accurate and efficient automated weld trajectory extraction.