Abstract: To achieve high-precision automatic classification of multi-type ship welds, addressing issues of multi-source noise interference, difficulty in modeling thickness features, and blurred edge structures in laser scanning point clouds of weldments, a PWeldNet model for multi-type weldment classification is proposed. Based on POINTNET + + , the model incorporates a self-attention-guided denoising module, a thickness prediction module, and an edge thickness guidance module to enhance anti-interference capability, improve plate thickness information extraction, and strengthen boundary structure perception. On the constructed 22 types of simulated weldment point cloud datasets, PWeldNet achieved 98.68% instance accuracy and 95.45% classification accuracy, which is significantly better than the performance of the five mainstream point cloud classification models. Robustness tests further verify its maintained performance under low-density sampling conditions. Ablation experiments confirm the effectiveness of each module and their combined synergistic benefits. Finally, the measured results based on the laser scanning platform show that the PWeldNet can still achieve efficient and robust classification of weldments when facing multi-type welding point cloud data with multiple noise interferences and complex representations.