Real-time Monitoring of Weld Surface Morphology with Lightweight Semantic Segmentation Model Improved by Attention Mechanism During Laser Keyhole Welding

During the welding process, the molten metal continuously solidifies along the trailing edge of the molten pool to form the weld seam, so the change in the weld surface morphology can be monitored by the molten pool profile characteristics. In this study, an innovative weld surface morphology diagnosis strategy based on a lightweight semantic segmentation model improved by an attention mechanism is proposed. Considering the characteristics of molten pool morphology change, a semantic segmentation label automatic generation method is proposed, and a large number of high-precision training labels are quickly obtained. The constructed lightweight semantic segmentation model runs more than four times faster than the classical Unet, PSPnet, and Deeplabv3+. The molten pool segmentation accuracy of the constructed model can reach 94.95 % on the new dataset obtained from the test weld. The weld surface morphology reconstruction method is proposed, and the weld morphology size failure defect monitoring is realized based on the molten pool contour features. The validation results show that the constructed model has strong resistance to optical noise interference and generalization ability, and the reconstructed weld surface morphology is consistent with the actual morphological changes.