Abstract

The laser welding of magnesium alloys presents challenges attributed to their low laser-absorbing efficiency, resulting in instabilities during the welding process and substandard welding quality. Furthermore, the complexity of signals during laser welding processes makes it difficult to accurately monitor the molten state of magnesium alloys. In this study, magnesium alloys were welded using near-infrared and blue lasers. By varying the power of the near-infrared laser, the energy absorption pattern of magnesium alloys toward the composite laser was investigated. The U-Net model was employed for the segmentation of welding images to accurately extract the features of the melt pool and keyhole. Subsequently, the penetrating states were predicted using the convolutional neural network (CNN), and the novel approach employing Local Binary Pattern (LBP) features + a backpropagation (BP) neural network was applied for comparison. The extracted images achieved MPA and MIoU values of 89.54% and 81.81%, and the prediction accuracy of the model can reach up to 100%. The applicability of the two monitoring approaches in different scenarios was discussed, providing guidance for the quality of magnesium welding.