Online welding quality diagnosis based on molten pool behavior prediction

Abstract

In molten pool vision, the prediction of hump and penetration can provide effective alerting for welding, and help to improve welding quality. A molten pool on-line monitor model (MPOM) based on a deep predictive network is proposed to monitor the welding process, including prediction and classification networks. According to the characteristics of molten pool images, a network improvement scheme combining structural similarity index (SSIM) and perceptual loss function are proposed to improve the predicted results of the molten pool morphology of the prediction network (MP-PredNet). The molten pool classification network (MP-ClassNet) is used to judge whether the molten pool image predicted by MP-PredNet network has the trend of hump formation or penetration, which provides basic data for the study of hump and penetration control. The MPOM can predict the change of molten pool morphology 10-time intervals, and each interval includes five frames, in advance, and the accuracy can meet our needs.