Fusion of Static and Dynamic Features of Vision Signal Based on Transformer for Penetration Depth Forecasting in Laser Beam Oscillation Welding

The penetration depth is a key factor that determines the load-bearing capacity of welded joint, so its monitoring is particularly important. Compared with image-based deep learning, the long-sequence time-series image deep learning can extract richer information to more accurately predict welding penetration depth. However, processing long-sequence image signals confronts a tradeoff between computational demands and computational efficiency. In this article, a patch time-series transformer penetration depth forecasting (PatchTST-PDF) method is proposed to improve computational efficiency. A machine vision system is constructed to collect the passive visual signal. Meanwhile, a low-dimensional multivariate dataset consisting of single-frame static and interframe dynamic features is constructed to mitigate the computational costs. Subsequently, PatchTST-PDF with multiparallel inputs is utilized to extract and fuse abstract spatiotemporal information from each independent feature and precisely capture the long-range dependencies associated with penetration depth. The PatchTST-PDF is optimized and achieved a mean-squared error of 0.032 and 0.048 for the training and validation set, respectively. Its forecasting error and response speed outperform other state-of-the-art temporal models. The visualization results of attention scores during the decision-making process demonstrate that the proposed approach can independently extract the information from each feature and accurately track their dynamic evolutions over time.