| Abstract Scope |
This study aims to extract essential information from the weld pool in gas metal arc welding (GMAW) utilizing a small, inaccurately labeled dataset. Weld pool views contain an abundance of information for determining welding status and thus achieve quality control. However, disturbances inherent in GMAW, such as strong arc light, spatter, and smoke, significantly impede the precise labeling of data by humans for subsequent deep learning model training. We introduce a deep learning strategy aimed at segmenting the critical boundaries within the weld pool, specifically the wire, arc, and pool boundary, which also addresses the challenges posed by limited dataset sizes and inaccurate labeling. This approach begins with deploying AOD-Net for defogging, followed by utilizing a YOLOX network to identify the region of interest. Subsequently, a timing segmentation network that incorporates an LSTM mechanism into a U-Net is developed for the segmentation of weld pool images in the dynamic welding process. Experimental results verified that the trained network could extract the critical boundaries accurately under various welding conditions. |