Characteristic classification and extraction of robotic multi-layer multi-pass MAG welding pool—An extended UNet network implementation based on transfer learning

The real-time and accurate acquisition of weld pool visual features during robotic multi-layer and multi-pass welding (MLMPW) of medium-thick plates is essential for controlling weld quality. To address the challenge of extracting pool information in complex welding environments, this study proposes a novel method for acquiring pool contours using the ResNet101-UNet architecture, with ResNet101 serving as the backbone. First, a custom dataset of MLMPW pool images (comprising seven different pool types) and their corresponding edge labels was used to train the network. Second, a comprehensive evaluation of different semantic segmentation models was performed, taking into account the inclusion of pre-trained modules from the ImageNet dataset. Experimental results demonstrated that the improved segmentation method can efficiently and effectively extract pool contours from 2D images captured by welding visual sensors. The designed ResNet101-UNet network architecture achieved an effective Mean Intersection over Union (MIoU) of 96.14 % and a Dice coefficient of 98.06 % on the self-constructed pool dataset. By defining the characteristic parameters of MLMPW molten pools and conducting statistical analyses on these parameters, seven classification standards for molten pools were established, including triangular (Type 1), trapezoidal (Types 2, 3, and 4), and parallelogram-shaped (Types 5, 6, and 7) weld formations. The MLMPW pool feature classification and extraction method presented in this paper can acquire richer pool visual features, thereby providing a data foundation for developing automated and intelligent models in the welding process of medium-thick plates.