Welding Journal Research Supplement

Abstract

In multilayer, multipass (MLMP) welding, the first pass in each layer establishes the reference elevation for subsequent passes and determines the layer height. On the other hand, the last pass, which completes the layer, critically affects its filling quality. In such an equal-height path-planning strategy, the online identification of the profiles of the just-finished beads is of great significance for real-time path adjustment in MLMP welding. However, due to the special positions of the first and last weld beads, identifying their profile features poses multiple challenges, including light occlusion, welding spatter, silicide precipitates, and brightness variation. To address the difficulty of identifying the first and last as-welded beads in MLMP welding, this study proposes an improved image segmentation method based on a Dual-UNet. In addition to the passive vision image, the edge mask map extracted from the active vision images is employed as the second input to the model. An Edge-Attention module and Edge-Aware loss are incorporated into the network architecture, along with channel-wise feature concatenation and a dynamic attention mechanism, to enhance the model’s ability to perceive edge features. Then, a Dual-UNet model based on a dual-branch encoding structure and attention fusion is constructed. The experimental results demonstrate that the improved model achieves maximum errors of less than 0.93 mm in width and 0.38 mm in height for identifying the first and last as-welded beads in MLMP welding.