Abstract: Accurate detection of welding deviations is a prerequisite for automatic seam tracking and intelligent welding by welding robots.An improved ViT-based method for molten pool recognition and online detection of welding deviation was proposed. Firstly,the lightweight ViT model Segformer was used as the baseline model. The Shuffle Attention (SA) was embedded before mask segmentation to better capture the dependencies of feature information in both spatial and channel dimensions. Thus,the model's segmentation accuracy was enhanced.Secondly,a Context Broadcasting (CB) module was added to the Multilayer Perceptron (MLP) to improve the generalization capability while ensuring low parameters of model. Finally,based on the model segmentation results,a welding deviation calculation method was proposed to quantitatively describe the deviation detection accuracy. The experimental results show that,compared with the baseline model,the mean intersection over union and mean pixel accuracy of proposed model were increased by 2.67 % and 2.12 %,respectively,and it has good generalization for different preset torch offsets. The welding deviation accuracy was controlled between ±0.021 mm,which provided a basis for seam tracking in precision welding.

Key words: welding deviation, seam tracking, molten pool recognition, Vision Transformer (ViT), attention mechanism