Abstract: In the process of MAG Welding, the traditional Welding pool detection method is difficult to adapt to the intense Arc light due to the serious interference of arc light and splash light. A fusion pool de-noising detection method combining GAN and improved PSPNet is proposed to realize dynamic tracking and accurate detection of the fusion pool. Firstly, the image of welding pool is denoised by generating decoder, encoder and small step deconvolution module in adversity-network (GAN), and the image blur caused by shielding gas and arc light is solved. Subsequently, the formation regularity of the molten pool at different positions is obtained by analyzing the mechanical phenomena of the molten pool at different positions. Finally, combined with GAN network denoising and all-position molten pool change law, PSPNet network is improved. Multi-layer pyramid pooling module is used to extract information, global information is obtained by cavity convolution large field, and Swish activation function is introduced to dynamically segment molten pool. Experimental results show that the average pixel accuracy and average intersection ratio of the method are improved to 88.03% and 85.45%, and the training time of the network model is improved by 19.05% compared with that before the improvement. Dynamic tracking of weld pool is realized which is of great significance to of robotic intelligent welding.

Key words: metal active gas arc welding, GAN, improvement of PSPNet, weld pool dynamic segmentation