In the manufacturing process of bathyscaphe, the uneven magnetic field around the arc during welding of high-strength low-alloy steel is easy to induce magnetic bias, which damages the welding quality and affects the pressure resistance and service life of bathyscaphe. Therefore, the accurate identification of arc magnetic blow is the key to ensure the welding quality, and it is the premise of implementing effective control measures. However, the traditional image processing technology needs to adjust the parameters when processing images under different conditions, which limits its applicability in changeable environments. To solve this problem, this paper proposes a deep learning-based arc blow angle extraction algorithm to improve the accuracy and robustness of arc magnetic blow identification. First of all, a welding arc monitoring platform is built, and the arc image is collected by this platform, and the welding arc blowing image data set is constructed. Secondly, the semantic segmentation model of welding arc based on MobileNetv3 improved UNet network is established, and an algorithm of blow angle extraction based on semantic segmentation output is proposed. Finally, interpretable analysis is carried out to verify the validity of the model decision. The experimental results show that the proposed method can effectively identify the welding arc deflection Angle, achieving 6.24FPS and 94.7%mIoU. The proposed method lays a foundation for the on-line correction of arc magnetic blow, and has important practical significance for ensuring the manufacturing quality of bathyscaphe.