Real-time estimation model for magnetic arc blow angle based on auxiliary task learning

Abstract

Magnetic arc blow during the welding of submersible pressure hulls can severely impact the weld quality. Real-time estimation of the arc posture by the teaching-replay welding robot is a crucial and challenging step for further correction of the magnetic blow. This paper proposes a lightweight deep learning model capable of real-time estimation of the welding arc posture. The model takes arc images as input and outputs the angle of arc magnetic blow end-to-end. Specifically, to enable the model to learn the prior knowledge of human perception of arc magnetic blow angles, a training strategy utilizing keypoint detection as an auxiliary task has been proposed. This approach enhances the estimation accuracy of the model without incurring additional inference costs. Additionally, an efficient multi-scale attention (EMA) mechanism was integrated into the angle prediction branch to facilitate the learning of long-range feature dependencies. To confirm the effectiveness of the model, an arc magnetic blow image dataset was constructed for training and testing. The experimental results show that the model achieves a cumulative score (${\mathrm{CS}}_3$) of 98.47 % and a mean absolute error (MAE) of 0.9985° during testing. The model achieves an inference speed of 70.49 FPS on an Intel® Core™ i7-8750H CPU, which satisfies the criteria for real-time monitoring.