Dual angle magnetic arc blow estimation in keyhole tungsten inert gas welding using high dynamic range imaging and a lightweight vision transformer network with coordinate attention and multiple auxiliary branches

Abstract

In high current Keyhole Tungsten Inert Gas (K-TIG) welding, magnetic arc blow frequently causes severe defects such as lack of fusion and undercut, which seriously affect weld formation quality. Conventional visual sensing systems are limited by dynamic range, making it difficult to capture arc morphology, while single angle descriptors fail to represent nonlinear deflection and lightweight convolutional models struggle with long range dependencies. To address these challenges, this study employs a High Dynamic Range (HDR, 120 decibel [dB]) imaging system to capture detailed arc variations and proposes a lightweight Vision Transformer (ViT) network with embedded Coordinate Attention (CA) and multiple auxiliary branches for real time angle estimation. A custom magnetic excitation system enables controllable arc blow simulation and consistent data acquisition. The method introduces a dual angle representation, namely the maximum curvature angle (${\theta }_{\mathrm{curv}}$) and the equivalent deviation angle (${\theta }_{\mathrm{eq}}$), to comprehensively describe arc geometry. The Artificial Intelligence (AI) framework integrates segmentation, keypoint localization, and regression tasks to improve accuracy and robustness. Trained on a self constructed HDR dataset containing 3,191 annotated images, model achieves a mean absolute error (MAE) of $1.12°$, a root mean square error (RMSE) of $2.84°$, a determination coefficient ($R^2$) of 0.96, and a per frame inference latency of 12.96 ms (ms) on an NVIDIA RTX 2080Ti graphics processing unit (GPU). These results demonstrate that AI based methods combined with HDR imaging cannot only achieve accurate monitoring of welding arc states, but also provide potential support for closed loop control in all position welding applications.