Abstract
<title>Abstract</title> <p>In keyhole tungsten inert gas (K-TIG) welding, accurate localization of the tungsten electrode tip is essential for torch alignment, seam tracking, and closed-loop process control. This task is challenging because the tip is extremely small, weakly textured, and highly susceptible to arc glare, spatter, and appearance drift. To address this problem, this paper proposes a Consistency-Gated Cascade (CGC) for robust tip localization. The framework combines a YOLOv11-nano-pose detection branch and a ResNet18 coordinate regression branch with a parameter-free consistency gate. Under standard validation conditions, the regression branch provides higher precision, whereas the detection branch shows better stability under distribution shift. The proposed CGC uses the coordinate deviation between the two branches as a failure signal: when the deviation is small, the regression output is adopted; otherwise, the system falls back to the detection output. Experimental results show that CGC achieves a mean error of 1.75 px and 100% PCK@5px on the validation set. On ten consecutive frames from a held-out welding condition, the standalone regression branch collapses to an almost constant output with a mean error of 254.04 px, whereas CGC reduces the mean error to 4.22 px and achieves 100% PCK@10px. These results demonstrate that detection-regression consistency provides a simple and effective mechanism for balancing high precision and robustness in intelligent K-TIG welding vision.</p>