| Abstract Scope |
Many industrial welding applications heavily rely on post-weld NDE techniques for quality assessment, where process modification and anomaly identification are limited by skilled welder and machine operator availability. In-situ data collection of weld pool morphology (length, width, aspect ratio, etc) can address these shortcomings by providing quantitative metrics for process anomaly detection and first-time-quality improvement. The current leading technology for monitoring weld pool behavior uses side-view or high-angle camera configurations to balance movement obstruction with process field of view. Weld pool imaging is commonly carried out with either standard optical CCD sensors coupled with narrow-band bypass filters or thermal imaging using short-wave infrared (SWIR) sensors. Thermal imaging typically employs temperature thresholding to detect melt pool edges, while visible-band monitoring uses machine learning-based segmentation approaches such as YOLO or edge detection algorithms such as Sobel. While both approaches have been proven in literature, there remains a need for an apples-to-apples comparison of measurement accuracy, robustness, and reliability between each modality for weld pool analytics. Bead-on-plate gas tungsten arc welding (GTAW) experiments were performed using an OPTRIS 08M camera for thermal analysis and a visible-band Xiris XVC-1000 camera under varying welding conditions to compare pool measurements across wire-feed-to-travel-speed ratios and heat input conditions. Synchronized optical and thermal measurements are evaluated to identify where each sensing method is most sensitive to process variation, arc interference, emissivity effects, and segmentation uncertainty. These results contribute to an analytical relationship between process condition and pool morphology while establishing practical guidance for sensor selection in automated welding monitoring. The findings also inform future work toward coaxial weld vision capable of omnidirectional measurement, in-situ anomaly detection, and closed-loop process correction. |