Abstract Surface defects are basically early indicators of welding quality that provide insights into potential issues in the welded joints. In arc welding, particularly for the fusion of thin metal sheets, weld defects such as cracks and burn-throughs occur quite often because the material and its thickness are prone to stresses caused by heat distortion and rapid cooling. Manual detection of welding defects through the different inspection methods is resource consuming. Automatic detection is, on the other hand, challenging due to the dynamic nature of the welding process and the subtle feature-based differences between acceptable and defective welds. To address this, a thermography camera is employed in this research to capture thermal variations during the welding process. The study uses thermographic imaging during the process of TIG welding and observes four different families of deep-learning detectors (YOLOv5, YOLOv8, YOLO11, and RT-DETR) across all size variants for each family. A singular thermal dataset was generated that represented the signs of burn-throughs, cracks, and burn spots, which were all consistently annotated without bias. Each model was trained and evaluated using precision, recall, mean average precision (mAP), and inference latency. YOLOv8 demonstrated comparable performance levels in the evaluation. YOLOv5 also performed robustly, but at a slower pace, RT-DETR had better feature extraction but also had the highest inference latency, and Yolo11 mistook and missed many small low-contrast defects. The results validate that thermography combined with modern deep-learning detectors is viable for automated detection and monitoring of weld quality.
Parmar et al. (Fri,) studied this question.