SAN-YOLO for Boiler Weld Defect Detection in Phased-Array Ultrasonic S-Scan Images

Abstract

Weld defect detection is critical for ensuring welding quality, and object detection has become an effective approach for localizing weld-seam defects. To reduce missed detections of small defects and address the significant scale variation in defects in ultrasonic phased-array S-scan images of boiler welds, this paper proposes SAN-YOLO, an enhanced model based on YOLOv8n. An SPD-Conv module is adapted to the backbone to preserve fine-grained features of minute defects. In addition, an Adaptive Scale Fusion (ASF) module is adapted to the neck to integrate Scale Sequence Fusion, Triple Feature Encoding and channel-and-position attention, thereby enhancing multiscale defect perception. Furthermore, a Morphology-Guided Normalized Wasserstein Distance (MG-NWD) loss is proposed to dynamically balance geometric and NWD-based regression constraints for each foreground sample according to its matched defect category, bounding-box scale and elongation, training progress, and current localization quality. Experiments on a proprietary boiler-weld dataset show that the proposed SAN-YOLO achieves a precision of 92.1%, a recall of 94.0%, and an mAP@0.5 of 88.2%, representing improvements of 0.7%, 7.0%, and 6.8%, respectively, over YOLOv8n. These results demonstrate the feasibility and potential of SAN-YOLO for automated defect detection in boiler-weld PAUT S-scan images.