The detection of welding defects in cranes presents challenges such as multi-scale analysis, task coupling, and complexities in learning from samples. This study introduces the YOLO-Rbw model, which is built upon YOLOv8 and incorporates three key innovations. The RepELAN module employs a multi-branch cascade and reparameterization to balance efficiency and characteristics. The BiConvHead dual-branch detection head mitigates task coupling, while the morphologically aware WIoU loss dynamically adjusts weights to enhance efficiency in handling challenging samples. Experimental results demonstrate that the proposed model achieves an mAP50 of 92.7%, mAP50-90 of 54.9%, and a recall rate of 88%. These metrics represent improvements of 18.3%, 18.4%, and 17.9% over YOLOv8, and 16.4%, 13.3%, and 14.8% over YOLOv11, respectively, while maintaining a processing speed of 189 frames per second and a parameter count of 9.3 million to meet industrial requirements. This research addresses issues related to minor missed detections and implicit complexities in identification, offering a versatile framework for integrating high-precision and lightweight defect detection in industrial settings.