Weld Defect Identification and Quality Evaluation for Modular Steel Construction Via YOLO with Loss Analysis

Modular steel construction is gaining widespread adoption due to its efficiency, cost-effectiveness, and reduced construction time compared to traditional methods. However, ensuring weld quality remains a critical challenge for structural integrity. This study presents a defect detection system for modular steel welds using the YOLOv10 object detection model. A publicly available dataset from Roboflow Universe, enriched with augmentation techniques, is used to enhance model training. To address class imbalance, a Weighted Dataloader framework is implemented. In addition, the impact of multiple IoU-based loss functions, including Generalized IoU (GIoU), Distance IoU (DIoU), Complete IoU (CIoU), Efficient IoU (EIoU), SCYLLA IoU (SIoU), Minimum-Point-Distance IoU (MPDIoU), and Focal-EIoU, is investigated. Experimental results show that integrating the Focal-EIoU loss with YOLOv10s significantly improves performance, achieving a precision of 77.8 % and a mean Average Precision (mAP) of$\mathbf{8 0. 4 \%}$. The proposed system outperforms YOLOv8 and YOLOv9 in ablation studies. This work advances automated weld inspection and supports future application development for weld quality assessment in modular steel structures.