Bolted connections are widely adopted as primary structural joints in engineering infrastructure. However, conventional manual inspection remains labor-intensive and time-consuming. Deep learning–based automated defect detection faces significant challenges due to pronounced image variability induced by variable camera angles, lighting conditions, partial occlusions, and complex backgrounds. This study addresses these issues by constructing a diverse bolt image dataset compiled from three sources: on-site acquisitions from suspension bridges, field photography of steel transmission towers, and controlled laboratory imaging of a custom-fabricated bolt assembly model. To enhance data robustness, we employed image enhancement techniques and generative adversarial networks (GANs) for data augmentation. A comparative analysis was conducted among three mainstream object-detection models—YOLOv5, YOLOv8, and YOLOv10—using recall, precision, and mean average precision (mAP) as evaluation metrics. Building upon the superior performance of YOLOv8 (mAP = 0.91, recall = 0.85, precision = 0.9), we proposed an enhanced architecture integrating a Swin-Transformer backbone and a novel Multi-Scale and Detail-Enhanced Module (MEDM) to specifically improve missing-bolt detection in challenging visual contexts. The improved model demonstrated consistent accuracy across diverse scenarios: 100% at 15°, 30°, and 45° viewing angles; > 94% under 30%, 50%, and 100% illumination levels; and > 97.2% for colored coatings (blue, red, white) against complex grassy and mixed backgrounds. In practical engineering deployment, the model achieved a 98.94% detection rate across 12,772 bolt sets, successfully identifying one instances of missing bolts. These findings validate the proposed approach’s effectiveness for real-world structural health monitoring.
Gu et al. (Mon,) studied this question.