Bolt Positioning Detection Based on Improved YOLOv5 for Bridge Structural Health Monitoring.

Area of Science:

• Civil Engineering
• Computer Vision
• Structural Health Monitoring

Background:

• Bridge stability is crucial, with bolts identified as key failure points causing symmetry issues.
• Detecting small-scale bolts in complex bridge imagery presents challenges due to limited feature expression.

Purpose of the Study:

• To develop an improved YOLOv5-based bolt positioning detection system for enhanced bridge structural health monitoring.
• To address challenges in bolt detection, including scale, background complexity, and feature representation.

Main Methods:

• Optimized anchor box calibration using K-means++ clustering for improved object detection.
• Implemented hypercolumn (HC) technique to fuse multi-scale features for better detection accuracy.
• Created a dedicated bridge bolt dataset comprising 1494 images for training and validation.

Main Results:

• The improved YOLOv5x model achieved a precision (P) of 87.3% and an average precision (AP) of 86.3%.
• These results represent a significant improvement, with precision increasing by 6.5% and AP by 5.9% compared to the original YOLOv5x.
• The enhanced method demonstrates superior performance in detecting bridge bolts.

Conclusions:

• The proposed improved YOLOv5 method effectively enhances bolt detection accuracy and speed for bridge structural health monitoring.
• The integration of K-means++ and hypercolumn techniques offers a robust solution for identifying critical structural components.
• The developed dataset and validated model provide a valuable resource for future research in infrastructure monitoring.