Vision-Based Detection of Bolt Loosening Using YOLOv5

Area of Science:

• Structural Engineering
• Machine Vision
• Deep Learning

Background:

• Bolted connections are critical in engineering structures but prone to loosening under cyclic loads.
• Detecting bolt loosening is essential for structural integrity and safety.
• Existing methods may lack accuracy or robustness in diverse environments.

Purpose of the Study:

• To propose and validate a novel method for detecting bolt loosening using deep learning and machine vision.
• To accurately measure the rotation angle of nuts relative to bolts to identify loosening.
• To assess the method's robustness across various shooting conditions.

Main Methods:

• Developed a bolt loosening detection system utilizing the YOLOv5 (You Only Look Once, version 5) deep learning model.
• Integrated machine vision by attaching distinct circular markers to the bolt and nut.
• Calculated nut rotation angles based on the center coordinates of detected markers.

Main Results:

• Achieved high precision (99.8%) and recall (100%) in detecting markers and assessing bolt status.
• Demonstrated the ability to detect minimal loosening angles as small as 1 degree.
• Verified robustness across different shooting distances, angles (up to 45° tilt with 5.91% error), and lighting conditions.

Conclusions:

• The proposed YOLOv5-based method offers a highly accurate and robust solution for early-stage bolt loosening detection.
• This machine vision approach effectively identifies subtle nut rotations, crucial for preventing structural failures.
• The method's performance in varied environments highlights its practical applicability in real-world engineering scenarios.