Bolt loosening angle detection using CNN and transformer fusion network with a single inference stage
Zhi Shan, Haolei Dou, Fanghao Wu, Zhiwu Yu · Smart Materials and Structures · 2025
Abstract Bolt looseness detection plays an indispensable role in safeguarding structural safety. Vision-based looseness detection methods have attracted extensive attention due to non-contact operation and low cost. However, existing methods typically involve multiple steps, leading to inefficiency and compromised real-time performance. This paper proposes an innovative bolt loosening angle detection method with a single inference stage. The method introduces a newly defined bolt reference angle (indicating rotational state) as an additional regression parameter into object detection via two innovations: (a) the BoltTransNet with multi-head self-attention for enhanced feature extraction, and (b) the first-ever bolt reference angle regression method. The bolt loosening angle is determined by the change in bolt reference angle. The method eliminates the need for multiple steps by directly outputting bolt locations and bolt reference angles in one pass (0.015 s per image). The accuracy of the method is verified by detecting loosening bolts in the lab. The robustness of the method is demonstrated by testing bolts under various capturing perspectives, lighting conditions, and backgrounds. The average error of bolt loosening angle detection was 4.57°. Hence, the method holds significant potential for autonomously detecting bolt loosening angles.