Deep learning‐based bolt loosening detection for wind turbine towers. Issue 6 (18th February 2022)

• Record Type: Journal Article
• Authors: Yang, Xinyue; Gao, Yuqing; Fang, Cheng; Zheng, Yue; Wang, Wei
• Is Part Of: Structural control and health monitoring. Volume 29:Issue 6(2022)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2022-02-18
• Subjects: bolt loosening -- computer vision -- deep learning -- synthetic datasets -- wind turbine tower; Structural engineering -- Periodicals; Structural control (Engineering) -- Periodicals; Automatic data collection systems -- Periodicals; Detectors -- Periodicals; 624.17
• DOI: 10.1002/stc.2943
• Languages: English
• ISSNs: 1545-2255

Abstract

Summary: Bolt loosening is a critical factor that triggers collapse of wind turbine tower structures, and a fast and accurate bolt loosening detection is of significant importance. This paper proposes a two‐stage detection framework, which combines the traditional manual torque method commonly used in engineering with the deep learning model to reduce the cost of manual inspection and the rate of missed inspection. A graphical model with bolt loosening marks was used to collect synthetic datasets, reducing the time required to collect training and test sets in real environments. The You Only Look Once (YOLO)‐based deep learning models (YOLO v3, YOLO v4, YOLO v4‐Tiny) were trained and tested using the same synthetic datasets. The test results showed that the average precision of the YOLO v3 was 0.9571, and the detection time of a single image was 0.024 s, outperforming other versions of YOLO models in bolt loosening detection. Subsequently, a test specimen was manufactured, and images obtained from different capture distances and perspective angles, and considering different resolutions, lighting conditions, and bolt stain conditions, were identified separately. The results demonstrated that the recommended deep learning model could identify bolts with only 2° of looseness under different environmental conditions. In addition, the method was verified by a video shot recorded by a smartphone to examine the autonomous detection ability of the YOLO‐based models. The study shows that the proposed detection framework has advantages in detection accuracy, detection speed, and generalization ability and has strong application advantages.