Automated vision-based bolt loosening detection and localization for operating hydraulic turbine rotors

Abstract

Turbine rotors are susceptible to bolt loosening and detachment due to factors such as vibration, stress concentration, and material degradation during the prolonged operation of hydropower units. These issues can result in significant safety incidents and economic losses. This article presents a novel vision-based automatic detection method for identifying loose bolts in turbine rotors by analyzing the angle between marking lines on the top surfaces of studs and nuts. The proposed method integrates enhanced you only look once version 8 (YOLOv8) for object detection, DeepSORT for multitarget tracking, DeepLabV3+ for image segmentation, and Theil-Sen estimation for line fitting. This approach effectively locates loose bolts on the rotor and determines the loosening angle. Experiments were conducted during the commissioning of a large hydroelectric unit, yielding results that demonstrate an average loosening angle calculation error 0.633°. With a loosening threshold established at 2°, the method achieves a detection accuracy of 98.88%. The proposed method does not depend on the historical states of the bolts and provides precise identification of bolt loosening angle up to 180°, thereby making it suitable for application to operating turbine rotors.