Automatic detection of arbitrarily oriented fastener defect in high-speed railway

Abstract

In high-speed railways, contact forces between the pantograph and the overhead catenary system (OCS) are substantial. Vibration and excitation propagated from the vehicle-track and the pantograph-OCS interactions would progressively damage fasteners, including but not limited to, loose bolts, cracked components, and missing parts. Existing automatic detection methods typically rely on a three-stage approach, of which the first two stages focus on locating joints and fasteners while the last stage focuses on the detection. Due to the nature of the three-stage detector, the computational cost is high, and the inspection speed is low. This study proposes an innovative two-stage method with two improved convolutional neural network (CNN)-based networks, cascade YOLO (You Only Look Once) and Rotation RetinaNet (RRNet). The proposed method was compared to traditional horizontal anchor-based methods and other methods. The results demonstrate the proposed method outperforms other methods in terms of accuracy, while maintaining a reasonably high processing speed.