Defect Detection of Axle Box Cover Device Fixing Bolts in Metro Based on Convolutional Neural Network

Due to the heavy load and dark, humid working environment of the metro, its components are prone to failure, which affects the safety of operation. In this paper, a set of fault detection algorithms based on the combination of Faster R-CNN and one-class convolutional neural network (OC-CNN) is designed for the axle box cover device fixing bolts (ABCDFBs) of metros. First, the ABCDFBs are located by Faster R-CNN algorithm, and then a classification model trained only with positive samples of bolts is used to determine whether the bolt image is normal or not. This algorithm solves the problem of insufficient negative samples in the actual scene, and has good robustness to complex environmental conditions such as dust, water stains, and trial-varying light. Based on the images of the axle box cover device collected by the high-precision linear array camera on the scene, experiments show that the algorithm performs well in speed and accuracy.