Fault Diagnosis of Train Clamp Based on Faster R-CNN and One-class Convolutional Neural Network

Abstract

In recent years, Chinese high-speed railway ushered in a great development. With the high-speed railway operation gradually gets busy, the traditional method of relying on manual inspection of train fault has been unable to keep pace with the pace. As a key part of the train, the rod and spring components of clamp is essential for the safe and smooth operation of the train. In this paper, a novel method combining Faster R-CNN and One-class Convolutional Neural Network (OC-CNN) is proposed for fault diagnosis of the clamp part on train. Firstly, the rod and spring on the clamp part are located by Faster R-CNN, and the rod component is detected to determine whether there is any abnormality. Meanwhile, the spring area is cropped from the clamp part picture and resized as a fixed size. Then, the image contains spring area is feeded into the OC-CNN algorithm which is trained by positive samples and fine tuned by negative samples to determine whether there are cracks in the spring. Through specific experiments, the conclusions show that this method is effective and it surpasses the other three types of combined methods, namely You Only Look Once version-4-tiny(YOLOv4-tiny) and OC-CNN, Single Shot Multibox Detector 512 (SSD512) and OC-CNN, as well as Nanodet and OC-CNN.