Fault Detection Of EMUs Based On Deep Learning

With the increase of the number and speed of high-speed multiple units,the task of fault detection of EMUs becomes more and more serious,and the hidden safety problems become more and more obvious.Therefore,it is necessary to develop an intelligent fault detection system for EMUs.Nowadays,people can use high-speed cameras to obtain high-definition and stable images of visible parts on the surface of motor vehicles,but a large number of staffs are required to find faults from these images.Target detection is an important research direction in the field of computer vision.Its main task is to identify and locate interested objects from images or videos.By using this technology,faults can be detected and classified from the image of the visual part of the multiple units.However,the traditional target detection algorithm has a large number of missed detection and false detection,so the technology has not been applied in practice.In recent years,deep learning has developed unprecedentedly,and target detection algorithm based on deep learning has begun to rise gradually.Compared with the traditional target detection algorithm,the deep learning method does not need to extract features manually,but learns to extract features from a large number of sample data.The features obtained are richer and more robust.It is far beyond the traditional target detection algorithm in terms of detection accuracy and speed.Therefore,in-depth study of target detection algorithm based on deep learning,and then applying to fault detection of EMUs is a very meaningful work.The purpose of this paper is to apply the current mainstream algorithm based on deep learning to the fault detection of EMUs,and then improve and optimize the algorithm according to the actual experimental results to improve the detection accuracy.The main work is as follows:(1)To sort out the existing target detection algorithms based on deep learning,it can be roughly divided into two ideas: one is based on candidate regions,which first extracts candidate regions,then classifies and locates them;the other is based on regression,which directly obtains the target category and location information.The latter has more advantages in detection speed,the former has higher detection accuracy and is more suitable for this topic.(2)Faster R-CNN,a target detection algorithm based on candidate regions,is deeply studied in this paper,and the training experiment is carried out on the self-made automobile fault image data set.It is found that the improved basic feature extraction network can optimize the algorithm.Inception network and deep residual network are the best performance feature extraction network at present.This paper combines the ideas of the two networks and designs a new basic network.By fusing it with Faster R-CNN,the detection accuracy is further improved.(3)To improve the accuracy of small target detection,multi-level feature fusion,multi-scale input and multi-anchor prediction are introduced to optimize the algorithm.