Research On State Detection Of Fastener Springs Based On Convolutional Neural Network

With the rapid development of China's economy and the growing demand of people's traveling,the operating mileage of China's railway increased from 52000 km of 1978 to 132000 km by the end of 2018.By 2019,the operating mileage has increased to 139000 km.In the process of China railway's rapid development,it is particularly significant to detect the status of infrastructure for the sake of ensuring railway's safe operation.As the connector of fixed rail and sleeper,the status of fastener directly affects the driving safety,and that's why the detection is one of the most important of the whole infrastructure safety detection.However,the detection method of railway fastener status in China mainly depends on the manual inspection of railway workers.With China's increasing railway mileage,the disadvantages of manual detection become obvious: pretty low detection efficiency and high labor intensity.Owing to heavily depending on the technical proficiency of the test workers which with high rate of missed inspection,the reliability of the results is not authentic.In order to highly improve the efficiency of detection and efficiently make use of the early work,the thesis proposes to establish the railway fastener state detection model based on convolution neural network combined with deep learning technology.Grounded on the convolution neural network detection method,the network can automatically learn the features of fastener image data instead of manual work supporting with enough fastener image data.This method not only effectively saves time,but also greatly improves the accuracy of detection.The main work of this thesis is as follows:1.Designing the fastener image acquisition system that meets the requirements of experiment.During this period,the thesis compared the advantages and disadvantages of the camera components required in the acquisition system and selected the right type of linear array light source for auxiliary lighting,in order to make the acquisition vehicle adapt to different environments(open air and in tunnel).Finally,through combining encoder and the wheel axle of collecting vehicle,the correct camera trigger scheme was determined.2.Data collection and preprocessing: a large number of fastener image data are collected by collecting vehicle.According to the principle that the fastener in a single photo must be taken completely and the fastener position shall not be at the edge of the image,3150 fastener photos meeting the requirements are screened.The source of fastener image acquisition comes from a track line in Shijiazhuang.Among the 3150 fastener images,2268 show images of normal fastener and 882 show abnormal fasteners.In case of the problem of category imbalance of data set,later we expanded the fastener images to 5300.3.In the light of fastener feature extraction,several common image feature extraction technologies are compared and analyzed,including the feature extraction based on hog operator,LBP Operator and convolution neural network R-CNN?Fast R-CNN?Faster R-CNN?YOLO).By comparison and combining the characteristics of fastener photos studied in this thesis,the appropriate feature extraction method is finally selected.Based on the previous research,this thesis determines the feature extraction method of fastener image.It starts with introducing the network structure of the Yolo algorithm,and then designs the network structure of the fastener state detection algorithm based on the Yolo.Later Yolo network is adopted for training and later detection.At last,the test results of Yolo network and other two networks are compared by experiments.4.On this basis of previous study,the Yolo network is improved and optimized mainly by adding multi-scale feature fusion and feature channel weight separation to make feature extraction more reasonable.In order to improve the detection speed of the algorithm,this thesis makes a pre-clustering analysis of the data set.Finally,through the experimental comparison,it comes to the conclusion that the optimized network further improves the performance of fastener detection.The data shows that the optimized Yolo network detection results show that the map reaches 96.7%,the precision rate reaches 96.3%,the recall rate reaches 95.2%,which is 3.9%,2.2% and 11.9% higher than that before the network improvement.