Research On Image Detection Of Hidden Dangers In Railway Passenger Car Bogies Based On Deep Learning

The railway passenger car bogie,as the main component of the carrying car body,is a key factor to ensure the safety of passenger car operation,so it is particularly important to carry out dynamic hazard detection on it.At present,my country has established a visualized TVDS system for passenger cars covering the entire railway network with the help of high-speed cameras.Train inspectors need to find hidden dangers from the captured images.This method requires them to complete thousands of hidden danger detection tasks in a short time.Train inspectors have extremely high requirements for their business capabilities and long-term manual operations cannot guarantee the detection accuracy.Therefore,it is a very meaningful work to use the target detection technology in computer vision to automatically detect hidden dangers in captured images.In this paper,five types of hidden dangers are obtained by sorting the collected TVDS images: bolt loss,cotter pin loss,disconnection of connecting line,component oil leakage and entrainment of foreign objects.This is the research object combined with deep learning target detection algorithm to achieve hidden danger images.automatic detection.The main work is as follows:(1)A large number of samples was required when training deep learning algorithm.In this paper,data enhancement was used to expand the hidden danger samples to meet the needs of model training.The expanded 2226 hidden danger samples were made into Pascal VOC format and named TL＿Dataset data set.Through the experiment,the Faster R-CNN algorithm has a mean average precision of 74.94% for the average of the five types of hidden components,which is 17.35% higher than the model Map value trained in the original data set.(2)In view of the poor detection effect of the two types of small targets,bolts and cotter pins,the Faster R-CNN algorithm was optimized from three aspects.Firstly,the reference network in the algorithm was replaced with Det Net-59 and the output of the last 4 layers of the network was fused into a multi-scale feature map;Secondly,the K-means++ algorithm was used to cluster the labeled boxes and the results was obtained as the new anchor box scheme;Finally,Pr ROI pooling was used as the pooling operation of the region of interest.The optimized algorithm in this paper was used to experiment on the test set,and compared with the original algorithm,the detection average precision of bolts and cotter pin hidden dangers increased by 38.02% and8.26%,respectively,the mean average precision of the five types of hidden dangers increased by 11.2%.(3)Based on the optimized Faster R-CNN algorithm,the Django framework is used to design a detection and early warning system for passenger car bogies.According to the research situation,the system is designed with 5 main modules and a "double confirmation" mechanism to realize the automatic detection and hidden danger management of the images taken by TVDS.Finally,according to the relevant regulations,the hidden danger level is divided,and the corresponding hidden danger level treatment plan is proposed.This paper uses the Faster R-CNN target algorithm based on deep learning to realize the automatic detection of 5 types of hidden danger components in the area of passenger car bogies: bolt loss,cotter pin loss,disconnection of connecting line,component oil leakage and entrained foreign body,and rely on the optimized algorithm A detection and early warning system for hidden dangers of passenger car bogies has been built,which provides research ideas and technical means for the subsequent improvement of automatic and intelligent detection of hidden dangers of railway passenger cars.