Research On Anomaly Detection Of Electric Locomotive Bottom Based On Deep Learning

With the continuous updating of China’s railway electric locomotive(hereinafter referred to as locomotive)models,the continuous improvement of locomotive performance and the increase of locomotive operation density,it is increasingly difficult to detect locomotive safety.In view of this practical problem,China has independently developed the train coach machine vision detection system for locomotive fault.The application of this system greatly improves the efficiency of locomotive safety detection,but the locomotive fault detection system currently put into use in China can only realize the image acquisition and transmission of the locomotive,and ultimately requires manpower to detect and analyze all parts of the locomotive,which costs a lot of cost and has low detection efficiency.In order to realize the automatic detection of locomotive bottom anomaly,this paper proposes a method of locomotive bottom anomaly detection based on deep learning,which takes the crack and loss of components as the object of anomaly detection,the main work and research contents are as follows:At present,there is no public abnormal data set of locomotive bottom in the industry,and the method based on deep learning needs a large number of data samples to train the network model.Therefore,this paper uses some pictures of locomotive bottom provided by the local railway bureau and relevant pictures searched from the Internet to build the data set,then preprocesses the pictures to improve the data quality,and then uses them to improve the data quality,at last,the dataset is expanded by techniques such as translation and mirror image.On the basis of theoretical analysis and a large number of experimental comparison of a variety of detection models,this paper chooses to use retinanet as the basic network model.At the same time,in order to further improve the detection accuracy of retinanet network and make it meet the actual detection needs,this paper puts forward the improvement of retinanet network.Firstly,two branches are introduced into the residual module of the original feature extraction network to strengthen the feature transmission and the feature extraction ability of the backbone network,to make the model more effective use of shallow features,and the introduction of attention mechanism to enhance the weight of important features;secondly,through the introduction of new loss function and anchor box optimization method to further improve the detection effect of abnormal targets;finally,to verify the detection performance of the improved network model,the introduction of new data sets for training and testing.Finally,the experimental results show that the network model used in this paper performs well.On this basis,the detection accuracy of the improved network is better than the current mainstream model,which has a certain application value.