Deep Learning And Fault Detection For Operation And Maintenance Of Bullet Train

With the continuous increase of the number of EMU,the task of fault detection of EMU is becoming larger and larger,and the difficulty and requirements of detection are also becoming higher and higher.Therefore,it is necessary to develop a set of unmanned or less manned fault detection system for EMU.At present,the process of EMU storage maintenance requires the staff to rely on the traditional inspection methods such as eye observation,flashlight auxiliary lighting industrial camera recording,ruler measurement and so on to detect the faults of EMU.Due to the influence of uncontrollable factors such as position change,human fatigue and illumination conditions,there are a lot of false detection and missed detection.In recent years,target detection algorithms based on deep learning have achieved unprecedented development,and the application of deep neural network target detection algorithms such as Faster R-CNN to EMU fault detection is also rising gradually.Compared with traditional target detection algorithms such as SVM vector machine,Faster R-CNN algorithm can extract more abundant features from sample data,and the network model obtained after training is more efficient and robust for fault detection.In order to make the Faster R-CNN algorithm meet the requirements of EMU fault detection application scenarios,this paper makes some improvements.In order to solve the problems of low efficiency and long time in the training process of Faster R-CNN algorithm in a single GPU,multi-GPU data parallelism was introduced to accelerate the training,and the asynchronous random gradient descent algorithm was improved by adding parameter copies in the CPU.On this basis,data transmission between GPUs was managed by data parallelism.In addition,a thread responsible for transmission and computation is created in the GPU to realize parallel tasks.Finally,the acceleration effect of the thread on the training process is verified by comparing the single GPU with the double GPU.In order to solve the problem that Faster R-CNN algorithm is not ideal for fault detection of EMU images,a method based on Inception-Res Net basic feature extraction network and small target optimization is proposed.First,the convolution kernel of different sizes is used to calculate the input from the input layer,and then the resulting calculation results are output jointly with the input from the input layer.Finally,the output of each Info-Res Net module is deconvolved to produce a multi-feature graph of the same size,which is input to the ROI layer and the RPN network.The time-consuming comparison experiment of network model training on PyTorch deep learning framework and the comparison experiment of fault image detection effect after algorithm improvement show that the optimization and improvement proposed in this paper has better detection effect than before.