Design And Implementation Of Battery RUL Prediction Model For EMU Based On GAN

The introduction of EMUs in China has a history of more than a decade.With the continuous increase trains amount and the maintenance costs,now staffs have focused on the "economic affordability".Therefore,the basic distinction between health status based on real-time operating data of equipment has become the focus of attention of enterprises.In actual maintenance,over-maintenance of some parts causes waste,and inadequate repair of some parts is prone to safety risks.Even in different seasons and provinces,the situation is different.Therefore,effective equipment health assessment can not only improve equipment utilization but Extending the service life can also avoid wasting resources and reduce unnecessary losses.The battery of the auxiliary power supply system of the EMU is one of the key parts of the EMU.Due to the characteristics of the battery,the battery is also one of the frequentlyoccurring components.Accidents of lifting pantographs occur from time to time.Therefore,the prediction of remaining useful life(RUL)for EMUs is of great significance for the repair and optimization of EMUs.There are various problems with the current general battery health detection methods.Although the method based on electrochemical modelling is accurate in the lab,it is not accurate enough in practical applications due to the continuous change of ambient temperature.The method using the Kalman filter is generally applied to estimate the health of the battery,it is not possible to predict the number of RUL cycles of the battery well.The method of using a single neural network model has good generalization ability.To predict the number of RUL cycles,this thesis uses a data-driven method to build a RUL prediction model and uses many neural network models to get the RUL.In general,the research points of this thesis are:(1)Processing of real-time operation and maintenance data of the battery of the EMU.This thesis is mainly based on the CRH2 EMU and type of EMU can detect the voltage and temperature of the battery in real-time,providing data for the quasi-real-time RUL prediction system propose.At the same time,according to the characteristics of real-time battery charge and discharge data,the advantages and disadvantages of various classical algorithms in PHM(such as electrochemical method,Kalman filter method,differential integrated mobile autoregressive method and neural network method)are studied.Research program.(2)This thesis applies a generative adversarial network(GAN)for image and speech generation to the field of PHM.Using the idea of data augmentation,historical charge and discharge data are used to generate new "future" charge and discharge data to determine the Generate the data,get the evaluation of the next several rounds of charge and discharge data,and get the RUL.(3)This thesis uses multiple neural networks.First,a GAN network composed of a bidirectional long-term and short-term memory network(Bi LSTM)and a convolutional neural network(CNN)is used.Bi LSTM is used to generate data.CNN supervises Bi LSTM based on existing data.The quality of the generated data makes the generated data have the characteristics of the original charge and discharge data.The second step uses known data to train a long-term and short-term memory network(LSTM)to estimate the battery state.Use this network to estimate the battery state of the data generated in the previous step.Repeat this process to obtain the predicted battery data and calculate based on the predicted battery state RUL.(4)To apply the algorithm model proposed in this thesis to the actual production process,a quasi-real-time RUL prediction system is proposed.The main idea is to transfer the battery data to the ground server through the train’s network module or to manually import the train record data into the server for calculation after each run.The distributed file system Hadoop is used to store the historical battery data and the trained model,and the distributed training environment based on Tensorflow is used to train the model in an offline environment,and the discriminative model uses the historical battery data and the data generated by the EMU in real-time The combined time-series data is used to determine the RUL of the battery.The design of this system uses the current mainstream storage engine and training inference engine,taking full account of expansibility,other parts of the train can also use this system to do some necessary calculations.