Research And Application On State Of Health And Remaining Useful Life Prediction In Energy Storage Devices

With the rapid promotion of science and technology,tremendous new electronic and electrical products have emerged,and various kinds of daily electronic equipment have exploded,significantly increasing the demand for energy storage devices.That leads to energy storage devices’ application scope expanding to new energy vehicles,airplanes,high-speed railways,buses and other high power transportation equipment with the growth in energy density,safety and service.However,the usage and abandonment of energy storage devices will affect the non-renewable resources and the environment.It is inevitable that when in use and storage,the internal physical and chemical characteristics will be changed,which might cause leakage,spontaneous combustion,explosion,and other safety accidents.The remaining useful life(RUL)of energy storage devices and fault prediction has provided critical technologies to the battery management system(BMS).As a result,the growing deployment requirements of BMS have been raising the cost of data acquisition,the storage of control systems,and the cost of data analysis due to the massive amount of data generated.However,the conservative maintenance strategy of the low accuracy of medium and long-term life predictions results in high maintenance costs.Therefore,the primary purpose of this study is to accurately predict the health status,failure of energy storage devices,provide support for scale monitoring and echelon utilization of energy storage devices by using the minimum parameters and data of energy storage devices.Simultaneously,the high precision prediction for the long and medium life can ensure the safe operation of the system and reduce the maintenance cost.Given some problems existing in current research methods and applications of lithium-ion battery RUL and faults prediction,this dissertation starts to solve them from the following aspects:(1)Given the current data-driven prediction method,it is necessary to use a variety of measurement parameters of energy storage devices to predict their health status.In this dissertation,the charging voltage data,which can represent the degradation of energy storage devices,is selected as the research object through data cleaning of the parameter data of energy storage devices.The feature matrix of parameter reconstruction is analyzed by deep learning methods with a single model and a composite model,including one-dimensional and two-dimensional feature extraction.The feature matrix is constructed to maximize the time series’ s depth and surface features.The matrix overcomes the problem that the health state of energy storage devices cannot be predicted by using the deep learning method with a single parameter and realizes the high-precision health state prediction of energy storage devices.(2)At present,the model training leads to massive data and time consuming,which is not conducive to healthy deployment and application.The characteristic periodic replacement components are found by analyzing the eigenmatrix’s periodicity and variability.The principal component analysis is performed to find the characteristic principal component replacement matrix.After data fusion of the two maximal feature sequences shows that the prediction accuracy is higher than the original data,and the convergence speed of data training is greatly improved.After data fusion,only 6.7% of the data is used to achieve the prediction accuracy of the original data,while 94% of the training time is reduced.(3)For long-term life prediction,the particle filter and the unscented Kalman filter rely on pre-trained models and parameters,which are prone to error accumulation in long-term data prediction.However,long and short-term memory(LSTM),which has advantages in time series prediction in deep learning,has a poor prediction effect due to the large entropy of data information and wide data fluctuation range.In this dissertation,the UIDS-GRU prediction method,which integrates the advantages of the three methods,is proposed to reduce the data uncertainty and control the prediction trend.This method does not rely on the degradation model and can predict the mid-term and long-term degradation data accurately by extracting historical data features with high accuracy.The universality of the algorithm is verified on different materials and different capacities of energy storage devices.(4)To prove the feasibility of the previous research on high-speed trains and provide a research basis for the intelligent detection of train energy storage systems.In this dissertation,short-term health status prediction and medium and long-term RUL prediction methods of energy storage devices are deployed to edge server and cloud server.The combination of two approaches can predict the remaining service life of the energy storage device in advance for targeted maintenance and give early warning when abnormal behavior occurs,ensuring the safe and reliable operation of the system.The comparison demonstrates that the short and long-term combination forecasting approach can shorten the transfer time,reduce the memory storage,save the computing resources and verify the advantage at the edge of the calculation.