Research On The Estimation Of The State Of Health Of The Vehicle Power Battery Based On Deep Learning

In the working process of power batteries for vehicles,the internal substances,including cathode and anode materials,electrolyte and others,have different degrees of irreversible chemical and physical changes,which are reflected in state of health deterioration outside the batteries.Quick and accurate estimation of the state of health of power batteries is of far-reaching significance for tapping the potential of battery performance and ensuring the work and safety performance of batteries and vehicles.Based on the experiments of cyclic life,hybrid dynamic working condition and AC impedance of vehicle power battery,this paper analyses the characteristics of battery decay process,regarding the state of health estimation for battery as a time series prediction problem,and proposes a health state estimation model for battery based on long short term memory neural network.After that,the updating process of backpropagation parameters of standard network model is optimized effectively,and trained and validated by different battery experimental data.The results show that the long short term memory neural network based on gradient descent learning rate self-adaptive optimization has a significant effect in the state of health estimation,under 5%,which can well meet the actual needs for battery health state estimation.The main contents are as follows:(1)To analyze the practical significance of battery state of health estimation,summarize the advantages and disadvantages of current battery modeling and health state estimation methods,and analyze the feasibility of applying deep learning algorithm to battery health state estimation.The working principle and characteristics of power batteries are described in detail,and the internal and external factors affecting the deterioration of batteries are analyzed.(2)Complete the power battery decay experiment and carefully analyze the characteristics of the experimental data.Battery experiments include battery cycle charging and discharging experiments,hybrid power pulse characteristics tests,battery constant capacity experiments,hybrid dynamic working conditions experiments and AC impedance experiments.Based on the results of battery experiments,the characteristics of power batteries during decay were analyzed.(3)Establish the long short term memory neural network model for health state estimation of batteries.Long short term memory neural network belongs to the branch of deep learning algorithm model.As an evolved recurrent neural networks,the internal " memory block " function of the long short term memory neural network is the core unit to store the historical information of the input data.It solves the "long-term dependence" problem of recurrent neural network and performs well in time series prediction.In this paper,the long short term memory neural network is applied to the prediction of battery health status,and its prediction performance is verified by experimental data.(4)The long short term memory neural network prediction model based on gradient descent learning-rate adaptive optimization is proposed.This paper briefly analyses the problem of constant learning rate in the process of back-propagation updating parameters of standard long short term memory neural network model.The original updating process is changed into an adaptive optimization process,which makes the optimized long short term memory easier to find the global optimal solution.The predictive performances of three kinds of self-adaptive optimized long short term memory neural network models are compared and analyzed,which are verified by the experimental data of batteries.The results prove that it is suitable for battery health estimation.