State Of Health Estimation For Lithium-ion Batteries With Gaussian Process Regression

As lithium-ion batteries play a more and more important role in power,consumption and energy storage,the estimation of their state of health(SOH)has attracted more and more attention.An accurate estimation of SOH is of great significance for battery maintenance,electric vehicle range assessment,cell balance in battery pack,and so on.In order to obtain an accurate SOH estimation of lithium-ion battery,the following three parts are performed in this thesis.Firstly,this thesis built an experimental platform for battery degradation.Three data sets with different working conditions are used in this thesis,including the battery degradation data collected at room temperature,constant temperature,and a public data set with diverse depths of discharging.In detail,the three data sets have violent capacity fluctuation phenomenon caused by diurnal temperature difference,a uniform capacity degradation phenomenon under constant temperature,and capacity regeneration phenomenon,respectively.Therefore,the estimation effectiveness of our method can be verified under various working conditions.Besides,for the problem that the cycle numbers is not suitable as inputs of the model,Two features are extracted based on the constant current charging curve,one of which is based on the dynamic time warping algorithm.These two features not only require small data,but also are simple to extract and easy to expand,which can avoid the subjectivity of feature extraction.Then,based on these two features,the effectiveness of Gaussian process regression for SOH estimation are compared on several kernel functions,and the dot-product square kernel is finally used for Gaussian process regression to estimate the SOH of lithium-ion batteries: the relative error is less than2% on most batteries.Finally,due to the complexity of traditional kernel function determination and poor SOH estimation effectiveness with small training samples,a method combining basic kernel and deep kernel learning is proposed,and the strong fitting ability of neural network is used to automatically learn kernel function of Gaussian process regression.The SOH estimation results show that the combination of deep kernel learning and gaussian process regression not only has a SOH estimation effect close to that of traditional Gaussian process regression when the number of training samples is sufficient(50%),but also has a satisfied SOH estimation performance when the number of training samples is small(25%): the root mean square error is less than 0.025.This thesis collected lithium-ion battery degradation data at room temperature and constant temperature,and then extracted a feature based on dynamic time warping.Finally,the method of gaussian process regression combined with deep kernel learning is used to perform SOH estimation.Compared with the fixed kernel gaussian process regression method,this method has achieved better SOH estimation results under low training threshold.