Remaining Useful Life Prediction Of Lithium-ion Batteries Based On Gaussian Process Regression Model

The state of health and remaining useful life of lithium-ion battery is very important for the safety and reliability of system.Achieving accurate and reliable remaining useful life of lithium-ion battery is of great significance to prevent sudden failure of battery,provide maintenance suggestions and prolong the service life of battery.Due to the time-varying,dynamic and nonlinear characteristics of lithium-ion battery system,this thesis carries out the remaining useful life prediction by using the Gaussian process regression model with the ability of uncertainty expression,which will contain the aspects of prediction accuracy,degradation state identification and online applicability.At first,the establishment process and prediction principle of the Gaussian process regression model are elaborated in terms of the weight space theory and the function space method,respectively.The determination of mean function and covariance function is the key to the establishment of the Gaussian process regression model.Then,the posterior distribution can be determined by optimizing the hyper-parameters based on the maximum likelihood estimation and conjugate gradient method.Besides,the establishment procedure of Gaussian process regression model through training data and test data is introduced.Secondly,in order to simulate the actual use of battery,the li-ion battery aging test using randomized discharging current has been carried out.Gaussian process regression model with uncertainty expression ability is proposed based on data-driven methods.After selecting the kernel function,the forecast model is established by training data to optimize hyper-parameters.The data set of charge and discharge tests of li-ion battery under randomized use is used to verify the prognosis results.Then,aiming at solving the difficulty in measuring capacity directly and capacity regeneration during state of health estimation and remaining useful life prediction for lithium-ion battery,a new method is proposed based on time interval of equal charging voltage difference.According to the data sets of charge and discharge tests of lithium-ion battery,the time interval of equal charging voltage difference is extracted during the constant current charge process of lithium-ion battery.With different steps ahead,the state of health estimation is carried out under different initial conditions.The Gaussian process regression model is optimized by using combined kernel functions and particle swarm optimization.The time interval of equal charging voltage difference can act as a health indicator for remaining useful life prediction of lithium-ion battery.The verification experiments are carried out.The results show that the proposed method can predict nonlinear degradation of capacity well and have high prediction accuracy and online remaining useful life prediction ability for lithium-ion battery.Finally,a novel method which combines indirect health indicator and Gaussian process regression model is presented for remaining useful life forecast.Three health indicators are extracted in constant current and constant voltage charge process.Both Pearson and Spearman rank correlation analytical approaches show that the correlations between health indicators and capacity are good.Then,the Gaussian process regression model is optimized with combined kernel functions to improve the ability of predicting the capacity regeneration.Next,based the online health indicator versus cycle number data,three Gaussian process regression models are built and the health indicators prognosis results are achieved at single point.The health indicators prediction results are added in the multidimensional Gaussian process regression model which is accomplished by using health indicators and capacity as input and output,respectively.The predicted capacity is used to compare with the threshold to acquire remaining useful life prediction results.The approach is validated by the two different aging test datasets.Results indicate that an accurate and reliable remaining useful life forecast of lithium-ion battery can be realized by using the proposed approach.