A Novel Data-driven Method For Remaining Useful Life Prediction Of Batteries

With the increase in the global population and economy,the demand for energy is increased.Because people are facing the great challenge of pollution,the lithium-ion batteries start to play a vital position in the new energy development.However,the state of health of batteries deteriorate rapidly.As a result.in order to increase the reliability of battery system,it is necessary to estimate the state of health of battery to ensure the safety of battery system.In this paper,we propose a data-driven framework of remaining useful life prediction of batteries.First,we analyze the degradation curves of lithium-ion batteries.A novel outlier detection method is introduced to solve the problem that there are many outliers in the original data.The result of experiment shows that the proposed model can remove the outliers in the original data and improve the quality of data.Secondly,we propose three metrics to quantify the randomness in the degradation process of lithium-ion batteries.Previous research mainly focus on the establishment of models,not the source of the error.Therefore in this paper we analyze this phenomenon.The result shows that degradation process of lithium-ion batteries is random to some extent.It is hard to predict the remaining useful life precisely.Thirdly.we use the Bayesian model evidence to avoid the problem of over-fitting.By estimating the Bayesian model evidence of each potential model,we can select the best model and thus increase the performance.Finally.we use the decision theory to deal with the randomness in the degradation process of models.