Research On State Of Health Estimation And Remaining Useful Life Prediction Based On Relevance Vector Machine For Lithium Batteries

As an important part of energy storage system,lithium-ion batteries have been applied increasingly widely in many industrial fields due to their superior performance.In order to ensure the safe and stable operation of lithium-ion batteries,it is necessary to utilize a battery management system(BMS)for prognostics and health management.The state of health(SOH)estimation and remaining useful life(RUL)prediction are the core functions of BMS.The lithium-ion battery degradation is a complicated nonlinear process,which is closely related to both environmental and load variations and their internal physiochemical reactions and thermal effects.However,in applications,battery capacity and impedance which can reflect SOH are difficult to measure directly.Based on relevance vector machine(RVM),the research on SOH estimation and RUL prediction is carried out in this paper to provide reference for battery maintenance and health management.The main research work of this paper includes:First,the battery indirect health feature which can reflect SOH is discussed and a clear and easy-to follow charging health feature extraction method is proposed.From the charging current,voltage and temperature curves,14 health features are picked out and grey relational analysis is used to numerically analyze the relevance of health features and capacity and select the features with relatively higher relevance to construct health feature matrix.Then,principal component analysis is applied to optimize the feature matrix.Finally,with the advantages of not being influenced by the battery discharge conditions and considering the temperature factor,optimized health feature is extracted.Second,a SOH estimation method based on adaptive RVM is presented.The indirect health feature is utilized as input to train the RVM regression model and an improved particle swarm optimization algorithm is adopted to optimize the kernel of RVM in order to improve its robustness and accuracy.The experiments of various discharging conditions,environmental temperatures and training data set show that the method proposed show good generalization and performance.Finally,based on nonlinearity,fluctuations and uncertainty of battery capacity degradation,this paper presented a probabilistic RUL prediction method based on Grey-Markov model and RVM.The optimized discrete grey model is used to track the battery overall degradation trend and Markov chain based on probability summation is implemented to model fluctuations of curve improving accuracy.Finally,relevance vector regression is utilized to provide the RUL prediction interval.The experiment results show that RVM not only improves prediction accuracy compared with the single Grey-Markov model but also offer more comprehensive and effective information thorough RUL prediction interval.