Online Remaining Useful Life Prediction Of Lithium-ion Batteries Based On Relevance Vector Machine

Lithium-ion batteries have many desirable characteristics,such as high working voltage,high energy density,no memory effect,low self-discharge rate,low loss of energy,long cycle life and low pollution,and so on.Lithium-ion batteries have been widely used both in home applications and military equipment.With the fast development of lithium-ion battery industry,many researchers have drawn attention on the prognostics and health management of lithium-ion batteries,especially on the online state monitoring and remaining useful life(RUL)prediction.These studies are significant for reliability evaluation,maintenance optimization and scientific management of lithium-ion batteries.Considering the real working situation of lithium-ion batteries,the research in this paper mainly focuses on the following aspects:First,indirect health indicator(HI)extraction of lithium-ion batteries is discussed.One of the most difficulties in online RUL prediction is how to select the HIs.The direct HI,capacity,is hardly measured during batteries’ working period.Thus,considering the feasibility of feature extraction,five indirect HIs,which can indicate the state of heath of lithium-ion batteries,are extracted from the curves of voltage,current and temperature in each charge-discharge cycle.Second,online capacity estimation of lithium-ion batteries is investigated.Considering the capacity is the direct HI of lithium-ion batteries,the five indirect HIs are utilized as the input of a data-driven model to estimate the online capacity.In order to increase the estimation accuracy,robustness and generalizability of the data-driven method,a dynamic ensemble model based on adaptive multi-kernel relevance vector machine(MKRVM)and weighted Euler’s distance similarity is proposed in this paper for capacity estimation.MKRVMs are used as the sub-models to enhance the generalizability of the ensemble model,and the accelerated particle swarm optimization is applied to adaptive optimization of MKRVM’s kernel parameters.Then,the weights of sub-models can be updated by using online data,in order to obtain a dynamic combination of sub-models’ outputs,which can further increase the accuracy of capacity estimation.Model comparison and cross-validation process are conducted to prove the effectiveness of the proposed model.Finally,capacity degradation tendency predict and online RUL estimation of lithium-ion batteries are studied.First,phase reconstruction techniques are used to deal with the original estimated capacity values.Mutual information algorithm and CAO method are used to calculate the delay time and embedded dimension,respectively,to construct the training dataset.Then,an online leaning MKRVM model is proposed to predict the online RUL of lithium-ion batteries.This model can learning from the online data collected while batteries operate,and the RUL predictions become more accurate with the increasing size of available online data.The proposed model can carry out a better long-term prediction with high computation efficiency and low computer memory.Experiment and comparison results show the higher accuracy of the online learning MKRVM for online RUL prediction.