Research On Remaining Useful Life Prediction Of Lithium-ion Battery With Particle Filter

One of the research focus and challenges is remaining useful life (RUL)prediction of the electronic system in the field of Prognostics and HealthManagement (PHM). Accurate RUL estimation and its uncertainty expression willimprove the system reliability, and achieve condition based maintenance (CBM),which has important research and practical value.Particle filtering (PF) is an algorithm using Monte Carlo to solve the Bayesianestimation problem for any form of state space model, in which, particle set is usedto represent the probability. What is more, PF algorithm could give the uncertaintyrepresentation of the results. RUL prediction metgod based on PF will be studiedfor lithim-ion battery.Firstly, based on cycle life degradation data of lithium-ion battery, lifedegradation process of the lithium-ion battery is analyzed and the empiricaldegradation model is adopted. PF algorithm is used to predict RUL and also givesthe uncertainty representation in form of probability density distribution (PDF). Asuitable resampling algorithm is selected by comparing four resampling algorithms.Secondly, to solve the problem of poor adaptation of prediction method based onempirical model, a new method for RUL estimation combining PF algorithm andtime series analysis is proposed, in which, the prediction results with AR model areused to be the observation sequence for PF, and the RUL and PDF are outputted byimproved PF algorithm. Focusing on the uncertainty expression of the RULprediction, the regularization particle filtering is introduced to the frame work,which solved the problem of particle diversity degradation for the standard PF, andimproved the uncertainty expression accuracy. Finally, the proposed method isevaluated by the PHM metrics; performance is improved by comparing with otherprediction method. At last, we adopt the PHM metrics to evaluate the predictionmethod and perform the comparison with other present prediction methods withuncertainty representation to validate the advantages of our proposed method. Wealso implement the quantification of the prediction results uncertainty to supplymore valuable reference information of system maintenance.Battery Data Set is from NASA PCoE and CACLE of university Maryland,and the experiment results show that the prediction method can effectively predict the cycle life of lithium-ion battery, and give its uncertainty expression in the formof PDF, which has high computational efficiency and uncertainty representationability. Moreover, for the “accelerated” nonlinear degradation feature of the batterydata, an improved AR model is proposed and the experiment results proved that itcan achive more accuracy RUL prediction.