Physics-Based Model Informed Smooth Particle Filter for Remaining Useful Life Prediction of Lithium-Ion Battery

The remaining useful life (RUL) is critically important for prognostic health
management (PHM) of lithium-ion batteries (LiBs) to provide early warning to ensure the reliability and safety of host devices. Recently developed methods in the literature for RUL prediction face two challenges. First, most approaches are mainly developed based on traditional empirical degradation models without considering degradation mechanisms. Second, the stability of the standard particle filter (PF) method is strongly constrained by the issue of a lack of particles and the uncertainty in the degradation model parameters, which are constrained by the availability of sufficient and reliable data. Consequently, this can lead to inaccurate RUL prediction. Therefore, this work proposed to develop a framework for RUL prediction based on the estimation of parameters of a reduced-ordered physics-based model of LiBs by extracting three main
degradation mechanisms directly correlated with the RUL of the LiB. These degradation mechanisms are active material (AM) loss in positive and negative electrodes and loss of lithium inventory (LLI). Unlike the traditional prognostic framework that primarily depends on the empirical degradation model trend, the proposed framework utilized the SPM to estimate the degradation parameters from the voltage and capacity data, hence quantifying the degradation mechanisms. Then the capacity fade trend is predicted based on the estimated degradation parameters. Following that, the estimated capacity resulting from the estimated SPM parameters is utilized to develop an RUL predictor
based on a smooth particle filter (SPF) to overcome the PF algorithm problems.
Comparing with the conventional capacity-based methods, such as the SPM-based particle filter (SPM-PF), the proposed physics-based method, produces a more accurate RUL prediction. The results demonstrate that the proposed framework predicting is relatively small. At the prediction starting point of 𝟐𝟎𝟎𝟎 cycles, the best-case RUL prediction is 𝟐𝟒𝟎𝟐cycles. Additionally, the minimum relative error is found to be around 𝟎.𝟎𝟖𝟗%, and the relative error of the traditional framework is approximately 𝟎.𝟖%. Furthermore, LiBs data including Gaussian white noise and dynamic discharging profiles, have been utilised to demonstrate the dependability and robustness of the proposed framework.