Efficient stochastic parametric estimation for lithium-ion battery performance degradation tracking and prognosis

Because of lithium-ion batteries’ wide applications in our daily life and industrial sectors, understanding their performance degradation mechanisms and improving their health management are essential to improve their durability, reliability, and sustainability. However, Li-ion batteries exhibit complex performance degradation behaviors, typically in a combination of nonlinear gradual degradation with time-varying deterioration rates and abrupt performance changes (e.g., sudden capacity drops or regenerations), posing significant challenges to accurate and reliable degradation tracking and prediction. This study tackles this challenge from two perspectives: an advanced stochastic model to describe complex degradation patterns and a generalizable Bayesian inference neural network for efficient parametric estimation of the stochastic model. Specifically, the stochastic model employs a rational polynomial term for tracking gradual battery degradation and a compound Poisson process term for capturing abrupt capacity changes. To estimate the model parameters related to degradation rates and scaling, a novel Conditional Invertible Neural Network (CINN) architecture is investigated. CINN can comprehensively evaluate the degradation likelihood (i.e., dependencies of capability observations on various battery degradation patterns) by leveraging extensive simulation data during the training phase, and then through its unique inverse calculation capability, efficiently and probabilistically estimate the posterior density of model parameters conditional on capacity observations in the real-world applications. The effectiveness of the proposed stochastic model and parametric estimation method, in terms of accuracy and generalizability, has been evaluated using simulation data and run-to-failure tests provided in NASA's lithium-ion battery dataset. Experimental studies and comparisons reveal that the CINN-based parametric estimation substantially outperforms two commonly adopted Bayesian inference methods, Particle Filtering (PF)-based step-by-step estimation and Markov Chain Monte Carlo (MCMC)-based batch estimation, on both accuracy and computational efficiency.