Robust survival model for the prediction of Li-ion battery lifetime reliability and risk functions

Abstract

Single-value prediction such as the End of Life and Remaining Useful Life is a common method of estimating the lifetime of Li-ion batteries. Information from such prediction is limited when the entire degradation pattern is needed for practical applications such as dynamic adjustment of battery warranty, improved maintenance scheduling, and battery stock management. In this research, a predictive, semi-parametric survival model called the Cox Proportional Hazards is proposed for the prediction of cell degradation in the form of survival probability (battery reliability) and cumulative hazard (battery risk) functions. Once this model is trained, the two functions can be obtained directly for a new cell without having to predict several cogent points. The model is trained on the first 50 cycles of only the voltage profile from either the charge or discharge data regime, implying that our methodology is data region agnostic. The signature method with both desirable mathematical and machine learning properties was adopted as a feature extraction technique.

The developed models are tested rigorously using application-driven strategies involving model robustness to the number of cycles of data required for model training and prediction, different fractions of training samples, and systematic data sparsity. The codes for modeling and testing are publicly available.