A Physics-Informed Bayesian Deep Learning Framework for Explainable and Probabilistic Lithium-Ion Battery Health Prognostics

Abstract

To ensure the safety and efficiency of battery-powered applications, it is essential to accurately estimate the State of Health (SOH) and Remaining Useful Life (RUL) of the batteries. Although physics-based models are available to represent the degradation of the batteries, these models may have difficulties modeling the nonlinear dynamics of the ageing of the batteries. Furthermore, data-driven deep learning models that do not incorporate the physics of the batteries provide limited insight into the reliability of their predictions. Hence, a framework that combines the Physics-Informed Neural Network (PINN), Bayesian Neural Network (BNN), and SHapley Additive exPlanations (SHAP) can provide accurate, uncertainty-aware, and interpretable estimation of the SOH and RUL of the batteries. The PINN incorporates the degradation of the batteries into the framework, the BNN estimates the uncertainty of the predicted SOH and RUL of the batteries using the PINN, and the SHAP framework estimates the contribution of the individual features to the predicted SOH and RUL of the batteries using the PINN. The proposed framework is evaluated using the publicly available CALCE CS2_8 lithium-ion battery dataset. The results of the framework achieve an SOH prediction accuracy of 89.78% and an RUL prediction accuracy of 95.31%. Furthermore, the results exhibit the effectiveness of the proposed framework compared with conventional machine learning and deep learning models. Thus, the framework has the potential to be incorporated into next-generation Battery Management Systems (BMSs).