Fractional-order derivatives to elevate the physics-awareness of the data-driven battery health estimators

Abstract

Detection and identification of the degradation mechanisms is an invaluable merit for a battery management system enabling a reliable estimation of battery state-of-health (SOH) and preventive interventions. However, this remains challenging for conventional data-driven models, which often lack the physics-awareness needed to capture electrochemically meaningful voltage-capacity distortions associated with degradation. Here, we introduce fractional-order derivatives to extract physics-informative features from the open-circuit-voltage (OCV) profiles of lithium-ion batteries. These features are used to train support vector regression (SVR) models and showcased to result in superior SOH prediction for LiNixMnyCozO2 (NMC), LiFePO4 (LFP), and LiNixCoyAlzO2 (NCA) battery chemistries. The increased sensitivity of fractional derivatives enables the detection of subtle degradation signatures, improving correlation with internal modes such as loss of lithium inventory and loss of active material particles. Our results highlight the added value of fractional-order derivatives for the development of accurate and physics-aware data-driven models for battery diagnosis and prognosis.