This work proposes a novel hybrid framework for real-time and embedded-compatible estimation of lithium-ion battery State of Health (SoH), integrating electrochemical insights with machine learning and physics-informed filtering. The novelty of this work lies in formulating a complete IC-driven virtual Electrochemical Impedance Spectroscopy (EIS) mechanism that estimates impedance parameters directly from time-domain charge data, eliminating any need for traditional frequency-domain excitation or EIS hardware. The approach introduces and utilises a virtual EIS module that emulates impedance diagnostics using Incremental Capacity (IC) curve features. Key impedance parameters—electrolyte resistance (Re) and charge transfer resistance (Rct)—are inferred via a Support Vector Regression (SVR) model optimized using the Cuckoo Search Algorithm (CSA), thereby removing the dependence on expensive, bulky, and non-deployable impedance instruments. These virtual impedance estimates, combined with terminal voltage and temperature readings, are utilized by an XGBoost regressor to predict SoH in a data-driven yet electrochemically grounded manner. To strengthen reliability, the proposed pipeline additionally incorporates quantitative performance validation, achieving Re MAE = 0.012 Ω, Rct MAE = 0.015 Ω, and SoH estimation with R 2 consistently above 0.95 across multiple NASA PCoE cells. Further to ensure temporal consistency and robustness under dynamic degradation conditions, an Enhanced Extended Kalman Filter (EKF) is applied, treating the ML output as a pseudo-measurement and refining it through recursive state estimation. This refinement reduces RMSE by up to 50% and improves R 2 score upto 0.965, ensuring degradation patterns that align with physical aging behavior. The pipeline is lightweight, modular, and suitable for real-time deployment in embedded Battery Management Systems (BMS). Since the proposed method require trivial measurements (voltage, temperature) and IC-derived features, it is well suited for microcontroller-based BMS hardware with limited computational resources, broadening its deployability across first-life EV packs, second-life stationary applications, and scalable onboard systems.
Khurana et al. (Thu,) studied this question.