Lithium-ion batteries (LIBs) are the enabling technology for electric mobility and stationary energy storage systems. Conventional battery health metrics such as State of Health (SoH) quantify capacity degradation but provide limited insight into the underlying ageing mechanisms occurring inside the cell. Physics-based degradation indicators, such as Loss of Lithium Inventory (LLI) and Loss of Active Material (LAM), provide a more mechanistic description of battery degradation and are typically estimated through low-rate diagnostic tests combined with thermodynamic or electrochemical modelling. However, these procedures are time-consuming and difficult to implement in practical applications and battery management systems.
In this work, we present a data-driven framework that links in-operando electrochemical measurements to physics-based degradation indicators using machine learning. The approach combines electrochemical impedance spectroscopy (EIS), voltage and temperature signals from charge–discharge cycles, together with operating conditions, to estimate SoH, LLI, and LAM. The model is based on a Transformer neural network architecture employing multi-head self-attention to capture relationships among electrochemical signals and across different frequency and time domains.
The framework is trained on an experimental database of commercial 18650 lithium-ion cells subjected to a laboratory ageing campaign under multiple combinations of temperature, state-of-charge window, and current rate, designed to reproduce realistic operating conditions. Low-rate diagnostic tests combined with a quasi-OCV model enable the estimation of LLI and LAM over ageing, providing physics-based reference values for model training.
The proposed Transformer-based model accurately predicts all degradation indicators directly from in-operando measurements, achieving a mean absolute error below 0.02 and a Pearson correlation coefficient above 0.99 across the full ageing range. Explainability analyses based on attention mechanisms and SHAP attribution highlight the dominant contribution of electrochemical impedance spectroscopy to the prediction of SoH, LLI, and LAM, with specific frequency regions of the impedance spectra emerging as the most informative features associated with degradation processes. Furthermore, the analysis shows that incorporating operating conditions significantly improves model robustness and generalization, while testing under previously unseen ageing conditions increases the prediction error up to approximately 0.06 MAE, highlighting their fundamental role in mechanism-aware battery diagnostics.