TinyML models for SoH estimation of lithium-ion batteries based on Electrochemical Impedance Spectroscopy

As battery systems become more widespread, the need for accurate and fast estimation of State of Health (SoH) in lithium-ion cells is increasingly critical. This study analyzes two data-driven model methods that leverage Electrochemical Impedance Spectroscopy (EIS) measurements to capture the internal electrochemical dynamics underlying battery degradation and estimate the cell’s current SoH. The first approach, Method A, utilizes an Equivalent Circuit Model (ECM) from EIS data and uses it to train various state-of-the-art Deep Learning (DL) models, including LSTM, GRU, CNN-LSTM, and CNN-GRU. In contrast, Method B directly employs raw EIS data to train the same set of DL models, bypassing the need for ECM development. Both methods demonstrated strong performance, with the CNN-GRU model from Method B yielding the best results, achieving a Mean Absolute Error (MAE) of only 0.87% and a Root Mean Square Error (RMSE) of 1.20%. Additionally, both methods included an analysis of various input features, such as State of Charge (SoC), to evaluate their impact on model performance. Finally, the models of Method B were optimized for size and computational efficiency, making them suitable for deployment on low-power edge devices and applications requiring TinyML capabilities. The average latency and size reduction of the models were 99.61% and 88.61%, respectively. • Developed TinyML models for battery SoH estimation from EIS data. • Compared ECM-based and raw EIS data-driven approaches for model development. • Optimized models converted to TinyML for real-time edge device deployment. • CNN-GRU model achieved highest performance by using raw EIS data.