Hardware-Friendly Machine Learning Approach for Battery SoH Estimation with Different Cells Chemistries and Different Measurement Setups

Accurate estimation of the State of Health (SoH) of lithium-ion batteries is crucial to ensure safety, reliability, and optimal performance in modern energy systems, particularly in electric-vehicle applications. This study proposes a data-driven SoH estimation approach based on machine learning techniques and validated through two distinct experimental test benches. The first bench operates under ambient conditions and is used to characterize 18650 cylindrical lithium-Ion battery cells, whereas the second test bench features active temperature control via a recirculating oil bath and is employed to test a LiFePO4 prismatic cell. Two supervised learning models, namely the Decision Tree (DT) and the Support Vector Machine (SVM), are trained on the acquired datasets and comparatively evaluated in terms of estimation accuracy, computational complexity, and suitability for embedded system implementation.Experimental results demonstrate that both models provide accurate SoH predictions. However, temperature-controlled testing significantly improves estimation performance, leading to substantially lower root mean square error values. Moreover, compared with SVM models, DT models outperform in terms of prediction speed and model size, making them particularly suitable for real-time deployment on resource-constrained microcontrollers.