A novel procedure for real-time SOH estimation of EV battery packs based on Time Series Extrinsic Regression

One of the critical challenges posed by the spread of Lithium-ion Batteries (LIBs) within Electric Vehicles (EVs) is the real-time estimation of their State-of-Health (SOH), commonly regarded as the leading indicator of EV aging. However, SOH estimation is still challenging due to the electrochemical complexity of LIBs. This work proposes a novel, computationally-inexpensive, and chemically agnostic Machine Learning (ML) procedure for onboard real-time SOH estimation. The proposed methodology requires a narrow time window of voltage, current, and State-of-Charge battery data, collected while driving the vehicle. We defined a Simulink-based EV simulator, modeling a specific real-world EV, and we utilized it to generate a synthetic dataset by simulating multiple driving sessions of the EV to compensate for the lack of large-scale publicly available EV monitoring data. Then, we examined three feature extraction methods and three ML regression models, estimating the battery pack's SOH. We conducted a thorough comparison of the proposed feature extraction methods and ML models, training the ML model with processed synthetic data and inferring over real driving session monitoring data from the corresponding real-world EV model. The best synthetic-trained ML model achieves an MAE of 0.27% and 5.08%, and an RMSE of 0.37% and 5.92% over synthetic and real test data, respectively. Finally, we implemented transfer learning over the ML models, employing a portion of the available real data, reaching the lowest MAE of 1.97%, and an RMSE of 2.56% over the remaining real test set.