Transfer Learning-Based Lithium-Ion Battery State of Health Estimation With Electrochemical Impedance Spectroscopy

Abstract

Lithium-ion batteries are utilized as energy storage units in mobile devices, electric vehicles, and other fields. To ensure the safety and reliability of batteries, the prediction of the batteries’ state of health (SOH) is one of the key technologies. This article proposes a transfer learning-based lithium-ion battery SOH estimation method using explainable electrochemical impedance spectroscopy (EIS). EIS has advantages such as explainability, rapid response, and noninvasiveness. Benefiting these, the physical parameters are extracted as the battery aging features by fitting an equivalent circuit model with the EIS measurement. To increase the representational power of the model and capture the complex sequential styles, a spatiotemporal long short-term memory (LSTM) network model is built to extract the time series features. Finally, the battery degradation features are fit through a fully connected layer. To improve the model’s generalization, a transfer learning strategy is added to estimate the SOH of the target cell by fine-tuning the initial model parameters on different temperatures and different types of cells. The proposed method, TL-ST-LSTM, has been validated on two public datasets, with an overall root-mean-square error (RMSE) error controlled within 1.9%. Compared to the spatiotemporal LSTM (ST-LSTM) method without transfer learning, the accuracy has been improved by over 80%. In addition, it also demonstrates an improvement in accuracy compared to existing transfer learning methods, such as TL-CNN and TL-LSTM.