Fast SOH estimation of lithium-ion battery using electrochemical impedance spectroscopy and feature extraction Pohang University of Science and Technology

Abstract

As the popularity of electric vehicles rises, accurately estimating the state of health (SOH) and performance of used EV batteries becomes increasingly crucial. Electro- chemical impedance spectroscopy (EIS) is a promising method for this purpose, but it typically requires an extensive time frame for accurate results. Considering the ex- pected increase in retired lithium-ion batteries in the future, our goal is to reduce the EIS measurement time to quickly and accurately estimate the SOH. This paper in- troduces the application of neural networks to reduce the EIS data input size without substantially compromising the accuracy. By using the class activation map (CAM) to identify the frequency range highly correlated SOH, we can train the neural network using only feature-extracted points instead of full frequency points. Tests performed on two distinct battery datasets, including a 45 mAh coin cell data and 18650 cylin- drical cell datasets, demonstrated that the error remained below 3% even with a sig- nificant reduction in data size. The mean absolute percentage error (MAPE) for the 45 mAh coin cell datasets increased slightly after the feature extraction, from 1.23% to 2.77%, but the EIS measurement time was reduced by approximately 400 times. For our future work, we aim to improve the method’s accuracy and extend its applicability to other types of batteries.