Search

• The physical effects of battery correlated equivalent circuit model parameters. • Low and medium-frequency EIS sensitive to battery aging. • Combined circuit parameters and gated recurrent unit to estimate battery SoH. • Considering the effect of ambient temperature and SoC on EIS. Electrochemical impedance spectroscopy (EIS) is one of the critical techniques to characterize the state of health (SoH) of lithium-ion batteries, which has different performances under different SoH and dynamic working conditions. However, the identification parameters of full-frequency EIS will consume too much computing time. Considering that low and medium-frequency EIS plays a dominant role in battery aging, the present paper constructs a simplified equivalent circuit (SECM) based on low and medium-frequency EIS. Based on the fitted SECM parameters, three SECM parameters are selected as features by grey correlation analysis. Combined with the ambient temperature and SoC under dynamic conditions, the Gated recurrent unit neural network estimates the SoH of lithium-ion batteries. The proposed method has good accuracy and robustness for different SoC states and ambient temperature variations. This method's error values are less than 2 %, which proves the critical value of low and medium-frequency EIS signals in battery management systems. [ABSTRACT FROM AUTHOR]

• A method for estimating lithium batteries' state of health (SoH) is proposed. • Transformation of EIS data into 2D pseudo-images. • Unsupervised feature extraction using VGG16 neural network. • Integration of feature extraction and SoH in one framework. Electrochemical impedance spectroscopy (EIS) is a key technique characterizing the batteries' State of Health (SOH). The extraction of features from the limited EIS information for SOH estimation relies heavily on the researcher's prior knowledge. This study proposes a method to enhance the EIS feature information and perform unsupervised feature extraction to estimate the SOH. First, the EIS data is transformed into images using Gramian angular field, which enhances the data features. Next, the images were subjected to unsupervised feature extraction using the VGG16 neural network framework. Finally, the unsupervised feature extraction and SOH prediction were integrated into a neural network framework to achieve end-to-end training and prediction. The experimental results show that the proposed method's SOH estimation error is less than 2%, and its accuracy is improved by 55.6% compared to its benchmark model; the feasibility of unsupervised feature extraction is demonstrated, overcoming the drawbacks of artificially performed feature extraction. [ABSTRACT FROM AUTHOR]