Localized EIS Data for Capacity and SOH Prediction With Neural Networks

ABSTRACT Accurately estimating the battery's capacity over its cycle life is essential for ensuring its safety in applications, including transportation and the medical field, where specific power delivery is a key component for optimal output. Most research concerning lithium‐ion health prediction utilizes current‐voltage data or techniques that rely on modeling microscopic degradation. Acquisition of current‐voltage data directly builds up degradation within the cell, and physics‐based methods require high computational power. Recent research pivoted to using electrochemical impedance spectroscopy (EIS) for battery health prediction since it provides information‐rich data while non‐destructive to the cell. One major drawback of using EIS is the time it takes to acquire data, especially at lower frequencies where diffusion within the cell is prevalent. To address this, this investigation focuses on feature extraction, which creates a subset of data from a publicly available data set to contain the frequencies that are mostly correlated with degradation. Analysis shows that a simulated cell's state of health (SOH) can get as low as 0.94% MAPE using the two most correlated frequencies in the charge transfer (CT) region. This study provides a methodology to accurately predict the capacity and SOH while reducing the time needed to acquire EIS data by 93% for this case. This method also highlights the usefulness of a single‐cell model for battery test bench applications.