Evaluation of Deep Learning Models for State of Charge Estimation for Lithium Batteries Using Image-Encoded Electrochemical Impedance Spectroscopy Data

State of Charge (SoC) estimation in Lithium-based batteries is crucial for enhancing the performance, safety, and longevity of energy storage systems, particularly in Electric Vehicles (EVs) and portable devices. Traditional SoC estimation methods face significant challenges due to the nonlinear and time-dependent electrochemical behavior of batteries. Machine Learning (ML) and Deep Learning (DL) offer promising solutions by leveraging large datasets to capture complex patterns in battery behavior. This paper investigates the use of Electrochemical Impedance Spectroscopy (EIS) data, encoded as images using the Gramian Angular Field (GAF) method, for SoC estimation. Two convolutional neural network (CNN) models, Simple CNN and AlexNet, are evaluated on GAF-transformed impedance data collected for different SoC levels from Lithium Iron Phosphate (LFP) batteries. The results highlight the potential of DL-based methods for SoC estimation, showing that both models exhibit strengths and challenges in handling the complex electrochemical patterns in EIS data. Variability in performance across different battery units emphasizes the need for improved feature extraction techniques and the expansion of datasets to achieve more reliable and robust SoC estimation in real-world applications.