Accurate state estimation is a critical factor for the optimal management, safety, and longevity of modern energy storage systems. One promising method for achieving precise state estimation is electrochemical impedance spectroscopy (EIS), a technique that has been widely documented in the literature for its effectiveness in monitoring the internal states of batteries. Traditional methods of state estimation using EIS typically rely on selecting simple, direct features from impedance spectra, such as the real part, imaginary part, magnitude, or phase at specific frequencies. These features are often chosen based on feature importance metrics derived from machine learning models, including neural networks and regression-based approaches. [1, 2]

In contrast, this methodology, demonstrated for State of Charge (SoC) estimation, introduces a more sophisticated approach to feature generation and evaluation, complementing the classical feature approach by McCarthy et al. [1] and extending the initial attempt of more complex features as presented by Hackmann et al. [3] by a factor of four. A custom tool has been developed, with the automated workflow starting with the processing of the impedance data and feature generation. This feature generation process is based on curve tracing of the impedance spectra, with the analysis of stationary, inflection and interception points, plotted in various configurations and combinations. This first set of characteristic features are used further for the generation of additional interconnected and nested features within the set. This leads to an exemplary feature, such as the ratio between the frequency of the inflection point and the first maximum from the plot of the imaginary part over frequency, thus moving beyond the limited view of classical and known battery characteristic impedance features. With the feature set being complete, each feature is evaluated based on its effectiveness for SoC estimation, considering to changes in SoC and sensitivity against external influences as e.g. temperature. The tool does not only perform automated assessments but also provides a detailed review of the generated features, enabling users to identify the best feature combinations.

The key innovation lies in the ability to capture hidden patterns and interactions within the impedance data that may not be apparent when using simpler features. This enhances the sensitivity to SoC variations, while also improving robustness due to lower sensitivity to cross-influences such as temperature—a critical external factor known to highly affect the impedance and therefore the SoC estimation accuracy. By accounting for temperature variations, this method aims to minimize the risk of inaccurate SoC predictions caused by uncertainties in the temperature estimation. Compared to the state of the art feature set, this new feature approach improves SoC accuracy by factor 12, based on a temperature uncertainty of ±3K, using just the most accurate single feature identified by the method proposed in this poster. The most accurate simple feature, based on the work of McCarthy et al. [1], achieved an SoC accuracy of 25%, while the most accurate feature from the new approach reached an accuracy of 2%.

In summary, the new feature generation approach presented in this methodology represents a significant step forward in the field of SoC estimation using EIS. By moving beyond simple features and incorporating more complex features, this method enhances the accuracy and robustness of SoC predictions.

[1] Mc Carthy, K., Gullapalli, H., Ryan, K.M., Kennedy, T., 2021. Use of impedance spectroscopy for the estimation of Li-ion battery state of charge, state of health and internal temperature. J. Electrochem. Soc. 168, 080517.
[2] Babaeiyazdi, I., 2023. State Estimation of Li-ion Batteries Using Machine Learning Algorithms. York University.
[3] Hackmann, T., Esser, S., Danzer, M., 2024 Operando determination of lithium-ion cell temperature based on electrochemical impedance features. J. Power Soc. 615, 235036.