The paper proposes a new experimental measurement method for State of Charge (SoC) estimation able to optimize between measurement time and target Accuracy adoptable in Battery Management System (BMS) design where both these parameters are key parameters for the overall management performance. The method is applicable when the SoC is estimated in classes through Electrochemical Impedance Spectroscopy (EIS) and is based on two operating stages: i) the experimental characterization of the real behavior of the considered batteries, ii) a time-Accuracy optimization based on suitable feature selection and Machine Learning approaches. In detail, as for the ii) phase, fixed the number of classes in which the SoC is estimated, the proposed measurement method finds the minimum number of impedance spectrum frequency measurements useful for SoC estimation. This is a big issue for BMS designers since in SoC estimation performed by EIS the measurement time is typically greater than some minutes if no optimization is considered. A possible strategy to reduce the required time for SoC estimation could be using Feature Selection (FS) techniques. In our method the FS is implemented using different combinations of search algorithms and fitness functions. Since FS strongly depends on the experimental set-up, the uncertainty of the measurement system, and the classifier adopted for the data-driving evaluation model, the proposed method is flexible and customizable depending on the specific applications. For this reason, the method is divided into steps where the BMS designers define the requirements based on their needs and their hardware. Also the output FS could be adapted to the different exigencies because the selected features could be different if more emphasis is on increasing classification Accuracy or decreasing the measurement time. Hence, we suggest two application scenarios: in the first one, the only requirement is increasing the classification Accuracy rather than the measurement time optimization. The second scenario has both a classification Accuracy target and lowering the measurement time. We also noted that the proposed method is useful also for increasing the overall classification Accuracy because naturally excludes the features that deceive the classifiers. As an application for the proposed method, we developed an automatic measurement system and performed an experimental campaign on seven Lithium Iron Phosphate batteries. The experimental results show that the proposed approach, for all the considered combinations, significantly reduces the measurement time required for EIS while maintaining high SoC estimation Accuracy. Furthermore, for all the obtained solutions, the effectiveness of our approach was confirmed by the significant savings in measurement time. In particular, the best solution reduces from 4 minutes when using all features to about 1.5 s with the selected features.
