A Correlation-based Method for the Estimation of the State-of-Health of Lithium Batteries in Electric Vehicles Using Data-Driven Methods

This paper presents a correlation-based framework for feature selection to estimate the lithium batteries state of health (SoH) using data-driven methods. These features can be calculated online for partial charge/discharge (C/D) cycles in electric vehicles. The main contribution is to provide a framework to determine a subset of independent features with maximum impacts on SoH estimation. Despite statistical features, the selected features are independent of C/D profiles, without requiring a full cycle information. Using the correlation criterion reduces the computational burden in a data-driven model as required in an online battery health estimation method. Extracting features, considering battery electrochemical reactions, and reducing the number of selected features, using the suggested correlation criterion, increase the robustness of the model and avoid overfitting. NASA labeled battery aging dataset is used to evaluate the performance of the method under various partial and random C/D cycles. The results show that the correlation-based features in Gaussian data-driven model can estimate the battery SoH with an average error of 10% under random patterns in a partial C/D cycle.