Data-Driven Methods for Robust Battery Capacity Estimation based on Electrochemical Impedance Spectroscopy

To ensure accurate battery capacity estimation over the battery life time, it is important to extract those features from battery data sets that give a good indication of battery capacity degradation. Data obtained from electrochemical impedance spectroscopy (EIS) are a promising route for detecting different aging effects. Many present methods for extracting battery aging features from EIS data are unsuitable for cells that have very different aging behaviour, which leads to low robustness in the battery capacity estimation. To improve battery capacity estimation of cells with significantly different aging behaviour, two methods for feature detection and consistency analysis are proposed for finding the high aging-correlated features in EIS data of these cells. A novel feature-consistency coefficient is proposed to assess whether the detected features are suitable for use in capacity determination. Based on the two new features that are found using this approach on a published data set of 8 battery cells with significantly inconsistent aging behavior, a capacity estimation is subsequently carried out using several advanced machine learning (ML) techniques, using Gaussian process regression (GPR) and Support vector machine (SVM) models. It appears that a third ML method based on automatic feature extraction and capacity estimation using convolution neural networks (CNNs) gives the best, most robust capacity estimation result. All methods presented in this paper significantly outperform GPR-based estimations published in the literature.