Training Gaussian process regression through data augmentation for battery SOC estimation

Estimating the state of charge of a battery is generally a challenging operation, as it depends non-linearly on the parameters that describe the internal state of the battery under test. Electrochemical impedance spectroscopy provides useful data for estimating state of charge, however, there is no simple relation between impedance spectra and state of charge. This relationship can be determined through the use of Machine Learning techniques. In particular, Gaussian process regression is used in this work. For the method to be efficient, the training ensemble must be quite large, but spectroscopy measurements can be time-consuming. A Monte Carlo ensemble generated from a limited set of experimental data was then used to train the model. Two sets of input features were used for the Gaussian process: components of the impedance spectrum, and equivalent circuit parameters. Both features give the same accuracy in estimating the state of charge, even of batteries whose impedance spectra are not included in the training set. The first set of features has the advantage of not requiring model fitting. The second feature set has a significantly shorter training time.