Transfer learning based generalized framework for state of health estimation of Li-ion cells

Abstract

Estimating the state of health (SOH) of batteries powering electronic devices in real-time while in use is a necessity. The applicability of most of the existing methods is limited to the datasets that are used to train the models. In this work, we propose a generic method for SOH estimation with much wider applicability. The key problem is the identification of the right feature set which is derived from measurable voltage signals. In this work, relative rise in voltage drop across cell resistance with aging has been used as the feature. A base artificial neural network (ANN) model has been used to map the generic relation between voltage and SOH. The base ANN model has been trained using limited battery data. Blind testing has been done on long cycle in-house data and publicly available datasets. In-house data included both laboratory and on-device data generated using various charge profiles. Transfer learning has been used for public datasets as those batteries have different physical dimensions and cell chemistry. The mean absolute error in SOH estimation is well within 2% for all test cases. The model is robust across scenarios such as cell variability, charge profile difference, and limited variation in temperature.