Physics-informed Gaussian Process Regression Model for Battery Management

The accurate estimation of battery SoH has become a significant challenge due to the wide-spread usage of lithium-ion batteries for high-impact utilization such as electric vehicles and aerospace applications, as well as operation in everyday life: computers, cell phones, and other small electronics. Physics models and data-driven approaches have shown success for this application of health monitoring. However, singular data-driven approaches often require substantial amounts of reliable data and necessitates access to the full history of the system of interest; in addition, purely data-driven techniques, particularly machine learning, can be prone to overfitting. Hybrid methodologies that combine physics models and data-driven approaches have shown some success in recent literature, however these studies are relatively scarce. In this study, a physics-informed machine learning (PIML) technique is employed to estimate SoH of LIBs using only a partial charging segment. Essentially, the machine learning model is trained by generated data from a 1D finite-element physics model. This addresses the problem of requiring an abundance of data to train the model, as the data used to train the model is obtained by simulation. The machine learning model for estimation consists of a nested Gaussian Process Regression (GPR) model, where temperature and current are inputted, and a projected voltage charging curve for each SoH level are outputted; then the most similar curve is chosen for SoH estimation. Overall, the results show the method performs well in terms of accuracy of capacity estimation; and this work provides a new baseline model for general health management applications.