Remaining Useful Life Estimation for Lithium-Ion Batteries using Physics-Informed Neural Networks

초록

Lithium-ion batteries (LiB) are commonly used sources of power for autonomous vehicles, unmanned aerial vehicles etc. and hence prognostic studies on LiBs are of utmost importance to ensure safety and reliability. Model-based and Data-driven methods are the commonly used prognostic methods in literature, however these methods are severely limited by model and data uncertainties. Hence, we propose a physics-informed hybrid prognostic approach which leverages on the strengths of the conventional model-based and data-driven methods and addresses its limitations to improve the prognostic performance. This work is an extension of one of our earlier works which combined the data-driven neural network model with the model-based particle filter algorithm. The particle filter algorithm was used to train the neural network model and hence helped overcome the dependency on accurate physics-based/empirical degradation model as well as large amount of historical failure data representing the system's degradation phenomenon. However, the model parameter values go astray after a few iterations which leads to unrealistic and illogical prediction traces. To address the outlier issues, we propose to integrate a physics-based loss function into the neural network model based on SEI film formation in this work and the method was tested on both NASA and CALCE datasets.

키워드