Physics-based Machine Learning with Filtering for Failure Prognostics Partially Observable Dynamic Systems

This paper demonstrates a useful hybrid methodology for online capacity estimation of lithium-ion batteries. Essentially, a Kalman filter was embedded with a physics-informed neural network to optimize the connection between observable measurements and the hidden state. In this case, for the lithiumion battery application, the hidden state is capacity, and the observable measurements are voltage. Fundamentally, the physics-informed neural network is a residual model, where the data-driven neural network models the error between the physics model and the noisy measurements. Overall, this structure performs well and has significant improvements over traditional Kalman filter frameworks. The paper is organized as follows: section 1 provides an introduction, section 2 explains the methodology, section 3 shows the results, and section 4 concludes with a discussion.