Neural extended Kalman filters for learning and predicting dynamics of structural systems

Accurate structural response prediction forms a main driver for health monitoring and control applications. This often requires the proposed model to adequately capture underlying dynamics of complex systems. In this work, we utilize learnable extended Kalman filter (EKF), named neural (neural EKF) throughout article, learning latent evolution physical The EKF is generalized version conventional EKF, where modeling process sensory observations can be parameterized by networks, therefore learned end-to-end training. method implemented under variational inference framework with conducting from sensing measurements. Typically, models are networks independent models. characteristic makes reconstruction accuracy weakly based on renders associated training inadequate. show that structure imposed Neural beneficial process. We demonstrate efficacy both simulated real-world datasets, results indicating significant predictive capabilities scheme.