Hybrid State Estimation : Integrating Physics-Informed Neural Networks with Adaptive UKF for Dynamic Systems
In this paper, we present a novel approach to state estimation in dynamic systems by combining Physics-Informed Neural Networks (PINNs) with an adaptive Unscented Kalman Filter (UKF). Recognizing the limitations of traditional state estimation methods, we refine the PINN architecture with hybrid los...
| Autores: | , |
|---|---|
| Tipo de recurso: | artículo |
| Fecha de publicación: | 2024 |
| País: | España |
| Institución: | Universidad Francisco de Vitoria |
| Repositorio: | DDFV. Repositorio Institucional de la Universidad Francisco de Vitoria |
| Idioma: | inglés |
| OAI Identifier: | oai:ddfv.ufv.es:10641/7621 |
| Acceso en línea: | https://hdl.handle.net/10641/7621 |
| Access Level: | acceso abierto |
| Palabra clave: | Physics-Informed Neural Networks Unscented Kalman Filter (UKF) state estimation Control and Systems Engineering Signal Processing Hardware and Architecture Computer Networks and Communications Electrical and Electronic Engineering Yes yes |
| Sumario: | In this paper, we present a novel approach to state estimation in dynamic systems by combining Physics-Informed Neural Networks (PINNs) with an adaptive Unscented Kalman Filter (UKF). Recognizing the limitations of traditional state estimation methods, we refine the PINN architecture with hybrid loss functions and Monte Carlo Dropout for enhanced uncertainty estimation. The Unscented Kalman Filter is augmented with an adaptive noise covariance mechanism and incorporates model parameters into the state vector to improve adaptability. We further validate this hybrid framework by integrating the enhanced PINN with the UKF for a seamless state prediction pipeline, demonstrating significant improvements in accuracy and robustness. Our experimental results show a marked enhancement in state estimation fidelity for both position and velocity tracking, supported by uncertainty quantification via Bayesian inference and Monte Carlo Dropout. We further extend the simulation and present evaluations on a double pendulum system and state estimation on a quadcopter drone. This comprehensive solution is poised to advance the state-of-the-art in dynamic system estimation, providing unparalleled performance across control theory, machine learning, and numerical optimization domains. |
|---|