.png){kind=icon}
.png){kind=icon}
Return
Deterministic Learning based Extended Kalman Filter for Nonlinear Dynamics Modeling
.png){kind=icon}
PRE
AI
DOI：10.1016/j.jfranklin.2025.108175.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Recent deterministic learning methods have achieved locally-accurate neural network (NN) approximation and true/ideal weights convergence under the persistent excitation (PE) condition. This locally accurate dynamics knowledge is well exploited in many downstream dynamical tasks, such as control, recognition, and fault diagnosis. However, most existing deterministic learning-based approaches are developed in noise-free environments, whereas real-world data are often corrupted by noise. The extended Kalman filter (EKF) is a widely used tool for handling noise and uncertainty in nonlinear dynamical systems, but it requires a priori knowledge of the system’s dynamics. To bridge the gap between accurate dynamics identification and state estimation in noisy environments, this article proposes a deterministic learning-based extended Kalman filter (DL-EKF) method. First, radial basis function (RBF) neural identifiers are designed to estimate the unknown system dynamics, with the prior and posterior estimation processes of the EKF integrated into the identifier’s framework. The errors between the prior estimation and posterior estimation are utilized to update the RBF neural weight parameter. Second, a rigorous convergence analysis is provided to demonstrate that both the neural weights converge to their true values and the system dynamics are accurately learned. Finally, simulation studies are presented to illustrate the effectiveness of the proposed method.
Journal
J
IF:
4.2
Papers:
822
Citations:
0