Research Area:  Machine Learning
This paper presents a novel technique, combining neural network and Kalman filter, for state estimation. The proposed solution provides the estimates of the system states while also estimating the uncertain or unmodeled terms of the process dynamics. The developed algorithm exploits a Radial Basis Function Neural Network that outputs an estimate of the disturbances that are included in the prediction step of an Adaptive Extended Kalman Filter. A recursive form of adaptation is used to limit the computational burden. The proposed solution is compared to classical navigation filter implementations. A realistic spacecraft relative navigation scenario is selected to test the filter performance. Simulations are performed with accurate tuning and also in off-nominal conditions to test the filter robustness.
Keywords:  
Author(s) Name:  Vincenzo Pesce, Stefano Silvestrini, Michèle Lavagna
Journal name:  Aerospace Science and Technology
Conferrence name:  
Publisher name:  ELSEVIER
DOI:  https://doi.org/10.1016/j.ast.2019.105527
Volume Information:  Volume 96, January 2020, 105527
Paper Link:   https://www.sciencedirect.com/science/article/abs/pii/S1270963818328785