An Adaptive Nonlinear Filter for Integrated Navigation Systems Using Deep Neural Networks.

Fei Yan,Sheng Li,Enze Zhang,Jian Guo,Qingwei Chen

DOI: https://doi.org/10.1016/j.neucom.2021.03.046

IF: 6

2021-01-01

Neurocomputing

Abstract:This paper presents a novel nonlinear adaptive sensor fusion method for integrated navigation systems with varying noise parameters. The innovation is utilizing deep neural networks to mine the noise-related patterns of specific sensors and combining it with conventional nonlinear filters. This hybrid approach improves the feasibility and robustness of adaptive filtering by achieving an effective estimation of the originally weakly observable noise parameters. The specific sensors are defined as α-type sensors whose errors are entirely generated by themselves. The mathematical model for analyzing α-type sensors output sequence and the deep neural network for mining the patterns of interest are established. All adaptive filtering systems using α-type sensors can benefit from this paper. Specifically, it is applied to inertial and satellite integrated navigation system. The numerical experiments indicate that the proposed filter achieves promising accuracy and robustness improvement as compared to conventional nonlinear filters. And the comparisons between different nonlinear approximation algorithms indicate that the first-order approximation is accurate enough for our application.