Variational Bayesian-Based Generalized Loss Cubature Kalman Filter
Wenxing Yan, Shanmou Chen, Dongyuan Lin, Shiyuan Wang
Abstract
Kalman filters equipped with both adaptivity and robustness have been developed to handle both unknown measurement noise and non-Gaussian noise affected by outliers. However, in the presence of complex non-Gaussian noise, the estimation performance of these filters tends to deteriorate. To address this issue, we propose a variational Bayesian-based generalized loss cubature Kalman filter (VB-GLCKF), which introduces a generalized loss (GL) in robust information learning to combat the effects of complicated measurement outliers. Unlike other robust loss functions, the GL modifies the shape of the function by adjusting the shape parameter. More importantly, to avoid the manual selection of the shape parameter, VB-GLCKF first establishes the linear regression model for a residual error vector and then introduces the negative log-likelihood (NLL) of the GL function for automating parameter optimization. Simulations on reentry vehicle tracking (RVT) confirm that VB-GLCKF can effectively estimate the shape parameter and achieve significant accuracy improvement compared to existing filters when dealing with complex noise scenarios involving both unknown measurement noise and outliers.