Mixture Variational Adaptive Filter with Uncertain and Non-Gaussian State Propagation

In this paper, a novel mixture variational adaptive filter (MVAF) is proposed to deal with the nonlinear state estimation problem with uncertain and non-Gaussian state propagation. The main idea of MVAF is to describe the state propagation by Gaussian mixture model, different from traditional one, the mean of which is composed of certain and uncertain two parts. In the certain part, the nonlinear state function is used directly without approximation so that the error caused by linearization can be avoided. In the uncertain part, variational parameters are introduced, which can capture the feature of unknown process noise. Then, based on the variational Bayesian framework, through iteratively and alternatively achieving the fitting of the state propagation and the approximation of process noise, the estimation accuracy and adaptiveness can be enhanced gradually. The superior performance of MVAF is demonstrated in two simulations.