Fully distributed variational Bayesian non-linear filter with unknown measurement noise in sensor networks

Abstract

In practical applications, the measurement noise statistics is usually unknown or may change over time. However, most existing distributed filtering algorithms for sensor networks are constructed based on exact knowledge of measurement noise statistics. Therefore, under situations with measurement uncertainty, the existing algorithms may result in deteriorated performance. To solve such problems, a distributed adaptive cubature information filter based on variational Bayesian (VB-DACIF) is proposed here. Firstly, the predicted estimates of interest from inclusive neighbours are fused by minimizing the weighted Kullback-Leibler average, in which the cubature rule is utilized to tackle system nonlinearity. Then, the free form variational Bayesian approximation is applied to recursively update both the local estimate and the precision matrices of sensing nodes. Finally, the posterior Cramer-Rao lower bound is exploited to evaluate performance of the proposed VB-DACIF. Simulation results with a maneuvering target tracking scenario validates the feasibility and superiority of the proposed VB-DACIF.