Centralized and distributed adaptive cubature information filters for multi-sensor systems with unknown probability of measurement loss

Abstract

This paper studies the centralized and distributed state estimation problems for nonlinear multi-sensor systems with unknown probability of measurement loss. Based on the variational Bayesian (VB) method and the cubature information filter (CIF), two adaptive filters, namely centralized adaptive CIF (CA-CIF) and distributed adaptive CIF (DA-CIF), are proposed respectively. First, a series of Bernoulli random variables are introduced as the indicators of measurement loss, and the loss probabilities are modeled as Beta distributions. By transforming all distributions into exponential family form, the CA-CIF can approximate the posterior distributions of the state, loss indicators and loss probabilities by the Expectation-Maximization (EM) algorithm. To reduce the computational cost and improve the reliability, the DA-CIF is further derived by utilizing the average consensus algorithm. Compared with the existing results which rely on the knowledge of the loss indicators or the loss probabilities, the proposed two filters require no information about the measurement loss, and are thus more applicable for practical scenarios. Finally, two examples of target tracking and permanent magnet synchronous motor are included to validate the feasibility and superiority of the proposed two filters.