Consensus Based Distributed Fusion In Sensor Networks

The fully distributed sensor network,which has neither fusion center nor all-to-all connection,has promising prospects in both civil and military areas.This is due to its attractive features such as scalability,selforganization and robust to nodes/links failure.Considerable research interests have been attracted to this field.However,the network structure of the fully distributed sensor network,on the other hand,makes the fusion task challenging.Recently,the consensus algorithm has been proposed and provides a powerful tool for the fusion problems in a fully distributed sensor network.Under the framework of the consensus algorithm,some achievements have been reported.Although the relevant research results are quite vast,there are still some interesting topics which worth further investigation.For example,the registration problems,the parameters uncertainty,the measurement uncertainty,and different field of view for different sensors.All of them are essential issues to be solved in practical scenarios.In this dissertation,we focus on the problems mentioned above under the framework of the consensus algorithm.The contributions of this dissertation are summarized as follows.1.A consensus sensor registration algorithm based on the expectation maximization is proposed.It is designed to deal with the fixed bias in the sensor network distributedly.Moreover,for the stochastic error such as the position uncertainty of the sensors,a variational Bayesian consensus algorithm for sensor location refinement is proposed for both linear and nonlinear measurement model.The simulations demonstrate that the proposed methods perform well and can approach the centralized algorithm when the number of consensus iteration is sufficiently large.2.A variational Bayesian consensus Kalman filtering algorithm is proposed to handle the uncertainty of the measurement noise covariance.In this algorithm,under the framework of the consensus algorithm,the variational Bayesian approximation is adopted to iteratively estimate the sufficient statistics of the measurement noise covariance on each step.Moreover,a variational Bayesian consensus Kalman filtering algorithm based on the Student-t distribution is proposed to deal with the measurement outliers.The simulations show that the proposed algorithms are robust to the time-varying measurement noise covariance and outliers.3.A consensus based labeled multi-Bernoulli filtering algorithm for sensors with different field-of-view is proposed.This is achieved by constructing an extended label space,modeling the undetected multi-target probability density and reconstructing the fusion weights based on the Dempster-Shafer evidence theory.The simulations demonstrate that using the proposed algorithm,each sensor in the sensor network can track multi targets accurately in the global field-of-view of the sensor network.