DECENTRALIZED FUSION IN NETWORKS WITH ARBITRARY TOPOLOGY

Abstract

To increase scalability, modularity, and robustness of modern sensor networks, decentralized approaches become more and more important. Especially in applications with low bandwidth and communication rate, or high demands in power consumption, distributing the workload over a network can have advantages over centralized topologies. These advantages come with the need for more sophisticated algorithms to deal with common system noise, possible double counting of information as well as strategies to cope with asynchronous communication. In order to investigate new strategies for distributed estimation a exible and versatile simulation framework is needed. During the course of this dissertation, a toolbox for decentralized estimation is further developed, so that state of the art fusion algorithms (Optimal, Covariance Intersection, Ellipsoidal Intersection, Inverse Covariance Intersection, Naive, Sample based Fusion and Sample based Fusion with constant number of samples), can be tested quickly and e ciently. The goal is to make the toolbox more user-friendly, modular and an e cient simulator using which further research in decentralized fusion can be carried out. Sample based Fusion gives optimal results however the number of samples to be generated, stored and communicated to the fusion centre increases linearly with time, adding to huge communicational overhead. To counter this Sample based Fusion is carried out with constant number of deterministic samples. However this algorithm may lead to inconsistency. A method to preserve consistency of the estimates by bounding the covariances of the rejected samples is developed. Its performance is evaluated and compared against the state of the art algorithms. A `Two node' approach, where any type of sensor network with any number of nodes is reduced to a two node sequential fusion problem, is followed to extend the applicability of all the above fusion algorithms to networks which are hierarchical like trees and nonhierarchical like rings. This also adds to the modularity and reliability of the fusion algorithm. However the results obtained are not optimal. iii In a decentralised network there is a possibility of ring structures forming inadvertently. Data once incorporated into the fusion result is fused multiple times. This is called double counting and leads to biased results . To tackle this measurement noise samples are also generated and communicated with the fusion centre to check for independence of measurements. The structure of the samples used in Sample Based Fusion signi cantly in uences the computational and communicational performance of the algorithm. An alternate sample structure called back triangle is proposed instead of the simplex structure that was initially used so as to facilitate greater control over the samples and increased accessibility to its various components. Reduction of communicational overhead is also an added advantage.