Distributed information fusion in sensor networks

This thesis addresses the problem of design and analysis of distributed in-network signal processing algorithms for efficient aggregation and fusion of information in wireless sensor networks. The distributed in-network signal processing algorithms alleviate a number of drawbacks of the centralized fusion approach. The single point of failure, complex routing protocols, uneven power consumption in sensor nodes, inefficient wireless channel utilization, and poor scalability are among these drawbacks. These drawbacks of the centralized approach lead to reduced network lifetime, poor robustness to node failures, and reduced network capacity. The distributed algorithms alleviate these issues by using simple pairwise message exchange protocols and localized in-network processing. However, for such algorithms accuracy losses and/or time required to complete a particular fusion task may be significant. The design and analysis of fast and accurate distributed algorithms with guaranteed performance characteristics is thus important. In this thesis two specific problems associated with the analysis and design of such distributed algorithms are addressed.;For the collaborative signal and information processing methodology a number of theoretical performance guarantees is obtained. The collaborative signal and information processing framework consists in activating only a cluster of wireless sensors to perform target tracking task in the cluster head using particle filter. The optimal cluster is determined at every time instant and cluster head hand-off is performed if necessary. To reduce communication costs only an approximation of the filtering distribution is sent during hand-off resulting in additional approximation errors. The time uniform performance guarantees accounting for the additional errors are obtained in two settings: the subsample approximation and the parametric mixture approximation hand-off.;For the distributed average consensus algorithm a memory based acceleration methodology is proposed. The convergence of the proposed methodology is investigated. For the two important settings of this methodology, optimal values of system parameters are determined and improvement with respect to the standard distributed average consensus algorithm is theoretically characterized. The theoretical improvement characterization matches well with the results of numerical experiments revealing significant and well scaling gain. The practical distributed on-line initialization scheme is devised. Numerical experiments reveal the feasibility of the proposed initialization scheme and superior performance of the proposed methodology with respect to several existing acceleration approaches.