The objectives of this research programme were :(i) to develop truly-distributed signal processing algorithms for distributed sensor networks based on the diffusion mode of operation; these algorithms will be provably convergent, offer good compromises between complexity and performance, and minimize communications overheads; (ii) to test and confirm the theoretical performance of these algorithms through computer simulation in an environment of changing network topologies (i.e. losing and adding sensors, moving sensors, changing communication paths) (iii) to apply the algorithms to source location and multi-lateralization problems in distributed networks.
In this study we have developed distributed signal processing solutions to the distributed sensor problem of acoustic source localization. In many such distributed systems, the objective is to reach agreement on values acquired by the nodes in a network. A common approach to solving such problems is the iterative, weighted linear combination of the neighbouring values to which each node has access. Methods to compute appropriate weights have been extensively studied, but the resulting iterative algorithms still require many iterations to provide a fairly good estimate of the consensus value. In this project we have shown that a good estimate of the consensus value can be obtained within a few iterations of conventional consensus algorithms by filtering the output of each node with an adaptive filter, i.e. a filter that learns from the data. This appears to be a new application for adaptive filters. The resultant algorithms do not require knowledge of the network topology and can handle networks that change with time.