Distributed Filtering with Wireless Sensor Networks

We investigate an ´inference first´ (IF) approach to information retrieval from a wireless sensor network (WSN). In this method, statistical estimation pertinent to the user´s application is implemented within the network (in-situ) and only the relevant sufficient statistics are exported. We formulate this procedure as a delay-free filtering problem on a spatio-temporal hidden Markov model (HMM), and propose a scalable approximate distributed filter. The algorithm is a novel application of the idea of iterated decoding, where we iteratively marginalize the joint distribution of the state of the HMM at two consecutive time epochs. We compare and contrast algorithms like the Gibbs sampler (GS), mean field decoding (MFD) and broadcast belief propagation (BBP), and discuss their suitability for in-situ marginalization. A simplified analysis of the energy gain achievable by the IF approach, relative to centralized processing, is provided.