Distributed Evidence Filtering in Networked Embedded Systems

Abstract

Networked embedded sensor systems use the evidence gathered by spatially distributed heterogeneous sensor nodes possessing partial and different scopes of expertise, to make inferences on the scenario under observation. For such system to make an accurate collective decision, the partial and incomplete evidence provided by nodes must be processed in a simple and straightforward manner during the information exchange and fusion process. To achieve these objectives, we present a novel, unified approach named “non-recursive evidence filtering” based on the Dempster-Shafer (DS) formalism for evidence representation. It is capable of selectively fusing partial evidence in such a network to directly infer on events of interest such as threats occurring with a certain temporal distribution, while accommodating the varying reliability and accuracy of information sources. Certain restrictions on the coefficients impose several challenges in the design of such evidence filters. We show that the gain of these evidence filters at frequency zero is always equal to one and all coefficients must be non-negative. This suggests that arbitrary frequency shaping is not possible and a pure bandpass evidence filter is not realizable. A method to design a simple FIR evidence filter to detect periodic events of interest is presented. Multi-dimensional spatio-temporal evidence filters are discussed as a direct extension to the above, along with a low signature threat detection example that clearly illustrates the effectiveness of the evidence filtering concept in distributed networked embedded systems.