Skip to Main Content
Distributed in-network inference plays a significant role in large-scale wireless sensor networks (WSNs) in applications for distributed detection and estimation. Belief propagation (BP) holds great potential for forming an essential and powerful underlying mechanism for such distributed inferences in WSNs. However, it has been recognized that many challenges exist in the context of WSN distributed inference. One such challenge is how to systematically develop a graphical model of WSN, upon which BP-based distributed inference can be effectively and efficiently performed, rather than ad hoc. This paper investigates this challenge and proposes a general and rigorous data-driven approach to building a solid and practical graphical model of WSN, given prior observations, based on graphical model optimization. The proposed approach is empirically evaluated using real-world sensor network data. We show that our approach can significantly reduce the energy consumption in BP-based distributed inference in WSNs and also improve the inference accuracy, when compared to the current practice of distributed inference in WSNs.