Distributed in-network inference plays a significant role in large-scale wireless sensor networks (WSNs) in various applications for distributed detection and estimation. While belief propagation (BP) holds great potential for forming a powerful underlying mechanism for such distributed in-network inferences in WSNs, one major challenge is how to systematically improve the energy efficiency of BP-based in-network inference in WSNs. In this paper, we first propose a systematic and rigorous data-driven approach to building information models for WSN applications upon which BP-based in-network inference can be effectively and efficiently performed. We then present a wavelet-based BP framework for multiresolution inference, with respect to our WSN information modeling, to further reduce WSNs' energy. We empirically evaluate our proposed WSN information modeling and wavelet-based BP framework/multiresolution inference using real-world sensor network data. The results demonstrate the merits of our proposed approaches.
