A systematic probabilistic approach to energy-efficient and robust data collections in wireless sensor networks

We propose a systematic approach, based on probabilistic graphical model, to infer missing observations in Wireless Sensor Networks (WSNs) for sustaining environmental monitoring. This enables us to effectively address two critical challenges in WSNs: (a) energyefficient data gathering through planned energy-saving sleep cycles and (b) sensor-node failure tolerance in harsh environments. In our approach, we model the spatial correlations in a sensor network as a pairwise Markov Random Field (MRF). Our MRF model is constructed from historical data using Iterative Proportional Fitting (IPF). Then Loopy Belief Propagation (LBP) is employed to estimate the missing data at the data sink. We demonstrate our approach using realworld sensed data on 32 × 32 grids. Empirical results show our approach can achieve the high rates of estimation accuracy (e.g. 65–90% for soil moisture data), even when the unobserved nodes consist of more than 50% of the total sensing nodes.