Blind learning of the optimal fusion rule in wireless sensor networks

Abstract

This work presents a general framework for blindly estimating the sensor parameters of decision-fusion systems over wireless sensor networks (WSNs). The sensors report their binary decisions to a fusion center (FC) through parallel binary symmetric channels. Then, the FC makes the final decision by combining the noisy sensor decisions according to a certain fusion rule. We present an algorithm for the FC to blindly estimate the sensor parameters from the noisy sensor decisions received after a number of sensing periods. The algorithm covers a wide variety of situations that may arise in WSNs. For example, the algorithm is applicable when the FC knows in advance some of the parameters of some sensors, when it knows the true hypothesis for a subset of sensing periods, or when only a subset of sensors communicates their decisions in each sensing period. Based on the estimates of the system parameters, optimal channel-aware fusion rules are derived considering the minimum Bayes risk criterion. Simulation results show that, after sufficient sensing periods, the estimates of the WSN parameters are accurate enough for the fusion rule to exhibit near-optimal detection performance.