System design in wireless sensor networks for parameter estimation and dynamic event region detection

In this dissertation, the practical system design issues are studied for statistical inference in wireless sensor networks (WSNs). First, the problem of distributed estimation of an unknown parameter corrupted by noise is studied. Imperfect data transmission between local sensors and a fusion center is considered and modeled as a Rayleigh fading channel. The conventional maximum likelihood estimation usually involves high computational complexity and is not suitable for resource-constrained WSNs. Efficient estimators are designed for different receiver models and their efficiency is shown both theoretically and through experiments. The distributed parameter estimation performance also depends on the selection of local quantization thresholds. Therefore, different threshold schemes are investigated under the minimax criterion. Two quantizer structures (sinusoid function and raised cosine function) are proposed. Simulation results show that the simple sinusoid structure outperforms the intuitive uniform structure and the raised cosine structure achieves near optimal performance. The problem of dynamic event region detection in WSNs is studied next. To provide detection results at each time step, a distributed event region tracking algorithm is proposed. The system dynamics (modeled by a dynamic Markov random field) and information collected from neighbors are used to predict the underlying hypothesis at each sensor node and its local observation is used for update. The performance of the proposed algorithm is analyzed both theoretically and through simulations. Detecting and reconstructing critical dynamic event regions at a control center is an important application of bandwidth-limited WSNs. This problem is studied with emphasis on adaptive bandwidth allocation for sensor data transmission. To meet the stringent bandwidth and energy constraints in WSNs, only a few selected sensors are allowed to transmit compressed data to a control center. To reconstruct and track the field state map at each time step, a processing framework including sensor selection, local and central processing is proposed. Adaptive bandwidth allocation is obtained by solving a conditional entropy based optimization problem. The overall communication costs in terms of bandwidth and energy consumption of the proposed framework are evaluated by considering all possible overheads in a practical communication protocol.