| Wireless sensor networks (SNets) is a cost-efficient technology that is typically comprised of many low-power, low-cost sensors. Current and potential applications of SNets include tracking, automation, control, surveillance, reconnaissance, security, and monitoring. All of these applications require some form of intelligent signal processing and decision making algorithms. Proposed algorithms involving fusion-center-based architectures do not scale well with increasing number of sensors. Instead, scalable, robust and power-efficient distributed signal processing and decision making algorithms are required in order to realize the full potential of SNets. Therefore, in this thesis, we propose and analyze power-efficient in-network signal processing and decision making algorithms for applications that are modeled as distributed classification, data aggregation and symmetric function computation problems.;The work in this thesis is a step towards solving distributed signal processing and decision making problems in low-power distributed sensor networks in a cost-efficient manner. The low-power, constrained communication architecture of emerging sensor network technologies demand scalable, asynchronous, energy-efficient and robust sensor network architectures and algorithms. We address these concerns in this thesis by declaring the proposed algorithm as: (i) Scalable because communication is restricted to immediate one-hop neighbors. (ii) Asynchronous because no synchronization between sensor nodes or order of transmission of messages are required. (iii) Energy-efficient in terms of reduced total power and average power consumed. (iv) Robust to node and edge failures.;In this thesis we propose our solution in two main categories. (i) We propose and analyze pair-wise message passing algorithms, where sensors achieve the optimal classification performance. (ii) We propose and analyze algorithms in order to calculate the measurement statistics of certain functions. Under both these problems and proposed solutions, we assume each sensor initially has immediate access only to its observation, each communication takes place between sensor and its immediate neighbors, communication is not reliable, and node and edge failures occur.;The proposed distributed classification algorithm is inspired by Belief Propagation algorithm. Our contributions include the calculation of convergence time and energy expenditure of the algorithm for certain sensor network topologies. The distributed function computation algorithm is inspired by coalescing random walks. We prove the convergence time and message count of the algorithms for certain sensor network topologies. We also present detailed comparisons of the proposed function computation algorithm with existing gossip spreading algorithms. In the last part of our work, we generalize the studied communication strategies in order to achieve reduced power consumption. By introducing power-control to the sensor network, we seek for trade-offs between average power consumed and total convergence time. Finally, we show how our work can be applied to realistic sensor network scenarios. |
