Decentralized Detection In Wireless Sensor Networks

Wireless sensor networks consist of small nodes which are capable of sensing, computation, and communication. The initial goal of a sensing system is to detect the events of interest. In this project, we study decentralized detection in sensor networks. First, we will review the classical decentralized detection framework, in which a set of spatially distributed sensor nodes receives information about the state of nature. Because of the channel bandwidth constraint, each node sends a quantized function of its observations to a central node, called fusion center. The fusion center then decides on one of the alternative hypotheses. Designing the local functions of individual sensor nodes and also the decision rule at the fusion to minimize the probability of error is of interest in this context [1, 2, 3]. The classical framework does not adequately take into account important features of sensor network architecture and of the wireless channel. Next, we will review some works that consider these features. The type of channel imposes some restrictions on the messages from sensors to fusion. Two different types of wireless channel between the sensors and the fusion center is studied here: multiple access channel [4] and broadcast channel [5]. In multiple access channel the total amount of data sent by all sensors is bounded by the channel capacity [4], and in broadcast channel, there is no central node and all nodes try to reach a consensus by successive retesting and rebroadcasting of the updated decision [5]. A common architecture for wireless sensor networks is the tree configuration. Here, we will review the optimal strategies for sensor nodes (i.e. local functions, which are based on the messages received from predecessors and own observations) and also the decision rule at fusion in terms of optimizing the error exponent [6]. In the end, as the novel part of this project we propose a totally decentralized detection method which is based on gossiping. In this method, we assume that nodes are only capable of short-range communications. In each round of the algorithm. sensors communicate twoby-two (locally) and update their likelihood ratio functions. We will prove that the proposed algorithm converges and its solution is exactly the same as the centralized scheme.