Incremental Distributed Identification of Markov Random Field Models in Wireless Sensor Networks

Wireless sensor networks (WSNs) comprise of highly power constrained nodes that observe a hidden natural field and reconstruct it at a distant data fusion center. Algorithmic strategies for extending the lifetime of such networks invariably require a knowledge of the statistical model of the underlying field. Since centralized model identification is communication intensive and eats into any potential power savings, we present a stochastic recursive identification algorithm which can be implemented in a fully distributed and scalable manner within the network. We demonstrate that it consumes modest resources relative to centralized estimation, and is stable, unbiased, and asymptotically efficient.

I. Introduction

A wireless sensor network (WSN) is an ad hoc network of battery powered motes, which measures a physical field like the presence of a toxic chemical [1], and reconstructs it at a distant fusion center (FC). The transmission of all raw data out of the network is an energy-intensive procedure that quickly drains the batteries of the motes, and severely curtails the lifetime of the WSN [1]. Hence, much research has been recently directed towards the exploitation of the spatiotemporal dependencies in the sensor data to improve the power efficiency of the network via strategies like distributed source coding [2], correlated data gathering [3], source-channel decoding [4], distributed detection [5], [6], distributed filtering [7], distributed learning [8], and energy aware routing [9]. In every case, some knowledge of the statistical model of the field is a prerequisite, e.g., to design optimal codes [2] or routing tables [9].