Data-driven processing in sensor networks

Adam Silberstein, Duke University
Rebecca Braynard, Duke University
Gregory Filpus, Duke University
Gavino Puggioni, Duke University
Alan Gelfand, Duke University
Kamesh Munagala, Duke University
Jun Yang, Duke University

Abstract

Wireless sensor networks are poised to enable continuous data collection on unprecedented scales, in terms of area location and size, and frequency. This is a great boon to fields such as ecological modeling. We are collaborating with researchers to build sophisticated temporal and spatial models of forest growth, utilizing a variety of measurements. There exists a crucial challenge in supporting this activity: network nodes have limited battery life, and radio communication is the dominant energy consumer. The straightforward solution of instructing all nodes to report their measurements as they are taken to a base station will quickly consume the network's energy. On the other hand, the solution of building models for node behavior and substituting these in place of the actual measurements is in conflict with the end goal of constructing models. To address this dilemma, we propose data-driven processing, the goal of which is to provide continuous data without continuous reporting, but with checks against the actual data. Our primary strategy for this is suppression, which uses in-network monitoring to limit the amount of communication to the base station. Suppression employs models for optimization of data collection, but not at the risk of correctness. We discuss techniques for designing data-driven collection, such as building suppression schemes and incorporating models into them. We then present and address some of the major challenges to making this approach practical, such as handling failure and avoiding the need to co-design the network application and communication layers.