NeTS-NOSS: SNI: A General and Robust Networking Architecture for Distributed Data Processing in Sensor Networks

Carnegie Mellon University, Pittsburgh PA

Investigators

Abstract

Sensor Network Inference (SNI) architecture is the first general and robust networking architecture developed specifically for inference in sensor networks that enables the rapid deployment of a wide range of complex large-scale querying, data processing and actuation tasks on a low-cost wireless sensornet. Unlike most previous approaches that focus on individual examples of such inference tasks (e.g., tracking or contour finding), our infrastructure is leveraged by a powerful abstraction of such tasks, Junction Trees, which enables the efficient solution of many inference problems, including probabilistic inference (e.g., sensor calibration and target tracking), regression (e.g., data modeling and contour finding), and optimization (e.g., actuator control, decision-making, and pattern classification). SNI is general and easy to deploy: the effective abstraction for a wide range of complex tasks enables the rapid deployment of novel sensornets applications. Furthermore, our approach is resource aware, efficient and adaptive: nodes have limited computational, communication and power resources, thus SNI automatically optimizes its communication pattern to reduce resource usage; this optimization is data driven, since the complexity of the high-level task is greatly dependent on the current state of the monitored phenomena. Robustness to node and communication failures is a fundamental element of SNI: Low-cost sensornets are prone to lossy communication, sensor and node failures; our architecture seeks to provide both theoretical and empirical robustness guarantees. Our evaluation process includes thorough testing on two different testbeds, with different hardware, in different locations. This evaluation is coupled with our education plan by using SNI in undergraduate and graduate classes.