Data Fusion Architectures

Massachusetts Institute Of Technology, Cambridge MA

Investigators

Abstract

This research deals with sensor networks consisting of several sensors (nodes), each of which makes (generally noisy) observations related to a phenomenon of interest. The sensors use their observations, as well as messages received from other sensors, to form and transmit their own messages. The messages propagate through the network until, eventually one (e.g., a fusion center) or multiple sensors make a final decision. The broad objective is to address scalability issues and to derive important lessons on the merits of different sensor network architectures, thus providing insight to a system designer as opposed to the narrower task of optimizing a given sensor network Intellectual Merit: . The objective is to perform a detection task and decide between alternative hypotheses on the state of the environment, detect a target, based on noisy measurements, and in the face of severely limited (and possibly noisy) communications. These sensors may either communicate directly to a fusion center, or indirectly through a more general network topology. We will study the performance of different sensor topologies (trees, directed acyclic graphs, general graphs), and develop an understanding of the tradeoff between performance and the architecture of the sensor network. Broader Impact: The broader impact of this work will be twofold. On the application side, data fusion and sensor networks play a prominent role in a vast range of contexts, from environmental monitoring to target recognition. However, very few guiding principles are currently available to engineers. New insights and architectural principles obtained will be valuable to system designers, resulting to reduced development time, and higher system efficiency. On the intellectual side, this work will provide a new paradigm for the elusive problem of designing "system architecture," yielding useful insights potentially transferable to broader classes of information processing systems. This project will contribute to education, human resource development, and training through teaching and mentoring graduate students, and through the dissemination of the results through articles, chapters, conferences, seminars, and short/tutorial courses.