Department: Electrical & Computer Engineering
Name: Periklis Liaskovitis
Email: pliaskov @ ucsd.edu
Grad Year: 2008
Sensor networks comprise of densely deployed devices (sensors) with sensing and wireless communication capabilities. The sensors are envisioned to be able to work autonomously over long periods of time, thus requiring energy efficient network operation. An important class of applications for such networks is to use sensors to estimate the spatiotemporal behavior of a physical phenomenon, such as temperature variations over an area of interest. The network thereby essentially acts as a distributed sampling system. It has been realized for such scenarios that sensors lying in close physical proximity are likely to generate correlated readings, thus creating redundancy in the data output by the network. Such redundancy can be exploited to increase network lifetime. The basic premise is that only a subset of all available sensors is kept active at each point in time while the others are in an energy efficient sleep mode. With appropriate choice of the active set of sensors, the negative impact on sensing quality can be kept to a minimum. If multiple such 'adequate' sets of sensors are found and activated sequentially there can be substantial increase in the overall network lifetime. In this work we present a novel approach for scheduling which sensors report at each point in time. Our approach adaptively selects the smallest set of reporting sensors to act as sampling points. By projecting the sensor space onto an equivalent Hilbert space, the method ensures sufficiently accurate sampling and interpolation, online, without a priori knowledge of the statistical structure of the physical process. Results are presented using both synthetic and real physical sensor data and show significant reductions in the number of active sensors.
« Back to Posters or Search Results