OVERVIEW

Objectives – Methods – Accomplishments

A novel entropy-based method of sensor selection for information fusion in distributed sensor networks is considered. Our method is illustrated in the context of localization applications. Given 1) an intermediate probability distribution of the target location based on information from some initial sensor readings, and 2) the locations and the sensing characteristics of a set of additional sensors, our method selects an additional "optimum" sensor for fusing with the current target location distribution. The selected sensor is optimum in the sense that fusion of measurements from that sensor with existing information would yield the most entropy reduction of the target location distribution. In addition, the set of candidate sensors are evaluated for selection without obtaining the real measurements of the set of candidate sensors. We thereby avoid the significant energy expenditure that would be required to retrieve the measurements of all candidate sensors. Our method has the advantage of maximizing information gain while keeping communication cost low. The method is not only applicable to localization using single-modal sensors such as bearing sensors, range sensors and time difference sensors separately, but it is also applicable to localization using a mixture of multi-modal sensors. We have also proposed using the Cramér-Rao bound (CRB) to analyze the source and sensors' location influence on the source localization error. These results show that better source location estimate can be obtained when the source is inside the convex hull of the sensors. Various analytical studies and simulations were performed to conclude this simple heuristic sensor placement methodology to be valid and useful for practical applications. Numerous journal and conference publications have been reported.

APPROACH

From 2003:

A novel entropy- based method of sensor selection for information fusion in distributed sensor networks is considered. Our method is illustrated in the context of localization applications. Given 1) an intermediate probability distribution of the target location based on information from some initial sensor readings, and 2) the locations and the sensing characteristics of a set of additional sensors, our method selects an additional "optimum" sensor for fusing with the current target location distribution. The selected sensor is optimum in the sense that fusion of measurements from that sensor with existing information would yield the most entropy reduction of the target location distribution. In addition, the set of candidate sensors are evaluated for selection without obtaining the real measurements of the set of candidate sensors. We thereby avoid the significant energy expenditure that would be required to retrieve the measurements of all candidate sensors. Our method has the advantage of maximizing information gain while keeping communication cost low. The method is not only applicable to localization using single-modal sensors such as bearing sensors, range sensors and time difference sensors separately, but it is also applicable to localization using a mixture of multi-modal sensors.

FUTURE DIRECTIONS

In area (C), we propose to refine the minimum entrophy sensor selection technique. We have also initiated a generalization of the CRB technique to the Bayesian error bounding methodology which incorporates known constraints signal and geometrical constraints on the source. Preliminary results show these newer bounds to provide indeed tighter estimation bounds on sensor locations. Of course, the goal of this study is to use the results obtained here to practical optimum/near optimum placement of acoustical, seismic, and imaging sensors. Results obtained in area (c) will be used in the research in areas (a) and (b).

PEOPLE

FACULTY

K. Yao
D. Estrin
G. Pottie

GRADUATE STUDENT

Hanbiao Wang
L. Yip