Distributed Sensor Fusion Networks

This paper introduces a framework for processing sensor measurements with respect to their dependability. Each measurement is represented as an observation, i. e., a compound of a name, a measurement instant, the measured value, and a confidence marker indicating the dependability of value and instant. Sensor observations are processed by a network of fusion nodes in order to produce data with higher confidence. The confidence of each measurement is attached to the transmitted data by the confidence marker. The sensor fusion algorithms use a probabilistic model for sensor readings where the expected variance of a measurement corresponds directly to its confidence. The presented approach supports a modular system development, because the fusion algorithms are implemented only with respect to the interface specification and thus are independent of the actual control application. Second, it is possible to extend existing sensor applications with fusion tasks. Given that the necessary timing constraints can be satisfied, the modified application will show the same temporal behavior as the original application.