ABSTRACT
In distributed sensor networks (DSNs), the fusion problems naturally arise when overlapping regions are covered by a set of sensor nodes. The sensor nodes typically consist of specialized sensor hardware and/or software, and consequently their outputs are related to the actual object features in a complicated manner, which is often modeled by probability distributions. While the fusion problems have been solved for centuries in various disciplines, such as political economy, the specific nature of fusion problems of DSNs require nonclassical approaches. Early information fusion methods required statistical independence of sensor errors, which greatly simplified the fuser design; for example, a weighted majority rule suffices in detection problems. Such solution is not applicable to DSNs since the sensors could be highly correlated while sensing common regions or objects, and thereby violate the statistical independence property. Another classical approach to fuser design relies on the Bayesian method that minimizes a suitable expected risk. A practical implementation of this method requires closed-form analytical expressions for sensor distributions to generate efficiently computable fusers. Several popular distributed decision fusion methods belong to this class [58]. In DSNs, the sensor distributions can be arbitrarily complicated. In addition, deriving closed form expressions for sensor distributions is a very 388difficult and expensive task since it requires the knowledge of a variety of areas such as device physics, electrical engineering, and statistical modeling. Furthermore, the problem of selecting a fuser from a carefully chosen function class is easier in an information-theoretic sense than inferring a completely unknown distribution [57].
