| A parallel distributed detection system consists of multiple local sensors/detectors that observe a phenomenon and process the gathered observations using inbuilt processing capabilities. The end product of the local processing is transmitted from each sensor/detector to a centrally located data fusion center for integration and decision making. The data fusion center uses a specific optimization criterion to obtain global decisions about the environment seen by the sensors/detectors. In this study, the overall objective is to make a globally-optimal binary (target/non-target) decision with respect to a Bayesian cost, or to satisfy the Neyman-Pearson criterion. We also note that in some cases a globally-optimal Bayesian decision is either undesirable or impractical, in which case other criteria or localized decisions are used. In this thesis, we investigate development of several fusion algorithms under different constraints including sequential availability of data and dearth of statistical information. The main contribution of this study are: (1) an algorithm that provides a globally optimal solution for local detector design that satisfies a Neyman-Pearson criterion for systems with identical local sensors; (2) an adaptive fusion algorithm that fuses local decisions without a prior knowledge of the local sensor performance; and (3) a fusion rule that applies a genetic In addition, we develop a parallel decision fusion system where each local sensor is a sequential decision maker that implements the modified Wald's sequential probability test (SPRT) as proposed by Lee and Thomas (1984). |
