Study Of Multisensor Decision Fusion

Decision-level data fusion methods are studied both for hard-decisions and soft- decisions. Hard-decision is a single-hypothesis about the sensor抯 observation and soft-decisions are multiple hypotheses with quantitative measure of confidence or uncertainty for each hypothesis. After a review of the main techniques based on probability, probability intervals and frizzy sets, we propose an new decision fusion method based on fuzzy measures and fuzzy integral, using fuzzy measures to present the belief level to the local decisions and fuzzy integral to act as decision fusion operators. When applied to hard-decision fusion situations, we proved that fuzzy integral satisfies the prerequisite of legal decision fusion rule, and equals to the mostly used voting strategy ?K out of N fusion rule when all the beliefs to the local decisions are at the same level. If the beliefs to local decisions are different, it was shown that the fuzzy integral decision fusion method is superior to KIN rule. The nonlinear weighted fuzzy expectation characteristic of fuzzy integral was analyzed. Based on this characteristic we proposed to apply fuzzy integral to soft-decision fusion, using as a special kind of trade-off operators. Numerical analyses and the multi-model classification experiment results validated the feasibility of the fuzzy integral decision fusion method we presented. When applying fuzzy integral to decision fusion, the fuzzy measures representing the beliefs to local decisions are determined by fuzzy densities, which effect the fusion system performance prominently. We present a hybrid learning method to determine the optimal fuzzy densities from training sample data by using genetic algorithms and Powell directional acceleration method. Numerical experiment results showed the feasibility of the learning method. The optimal fuzzy densities determined by the learning method reflect the decision characters of the sensors perfectly even with a small number of training sample data. As the typical application of hard-decision fusion, distributed detection system design and performance analysis are of great importance. We presented a hybrid design method for distributed CFAR detection systems using evolutionary computation and Powell method. Compared with the mostly used Newton-Raphson method by numerical experiments, Our hybrid method needs not to calculate the gradient of the object function and more constraints can be added. Both the convergency and precision of the hybrid method are superior to Newton-Raphson method. The conclusions we got from the analyses of the relationship between the fusion performance of distributed detection systems with the decision fusion rule and local signal-to-noise ratios are of great importance to the optimal design of distributed radar networks.