Research Of Uncertain Information Fusion Based On Belief Function Theory

The thesis mainly researches the uncertain information fusion based on the belief function theory. Along with the development of modern science and technology, more and more information sources are emerging. The information comes from these sources is usually uncertain, even deceived and disturbed. For the advantages of belief function theory in uncertain information representing and reasoning, it has been applied widely in the domain of uncertain information fusion. Because counterintuitive results are obtained when combing highly conflictive evidences, the classical belief function theory suggested by Dempster and Shafer suffers from impugnation till now. The scholars dedicating in the research of belief function theory can be divided into two groups. One group defends to modify Dempster's rule of combination to solve its irrational redistribution of conflicting mass, the other one argues that such problem is not due to the rule but due to the unreliability of the evidence to be combined and it must be modified. This thesis focuses on the measurement of conflict and the combination of highly conflictive evidences, as well as the reliability evaluation of unreliable evidence and their combination.Firstly, the measurement of the conflict is investigated. After summarizing several schemes in conflict measurement in existence, a method for measuring generalized conflict has been developed. This measurement not only takes into account the uncommitted belief mass in the conjunction processing, but also involves the potential conflict. By comparing with former ones in numerical examples, the rationality and effectiveness of this method can be demonstrated. Secondly, many modified combination rules are classified into four class and the properties of an ideal rule are indicated. Then a global conflict proportionally redistributing rule of combination is proposed based on the generalized measurement of conflict. This rule is commutative and quasi-associative. It can combine the evidences which are highly conflictive and determine the weights of focusing downwards according to the cardinality of focal sets to be combined. When the number of evidence increasing, the result of the combination tends to their average. Analysis and simulation demonstrate this new rule can avoid most paralogisms of existing rules. Finally, we studied the reliability evaluation of evidence and the combination of unreliable evidences. In the context of decision level recognition fusion, an algorithm for combining unreliable evidences has been presented. This algorithm considers both the static and dynamic information to modify the quantitative model of evidence, and then fuses them by Dempster's rule of combination. The result of an experiment demonstrates the effectiveness of this algorithm.