Research On Evidence Derivation And Combination In Evidence Theory

The technology of information fusion can realizes representation, reasoning, combinationand decision-making of multi-sources information, however, the observed information arealways uncertain due to effects of sensors’self-performance and environmental noises.Dempster-Shafer evidence theory has unique advantage in representation and combination ofuncertain information, the way of solving problem using this theory is that deriving evidencefrom uncertain information described the system at first, then using Dempster’s rule ofcombination to fuse these derived evidences. In real applications, however, evidence theoryhas two primary problems: the first one is how to derive evidence from a variety of uncertaininformation; the second one is how to combine conflicting evidences reasonably.To overcome the problems mentioned above, using Dempster-Shafer evidence theoryand fuzzy set theory, this thesis studies the problems of evidence derivation and combinationfrom a variety of uncertain information. The main works in the thesis are as follows:(1) The method of deriving evidence based on similarity measure of generalized fuzzynumbers. A new similarity measure between generalized trapezoidal fuzzy numbers ispresented; it combines the concepts of exponential distance, the perimeter and the area ofgeneralized trapezoidal fuzzy numbers for calculating the degree of similarity. Then 12 typicalexamples of generalized fuzzy numbers are given to illustrate the effectiveness of thepresented method. The diagnostic examples of machine rotor system demonstrate thepresented method can utilize more information to derive evidence.(2) The method of deriving evidence based on similarity measure of random-fuzzyvariables. The random-fuzzy variable is used to represent observed data affected byrandomness and fuzziness, a similarity measure between random-fuzzy variables is given toderive evidence. The diagnostic examples of machine rotor system demonstrate theeffectiveness of the proposed method.(3) The method of deriving evidence based on inverse pignistic transformation fromlinguistic information. A new method for the transforming fuzzy membership functions intoevidence with non-consonant structure based on pignistic transformation and its inversetransformation. Examples of target recognition illustrate that the non-consonant evidenceprovides less degrees of nonspecificity than the consonant case.(4) The method of combining conflicting evidences based on multi-objective optimization. Adopting the idea of discounting, an optimal model to learn discounting factorsbased on evidence distance criterion that considers improving focusing degree and reducingconflict simultaneously. Examples of target recognition illustrate that presented method ismore reasonable than some existing methods in both reducing conflict and focusing degree.