Study On Reliable Evidence Combination Methods In High-level Information Fusion

High Level Information Fusion (HLIF) corresponds to the advanced stage of InformationFusion which displays more intuitive results and has flexible methods. HLIF can better reflectthe ability of fusing system than Low Level Information Fusion (LLIF). HLIF demands for morereliable information fusion methods because of the uncertainty due to more complex processingand information. Evidence theory has significant advantages in information representation.However, related research of evidence theory has paid little attention to the reliability ofevidence combination, which is preventing it from meeting the serious demand of reliability.Therefore, the research of reliable evidence combination is not only can promote thedevelopment of evidence theory and HLIF, but also has important theoretical significance andpractical value for their generalization in other application areas.Based on a key project of national defense, the thesis charges for the information fusionproblem of integrated information system, aims at the evidence theory of information fusionmethod, takes the improvement of reliability of evidence combination results as the mainline,mainly studies the evidence distance measurements, the evidence conflict measurements and theconflict evidence combination methods, specific content as follows:1. It is very difficult to define the rigorous evidence distance directly. However, the indirectevidence distance measurement based on BPA probability transformation is simple and useful.Therefore, a weighed BPA probability transformation method based on uncertainty for theindirect method is studied. The suggested method which takes Pignistic transformation, PPTmethod and DSmP method as the basis chooses cognitive degree as the indication ofconservation or optimism, and the transformation process can adjust adaptively. Takingspecificity as the uncertainty weight, we can get an example of BPA probability transformationmethod based on uncertainty. The “hypothesis and verification” method is proposed to get boththe most specific BPA and transformation probability for their contradiction in calculation flow.The test results indicate that the method is reasonable, and can be used in indirect evidencedistance measurement based on BPA probability transformation.2. Considering the accuracy of measure would descend if the information content isdisaccord before and after the probability transformation, a BPA probability transformationmethod is proposed for the measurement accuracy of evidence distance based on BPAprobability transformation. For the difficulty to get the analytical solution of the transcendentalequation, a fast solving algorithm of numerical solutions is proposed based on the certification ofthe existence and uniqueness of the equation’s solution, as well as its monotonic and boundedproperties. The comparison between the2-norm distance of probability vectors and the commonly used evidence distance measurement shows that our indirect evidence distancemeasurement is reasonable and effective in measuring evidence distances.3. Aiming at the problem that the existing methods can not reflect the difference ofconflicting degrees among each other, two asymmetric evidence conflict measures are proposed.The proposed methods are based on the asymmetric character consideration and redefinition ofevidence conflict, and make good use of intersect and inclusion information of focal elementsbetween evidences synthetically. The first measure concerns simple calculation and smalldynamic range, while the second measure entails more complex calculation and larger dynamicrange. Finally, the experimental compare with the existing evidence conflict measurementsindicate that our methods, which can present the variant degrees of conflict among variousevidences, are more reasonable in presenting conflict degrees among evidences.4. Two kinds of conflict evidence discounting combination are proposed which can solvethe Dempster’s counterintuitive conclusions. The Dempster’s rule produced inconsistent resultswith people’s intuition, implying its processing of evidence differs from people’s cognition. Thethesis proposes two evidence importance models and two evidence combination methods:classified discounting and composite discounting for the static and dynamic evidencecombination respectively. To achieve this, we use the information processing method in people’sdaily life, and people’s reaction model to different evidences in biology research for referenceand construct the discount factor in Shafer’s discounting method considering the evidenceimportance and evidence reliability. The test results indicate that our methods are consistent withpeople’s cognitive process, and can solve the Dempster’s counterintuitive problems with morereliable combination results.