【作者】 周峰；

【导师】 缑锦；

【作者基本信息】 华侨大学 ， 计算机应用技术， 2012， 硕士

【摘要】 随着科学技术的发展，各种面向复杂应用背景的传感器系统大量涌现。人们迫切需要一种理论工具来解决多源信息融合问题。D-S证据理论在不确定性表示与处理上的优势，使其在信息融合领域得到广泛发展。在实际应用中，D-S证据理论只适合于证据间低冲突情况，当证据间存在强冲突时，直接采用D-S合成规则合成会导致合成结果出现悖论。针对冲突悖论问题，国内外学者分为两派：一派认为应该修改D-S(Dempster-Shafer)合成规则，另一派认为应该修改证据源。本文认为D-S合成规则本身没有问题，具备完备的数学性质，在合成前应该对冲突证据进行预处理，即修改证据源。修改证据源又分为两种：一种是证据折扣修正法；另一种是修改证据源模型法。本文选择前者，因为冲突证据本身带有不确定性，合成前应该对证据进行折扣修正。本文研究内容如下：提出基于可信度证据折扣修正和基于相似度动态调整的证据折扣修正两种方法。基于可信度证据折扣修正方法通过距离法和向量余弦法度量证据冲突，然后计算得出每条证据的可信度，以可信度作为折扣系数，对证据折扣修正；基于相似度动态调整的证据折扣度方法通过参考证据对证据源在大小和方向上进行不确定性判定，得出证据源的每条证据和参考证据的大小相似度αi和方向相似度βi，然后建立一个相似度动态调整模型，以这个动态调整模型的结果作为折扣系数，对证据进行折扣修正，得出多组证据，找出多组证据中，冲突最小的那组，直接采用D-S合成规则，作为合成结果。在实际应用中，冲突证据可能只是证据源里面的极少一部分，提出基于Jousselme距离的凝聚层次聚类方法。即通过聚类把证据源分为若干类，类之内的冲突小，直接采用D-S合成规则合成，类间采用证据折扣修正后，再采用D-S合成规则合成。聚类方法减少了需要预处理的证据，同时实验验证了合成结果的准确性和合理性。更多还原

【Abstract】 With the development of science and technology, all kinds of sensor systemfacing complex application background are springing up. A theoretical tool to solvethe multiple source information fusion is desperately needed. D-S evidence theory isdeveloped widely due to the advantages of the uncertainty in the representation andprocessing.In practical applications, D-S evidence theory is only suitable for low conflictsof evidences. While there are strong conflicts between evidences, D-S synthetic ruleswill result in synthesis paradox. For conflict paradox problems, views are divided intotwo kinds, one kind think that we should revise D-S (Dempster-Shafer) syntheticrules while other kind think that the source evidences should be modified. This paperthinks that D-S synthetic rules with the deepest mathematical properties have noproblems and the source evidences should be modified, the views in which aredivided into two kinds: one kind is evidence discount amendment method while theother kind is the method of modifying the evidence source model. This paper choosesthe former due to the uncertainty in conflict evidences, the main contents are includedas follows:Based on the credibility and the similarity of dynamic adjustment, two methodsof evidence discount correction are proposed. The former measures conflict evidenceby the methods of distance and vector cosine, then calculates the credibility ofevidence as the discount coefficient. The latter judge the uncertainty in size anddirection of evidence through the reference evidence to gain the size similarity αi andthe direction similarity βi of each evidence and reference evidence, and then build asimilarity dynamic adjustment model, the result of which will be the discountcoefficient.We can get many groups of evidence from similarity dynamic adjustmentmodel and find one whose conflict is smallest,using D-S synthetic rules to combine itas final synthesis result.In practical applications,conflict evidence may be just a very small part of thesource of evidences.A agglomerative hierarchical clustering method is proposed based on Jousselme distance in this paper. By Clustering, evidence source is dividedinto several classes, the conflict of evidences within the same class is small, whichcan directly combine by the DS combination rule.The conflict of evidences within thedifferent class is strong,so it need to discout evidence, and then using the DScombination rule synthesis.Clustering method can reduce the number of evidencewhich need topretreat.Experiments show that the synthesis results is accurate andreasonable.更多还原