【作者】 易正俊；

【导师】 黄席樾；

【作者基本信息】 重庆大学 ， 控制理论与控制工程， 2002， 博士

【摘要】 本篇论文把多源信息融合的概念推广，不仅仅限于多传感器的信息融合，而是把多个渠道，多方位采集的局部环境的不完整信息加以综合，消除多源信息间可能存在的冗余和矛盾的信息，加以互补，降低其不确定性，以形成对系统环境的相对完整一致性描述的过程，从而提高智能系统的决策、规划、反映的快速性和正确性，降低决策风险，是一个涉及到信息科学，计算机科学，自动化科学的交叉学科，是目前信息社会所必须研究的一个重要方向。 本文阐述了信息融合的现状，分析了目前融合中存在的问题及发展方向。综合所查阅的文献，把信息融合系统结构划分为集中式，分散式，无反馈和有反馈分级结构四种基本模式。用证据组合理论融合多源信息时，提出了在两种不同情况的融合算法：其一，多分类器输出结果的组合算法，取出同一对象的不同分类特征组，对所有的分类特征组分别设计出具有不同分类能力的神经网络分类器，并用遗传算法训练神经网络；将每个分类器的输出结果作为一条证据，然后对每一条证据给出一个信任度，这个信任度能够反映对不同类的输入及同一类的不同输入的信任程度，再用证据组合理论融合这些证据，并改进了现有对冲突信息的处理方法，可以有效的提高分类精度，并且能够避免由于某一个神经网络分类器遭到破坏，而使整个分类系统陷于瘫痪状态的不良后果。其二，在对有些待识别的对象在时间范围内允许多次测量时，可通过多次重复测量，每次测量经分类器分类得到一条证据，随着测量次数的增加，证据积聚增多，最后可通过证据组合理论融合这些证据信息，则可识别待识别的对象，采用这种融合算法，可以消除由于测量的误差和神经网络的偶然偏差所带来的误识，从而提高了模式识别的精度。用模糊积分融合不确定的多源信息时，可以分析和处理多源的不确定信息，它与传统的Bayes方法比较，不需要概率的先验信息及其概率分布，克服了证据组合理论融合方法中的证据难于获得，计算量大等问题，采用模糊积分的多源信息智能融合算法，如果配上一个专家系统，能够实时地调整融合结果，从而可显著地提高系统的智能度。本文根据此理论解决了一个基于模糊积分的高速公路上的智能型辅助驾驶系统中的多源信息融合算法，用模糊积分方法融合高速公路上智能辅助驾驶系统的多源信息，确定汽车应采取的安全运行模式，使之经过专家系统的推理，调整汽车运行中的可控参数值，进而可避免交通事故的发生，提高道路的通行能力。解决了特征输入—决策输出的信息融合问题，提高了系统不确定信息的处理能力和系统的智能度。<WP=5>粗糙集理论应用于多源信息的融合，可对不完整和不确定的数据进行分析，剔除相容信息，抽取潜在有价值的规则知识，获取最简，最快的信息融合算法，解决了数据超载和不完整的信息融合问题。模糊神经网络和证据组合理论，给出了多源信息在空间和时间上的两级数据融合结构，结合模糊积分理论和证据组合的优点，把模糊推理机制嵌入神经网络结构中，既能处理精确信息，又能处理模糊信息，而且克服了模糊规则和隶属函数不易确定的缺点，拓宽了神经网络处理信息的能力，使其既能处理精确信息，又能处理模糊信息，而且，训练好以后的神经网络无须任何额外信息，就可以融合多源信息。多元统计分析可用于多源信息的融合，本文用动态聚类和多组判别分析为经济发展建立了一个宏观经济预警模型，以某市的经济发展为背景，把经济系统在不同时期的经济运行状况划分为5种不同的经济运行模式，并对每类的模式建立边界识别函数，为经济预警提供直接的定量界限，给出的聚类方法克服了传统的经济预警系统在不该报警时而报了警的缺点。同时说明信息融合不仅仅限于多传器的信息融合，信息融合在现实社会中是普遍存在的。更多还原

【Abstract】 In this article.Multi-source information systems provide a purposeful description of the environment that single source information can’t offer Fusing several sources information increases the capability of intelligent system and yields more meaningful information, which is hard to be acquired by a single information source. Multi-source information fusion is a cross-science concerned with information science, computer science and automation science. The concept of information fusion is introduced at the beginning of this paper. the problems existing in fusion are analysed on the base of many references , and the future development of information fusion is predicted. The fusion structure is divided into four basic structures, which are centralized structure, distributed structure, hierarchical structure without feedback and hierarchical structure with feedback. Two fusion methods is proposed under different situations are proposed with evidence combination theory. One is the combination algorithm of results output by multi-classifier. The different classification feature of the same object is extracted and neural network classifier with different classification capability is designed separately,The genetic algorithm is applied to training the networks classifiers, output result of each neural network is thought of as evidence, Then the BPA of that evidence is determined. BPA must react the belief degree of different kind or same kind of different input. Evidence combination fusion is modified to deal with the conflicting information easily. The different capacity of each classifier is caused by different the classified feature. Feature input which cant be identified by one classifier may be identified by another. This paper says that model identification can be performed by multi-classifier, output result can be regarded as an evidence. Furthermore, we can determine the BPA of each classifier, then the precision of the model identification must be improved. Another is that multi-measurement can be made. An classified evidence by classifier can be obtained at each measurement, As the member of measurement increase4s, the accumulation of evidence increases, According to the object to be identified can be identified. A method of fusing uncertain information based on fuzzy integral is proposed. An intelligent fusion system is given, which can solve the FEI-DEO and DEI-DEO<WP=7>fusion problems. and is also effective for solving the problem of model identification for automatic highway operation by applying the intelligent fusion system to combining several feature index information. It enhances the capability of handling uncertain information. Two kinds of fuzzy neural networks, which are used to fuse multi-source information, are proposed by combing the advantages of neural networks and fuzzy inference, one is FMLPNN neural network, another is FBFNN neural networks. The training fuzzy neural networks fuses not only precious information but also imprecise and fuzzy data as well, which can solve the difficulty in getting fuzzy rules and membership function. The main advantage of fusing method with rough sets is that it does not require any prior or additional knowledge about the data. by analyzing uncertain, incomplete, and imprecise data, the fastest fusion algorithm is extracted, which can solve the problem of fusing over-loaded or incomplete information in multi-source information systems. The macroeconomic alarming model is established with statistical analysis, at first, the situations of economic operation of economic systems in different periods have been classified into five economic models by statistical analysis, then we have established the boundary cognition functions for each economic model, which provides the direct quantitative limit for the economic alarming.更多还原