【作者】 陈帅均；

【导师】 吴钦章； 蒋平；

【作者基本信息】 中国科学院研究生院（光电技术研究所） ， 信号与信息处理， 2014， 博士

【摘要】 针对当前专家系统知识自动获取困难、自学习机制不足，推理效率低下等问题，对基于MapReduce的知识抽取技术，基于范例推理、范例相似度计算、范例约简算法，和基于产生式推理的规则与事实的表示、推理机制、RETE匹配算法等进行了深入分析与研究，提出了一些优化算法与新的模型，解决在飞行器评估领域要求能处理模糊知识、实效性强的应用问题。飞行器的参数（属性）繁多，但仅有部分参数对模式分类和评估有重要作用，如何提取重要参数为专家后续分析、总结评估规则，具有重要意义。基于粗糙集的知识抽取技术便可以从历史数据中众多的参数里面抽取重要的参数，经典的算法是一次性将小数据集装入内存进行计算，但无法处理像飞行器历史数据这种大数据集。本文分析了粗糙集理论中知识抽取算法的可并行性，构建了一种基于MapReduce的知识抽取模型，用于并行计算基于正域、边界域、信息熵的参数重要性测度。最后在Hadoop平台上进行了相关实验，实验表明，该技术能高效地处理飞行器历史数据等海量数据集。飞行器评估规则不成熟，评估参数之间关系复杂，模糊性强，因此从参数到顶层事实的推理不能采用“非白即黑”的二值逻辑。为了计算顶层事实的状态及置信度，本文提出了基于范例推理的方法，即根据历史数据指导评估飞行器当前的状态。首先介绍了当前常用的范例相似度度量方法如枚举型距离、欧氏距离、基于本体论的语义相似度、曲线相似度等，针对KNN经典算法的不足，提出引入权值和概率分布的KNN改进算法。最后，针对经过多次范例推理后范例库不断增大而导致占用存储空间大、检索速度慢等问题，提出了基于效用值的范例库记忆算法和基于支持向量的范例约简算法。随着对飞行器评估研究的深入，专家总结的规则也不断增多，规则的增多会导致知识库的不一致和推理机效率的下降。因此本文研究了不确定性产生式专家系统的体系结构，介绍了产生式知识的表示、飞行器评估中不确定性的来源、不确定的传播、前向推理方式、知识库一致性维护、解释机制等，深入剖析了经典的RETE匹配算法及常用的改进策略，最后提出基于代价模型的RETE优化算法，减少了匹配过程中join结点的数目，降低RETE匹配过程中的空间复杂度和时间的复杂度，提高了推理效率。针对飞行器评估任务中数据种类多、数据量大、实时性强、评估过程复杂等问题，提出一种混合推理模式的飞行器评估系统。该系统包括多个功能独立的模块，能够完成从飞行器原始数据到最终评估结论的推理和寻因。为了满足实时性需求，本文提出评估树结构和触发点机制。根据飞行器评估的特点，提出知识库与推理机分离、计算与推理分离的三层结构的评估推理系统，在顶层事实获取阶段采用范例推理，充分利用其非精确推理和自学习的优点；在评估推理阶段，采用高效的精确的产生式推理机制。通过历史数据的仿真测试，该系统能够即时完成飞行器评估，评估过程全程自动化。更多还原

【Abstract】 The current Expert System is difficult of obtaining knowledge automatically, lackof learning mechanism, and inefficity to reason. As for these problem, this paperin-depth studied on related technology and algorithm, which included knowledgeacquisition based on MapReduce, Cased-Based Reasoning, Cases Similarity algorithm,Case-Reduction algorithm, the rule representation and fact representation based onRule-Based Reasoning, reasoning mechanism, and RETE Pattern Matching algorithm.In addition, this paper proposed some new model and improved algorithm, in order tosolve the problem in aerocraft assessment field, that included difficulty for processingfuzzy knowledge and poorness of timeliness.The number of aerocraft parameters(attributes) was huge, but few parameters wereimportant in the pattern classification and assessment, therefor it was meaningful tostudy how to extract key parameters from historical data, so as to analysis andsummarize assessment rules. The knowledge extraction techniques based on Rough Sethad ability to acquire key parameters from numerous parameters, and the classicalalgorithm can only load small data set into memory to process at one time, however,which can not process huge massive data set likc aerocraft historical data. This paperargued that the knowledge extraction techniques based on Rough Set can be parallelcomputing, and this paper builded a knowledge extraction model based on MapReducein order to calculate the importance measure of parameters. Finally, this paper held arelated experment on the Hadoop, that showed this technique can efficiently processmassive data like aerocraft historical data.The aerocraft rules were immature, the relationship between parameters werecomplexity and ambiguous, as a result, the reasoning mechanism can not useNon-White or Black logic and Two valued logic from Remote Data to top facts. In orderto calculate the status and confidence of top facts, this paper proposed a method basedon Case-Based Reasoning, namely to assess the current status of aircraft according tohistorical data. Firstly, this paper introduced the current commonly used casessimilarity algorithms, such as enumeration type distance, Euclidean distance,ontology-based semantic similarity, the curves similarity. For the lackness of classicalKNN algorithm, this paper proposed improved KNN algorithm based on probability distribution and weights. The utility problem will occurs after the Case-BasedReasoning system runs many times, and this problem results in a decrease performance,such as a large storage space, a low retrieval rapid. To solve this proble, this paperproposed Memory Algorithm of Case Base Based on Utility Value and Cases ReductionAlgorithm Based on Support Vector.With the study on the aerocraft assessment, the number of rules expertssummarized become more and more, as a result, the knowledge base is inconsitent andthe reasoning is inefficient. Therefore, this paper studied the stucture of uncertaintyRule-Based expert system, and introduced some related key points, which included theknowledge representation, the source of uncertainty in the aerocraft assessment, thespreading of uncertainty, forward reasoning, the consistency maintenance of knowledgebase, explaination mechanism and so on. After analysising the classical RETE algorithmand commomly used improvement stragegies, this paper proposed improved RETEalgorithm based on Cost Model, which can automatically find the optimal RETEtopology, and reduces intermediate nodes, and greatly reduced RETE algorithm’s timecomplexity and space complexity.To solve the problem in the aerocraft assessment task, which includes the amountof data is huge, the assessment process is complex, and the conclusion of the assessmentis inaccurate, an aircraft assessment system based on hybrid reasoning model isproposed. The system comprises a plurality of independent module on function, and itcomplete reasoning and searching results from the original data of aerocraft to finalassessment conclusion. This paper presents assessment tree and trigger pointmechanism, in order to meet real-time requirements. According to the characteristics ofthe aircraft assessment, the top-facts are get using case-based reasoning to fully utilizeits non-precise reasoning and self-learning advantages; reasoning in the assessmentstage is use of accurate and efficient rule-based inference mechanism. Practical exampleusing historical data show that, the proposed assessment system performs very well, andcan real-time complete aerocraft assement, and the whole process is automatic.更多还原