【作者】 李爱；

【导师】 陈果；

【作者基本信息】 南京航空航天大学 ， 载运工具运用工程， 2013， 博士

【摘要】 航空发动机结构极复杂，工作在高温、高速的恶劣条件下，极易发生各种机械故障。据统计，在造成各类飞行事故的诸因素中，发动机故障原因所占比例一般在25％～30％，而航空发动机转子系统及传动系统中的齿轮和轴承磨损失效是航空发动机研制和使用过程中所出现的主要故障。由此可见，及时有效地诊断和预测出航空发动机的磨损故障，对于提高飞行安全，降低发动机维修成本，实施航空发动机视情维修，具有重要意义。然而，由于航空发动机的复杂性，各种磨损数据与磨损故障之间是一种模糊的、非线性、不确定的关系，传统方法已经不能满足磨损故障诊断的要求。鉴于此，本文将现代人工智能和模式识别技术引入航空发动机磨损故障诊断，围绕航空发动机磨损故障智能诊断若干关键技术进行研究，主要内容如下：1）不局限于正态分布假设的磨损界限值的制定。抛弃了传统油样数据正态分布假设，提出了基于支持向量机的磨损界限值制定方法。利用支持向量机从大量的油样分析数据中估计出概率密度函数，再依据估计出的概率密度函数得到航空发动机磨损界限值。该方法利用了支持向量机全局最优、良好泛化能力，以及解的稀疏性等优越性能，与传统统计方法相比，更具科学性和合理性。最后本文应用实际的航空发动机油样光谱数据对方法进行了验证分析，表明了方法的正确有效性。2）磨损趋势的组合预测方法。对油样分析数据进行数学建模，外推出未来发展趋势，对于航空发动机磨损状态的预测，尽早对故障的发展趋势进行预测和评估，从而避免重大事故的发生和及时安排维修工作，具有重要意义。鉴于此，本文提出基于最小二乘支持向量机的组合预测方法，首先利用灰色预测模型，神经网络预测模型和AR预测模型进行单项预测，然后利用最小二乘支持向量机方法实现组合预测，同时利用粒子群算法对支持向量机参数进行了优化。该方法解决了单一预测模型的信息源不广泛性，对模型设定形式敏感等问题。最后，利用实际的航空发动机油样光谱数据对方法进行了验证，表明了本文组合预测方法较单一预测模型方法大大提高了预测精度。3）磨损故障诊断知识规则的自动提取。为了解决目前航空发动机磨损故障智能诊断专家系统普遍存在知识获取能力弱，知识更新困难，知识适应性差等方面的缺陷，本文提出了基于支持向量机的数据挖掘技术，利用支持向量机进行了磨损规则的自动获取研究。在该方法中，首先利用遗传算法对样本数据特征进行选取，然后将特征选取后的数据样本映射到一个高维特征空间中，得到样本的最优分类超平面以及支持向量，利用支持向量机聚类算法得到样本的聚类分配矩阵，最后在聚类分配矩阵的基础上构建超矩形，得到超矩形规则，并利用规则合并、维数约简、区间延伸等方法对超矩形规则进行了简化。针对样本严重不平衡问题，本文采用过抽样算法中典型的SMOTE算法对故障样本进行重采样之后再进行规则提取，取得了良好的效果。同时，开发了专家系统与国外著名数据挖掘开源软件Weka的接口技术，利用Weka软件的数据挖掘算法实现了航空发动机磨损故障诊断专家系统的知识自动获取。最后，利用实际的航空发动机故障数据进行了验证，表明了本文方法的正确有效性。4）基于多Agent的磨损故障融合诊断方法。该方法综合运用各油样分析方法的冗余性和互补性，有效地利用各种油样分析方法的特点和优势以提高诊断精度。该多Agent诊断系统主要包括颗粒计数Agent、理化分析Agent、铁谱分析Agent、光谱分析Agent、总控Agent、调度Agent、通信Agent、融合诊断Agent、油样数据和知识规则库以及人机智能界面。本文根据飞机发动机磨损故障诊断的实际情况，给出了各Agent诊断规则，并用具体的油样分析数据进行了验证，表明了多Agent融合诊断的有效性。5）最后本文将所研究的若干智能方法运用于与成都飞机工业（集团）有限责任公司以及北京航空工程技术研究中心合作开发的航空发动机磨损故障诊断专家系统中，实现磨损界限值制定、磨损趋势预测、融合诊断以及专家系统的知识自动获取。应用结果表明，本文的研究工作大大提升了航空发动机磨损故障智能诊断专家系统的智能化和自动化水平。更多还原

【Abstract】 Aero-engine has extremely complex structure, and easily has broken down all kinds ofmechanical failure working in harsh conditions of high temperature and high speed. According tosome statistics, in the factors that cause various types of flight accident, the proportion of the enginefailure reason is generally in the range of25%to30%. Moreover, the gear and bearing wear failurein the aero-engine rotor system and transmission system is the main fault occurred in the study andapplication. Therefore, it is critical to diagnose and predict the aircraft engine wear fault timely andeffectively in order to elevate the flight safety, lower the engine maintenance cost, implementaero-engine condition based maintenance. However, because of the complexity of the aero-engine,the relationships between the various wear data and the wear failure is fuzzy, nonlinear and uncertainrelationship, and the traditional methods can’t meet the requirements of the wear fault diagnosis. Inview of this, this paper introduces the modern artificial intelligence and pattern recognitiontechnology into the aero-engine wear fault diagnosis and has commenced the study on some pivotalproblems about aero-engine wear fault intelligent diagnosis. Now the summary of main workingcontents in this paper is as follows:(1) The establishment of the wear threshold that is not limited to the normal distributionassumption. The establishment of the wear threshold based on Support Vector Machine is proposedwith abandoning the traditional normal distribution assumption of the sample data. The probabilitydensity is estimated from a large number of oil samples by using Support Vector Machine, and thenthe wear threshold is obtained according to the probability density. This method takes advantage ofserious advantages of Support Vector Machine such as the global optimal solutions, goodgeneralization ability, and the sparse solution. This method is more scientific and reasonable incontrast to the traditional statistical methods. Lastly, the verification analysis is done by using theactual Aero-engine spectroscopic data, and the results suggest the correctness and effectiveness ofthe method.(2) Combinational Forecast Method of the wear trend. It is critical to estimate the futuredevelopment trend through making Mathematical Modeling for oil sample data in order to predictthe aircraft engine wear trend, and forecast and evaluate the development trend of the fault as earlyas possible, so as to prevent major accidents and schedule maintenance work in time. In view of this,this paper proposed the combinational forecast method based on Least Square Support VectorMachine. The first step is using AR model、GM(1,1) model and BP neural network model to predictindividually, and the next step is combination forecast based on LSSVM, at the same timeoptimizing the parameters of SVM method by using particle swarm algorithm. This method solvessome issues such as comprehensive information of the single forecasting model, and sensitive to themodel form setting. Lastly, the verification analysis is done by using the actual Aero-enginespectroscopic data, and the results show that, compared with the individual forecast methods, thecombinational forecast method has greatly improved the prediction precision.(3) Automatic extraction of wear fault diagnosis knowledge rules. In order to solve the defectsof the current aircraft engine wear fault intelligent diagnosis expert system, such as the weakness ofcapability of knowledge acquisition, the difficulty of knowledge updating, poor adaptability of knowledge, and so on. In this paper, a data mining approach based on SVM is proposed to extractwear rules automatically. In this method, the first step is to choose the features of the sample data byusing Genetic Algorithm for improving the comprehensibility of the knowledge rules. Then the SVCalgorithm is adopted to get the Clustering Distribution Matrix of the sample data whose featureshave been chosen. Finally, hyper-rectangle rules are constructed on the base of the ClusteringDistribution Matrix. In order to make the rules more concise, and easier to be explained,hyper-rectangle rules are simplified further by using rules combination, dimension reduction andinterval extension. In addition, the SMOTE algorithm is adopted to resample fault samples in orderto solve the serious imbalance problem of samples. Meanwhile, the interface between the foreignwell-known open source software in data mining called Weka and expert system was researched,and the data mining method in Weka is used to extract knowledge automatically from aircraft enginewearing fault data. Lastly, the verification analysis is done by using the actual Aero-enginespectroscopic data, and the results suggest the correctness and effectiveness of the method.(4) Engine wear fault fusion diagnosis method based on Multi-Agent. This method improvesdiagnostic accuracy by using of characteristics and advantage and comprehensive use of redundancyand complementarity of various oil analysis methods. The Multi-Agent diagnosis System isconstituted by Particle Count Agent, Physicochemical Analysis Agent, Ferrograph Analysis Agent,Spectrometric Analysis Agent, General Control Agent, Scheduling Agent, Communication Agent,Fusion Diagnosis Agent sample data and knowledge rule database, and man-machine intelligentinterface. In this paper, according to the actual situation of aero-engine wear fault diagnosis, eachagent diagnosis rules are given. At last, the test results of the specific oil analysis data show theeffectiveness of the multi-agent fusion diagnosis.(5) Finally, the intelligent methods of this paper researched are applied to the developedaero-engine wear fault diagnosis expert system cooperated with Chengdu aircraft industrial (group)co., LTD and the Institute of Airforce Equipment. The establishment of the wear threshold, the weartrend, fusion diagnosis and automatic extraction of wear fault diagnosis knowledge rules are realized.The application results show that the work researched has greatly increased the intelligent andautomation level of the aero-engine wear fault intelligent diagnosis expert system.更多还原