【作者】 崔智全；

【导师】 钟诗胜；

【作者基本信息】 哈尔滨工业大学 ， 机械设计及理论， 2013， 博士

【摘要】 航空发动机是飞机的心脏，其健康状态对飞行安全有很大影响。发动机气路参数偏差值的求解和性能状态评价方法的建立是发动机状态监测的基础，也是国内各航空公司的迫切需要。本文在分析国内外航空发动机性能监控现状的基础上，从国内航空公司的实际需求出发，首先研究航空发动机报文解析技术和气路参数自变量筛选方法。之后分别提出了基于发动机性能基线的气路参数偏差值挖掘方法和参数核心变换支持向量机的偏差值回归求解方法。然后研究了基于偏差值的气路分析技术，主要研究气路分析技术中的偏差值预测、偏差值平滑、基于高维偏差值的发动机性能状态评价以及基于偏差值的气路故障诊断方法，最后将上述理论和技术方法应用到航空发动机性能监控系统的开发中。航空发动机气路参数偏差值求解之前对相关参数进行筛选处理，为此本文首先对与偏差值相关的自变量参数集合进行了筛选研究。为了消除多重相关性和重复属性对回归分析的影响，采用方差膨胀因子法对自变量参数之间的多重相关性进行诊断分析，采用航空发动机原理知识分析自变量参数之间的非线性关系。针对参数之间复杂的非线性关系，提出基于平均影响值和小波神经网络相结合的发动机自变量参数筛选方法，消除了无关自变量及弱相关自变量对回归效果的影响，为文中的回归分析实现了合理降维。针对机群发动机数量较大时，提出了基于基线的偏差值挖掘方法。根据气路参数偏差值的含义，将偏差值模型转化为参数标准化模型和性能基线模型的差值，并推导了气路参数标准化模型。考虑到大气环境温度、湿度以及传感器偏置对发动机性能的影响，提出了变指数的气路参数标准化修正模型，并根据性能基线的性质和函数逼近理论采用待定系数的多项式基线模型。根据偏差值定义将上述两个模型做差得到气路参数偏差值待定系数模型，将原厂家解析的历史数据代入偏差值模型进行回归分析，采用改进的高斯牛顿迭代法进行偏差值模型求解，从而准确挖掘出偏差值模型，并使用该模型进行偏差值的精确自主求解。论文还对机群发动机数量较少的小样本情况下气路参数偏差值求解方法进行了研究。提出一种航空发动机状态参数核心变换后的支持回归向量机的气路参数偏差值求解方法，该方法针对发动机测量参数之间的相似比例关系，通过增加优化参数实现输入集的局部降维，降低模型复杂程度，并提高模型的计算效率。同时该方法采用高维空间点到空间原点欧氏距离排列规则的训练集构造方法，提高支持向量机模型的泛化能力。该方法能够提高气路参数偏差值的求解精度求解速度。在上述偏差值求解的基础上，还对气路参数偏差值的预测技术进行了研究。针对气路参数偏差值时间序列中存在真实突变的特点，提出一种基于离散沃尔什变换的分式非线性聚合过程神经网络预测模型，由于有理函数具有更好的非线性逼近能力，采用非线性的有理分式空间聚合运算代替线性聚合运算。为避免离散数据在拟合过程中的精度损失，给出基于离散沃尔什变换Levenberg-Marquardt(LM)的网络学习算法，使用测试数据的离散沃尔什变换对内积运算代替过程神经网络中的积分算子，简化了计算过程，提高了计算速度。实际案例测试表明，文中提出的预测模型及网络学习算法能够提高预测精度和突变点处的预测灵敏度，具有更好的非线性逼近能力。在气路参数平滑处理技术方面，论文提出一种离群点分析与类指数平滑相结合的气路参数偏差值平滑处理方法，并以样本均方误差最小化为目标函数进行模型参数的优化，该方法在对偏差值序列进行合理平滑的同时保留了离群点数据。为了克服使用单维的气路参数偏差值进行发动机性能评估带来的片面性，文中提出基于离散Hopfield网络的发动机性能评估模型，采用能够表征发动机性能的高维参数对发动机性能进行评价，使评价结果更具科学性。为了克服相似故障区分难和使用故障指印图效率低且需要丰富专业知识的不足，文中提出基于故障指印图和自组织竞争网络相结合的气路故障诊断方法，该方法能够提高气路故障诊断的效率和准确性，并在增加新的故障模式时，能重新快速进行故障模式分类，实现准确快速的故障诊断。最后，以中国国际航空股份有限公司的发动机工程管理为应用背景，研究性能基线与气路参数偏差值在性能监控中应用的关键技术，并从该公司发动机工程管理的实际需求出发，开发了基于Web的航空发动机性能监控系统，实现了气路参数偏差值的自主求解，并为基于Web的航空发动机健康管理与维修决策支持系统其它模块提供数据支持，进而实现了状态参数自动报警、状态参数趋势分析以及发动机健康状态排队等功能。更多还原

【Abstract】 Aero-engine is the heart of aircraft, its health condition has a great impact onflight security. The gas path parameter deviation solution and performance statusevaluation method established are foundation of engine condition monitoring, andthey are also urgent needs of various airlines. Based on the analysis of domestic andinternational aero-engine performance monitoring research, according to the need ofairline, first the message parsing technology and gas path parameter independentvariables selection method are studied in this paper. Gas path parameter deviationmining method based on engine baseline and deviation regression solution based onparameter core transform SVM are proposed. Then the gas path analysis technologybased on the deviation is studied, mainly the deviation forecasting, deviationsmoothing, engine performance evaluation based on high dimension deviation andgas path fault diagnosis based on deviation are studied. Finally, the theoreticalmethods and calculation results are applied to the development of aero-engineperformance monitoring system.The relevant parameters must be selected before the aero-engine gas pathparameter deviation solving, so first the selected method of argument parameterssets relevant to deviation is studied in this paper. In order to eliminate the hazards toregression analysis of multiple correlation and repeat property, the variance inflationfactor method is used to diagnose multiple correlations between argumentparameters, the knowledge of aero-engine principle is used to analyze the non-linearrelationship between gas path parameter independent variables. According to thecomplex non-linear relationship between the argument parameters, the engineargument parameters selection method based on the combination of mean impactvalue(MIV) and wavelet neural network is proposed, which can eliminate unrelatedand tiny related independent variable effect to regression, and which reasonablyrealize the dimension reduction purpose for regression theory of this article.Then the deviation mining method is proposed. The deviation model is dividedinto difference of parameter standardization model and performance baseline modelaccording the meaning of gas path parameter deviation, and the gas path parameterstandardization model is derived. Taking into account the influence of ambienttemperature and humidity, and sensor offset, the corrected uncertain exponentstandardized model based on deviation is proposed. Uncertain coefficientspolynomial baseline model is used according to function approximation theory. Theuncertain coefficients gas path parameter deviation model is obtained using theabove two models to subtract each other, the regression analysis is done by inputting the original manufacturer historical data into the deviation model. The deviationmodel is mined accurately by using the improved Gauss-Newton iterative method tosolve the deviation model, and the deviation is accurately self-solved using thatdeviation model.The solution method of aero-engine gas path parameter deviation under smallnumber engine in the fleet is studied in this paper. Gas path parameter deviationsolution method based on aero-engine state parameter core transform support vectorregression machine is proposed, the input set local dimensionality reduction isachieved by increasing the optimization parameters according to the similarproportional relationship between the engine measurement parameters, thecomplexity of the model is reduced and calculation efficiency is improved. Thetraining set constructor based on Euclidean distance arrangement rule ofhigh-dimension space point to origin point is proposed in order to improve thegeneralization ability of support vector machines. The algorithm can improve thegas path deviation solution accuracy and solution speed of that model.Gas path parameter deviation forecasting techniques is studied on the basis ofthe deviation solving. A fractional nonlinear polymerization process neural networkprediction model based on discrete Walsh transform is proposed according to truemutation data exist in gas path parameter deviation time series, since the rationalfunction has better nonlinear approximation ability, nonlinear rational fractionalspace aggregation operations substitute for linear aggregation operations. In order toavoid the loss accuracy of the discrete data in fitting process, the network learningmethod is proposed based on Levenberg-Marquardt(LM) algorithm of discreteWalsh transform, the inner product of measurement data discrete Walsh transform isused to substitute the integral operator in process neural network, that simplify thecalculation process, improve the calculation speed. The conclusion can drawn frominstance that the forecasting model and network learning algorithm proposed in thispaper can improve prediction accuracy and prediction sensitivity on mutations dataand has better nonlinear approximate ability.In the gas path parameter smoothing technology, the outlier analysis andsimilar exponential smoothing combination of gas path parameter deviation smoothprocessing technique is proposed in the paper, and the parameters are optimizedusing sample mean square error minimum as objective function, the algorithm canreasonably smooth the deviation series and reserve the outlier data. In order toovercome the one-sidedness using one dimension gas path parameter deviation toevaluate engine performance, the engine performance evaluation model based ondiscrete Hopfield network is proposed, the high dimension parameter which cancharacterize engine performance is used to evaluate engine performance, whichmake the evaluation results has better scientific. In order to overcome difficult to distinguish similar fault, low efficiency of fault fingerprints figure used and requiresstrong expertise difficulties, the gas path fault diagnosis based on Self-OrganizingFeature Map(SOFM) under fingerprints figure is proposed in this paper, the methodcan improve the efficiency and accuracy of fault diagnosis, when new fault modelsis adding, it can quickly and easily achieve fault re-classification, to achieveaccurate and fast fault diagnosis.Finally, the key technology of performance baseline and gas path parameterdeviation in performance monitoring applications are studied based on engineproject management application of air china, and from the actual needs of thecompany’s engineering department, a Web-based aero-engine performancemonitoring system is developed, which can achieve self-solving techniques of thegas path parameter deviation, and provide data support to other modules ofWeb-based aero-engine management and maintenance decision support system,achieve auto alert and trend analysis of state parameter and engine health statusqueue function.更多还原