【作者】 栾圣罡；

【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， 2008， 博士

【摘要】 航空发动机是飞机的动力核心,其结构复杂、工作环境恶劣,是飞机的主要故障来源,其健康状态的好坏直接影响着飞行安全和航空公司的效益。本文在国家自然科学基金(项目号: 60572174)及中国国际航空公司(国航)科研基金的资助下,研究了基于气路参数的民用航空发动机健康状态监视方法及其系统。发动机气路参数样本本质上是一种混有较强随机噪声的时间序列,经常出现孤立突变数据和趋势突变数据等异常。小波变换是一种多尺度时-频分析工具,对于数据点突变和数据趋势突变等局部特征都能很好地描述,十分适合处理非平稳的发动机气路参数时间序列并识别其异常数据。为了提高发动机性能数据的可用性,及时识别发动机的异常状态并准确恢复原数据,论文提出了基于信号连续小波变换模极大曲线搜索的信号突变识别及重构方法,该方法可以在识别数据突变的同时有效地抑制随机噪声,进而有效地恢复真实信号。为了解决采用连续小波变换带来的运算速度问题,给出了基于傅里叶变换的信号连续小波变换与反演的快速算法。通过仿真算例研究了边沿效应及小波变换模极大曲线搜索过程中可能出现的伪模极大曲线等问题的处理方法,籍此提高突变数据识别及信号重构的准确性。发动机气路参数数据预处理的实例表明,所提出的方法能有效地恢复发动机性能参数数据,其信号重构效果优于一般的小波软阈值降噪方法。为解决复杂系统状态预测模型难以建立且预测精度难以保证的问题,提出了一种混合递归过程神经网络。设计了网络的拓扑结构,通过引入一组合适的正交基函数对网络的输入函数和连接权函数进行展开,实现了网络的简化。给出了一种基于弹性BP算法的混合递归过程神经网络学习算法。将网络与几种传统神经网络分别应用到Mackey-Glass混沌时间序列预测中,验证了网络的有效性。最后将上述网络分别应用到航空发动机气路参数预测的具体问题中,对比结果表明,混合递归过程神经网络在这一问题中有更好的工程可用性,是一种有效的发动机健康状态监视工具。针对基于最小二乘理论的传统数据拟合方法在求解双随机变量间的函数关系时的不适应性问题,提出了一种可同时考虑双拟合变量的随机噪声影响的欧氏距离最小二乘支持向量回归方法,给出了基于Matlab二次规划工具箱的实现算法。通过实例验证了该方法在发动机性能趋势分析中的有效性。另外,由于拟合不同的发动机性能参数的时间序列数据可能需要不同的拟合方法,简要地提出了基于多阶次多项式回归算法的发动机性能参数时间序列数据拟合方法。层次分析法是一种实用有效的多准则决策方法,但其判断矩阵的确定需进行一致性检验,同时,其对于评价指标观测值的评分准则一般由专家凭经验给出,带有很强的不确定性。为此,提出一种灰色层次分析法,该方法将判断矩阵的拟优一致化求解方法与灰色聚类评价理论相结合,有效地克服了层次分析法的上述缺点。针对发动机状态综合评价问题,建立了发动机状态综合评价指标体系;基于拟优一致判断矩阵求取了发动机状态的各评价指标综合权值;对发动机状态的各评价指标进行了灰类划分,并基于三角权函数对评价指标所表征的发动机实际状态观测值进行评分。最后给出了一个发动机状态综合评价问题的示例,结果表明了方法的有效性。从国航发动机健康状态监视及工程管理这一应用背景出发,进行了软件开发需求分析,系统功能模型、数据模型和体系结构的设计。综合本文提出的几种发动机性能监视和健康状态综合评价方法,开发了面向航空发动机全寿命管理的发动机健康状态监视软件系统。系统包括发动机维护及状态监视信息管理、发动机状态趋势分析、发动机状态参数超限自动报警、发动机状态综合评价及发动机拆换期预报等功能模块。该系统作为“发动机全寿命管理系统”的重要组成部分已被应用于国航的发动机健康状态监视工作中,测试结果表明了系统的有效性。更多还原

【Abstract】 Aeroengine is the kernel power supply of aircraft, which structure is complex and which operation environment is bad, so it is the main source of aircraft trouble, and its condition is the direct influencing factor of flight safty and airline company benefit. In this dissertation, aeroengine condition monitoring technique and software system based on aeroengine gas path parameters is studied under the supports of the National Natural Science Foundation of China (Grant No. 60572174) and the Air-China Scientific Research Foundation.Aeroengine gas path parameter sample is a non-stationary time series blending stochastic noise essentially, of which there are frequently singularity data such as isolated mutational data and trend mutational data, etc. Wavelet transform is a kind of multi-scales time-frequency analysis tool, which can extract the characteristics of isolated data point mutation and data trend mutation excellently, consequently, wavelet is adapt to process non-stationary aeroengine gas path parameter data and of which to identify the data singularity. In order to improve the usability of aeroengine performance data and identify the abnormal condition of aeroengine in time, and to reconstruct original data accurately, signal mutation identifying and reconstruction techniques are proposed based on modulus local maxima curves searching of continuous wavelet transform coefficients of signal, which can identify signal mutation and denoise at the same time, thus, the original signal can be resumed effectively. To resolve the problem that the computing speed of continuous wavelet transform method is very low, discrete Fourier transform based the continuous wavelet transform and inverse transform algorithms of signal are proposed. To further improve the accuracy of the mutational data identify and original data reconstruction, a simulation example is presented to research the processing methods of edge effect and false modulus local maxima curves which may come forth in the searching process of modulus local maxima curves of continuous wavelet transform coefficients. The data preprocess instance of aeroengine gas path performance parameter indicates the technique proposed can reconstructe efficientlly aeroengine performance parameter original data and the reconstruction effect is better than the effect of wavelet soft threshold denoising. A Hybrid Recurrent Process Neural Network is proposed to resolve the problems that building the prediction model of complex system condition is hard and the prediction precision is hard to be guaranteed. Concretely, the topological structure of this network is designed, and the network is simplified by introducing a set of appropriate orthogonal basis functions to expand the input functions and the connection weight functions of the network. And a learning algorithm base on resilient backpropagation is proposed for the network. The validation of this technique is proved by a benchmark of the Mackey-Glass chaos time series prediction using the network proposed and several tradition artificial neural networks. A practical utilization of the aeroengine gas path performance parameter data prediction by above networks demonstrates this point in terms of aeroengine condition monitoring too, the prediction results indicate Hybrid Recurrent Process Neural Network is an efficient aeroengine condition monitoring tool.Because tradition data fitting techniques based on least square procedure have inadaptabilities in fitting two aeroengine performance parameters which are both random variable, a Euclidean Distance Least Squares Support Vector Regression technique is proposed, which processes the random noise of two fitting variables at the same time, and a corresponding realization algorithm based on quadratic programming toolbox of Matlab is presented. The validity of the Euclidean Distance Least Squares Support Vector Regression technique is proved by a instance of aeroengine fleet trend analysis. In addition, A technique fitting aeroengine performance parameter time series is proposed summarily based on multi-order polynomial regression method to resolve the problem that fitting the time series of the different aeroengine performance parameter may need different fitting method.Analytic Hierarchy Process is a practical effective multiple criteria decision making method, But the consistency check is needed to determine its judgment matrix, and the grading criterions of its evaluating indicators observed values are generally determined by expert experience so that the non-determinacy comes forth. Therfore, a Grey Analytic Hierarchy Process based on Grey Cluster Evaluating Theory is proposed, and for aeroengine, the comprehensive evaluating indicators system of aeroengine condition is built. And the synthetic weight and the grey classification of every aeroengine condition evaluating indicator is computed or is present, and the triangle-type weight fuctioin based the parameters observed values of aeroengine condition evaluating indicators are graded. Finally, a instance of aeroengine condition comprehensive evaluation is presented.Orienting the application background of the aeroengine condition monitoring and the engineering management of Air-China, software system requirement is analyzed, and the function model and the data model and the architecture of the system are designed in turn. Subsequently, based on the aeroengine performance monitoring and condition evaluation techniques proposed, a aeroengine condition monitoring system orienting aeroengine life-cycle management is developed. The system includes serveral function modules such as the information management module of maintenance and condition monitoring of aeroengine; the aeroengine condition trend analysis module; the autoalarm module monitoring aeroengine condition parameters; the aeroengine condition comprehensive evaluation module; and the aeroengine removal time prediction module, etc. As one significant part of the aeroengine cycle-life management system, the aeroengine condition monitoring system has been applied to the project of aeroengine condition monitoring in Air-China, and the system validition is proved by the test run effects.更多还原