【作者】 章雷；

【导师】 丁刚；

【作者基本信息】 哈尔滨工业大学 ， 机械设计及理论， 2012， 硕士

【摘要】 为在保证安全的前提下，降低大型复杂装备的运营成本，基于状态监控的视情维修和预知维修等先进维修理念和模式已得到了越来越广泛的应用，所以迫切需要开展对其关键技术——装备性能衰退预测的研究，而装备性能衰退预测就是要预测可反映装备性能状态的参数数据变化趋势，本文就针对大型复杂装备的性能衰退主要受时间累积效应影响的特点，提出了两种以泛函逼近理论为基础的预测装备性能参数数据的方法——过程回归分析和基于卷积的过程神经网络，并应用于航空领域的实际问题，取得了满意的结果。传统回归分析是函数逼近工具，受输入同步瞬时限制，难以反映时间累积效应，预测装备性能数据精度不高，为了解决此问题，本文提出了一种过程回归分析方法，该方法输入为函数并有时间聚合算子，也因此成为一种泛函逼近工具，仿真实验验证了过程回归分析对数据的预测效果要优于传统回归分析方法。过程神经网络虽然已经是一种泛函逼近工具，但考虑到传统积分过程神经网络学习算法在进行时间累积聚合效应处理时使用了正交基展开输入函数和权函数，虽能大大简化训练过程中的运算，但由于展开函数的波动，更由于正交基的积分运算结果非零即一的特点带来了部分信息丢失的问题，从而对装备性能数据预测结果的精度造成一定影响。本文将信号处理领域中的卷积运算引入到过程神经网络模型的时间累积聚合运算算子中来，将其改进为基于卷积的过程神经网络，同时采用对连续函数的快速卷积来保证其计算速度。仿真实验证明，由于信息保留完整，卷积过程神经网络对数据预测的性能较积分过程神经网络好。本文以民用航空发动机为例，将所提出的过程回归分析和基于卷积的过程神经网络这两种方法应用于民用航空发动机排气温度的预测，并取得了满意的结果。更多还原

【Abstract】 In order to reduce the operation cost of large and complex equipment on thepromise of safe operation, condition based maintenance and predictivemaintenance based on condition monitoring has been applied more and more.Thekey to condition based maintenance and predictive maintenance is predictingperformance deterioration of equipment.Predicting performance deterioration isto predict the changing of data which can reflect the performance, consideringperformance deterioration of large equipment is mainly influenced by timecumulation effect, two process prediction methods: process regression analysisand process neural network based on convolution, based on functionalapproximation are given in this article and are applied in the practice, achieving agood result.A new process regression analysis is proposed based on traditionalregression analysis, do not like the traditional regression analysis, it is abecause of having a operator which represent time cumulation effect. Simulationexperiment indicates that the predicting accuracy of process regression analysisis higher than the predicting accuracy of traditional regression analysis.Though process neural network is a functional approximation tool, adisadvantage that makes information incomplete when expanding the inputfunction and weight function with orthogonal basis function is found, so a newprocess neural network based on convolution which can keep informationcomplete is given. In the calculation of process neural network based onconvolution, a kind of fast convolution is used to make the computation speedfaster. The experiment shows that the prediction accuracy of new process neuralnetwork is higher than the traditional one’s.Process regression analysis and process neural network based onconvolution proposed in this article are applied to the prediction of performancedeterioration of aero engine and get a good result.更多还原