【作者】 雷达；

【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】 民航发动机健康状态预测是制定合理的发动机调度和维修规划的基础，也是保障运营安全、提高经济性的重要支撑技术。作为无需先验假设的数据驱动模型，以神经网络为代表的智能学习模型可直接利用发动机健康状态数据建立预测模型，解决了发动机健康状态衰退解析模型难以建立的问题。但是，发动机健康状态衰退过程的复杂非线性时变特性给智能学习模型的模型选择和优化带来了较大困难。本文针对具有复杂非线性时变特性的民航发动机健康状态预测的建模和求解问题，对自适应数据降噪、单一全局预测模型和集成局域预测模型及模型预测结果不确定性评估等健康状态预测技术进行了系统研究。提出一种基于经验模态分解增强的奇异值分解健康状态信号降噪方法，采用经验模态分解从原始健康状态信号提取趋势分量，对信号剩余部分进行奇异值分解降噪，消除趋势分量对奇异值差分谱的干扰，实现信号重构奇异值的自适应选择。发动机健康状态信号降噪结果证明了该方法的有效性。针对目前连续函数输入型过程神经网络不能直接利用离散样本进行训练的不足，提出一种离散输入过程神经网络模型。离散输入过程神经网络以离散向量作为输入，以卷积和算子实现时间聚合运算，可以避免连续函数输入过程神经网络在函数拟合和基展开过程中存在的参数难以选择及信息丢失的问题。民航发动机健康状态预测实例表明，离散输入过程神经网络在预测精度不低于连续函数输入过程神经网络情况下，具有更好的操作性。针对单一全局预测模型存在模型复杂难以优化的问题，提出了基于改进AdaBoost.RT的静态权值和动态权值组合集成模型。对AdaBoost.RT算法的误差函数进行了改进，并采用自适应阈值调整方法在训练过程中自动调整分类阈值。分别以离散输入过程神经网络、传统人工神经网络和极端学习机为弱学习机构建静态权值和动态权值组合集成预测模型。其中，动态权值集成模型在训练过程中采用训练样本的近邻样本对弱学习机进行性能评估，根据学习机在测试样本近邻样本上的性能来动态赋予学习机组合权值，从而可以充分挖掘弱学习机的局部性能，进一步提高了预测效果。对民航发动机的健康状态预测结果表明，由具有简单结构弱学习机构建的集成预测模型的预测效果好于单一全局模型，且集成模型可以降低对弱学习机的性能要求。针对模型点预测结果存在的不确定性，采用Bootstrap预测区间估计方法对所提出的上述预测模型的预测区间进行估计，实现对模型预测结果可靠性和精确性的量化评估。基于上述研究成果，根据中国国际航空股份有限公司的实际需求研发了“民航发动机拆发日期预测系统”，成为“航空发动机健康管理与维修决策支持系统”的重要组成模块。该模块从系统获得发动机的健康状态信息实现发动机拆发日期预测，将预测结果提供给系统的“维修计划制定”模块作为维修计划制定的依据，实现了从发动机状态监控到维修决策的无缝集成。更多还原

【Abstract】 The health condition prediction of civil aircraft engine is the basis for makingreasonable scheduling and maintenance plans, and is the support technology forensuring operation safety and improving economy. As data-driven models withoutprior hypotheses, intelligent learning models represented by neural networks can betrained using the health condition data directly, which overcomes the difficulties ofestablishing exact mechanism models for engine health condition prediction.However, the complex nonlinear time-varying characteristics of the aircraft enginehealth degradation process bring about difficulties to the actual use of intelligentlearning models.In view of such problems and the requirements of civil aircraft engine healthcondition prediction, the adaptive noise reducing method, the single global modelingmethod, the ensemble local modeling method and the uncertainty assessing methodfor prediction results are studied in this paper.An engine health signal noise reducing method based on singular valuedecomposition (SVD) enhanced by empirical mode decomposition (EMD) isproposed. The trend component is abstracted from the original health signal usingEMD, and then the residual is de-noised using SVD. Since the disturbance of thetrend component is eliminated, the singular values for signal reconstruction can beselected adaptively according to the singlular value spectrum. The given method isvalidated through the noise reducing of the actual engine health signals.In view of the problem that the process neural network (PNN) with continuousfunction inputs can not be trained using discrete samples directly, a discrete inputprocess neural network model (DPNN) is presented. DPNN takes discrete vectors asinputs and utlize convolution sum to realize the time aggregate operation. Thus,DPNN can avoid problems such as parameter selection and information loss duringprocedures of function fitting and expanding required by PNN with continuousfunction inputs. The engine health condition prediction results show that, DPNNperforms comparable to PNN with continuous function inputs while with betteroperability.To overcome the difficulties of model optimization required by single globalmodels for engine health condition prediction, two ensemble prediction models with static and dynamic combining weights respectively based on the improvedAdaBoost.RT are proposed. The error function of AdaBoost.RT is improved, and anadaptive adjustment strategy is adopted to adjust the threshold during the trainingprocess. Then, DPNN, ANN and the extreme learning machine are utilized as weaklearners to construct ensemble models for engine health condition prediction. Forensemble models with dynamic combining weights, the weak learners are evaluatedusing the neighboring samples during the training process, and the dynamiccombining weights of the weak learners are generated according to their performanceon the neighboring samples of the testing sample in the training sample set. Since thelocal performance of the weak learners can be well mined, the ensemble models withdynamic combining weights are superior to their counterparts with static combiningweights. The prediction results of engine health conditions show that the ensemblemodels always perform better than single global models. And the performancerequirements for predictors can be reduced while being utilized as weak leaners.According to the fact that the point prediction results are accompanied withuncertainty, the Bootstrap based prediction interval estimation method is employed toestimate the prediction intervals of the presented models, thus to achieve quantitativeassessment of the reliability and accuracy of the prediction results.Based on such reseach, the “system for engine removal date prediction” isdeveloped according to the actual requirements of Air China. As a key subsystem ofthe “aircraft engine health management and maintenance decision support system”,the developed system acquires health condition information from its parent system toexecute engine removal date prediction, and the results are provided to themaintenance planning module for maintenance plan making, which implementsseamless integration from engine condition monitoring to maintenance decisionmaking.更多还原