【作者】 李大鹏；

【导师】 郭钧；

【作者基本信息】 武汉理工大学 ， 机械工程， 2024， 硕士

【摘要】 航空发动机作为飞机上的重要部件,其性能直接关系到飞机的安全和运营效率。对航空发动机剩余使用寿命的准确预测,对于保障航空安全和降低运营成本具有重要意义。然而,随着航空发动机日趋复杂化、精密化和智能化,传统的基于物理学和统计学的方法已难以满足现代发动机维护的需求。因此,本文针对发动机运行状态多变、故障模式多样及多传感器数据融合所造成的发动机剩余使用寿命预测可靠性差,预测精度低等问题,利用集成学习的方法,对基于数据驱动的发动机寿命预测方法开展了研究,主要工作如下:(1)考虑到传感器采集的监测数据参数较多,且各参数对预测模型性能的贡献存在差异性,无法构建匹配模型的数据集的问题。采用数据归一化技术对原始数据集进行尺度变换,以提升其对数据的处理效率;通过标签映射对数据集的标签进行校准;利用滑动时间窗构建训练采用的样本。为后续模型学习数据里的时间相关性进而达到精准预测提供了坚实的基础。(2)针对发动机运行状态监测数据的维度多、难以提取特征导致预测困难的问题,提出了基于k折同质集成学习的时域卷积网络(Temporal Convolutional Network,TCN)模型进行预测。通过多个TCN模型来学习不同的样本,得到不同的预测结果。之后,对多个预测结果再进行集成学习,从而减少预测的误差,提升预测的精度以及可靠性。最后通过对比实验,验证了所提模型的优越性。(3)针对复杂工况下多故障模式的发动机剩余使用寿命预测可靠性较低的问题,提出了一种基于stacking集成的多模型融合预测方法。设计了多滑动时间窗策略来捕捉数据中发动机的退化趋势;提出了基于TCN和卷积双向门控循环单元神经网络(Convolutional Neural Network Bi-directional Gated Recurrent Unit,CNN-Bi-GRU)的混合预测模型。首先组建两种基础模型,随后根据stacking集成的思路,将两个基础模型的输出作为特征序列来输入到线性回归(Linear Regression,LR)模型中,最终获得预测结果。实验表明,相较于其它方法,所提方法的表现优异且具有较强的泛化性和可靠性。更多还原

【Abstract】 As an important component of an aircraft,the performance of aircraft engines is directly related to the safety and operational efficiency of the aircraft.Accurate prediction of the remaining service life of aircraft engines is of great significance for ensuring aviation safety and reducing operating costs.However,with the increasing complexity,precision,and intelligence of aviation engines,traditional methods based on physics and statistics are no longer able to meet the needs of modern engine maintenance.Therefore,this article focuses on the problems of poor reliability and low prediction accuracy in predicting the remaining service life of engines caused by the variable operating conditions,diverse fault modes,and multi-sensor data fusion of engines.Using ensemble learning methods,a data-driven engine life prediction method is studied.The main work is as follows:(1)Considering the large number of monitoring data parameters collected by sensors and the varying contributions of each parameter to the performance of the prediction model,it is not possible to construct a dataset that matches the model.Using data normalization techniques to scale the original dataset to improve its processing efficiency;Calibrate the labels of the dataset through label mapping;Construct training samples using sliding time windows.This provides a solid foundation for the subsequent model to learn the temporal correlation in the data and achieve accurate prediction.(2)A temporal convolutional network(TCN)model based on k-fold homogeneous ensemble learning is proposed to address the problem of difficulty in predicting engine operating status monitoring data due to its multiple dimensions and difficulty in extracting features.Learning different samples through multiple TCN models to obtain different prediction results.Afterwards,ensemble learning is performed on multiple prediction results to reduce prediction errors,improve prediction accuracy and reliability.Finally,the superiority of the proposed model was verified through comparative experiments.(3)A multi model fusion prediction method based on stacking ensemble is proposed to address the issue of low reliability in predicting the remaining service life of engines with multiple fault modes under complex operating conditions.Designed a multi sliding time window strategy to capture the degradation trend of the engine in the data;A hybrid prediction model based on TCN and Convolutional Neural Network Bi directional Gated Recurrent Unit(CNN Bi GRU)was proposed.Firstly,two basic models are constructed.Then,based on the stacking integration approach,the outputs of the two basic models are input as feature sequences into the Linear Regression(LR)model to obtain the prediction results.The experiment shows that compared to other methods,the proposed method performs excellently and has strong generalization and reliability.更多还原