基于深度长短期记忆网络的发动机叶片剩余寿命预测

为了研究航空发动机转子叶片的剩余寿命预测问题,提出了一种基于多传感器信号融合的深度长短期记忆网络(DLSTM)预测模型。首先利用深度学习和长短期记忆的组合来构造DLSTM网络。然后,将多个传感器信号数据进行融合处理,从而通过深度学习发现各个传感器时序信号之间隐藏的长期依赖关系。进一步在给定网格搜索策略的情况下,通过自适应矩估计算法调整DLSTM的网络结构和参数,并且在DLSTM模型中引入了一种随机丢失策略,以缓解过度拟合问题并使预测模型规范化。最后利用CMAPSS涡扇发动机进行了实验验证,在一种故障模式和两种故障模式条件下,DLSTM网络预测模型相对于其他传统方法的评价指标Score分别下降了17.19%和14.37%,其他两个评价指标相对来说也较优,结果表明本文提出的方法具有更高的准确性以及稳定性。     

基于融合神经网络的航空发动机剩余寿命预测

航空发动机的性能退化是影响飞机飞行安全的重要因素。准确预测发动机的退化过程,对于飞机安全飞行具有重要意义。针对航空发动机剩余寿命预测问题,本文提出了一种将卷积神经网络和长短期记忆网络相融合的数据驱动模型。与常规使用单一的神经网络不同,所提出的融合模型结合了两种神经网络的优点,利用卷积神经网络提取数据中的空间特征并采用长短期记忆网络提取时间特征。实验结果证实,在寿命预测中,所提出的数据驱动模型与已有的方法相比,评分和均方根误差分别下降了32%和6.4%,因此,所提出的数据驱动模型可对数据中所包含的信息进行充分挖掘,其对航空发动机寿命预测精度较高,并具有良好的稳定性。     

Dynamic temperature prediction of electronic equipment under high altitude long endurance conditions

Unmanned Aerial Vehicle (UAV) is developing towards the direction of High Altitude Long Endurance (HALE). This will have an important influence on the stability of its airborne electronic equipment using passive thermal management. In this paper, a multi-node transient thermal model for airborne electronic equipment is set up based on the thermal network method to predict their dynamic temperature responses under high altitude and long flight time conditions. Some relevant factors are considered into this temperature prediction model including flight environment, radiation, convection, heat conduction, etc. An experimental chamber simulating a high altitude flight environment was set up to survey the dynamic thermal responses of airborne electronic equipment in a UAV. According to the experimental measurement results, the multi-node transient thermal model is verified without consideration of the effects of flight speed. Then, a modified way about outside flight speed is added into the model to improve the temperature prediction performance. Finally, the corresponding simulation code is developed based on the proposed model. It can realize the dynamic temperature prediction of airborne electronic equipment under HALE conditions.     

双油路离心喷嘴雾化特性的半经验预测方法

为了掌握主、副油路相互干涉的机制,进而获得双油路离心喷嘴雾化特性的预测工具,开展了双油路离心喷嘴的理论和试验验证研究,研究中所涉及的供油压差范围为0.05~3MPa。理论推导了双油路离心喷嘴液膜破碎长度和索太尔平均直径(SMD)的半经验预测公式,并使用了3个不同结构的双油路离心喷嘴进行了试验验证。结果表明:双油路离心喷嘴破碎长度和SMD的预测精度达到±20%。这些半经验关系式可以作为双油路离心喷嘴设计阶段的预测工具和仿真边界条件。      

某涡轮叶片热障涂层的寿命预测方法

基于NASA发展的CoatLife软件中推荐的应力 寿命建模方法,研究了某涡轮叶片热障涂层寿命预测方法。通过在寿命模型中引入氧化增质量,并考虑氧化动力学,实现了寿命模型与高温氧化效应的关联。由于寿命模型中的疲劳强度系数是时间和半径的函数,从而可以应用于实际结构中的不同几何形状,并考虑了时间相关的退化效应。计算结果表明,随着最高温度或疲劳强度系数的增加,涂层的循环寿命和时间寿命均会减少。通过对某涡轮叶片在设计温度场下的涂层寿命预测结果表明,循环时间为1h条件下叶片前缘1/2叶高处涂层剥落寿命大约336 h,与实际叶片涂层失效在300~400h之间吻合。      

基于动态集成算法的航空发动机气路参数预测

针对单一学习机对航空发动机气路参数预测困难的问题,提出了基于动态加权核密度估计（DWKDE)组合方法的集成预测算法,该组合方法选择测试样本的近邻样本,通过评估学习机在近邻样本的局部性能动态确定各学习机的权值,并基于该权值利用加权核密度估计实现数据序列的集成预测。该组合方法不易受离群值和样本不对称分布的影响,将该组合方法用于AdaBoost.RT和AdaBoost.R2算法,获得了改进后的集成学习算法。实验证明:相比于神经网络和原始集成学习算法,改进后的集成学习算法较好地提高了航空发动机气路参数序列的预测精度,方均根误差（RMSE)指标至少可降低27%。      

某型激光陀螺寿命分析与改进

我国环形激光陀螺 (RLG)技术发展和应用已经取得巨大成就,作为惯导系统的核心仪表,激光陀螺的寿命很大程度上决定了惯导系统的寿命。介绍了激光陀螺组成,分析了其内部工作气压稳定性影响寿命的机理,给出了决定寿命的关键因素,提出了提高陀螺寿命的多项措施和方法,对我国激光陀螺的寿命提升具有较大的指导意义。     

川航支线航空市场预测决策

本文从川航支线航空战略出发,分析长江上游支线航空的市场需求,结合航空公司的供应能力,运用市场分析与预测的方法,辅以成本分析,得出了相应的决策方案.     

Lifetime prediction based on Gamma processes from accelerated degradation data

Accelerated degradation test is a useful approach to predict the product lifetime at the normal use stress level, especially for highly reliable products. Two kinds of the lifetime prediction based on Gamma processes were studied. One was to predict the lifetime of the population from accelerated degradation data, and the other was to predict the lifetime of an individual by taking the accelerated degradation data as prior information. For an extensive application, the Gamma process with a time transformation and random effects was considered. A novel contribution is that a deducing method for determining the relationships between the shape and scale parameters of Gamma processes and accelerated stresses was presented. When predicting the lifetime of an individual, Bayesian inference methods were adopted to improve the prediction accuracy, in which the conjugate prior distribution and the non-conjugate prior distribution of random parameters were studied. The conjugate prior distribution only considers the random effect of the scale parameter while the non-conjugate prior distribution considers the random effects of both the scale and shape parameter. The application and usefulness of the proposed method was demonstrated by the accelerated degradation data of carbon-film resistors.