基于多尺度融合预测模型的航空发动机剩余寿命预测  

Residual lifetime prediction of aero-engines based on multi-scale fusion prediction models

作　　者：刘纳川 郭建胜[1] 张晓丰[1] 余稼洋 解涛 LIU Nachuan;GUO Jiansheng;ZHANG Xiaofeng;YU Jiayang;XIE Tao(College of Equipment Management and UAV Engineering,Air Force Engineering University,Xi’an 710051,China)

机构地区：[1]空军工程大学装备管理与无人机工程学院,西安710051

出　　处：《兵器装备工程学报》2023年第7期289-296,共8页Journal of Ordnance Equipment Engineering

摘　　要：针对大多数基于数据驱动的航空发动机剩余寿命预测方法未细分其退化过程与复杂输入数据之间的关系,无法准确识别和提取关键特征的问题。提出一种基于多尺度融合预测模型(MSF)的航空发动机剩余寿命预测方法。该方法利用静态协变量编码网络(SCCN)和变量选择网络(VSN)针对输入数据类型进行特征选择,将SCCN生成的静态协变量连接到模型的不同位置,以提升模型捕捉不同尺度时间特征的能力,并融入门控残差机制构建模型基本框架,既能提高模型的适应性也能保证信息在网络中传递的效率,采用分位数误差作为损失函数实现了多尺度的预测,有效的提高了预测的准确性。在CMAPSS涡扇发动机数据集上进行实验分析,FD002、FD004测试集的预测精度分别达到91.9%和92.4%,通过与其他深度学习方法进行对比,RMSE最优值分别提高15.54%和16.91%,Score最优值分别提高83.21%和78.78%。For most data-driven aero-engine remaining life prediction methods,the relationship between its degradation process and complex input data is not broken down to accurately identify and extract key features.In this view,this paper proposes a method for predicting the residual lifetime of aero-engines based on a multi-scale fusion(MSF) prediction model.This method uses the static covariate coding network(SCCN) and variable selection network(VSN) to select features for the input data type,and connects the static covariates generated by SCCN to different positions of the model so as to improve the model ability to capture the temporal features of different scales and integrate the gated residual mechanism to build the basic framework of the model.It can not only improve the adaptability of the model but also ensure the efficiency of information transmission in the network.Besides,it uses quantile error as the loss function to achieve multi-scale prediction,effectively improving accuracy of the prediction.The experimental analysis on the CMAPSS turbofan engine dataset shows that the prediction accuracy of FD002 and FD004 test sets reaches 91.9% and 92.4% respectively.Compared with other deep learning methods,the RMSE optimal values increase by 15.54% and 16.91% respectively,and the optimal Score values increase by 83.21% and 78.78% respectively.

关 键 词：剩余寿命 门控残差机制 静态协变量编码器 LSTM神经网络 多头注意力机制 分位数 

分 类 号：V263.5[航空宇航科学与技术—航空宇航制造工程]