现代制造工程 ›› 2026, Vol. 552 ›› Issue (9): 118-127.doi: 10.16731/j.cnki.1671-3133.2026.09.014

• 仪器仪表/检测/监控 • 上一篇    下一篇

航空发动机剩余寿命的多尺度时序注意力预测方法研究

邢荣鑫1,2, 高军伟1,2, 吴文凯1,2, 崔吉强1,2, 杨彦萍1,2   

  1. 1 青岛大学自动化学院,青岛 266071;
    2 山东省工业控制技术重点实验室,青岛 266071
  • 收稿日期:2025-07-23 出版日期:2026-09-18 发布日期:2026-10-22
  • 作者简介:邢荣鑫,硕士研究生,主要研究方向为模式识别与智能系统。高军伟,博士,教授,主要研究方向为模式识别与智能系统。吴文凯,硕士研究生,主要研究方向为模式识别与智能系统。崔吉强,硕士研究生,主要研究方向为模式识别与智能系统。杨彦萍,硕士研究生,主要研究方向为模式识别与智能系统。E-mail:xingrongxin2001@163.com
  • 基金资助:
    山东省自然科学基金项目(ZR2019MF063)

Research on multi-scale temporal attention prediction method for useful life of aero engine

XING Rongxin1,2, GAO Junwei1,2, WU Wenkai1,2, CUI Jiqiang1,2, YANG Yanping1,2   

  1. 1 College of Automation,Qingdao University,Qingdao 266071,China;
    2 Shandong Provincial Key Laboratory of Industrial Control Technology,Qingdao 266071,China
  • Received:2025-07-23 Online:2026-09-18 Published:2026-10-22

摘要: 针对现有航空发动机剩余使用寿命(Remaining Useful Life,RUL)预测算法中存在的特征提取不足、预测不准确等问题,提出了一种基于多尺度时序注意力剩余使用寿命预测模型。首先,外部注意力模块(External Attention Blocks,EAB)通过引入外部可学习的记忆单元,提取航空发动机运行数据中的关键特征,增强模型对全局信息的捕捉能力;随后,多尺度时间卷积网络(Multi-Stage Temporal Convolutional Network,MSTCN)捕捉时间序列数据中的多尺度时序特征,包括短期波动和长期趋势;然后,采用挤压-激励机制(Squeeze-and-Excitation networks,SE)对不同通道的贡献进行加权,提高模型的预测精度;最后,构造全连接层来融合加权特征以输出预测结果。利用美国航空航天局(National Aeronautics and Space Administration,NASA)提供的航空发动机退化仿真数据集(Commercial Modular Aero-Propulsion System Simulation,C-MAPSS)验证了该方法的有效性。与基于不同神经网络架构,如卷积神经网络(Convolutional Neural Networks,CNN)、循环神经网络(Recurrent Neural Network,RNN)及卷积双向长短时记忆网络(Convolutional Bidirectional Long Short-Term Memory Fusion Model,CNN-BiLSTM)等对比实验中的模型进行了比较,结果表明,该方法在航空发动机剩余使用寿命预测方面展现出更高的预测精度。

关键词: 航空发动机, 剩余使用寿命, 外部注意力模块, 多尺度时间卷积网络, 挤压-激励机制

Abstract: In order to solve the problems of insufficient feature extraction and inaccurate prediction in the existing Remaining Useful Life (RUL) prediction algorithms of aero engines,a prediction model based on multi-scale time series attention remaining service life was proposed. Firstly,External Attention Blocks (EAB) extracted the key features in the aeroengine operation data by introducing an externally learnable memory unit,so as to enhance the model's ability to capture global information. Subsequently,the Multi-Stage Temporal Convolutional Network (MSTCN) captured the multi-scale time series features in the time series data,including short-term fluctuations and long-term trends. Then,the Squeeze-and-Excitation networks (SE) mechanism was used to weight the contributions of different channels to improve the prediction accuracy of the model. Finally,a fully connected layer was constructed to fuse the weighted features to output the prediction results. The effectiveness of the method was verified by using the Commercial Modular Aero-Propulsion System Simulation (C-MAPSS) dataset provided by National Aeronautics and Space Administration (NASA). Compared with the models in the comparison experiment based on different neural network architectures, such as Convolutional Neural Networks (CNN), Recurrent Neural Network (RNN), Convolutional Bidirectional Long Short-Term Memory Fusion Model (CNN-BiLSTM), etc., the results show that the proposed method shows higher prediction accuracy in predicting the remaining useful life of aero engines.

Key words: aero engine, Remaining Useful Life(RUL), External Attention Blocks (EAB), Multi-Scale Temporal Convolutional Network(MSTCN), Squeeze-and-Excitation networks (SE)

中图分类号: 

版权所有 © 《现代制造工程》编辑部 
地址:北京市东城区东四块玉南街28号 邮编:100061 电话:010-67126028 电子信箱:2645173083@qq.com
本系统由北京玛格泰克科技发展有限公司设计开发 技术支持:support@magtech.com.cn