• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2026, Vol. 62 ›› Issue (18): 99-113.doi: 10.3901/JME.260820

• 特邀专栏:大模型驱动的高端装备智能运维 • 上一篇    

扫码分享

基于龙卷风优化诺顿蠕变引导神经网络的航空发动机剩余寿命预测

吴豪1, 李锋1, 王永明1, 汤宝平2, 罗远航3, 田大庆1   

  1. 1. 四川大学机械工程学院 成都 610065;
    2. 重庆大学高端装备机械传动全国重点实验室 重庆 400044 3. 四川大学数学学院 成都 610065
  • 收稿日期:2025-11-03 修回日期:2026-05-19 发布日期:2026-11-20
  • 作者简介:吴豪,男,2002年出生,硕士研究生。主要研究方向为机械设备状态监测、故障诊断与剩余寿命预测。E-mail:2024223025056@stu.scu.edu.cn;李锋(通信作者),男,1982年出生,博士,副教授,硕士研究生导师。主要研究方向为航空发动机转子动力学、机械设备故障诊断与剩余寿命预测。E-mail:lifeng19820501@163.com;王永明,男,1963年出生,博士,研究员,博士研究生导师。主要研究方向为航空发动机设计技术和研制。E-mail:gtekjb@mail.sc.cninfo.net;汤宝平,男,1971 年出生,博士,教授,博士研究生导师。主要研究方向为设备状态监测与故障诊断、虚拟仪器和无线传感器网络。E-mail:bptang@cqu.edu.cn;罗远航,男,2000年出生。主要研究方向为概率统计与数据科学。E-mail:3338067918@qq.com;田大庆,男,1971 年出生,博士,副教授,硕士研究生导师。主要研究方向为设备状态检测与智能诊断、工业系统智能化自动化。E-mail:tiandaqing@scu.edu.cn
  • 基金资助:
    中央高校基本科研业务费(2022CDZG-12)、四川大学遂宁市校市战略合作(2023CDSN-15)和四川省科技厅重点研发(2023YFS0234)资助项目。

Remaining Useful Life Prediction of Aeroengines Based on Norton Creep-informed Neural Network with TOC Optimizer

WU Hao1, LI Feng1, WANG Yongming1, TANG Baoping2, LUO Yuanhang3, TIAN Daqing1   

  1. 1. School of Mechanical Engineering, Sichuan University, Chengdu 610065;
    2. State Key Laboratory of Mechanical Transmission for Advanced Equipment, Chongqing University, Chongqing 400044 3. School of Mathematics, Sichuan University, Chengdu 610065
  • Received:2025-11-03 Revised:2026-05-19 Published:2026-11-20

摘要: 航空发动机剩余使用寿命(Remaining useful life,RUL)预测的精度与可靠性直接关系到航空飞行器的飞行安全与运营成本。然而,现有的两类主流航空发动机RUL预测方法存在以下挑战:数据驱动的RUL预测方法虽具备良好拟合能力,但缺乏物理可解释性,预测结果常违背工程机理;基于物理方程的预测方法虽机理明确,却难以刻画复杂工况下的航空发动机非线性退化行为。以上挑战导致现有主流航空发动机RUL预测方法的预测精度不足、可靠性受限。为此,提出一种新型物理信息神经网络——龙卷风优化诺顿蠕变引导神经网络(Norton creep-informed neural network with TOC optimizer,TOC-NCINN)用于航空发动机RUL预测。在TOC-NCINN中,首先设计了新型多尺度特征提取网络以通过时序卷积与注意力机制来融合传感器信号的短期局部特征与长期依赖关系,以解决在长序列预测中感受野有限和历史信息遗忘的问题;其次,建立了基于诺顿蠕变方程的物理约束机制,通过该机制将材料蠕变本构关系以残差形式嵌入损失函数,以构建数据驱动与物理机理的双重监督网络,从而形成具有物理可解释性、鲁棒性和外推泛化性的NCINN框架;最后在NCINN中引入龙卷风优化器(Tornado optimizer with Coriolis force, TOC),实现对物理约束权重、学习率等超参数的自适应优化,从而提升NCINN的训练效率并进一步提高NCINN的泛化能力和RUL预测精度。采用公开数据集C-MAPSS开展的航空发动机RUL预测实验验证了所提出的基于TOC-NCINN的RUL预测方法的有效性和优越性。

关键词: 航空发动机, 剩余使用寿命预测, 物理信息神经网络, 诺顿蠕变方程, 龙卷风优化器

Abstract: The accuracy and reliability of predicting remaining useful life (RUL) of aeroengines are directly related to the flight safety and operating costs of aircraft. However, the existing two mainstream RUL prediction methods for aeroengines face the following challenges: although the data-driven RUL prediction methods have good fitting ability, they lack physical interpretability, so their prediction results often violate engineering mechanisms; although the physical equation-based RUL prediction methods have clear physical mechanisms, it is difficult for them to characterize the nonlinear degradation behavior of aeroengines under complex operating conditions. The above challenges have led to insufficient prediction accuracy and limited reliability of existing mainstream RUL prediction methods for aeroengines. To this end, a novel physics-informed neural network, namely the Norton Creep-informed Neural Network with TOC Optimizer (TOC-NCINN), is proposed for RUL prediction of aeroengines. In TOC-NCINN, a new multi-scale feature extraction network is first designed to fuse short-term local features and long-term dependencies of sensor signals through temporal convolution and attention mechanism, in order to solve the problems of limited receptive field and forgotten historical information in long sequence prediction; Secondly, a physical constraint mechanism based on the Norton creep equation is established, through which the material creep constitutive relationship is embedded in the loss function in residual form to construct a dual supervised network of data-drivenness and physical mechanism, thus forming an NCINN framework with physical interpretability, robustness, and extrapolation generalization; Finally, the Tornado Optimizer with Coriolis Force (TOC) is introduced into NCINN to achieve adaptive optimization of hyperparameters such as a physical constraint weight and a learning rate, thereby improving the training efficiency of NCINN and further enhancing its generalization ability and RUL prediction accuracy. The effectiveness and superiority of the proposed RUL prediction method based on TOC-NCINN are verified through experiments using the publicly available dataset C-MAPSS for aeroengine RUL prediction.

Key words: aeroengine, remaining useful life prediction, physics-informed neural network, norton creep equation, tornado optimizer with coriolis force

中图分类号: