基于ODENet的航空发动机动态实时建模研究

伯丽欣; 李睿超; 刘渊; 苏三买

doi:10.13224/j.cnki.jasp.20240513

基于ODENet的航空发动机动态实时建模研究

doi: 10.13224/j.cnki.jasp.20240513

伯丽欣¹,
李睿超^1,*, ,,
刘渊²,
苏三买¹

1.
西北工业大学动力与能源学院，西安 710072
2.
中国航发湖南动力机械研究所，湖南株洲 412002

基金项目: 陕西省自然科学基金（2024JC-YBQN-0534）

详细信息

作者简介:
伯丽欣（2000-），女，硕士生，研究领域为发动机控制与仿真

通讯作者:
李睿超（1993-），男，副教授，博士，研究领域为发动机建模、仿真与控制。E-mail：liruichao@nwpu.edu.cn

中图分类号: V233.7
计量
- 文章访问数: 553
- HTML浏览量: 291
- PDF量: 56
- 被引次数: 0
出版历程
- 收稿日期: 2024-07-28
- 网络出版日期: 2025-02-09

Research on dynamic real-time modeling of aero engine based on ODENet

BO Lixin¹,
LI Ruichao^{1,*
, ,},
LIU Yuan²,
SU Sanmai¹

1.
School of Power and Energy，Northwestern Polytechnical University，Xi’an 710072，China
2.
Hunan Aviation Powerplant Research Institute，Aero Engine Corporation of China，Zhuzhou Hunan 412002，China

摘要

摘要:
提出一种基于神经常微分方程（ODENet）的航空发动机动态实时建模技术。首先基于发动机动态的先验知识确定了ODENet的结构框架；其次设计了幅度调制伪随机二进制序列信号，用以采集发动机部件级模型的输入输出数据，对ODENet模型进行训练，使其涵盖全包线范围内的动态特性；最后以带后涵道引射器的混排涡扇发动机为应用对象，将模型应用于开环和闭环仿真，以验证模型的实时性、精度和鲁棒性。结果表明：在开环仿真中，ODENet模型的运行速度相较于部件级模型提高了约15倍，并且ODENet模型在训练数据集和测试数据集上各参数的拟合度均不低于90%；在闭环仿真中，ODENet模型和部件级模型之间，输出变量的平均稳态误差和平均动态误差不超过4%。
- 动态实时模型 /
- 神经常微分方程 /
- 动态特性 /
- 模型预测控制 /
- 数据驱动
Abstract:
A new dynamic real-time modeling technique for aircraft engines based on Neural Ordinary Differential Equations (ODENet) was proposed. First, the structure of the ODENet model was determined based on the prior knowledge of engine dynamics. Then, an amplitude-modulated pseudo-random binary sequence signal was designed to collect input-output data from the component-level model of the engine, which was used to train the ODENet model, ensuring to cover the dynamic characteristics across the entire operating envelope. Finally, the model was applied to a mixed-flow turbofan engine with a bypass duct injector to validate its real-time performance, accuracy, and robustness through open-loop and closed-loop simulations. The results show that: in the open-loop simulation, the ODENet model’s running speed was approximately 15 times faster than the component-level model, and the fitting degree of each parameter on both the training and testing datasets was no less than 90%; in the closed-loop simulation, the average steady-state and dynamic errors of the output variables between the ODENet and component-level models did not exceed 4%.
- dynamic real-time model /
- neural ordinary differential equations /
- dynamic characteristics /
- model predictive control /
- data-driven

HTML

图 1 基于CLM的动态实时模型与基于ODENet动态实时模型的联系

Figure 1. Relationship between dynamic real-time model based on CLM and dynamic real-time model based on ODENet

下载: 全尺寸图片幻灯片

图 2 带后涵道引射器的混排涡扇发动机结构示意图

Figure 2. Structure diagram of a mixed-exhaust turbofan engine with a rear ducted ejector

下载: 全尺寸图片幻灯片

图 3 飞行包线划分示意图

Figure 3. Flight envelope division diagram

下载: 全尺寸图片幻灯片

图 4 ODENet模型训练框图

Figure 4. ODENet model training block diagram

下载: 全尺寸图片幻灯片

图 5 不可测输入信号

Figure 5. Undetectable input signal

下载: 全尺寸图片幻灯片

图 6 输入空间和输入空间边界线的转子转速

Figure 6. Input space and input space boundary rotor speed

下载: 全尺寸图片幻灯片

图 7 输入空间数据点分布

Figure 7. Input spatial data point distribution

下载: 全尺寸图片幻灯片

图 8 APRBS序列

Figure 8. APRBS sequence

下载: 全尺寸图片幻灯片

图 9 损失随神经元数量和隐藏层层数的变化

Figure 9. Loss varies with the number of neurons and the number of hidden layers

下载: 全尺寸图片幻灯片

图 10 ODENet模型输出与训练数据的对比

Figure 10. Comparison of ODENet model output with training data

下载: 全尺寸图片幻灯片

图 11 ODENet模型输出与测试数据的对比

Figure 11. Comparison of ODENet model output with test data

下载: 全尺寸图片幻灯片

图 12 ODENet-NMPC示意图

Figure 12. Schematic diagram of ODENet-NMPC

下载: 全尺寸图片幻灯片

图 13 ODENet-NMPC闭环仿真结果

Figure 13. ODENet-NMPC closed loop simulation results

下载: 全尺寸图片幻灯片

表 1 ODENet的超参数设置

Table 1. Hyperparameter settings of ODENet

参数	数值或说明
训练轮数	90
状态方程的隐藏层层数	2
状态方程的单层神经元数量	85
学习率	10⁻³
激活函数	tanh
$ {n_{{\text{minibatch}}}} $	100
输出方程的隐藏层层数	2
输出方程的单层神经元数量	85
优化器	adam
梯度衰减率	0.9

下载: 导出CSV

表 2 控制器参数设置和约束条件

Table 2. Controller parameter Settings and constraints

参数	数值
预测时域N_p	10
控制时域N_u	4
状态权重矩阵Q	diag[0.5,0,0,0]
控制输入权重矩阵R	diag[0,0.1,0.1]
高、低压转子转速N_l、N_h	≤100%
高、低压转子喘振裕度M_s,l、M_s,h	≥20%
燃油质量流量W_f	≤100%
涡轮前温度T₄	≤100%
后涵道引射器开度α₁₆₃/%	60～140
尾喷管喉部面积开度α₈/%	100～140

下载: 导出CSV

参考文献(22)

[1]	SIMON D L,LONG T W. Adaptive optimization of aircraft engine performance using neural networks[R]. Washington,US: Symposium on Advanced Aero Engines Concepts and Controls,1995.
[2]	黄劲东. 面向预测性维修构建航空发动机综合监控和健康管理系统[J]. 航空动力,2022(5): 74-78. HUANG Jindong. Constructing aero engine integrated monitoring and health management system towards predictive maintenance[J]. Aerospace Power,2022(5): 74-78. (in Chinese HUANG Jindong. Constructing aero engine integrated monitoring and health management system towards predictive maintenance[J]. Aerospace Power, 2022(5): 74-78. (in Chinese)
[3]	马帅,吴亚锋,郑华,等. 基于飞行过程数据的航空发动机故障诊断方法研究[J]. 推进技术,2023,44(5): 2208041. MA Shuai,WU Yafeng,ZHENG Hua,et al. Aircraft engine fault diagnosis based on flight process data[J]. Journal of Propulsion Technology,2023,44(5): 2208041. (in Chinese MA Shuai, WU Yafeng, ZHENG Hua, et al. Aircraft engine fault diagnosis based on flight process data[J]. Journal of Propulsion Technology, 2023, 44(5): 2208041. (in Chinese)
[4]	YAN Zhiqi,ZHONG Shisheng,LIN Lin,et al. Adaptive Levenberg-Marquardt algorithm: a new optimization strategy for Levenberg-Marquardt neural networks[J]. Mathematics,2021,9(17): 2176. doi: 10.3390/math9172176
[5]	KIDGER P. On neural differential equations[EB/OL]. [2024-04-10]. https://arxiv.org/abs/2202.02435v1.
[6]	GU Ziyu,PANG Shuwei,ZHOU Wenxiang,et al. An online data-driven LPV modeling method for turbo-shaft engines[J]. Energies,2022,15(4): 1255. doi: 10.3390/en15041255
[7]	CHEN Hongyi,LI Qiuhong,PANG Shuwei,et al. A state space modeling method for aero-engine based on AFOS-ELM[J]. Energies,2022,15(11): 3903. doi: 10.3390/en15113903
[8]	CHEN R T Q,RUBANOVA Y,BETTENCOURT J,et al. Neural ordinary differential equations[C]//Proceedings of the 32nd International Conference on Neural Information Processing Systems. New York: ACM,2018: 6572-6583.
[9]	RAHMAN A,DRGOŇA J,TUOR A,et al. Neural ordinary differential equations for nonlinear system identification[C]//2022 American Control Conference. Piscataway,US: IEEE,2022: 3979-3984.
[10]	BRADLEY W,BOUKOUVALA F. Two-stage approach to parameter estimation of differential equations using neural ODEs[J]. Industrial & Engineering Chemistry Research,2021,60(45): 16330-16344.
[11]	CHEE KONG Yao,ANI HSIEH M,MATNI N. Learning-enhanced nonlinear model predictive control using knowledge-based neural ordinary differential equations and deep ensembles[R]. Oxford,UK: Learning for Dynamics and Control Conference,2023.
[12]	LINOT A J,BURBY J W,TANG Qi,et al. Stabilized neural ordinary differential equations for long-time forecasting of dynamical systems[J]. Journal of Computational Physics,2023,474: 111838. doi: 10.1016/j.jcp.2022.111838
[13]	TRAN G,WARD R. Exact recovery of chaotic systems from highly corrupted data[J]. Multiscale Modeling and Simulation,2017,15(3): 1108-1129. doi: 10.1137/16M1086637
[14]	LUO Junwei,CANUSO V,JANG J B,et al. Machine learning-based operational modeling of an electrochemical reactor: handling data variability and improving empirical models[J]. Industrial & Engineering Chemistry Research,2022,61(24): 8399-8410.
[15]	GOYAL P,BENNER P. Neural ordinary differential equations with irregular and noisy data[J]. Royal Society Open Science,2023,10(7): 221475. doi: 10.1098/rsos.221475
[16]	张书刚,郭迎清,陆军. 基于GasTurb/MATLAB的航空发动机部件级模型研究[J]. 航空动力学报,2012,27(12): 2850-2856. ZHANG Shugang,GUO Yingqing,LU Jun. Research on aircraft engine component-level models based on GasTurb/MATLAB[J]. Journal of Aerospace Power,2012,27(12): 2850-2856. (in Chinese ZHANG Shugang, GUO Yingqing, LU Jun. Research on aircraft engine component-level models based on GasTurb/MATLAB[J]. Journal of Aerospace Power, 2012, 27(12): 2850-2856. (in Chinese)
[17]	HU Pipi. A note on the adjoint method for neural ordinary differential equation network[EB/OL]. [2024-04-10]. https://arxiv.org/abs/2402.15141v1.
[18]	NELLES O. Nonlinear dynamic system identification[M]//Nonlinear System Identification. Cham: Springer International Publishing,2020: 831-891.
[19]	MONTAZERI-GH M,RASTI A,JAFARI A,et al. Design and implementation of MPC for turbofan engine control system[J]. Aerospace Science and Technology,2019,92: 99-113. doi: 10.1016/j.ast.2019.05.061
[20]	BRUNELL B J,BITMEAD R R,CONNOLLY A J. Nonlinear model predictive control of an aircraft gas turbine engine[C]//Proceedings of the 41st IEEE Conference on Decision and Control. Piscataway,US: IEEE,2002: 4649-4651.
[21]	AKPAN V A,HASSAPIS G D. Nonlinear model identification and adaptive model predictive control using neural networks[J]. ISA Transactions,2011,50(2): 177-194. doi: 10.1016/j.isatra.2010.12.007
[22]	李睿超,虞超,姚竞豪,等. 基于冗余执行器的航空发动机推力快速响应控制方法研究[J]. 推进技术,2024,45(7): 2307039. LI Ruichao,YU Chao,YAO Jinghao,et al. Rapid thrust modulation control of aeroengines based on redundant actuators[J]. Journal of Propulsion Technology,2024,45(7): 2307039. (in Chinese LI Ruichao, YU Chao, YAO Jinghao, et al. Rapid thrust modulation control of aeroengines based on redundant actuators[J]. Journal of Propulsion Technology, 2024, 45(7): 2307039. (in Chinese)