航空学报 > 2026, Vol. 47 Issue (15): 633361-633361   doi: 10.7527/S1000-6893.2026.33361

航空发动机智能控制与健康管理专栏

基于对抗强化学习的组合动力发动机模态转换智能鲁棒控制方法

程梓昭1, 刘利军1,2(), 林君健1   

  1. 1.厦门大学 航空航天学院,厦门 361102
    2.厦门大学深圳研究院,深圳 518000
  • 收稿日期:2026-01-13 修回日期:2026-03-13 接受日期:2026-03-27 出版日期:2026-04-21 发布日期:2026-04-14
  • 通讯作者: 刘利军 E-mail:liulijun@xmu.edu.cn
  • 基金资助:
    中国航空科学基金(2023L039068002);厦门市自然科学基金(3502Z202673006);深圳市科技基础研究专项(JCYJ20250604122930040)

Intelligent robust control method for modal transition in combined power engines based on adversarial reinforcement learning

Zizhao CHENG1, Lijun LIU1,2(), Junjian LIN1   

  1. 1.School of Aerospace Engineering,Xiamen University,Xiamen 361102,China
    2.Shenzhen Research Institute of Xiamen University,Shenzhen 518000,China
  • Received:2026-01-13 Revised:2026-03-13 Accepted:2026-03-27 Online:2026-04-21 Published:2026-04-14
  • Contact: Lijun LIU E-mail:liulijun@xmu.edu.cn
  • Supported by:
    Aeronautical Science Foundation of China(2023L039068002);Natural Science Foundation of Xiamen(3502Z202673006);Basic Research Program of Science and Technology of Shenzhen(JCYJ20250604122930040)

摘要:

针对组合动力发动机在模态转换过程中存在的强非线性、多执行器耦合以及对安全性和鲁棒性要求高等问题,对模态转换阶段的智能鲁棒控制方法进行系统研究。围绕推力连续跟踪与安全约束协同满足的控制目标,构建基于深度强化学习的智能控制框架,并引入对抗训练机制提升控制系统对观测扰动和不确定性的适应能力。在部件级仿真模型基础上,针对不同模态转换过程设计多输入多输出控制策略和分阶段训练环境,通过对抗强化学习实现控制策略的自学习与鲁棒性增强。同时,针对发动机参数波动问题,采用多智能体强化学习方法构建分布式控制结构,并在集中式训练、分布式执行的模式下完成控制器训练。进一步通过硬件在环平台开展实时性验证,结果显示控制周期满足毫秒级实时要求。仿真结果表明:在典型模态转换工况下,所设计的控制方法能够实现推力误差小于1%的稳态跟踪性能,模态转换期间推力波动幅值较对比控制方法表现更优,在引入观测扰动和参数拉偏条件下仍能保持安全约束不被触发。该智能鲁棒控制方法能够有效提升组合动力发动机模态转换过程中的控制精度、安全性与鲁棒性,为宽速域组合动力系统的工程应用提供了一种可行的智能控制方案。

关键词: 组合动力发动机, 模态转换, 智能鲁棒控制, 强化学习, 对抗训练

Abstract:

To address the strong nonlinearity, multi-actuator coupling, and stringent safety and robustness requirements of combined power engines during modal transition, an intelligent robust control method for the modal transition process is investigated. Focusing on the coordinated satisfaction of thrust tracking performance and safety constraints, an intelligent control framework based on deep reinforcement learning is established, in which an adversarial training mechanism is introduced to enhance robustness against observation disturbances and uncertainties. Based on a component-level engine simulation modal, multi-input multi-output control strategies and stage-wise training environments are designed for different modal transition processes, enabling adaptive policy learning and robustness improvement through adversarial reinforcement learning. In addition, to cope with engine parameter variations, a distributed control architecture based on multi-agent reinforcement learning is developed, and controller training is carried out under a centralized training and distributed execution scheme. Furthermore, real-time performance is validated on a hardware-in-the-loop platform, showing that the control cycle meets millisecond-level real-time requirements. Simulation results demonstrate that, under typical modal transition conditions, the proposed control method achieves steady-state thrust tracking errors within 1%, while exhibiting superior performance in thrust fluctuation amplitudes during modal transition compared with conventional control approaches. Under observation disturbances and parameter deviations, safety constraints are consistently satisfied without violation. The results indicate that the proposed intelligent robust control method effectively improves control accuracy, safety, and robustness during modal transition of combined power engines, providing a feasible solution for intelligent control of wide-speed-range combined power propulsion systems.

Key words: combined power engine, modal transition, intelligent robust control, reinforcement learning, adversarial training

中图分类号: