航空学报 > 2025, Vol. 46 Issue (S1): 732259-732259   doi: 10.7527/S1000-6893.2025.32259

第二届空天前沿大会/第二十七届中国科协年会优秀论文

数据与模型驱动的航天器姿态控制系统故障诊断

寇容海1, 李文博2,3, 党庆庆1(), 谢进进4   

  1. 1.西北工业大学 民航学院,西安 710072
    2.南京航空航天大学 自动化学院,南京 211106
    3.北京控制工程研究所,北京 100190
    4.东南大学 仪器科学与工程学院,南京 210096
  • 收稿日期:2025-05-19 修回日期:2025-05-30 接受日期:2025-06-10 出版日期:2025-07-21 发布日期:2025-06-27
  • 通讯作者: 党庆庆 E-mail:dangqingqing@nwpu.edu.cn
  • 基金资助:
    国家重点研发计划(2021YFB1715000);国家自然科学基金(62022013);国家自然科学基金(62373046);国家自然科学基金(210096);国家自然科学基金(62273240);姑苏创新领军人才(ZXL2023177)

Fault diagnosis of spacecraft attitude control system driven by data and model

Ronghai KOU1, Wenbo LI2,3, Qingqing DANG1(), Jinjin XIE4   

  1. 1.School of Civil Aviation,Northwestern Polytechnical University,Xi’an  710072,China
    2.College of Automation Engineering,Nanjing University of Aeronautics & Astronautics,Nanjing  211106,China
    3.Beijing Institute of Control Engineering,Beijing  100190,China
    4.School of Instrument Science and Engineering,Southeast University,Nanjing  210096,China
  • Received:2025-05-19 Revised:2025-05-30 Accepted:2025-06-10 Online:2025-07-21 Published:2025-06-27
  • Contact: Qingqing DANG E-mail:dangqingqing@nwpu.edu.cn
  • Supported by:
    National Key R&D Program of China(2021YFB1715000);National Natural Science Foundation of China(62022013);Suzhou Municipal Science and Technology Bureau(ZXL2023177)

摘要:

针对航天器姿态控制系统在复杂太空环境中易发生多重故障的问题,提出了一种融合鲁棒观测器与支持向量机(SVM)的故障检测与隔离策略。首先基于物理模型构造的鲁棒观测器与系统动力学比较产生残差,利用训练好的神经网络设计动态阈值与残差相比较进行故障检测。再结合观测器生成的残差信号和系统的输入输出,利用SVM对数据进行分类,检测出多重故障类型。通过物理模型与数据融合的双驱动策略,可以实现对典型单故障、双重故障及三重故障进行分类识别,提升航天器姿态控制系统故障检出精度与分类准确率。动态阈值对不同故障模式具有普适性,能够有效提升检测性能,通过仿真对比静态阈值和神经网络训练的动态阈值,表明动态阈值方法较传统静态阈值在故障响应速度和准确率上有所提升。通过对比BP神经网络、随机森林和SVM故障分类结果,实验表明SVM分类准确率达99.26%,优于随机森林(98.52%)和神经网络(97.04%),展现出较高性能。结论表明,所提方法通过观测器与SVM融合,有效解决了高耦合故障的检测与隔离难题,显著提升了系统故障检测能力。

关键词: 航天器姿态控制系统, 故障诊断, 数据与模型驱动, 鲁棒观测器, 神经网络, 支持向量机

Abstract:

To address the problem of multiple faults in spacecraft attitude control systems in complex space environments, this paper proposes a fault detection and isolation strategy that integrates robust observers and Support Vector Machines (SVM). Firstly, a robust observer constructed based on a physical model is compared with system dynamics to generate residuals. These residuals are then compared with dynamic thresholds, designed using a trained neural network, for fault detection. Furthermore, combining the observer-generated residual signals with the system input-output data, SVM is used to classify and detect multiple fault types. Through the dual drive strategy of physical model and data fusion, typical single faults, dual faults, and triple faults can be classified and identified, improving the accuracy of fault detection and classification in spacecraft attitude control systems. The dynamic threshold has universality for different fault modes and can effectively improve detection performance. By comparing the static threshold and the dynamic threshold trained by neural networks through simulation, it is shown that the dynamic threshold method has improved the fault response speed and accuracy compared to the traditional static threshold method. By comparing the fault classification results of BP network, random forest and SVM, the experiment shows that the SVM classification accuracy reaches 99.26%, which is better than random forest (98.52%) and neural network (97.04%), demonstrating superior performance. The conclusion shows that the proposed method effectively solves the problem of detecting and isolating high coupling faults through the fusion of observer and SVM, significantly improving the system’s fault detection capability.

Key words: spacecraft attitude control system, fault diagnosis, data and model driven, robust observer, neural network, support vector machine

中图分类号: