| 引用本文: | 程玉强,邓凌志.Wasserstein距离在液体火箭发动机故障检测中的应用.[J].国防科技大学学报,2023,45(4):20-27.[点击复制] |
| CHENG Yuqiang,DENG Lingzhi.Application of Wasserstein distance in fault detection for liquid-propellant rocket engines[J].Journal of National University of Defense Technology,2023,45(4):20-27[点击复制] |
|
| |
|
|
| 本文已被：浏览 4708次 下载 3399次 |
| Wasserstein距离在液体火箭发动机故障检测中的应用 |
| 程玉强,邓凌志 |
|
(国防科技大学 空天科学学院, 湖南 长沙 410073)
|
| 摘要: |
| 健康监控技术能够切实提高液体火箭发动机的可靠性,针对液体火箭发动机健康监控中的故障检测问题,提出基于Wasserstein距离的方法,利用液氢液氧火箭发动机地面热试车数据进行验证。该方法的核心思想是利用Wasserstein生成对抗网络模拟正常数据的样本分布,利用其判别器计算测试样本与模拟分布间的Wasserstein距离,进而实现故障检测。结果表明:该方法能够克服故障数据不足的困难,有效检测稳态过程中的故障,没有发生误报警,且对早期异常有较高敏感性；在训练样本较少的情况下,当Wasserstein距离阈值为3σ时对启动过程的早期异常有较高敏感性,取5σ时仍可有效检测启动过程中的故障,误报警率为12.5%。 |
| 关键词: 液体火箭发动机 故障检测 Wasserstein距离 生成对抗网络 |
| DOI：10.11887/j.cn.202304003 |
| 投稿日期：2023-03-13 |
| 基金项目:国家自然科学基金创新研究群体资助项目(T2221002) |
|
| Application of Wasserstein distance in fault detection for liquid-propellant rocket engines |
| CHENG Yuqiang, DENG Lingzhi |
|
(College of Aerospace Science and Engineering, National University of Defense Technology, Changsha 410073, China)
|
| Abstract: |
| Health monitoring can effectively improve the reliability of liquid-propellant rocket engines. Aiming at the problem of fault detection in health monitoring for liquid-propellant rocket engines, a method based on Wasserstein distance was proposed and verified using ground hot test data of an LOX/LH2 rocket engine. The core idea was to use Wasserstein generative adversarial network to simulate the sample distribution of normal data, and used its discriminator to calculate the Wasserstein distance between the test sample and the simulated distribution, so as to achieve fault detection. The results show that the proposed method can overcome the difficulty of insufficient fault data, detect the faults in the steady-state process effectively without false alarm, and is sensitive to early anomalies. In the case of a small number of training samples, when the Wasserstein distance threshold is set to 3σ, the method is sensitive to early anomalies during the start-up transient. Starting fault can still be effectively detected with the threshold of 5σ, with a false alarming rate of 12.5%. |
| Keywords: liquid-propellant rocket engine fault detection Wasserstein distance generative adversarial network |
|
|
|
|
|
