北京航空航天大学学报 ›› 2019, Vol. 45 ›› Issue (7): 1329-1336.doi: 10.13700/j.bh.1001-5965.2018.0682

• 论文 • 上一篇    下一篇

基于生成对抗网络的航天异常事件检测方法

张克明1, 蔡远文1, 任元2   

  1. 1. 航天工程大学 研究生院, 北京 101416;
    2. 航天工程大学 宇航科学与技术系, 北京 101416
  • 收稿日期:2018-11-22 出版日期:2019-07-20 发布日期:2019-07-25
  • 通讯作者: 任元 E-mail:renyuan_823@aliyun.com
  • 作者简介:张克明 男,博士研究生,高级工程师。主要研究方向:异常事件检测、信号识别;蔡远文 男,博士,教授。主要研究方向:航天器测试与发射;任元 男,博士,副教授。主要研究方向:导航、制导与控制。
  • 基金资助:
    国家自然科学基金(51475472,61803383,51605489)

Space anomaly events detection approach based on generative adversarial nets

ZHANG Keming1, CAI Yuanwen1, REN Yuan2   

  1. 1. Graduate School, Space Engineering University, Beijing 101416, China;
    2. Department of Aerospace Science and Technology, Space Engineering University, Beijing 101416, China
  • Received:2018-11-22 Online:2019-07-20 Published:2019-07-25
  • Supported by:
    National Natural Science Foundation of China (51475472,61803383,51605489)

摘要: 航天环境复杂,技术难度大,风险高,安全可靠性要求苛刻。航天异常事件样本少,且难以获取,有针对性地开展异常事件检测(AED)很有必要。为预防航天事故,尽早发现可能导致故障的异常事件,深入研究了最新人工智能和生成对抗网络(GAN)技术,提出了一种基于生成对抗网络的航天异常事件检测方法。使用正生成对抗网络模拟生成正常事件样本,训练反生成对抗网络模拟生成异常事件样本,设计合理算法训练测试,计算输入事件与正生成对抗网络生成的模拟正常事件欧氏距离,以及输入事件与反生成对抗网络生成的模拟异常事件的欧氏距离差,实现对异常事件的精确检测。通过在美国国家标准与技术研究所数据库(MNIST)数据集全部使用正常数据训练,并对异常事件检测性能进行了试验验证,试验结果表明:在MNIST数据集下,精确率和召回率综合评价指标(F1)及精确率和召回率曲线下面积(PRC)等关键技术指标比变分自动编码器(VAE)方法相应指标性能至少分别提升了31%和11%。在真实环境下采集的模拟航天音频数据试验,异常事件检测性能良好,进一步证实了所提方法真实可用。

关键词: 生成对抗网络(GAN), 异常检测, 学习算法, 深度学习, 航天应用

Abstract: Anomaly events detection (AED) is quite important in space field for the complex space environment, difficult technology, high risk and strictly safe and reliable requirements. Since there are few space anomaly events samples and they are hard to obtain, it is necessary to carry out targeted AED. In order to prevent space accidents and find anomaly events that may lead to fault as soon as possible, a novel approach for space anomaly events detection based on generative adversarial nets (GAN) is proposed in this paper. Normal event samples are generated by normal GAN, anomaly event samples are generated by anomaly GAN. We proposed a reasonable algorithm to calculate the divergence of Euclidean distance between input events and simulated normal events generated by normal GAN, and Euclidean distance between input events and simulated abnormal events generated by anomaly GAN.As a result, abnormal events is detected accurately. The method is trained and tested using the Mixed National Institute of Standards and Technology (MNIST) database. The test results show that the key technical indexes, such as precision rate and recall rate of comprehensive evaluation index (F1) and precision recall curve (PRC), are at least 31% and 11% higher than the traditional variational autoencoder (VAE) method. In addition, we evaluated the method by collected data in real environment which simulated space audio data. The abnormal event detection performance is very good,which proved that the proposed method could detect anomaly event in real environments.

Key words: generative adversarial nets (GAN), anomaly detection, learning algorithms, deep learning, space application

中图分类号: 


版权所有 © 《北京航空航天大学学报》编辑部
通讯地址:北京市海淀区学院路37号 北京航空航天大学学报编辑部 邮编:100191 E-mail:jbuaa@buaa.edu.cn
本系统由北京玛格泰克科技发展有限公司设计开发