基于时序插补生成式对抗网络的卫星遥测参数异常检测方法

杜晓龙; 白萌

doi:10.11728/cjss2025.04.2024-0099

基于时序插补生成式对抗网络的卫星遥测参数异常检测方法

doi: 10.11728/cjss2025.04.2024-0099 cstr: 32142.14.cjss.2024-0099

杜晓龙^{1, 2,},
白萌^{1, 2,}

1.
中国科学院国家空间科学中心　北京　100190
2.
中国科学院大学　北京　100049

基金项目: 中国科学院空间科学战略性先导科技专项项目(XDA15040100)和中国科学院青年创新促进会项目(E1213A05)共同资助

详细信息

作者简介:

杜晓龙　男, 2000年6月出生于河南省南阳市, 现为中国科学院国家空间科学中心在读硕士研究生, 主要研究方向为卫星智能运控技术、卫星健康管理等. E-mail: duxiaolong22@mails.ucas.ac.cn

白萌　女, 1987年6月出生, 高级工程师, 中国科学院国家空间科学中心运控部副主任, 青促会会员. 全面主持空间科学卫星在轨运行管理, 长期深耕于智能运控技术研究. E-mail: baimeng@nssc.ac.cn

中图分类号: V474
计量
- 文章访问数: 228
- HTML全文浏览量: 45
- PDF下载量: 19
- 被引次数:
  0(来源:Crossref)
  
  0(来源:其他)
出版历程
- 收稿日期: 2024-08-03
- 录用日期: 2025-07-10
- 修回日期: 2024-12-12
- 网络出版日期: 2024-12-17

Anomaly Detection Method for Satellite Telemetry Parameters Based on Time-Series Imputation Generative Adversarial Networks

DU Xiaolong^{1, 2
,},
BAI Meng^{1, 2
,}

1.
National Space Science Center, Chinese Academy of Sciences, Beijing 100190
2.
University of Chinese Academy of Sciences, Beijing 100049

摘要

摘要: 为确保卫星的安全稳定运行, 及时进行遥测参数的数据挖掘、态势分析及异常响应至关重要. 鉴于现有方法在处理卫星遥测参数异常时存在的局限性, 本文提出一种基于时序插补和生成式对抗网络的异常检测方法. 该方法通过一维卷积神经网络提取时序特征, 并利用生成式对抗网络对遥测参数的分布进行建模, 创新性地采用基于插补的检测方式, 有效提高了异常检测的准确性和对复杂异常情况的适应能力. 基于真实卫星数据和公开数据集的测试结果表明, 与多种已有方法相比, 本文方法在多数数据集上获得了最高的F1分数, 并在不同异常浓度下显示出良好的稳定性. 这一研究成果为卫星任务的地面运控进行卫星态势分析和异常处置提供了有力的决策支持.
- 遥测数据 /
- 时序异常检测 /
- 一维卷积神经网络 /
- 生成式对抗网络 /
- 插补生成
Abstract: To ensure the safe and stable operation of satellites, it is of paramount importance to promptly conduct data mining, situation analysis, and abnormal response for telemetry parameters. Given the limitations of existing methods in effectively addressing satellite telemetry parameter anomalies, this paper introduces an innovative anomaly detection method that leverages temporal interpolation and Generative Adversarial Networks (GANs). The proposed method employs a one-dimensional Convolutional Neural Network (1DCNN) to extract temporal features from the telemetry data. These features are then used to model the distribution of telemetry parameters through a generative adversarial network, which consists of a generator and a discriminator that are trained simultaneously to learn the distribution of normal data. The method innovatively incorporates a detection approach based on interpolation, which significantly enhances the accuracy of anomaly detection and the capability to handle complex and subtle anomalies that may not be easily identified by traditional methods. The effectiveness of the proposed method is validated through comprehensive testing on real satellite data as well as established public datasets. Results demonstrate that, when compared with various existing anomaly detection methods, the proposed approach achieves the highest F1 scores on the majority of the datasets tested. This indicates a superior balance between precision and recall, which is crucial for reliable anomaly detection. Furthermore, the method exhibits good stability under different anomaly densities, suggesting its robustness in varying operational conditions. This research outcome not only enhances the understanding of satellite telemetry anomalies but also provides strong decision support for ground operation in satellite mission analysis and anomaly handling, thereby contributing to the overall safety and efficiency of satellite operations.
- Telemetry data /
- Temporal anomaly detection /
- One-dimensional convolutional neural networks /
- Generative adversarial networks /
- Interpolation generation

HTML全文

图 1 三种时序数据异常类型

Figure 1. Three time-series data exception types

下载: 全尺寸图片幻灯片

图 2 滑动窗口处理时序数据示意

Figure 2. Schematic diagram of sliding window processing timing data

下载: 全尺寸图片幻灯片

图 3 TiGAN异常检测算法框架

Figure 3. Framework of TiGAN anomaly detection algorithm

下载: 全尺寸图片幻灯片

图 4 一维卷积过程

Figure 4. Process of one-dimensional convolution

下载: 全尺寸图片幻灯片

图 5 平滑卷积检测示例

Figure 5. Smooth convolution detection example

下载: 全尺寸图片幻灯片

图 6 预测型方法的检测逻辑

Figure 6. Detection logic of predictive methods

下载: 全尺寸图片幻灯片

图 7 重构型方法的检测逻辑

Figure 7. Detection logic of reconstructed methods

下载: 全尺寸图片幻灯片

图 8 插补方法的检测逻辑

Figure 8. Detection logic of interpolation methods

下载: 全尺寸图片幻灯片

图 9 算法检测效果显示

Figure 9. Display of algorithm detection effect

下载: 全尺寸图片幻灯片

图 10 不同方法生成数据效果对比

Figure 10. Comparison of data generated by different methods

下载: 全尺寸图片幻灯片

图 11 各种算法检测结果的F1分数对比

Figure 11. F1 score comparison of detection results of various algorithms

下载: 全尺寸图片幻灯片

图 12 不同异常浓度对算法有效性的影响

Figure 12. Display of algorithm detection effect

下载: 全尺寸图片幻灯片

表 1 实验环境配置

Table 1. Setting of the experimental environment

软硬件平台	版本参数
CPU	AMD Ryzen 5700 X
GPU	NVIDIA Geforce RTX 4060
操作系统	Windows11
Python	3.9.18
Pytorch	2.1.1
CUDA	12.2

下载: 导出CSV

表 2 实验数据集数据分布情况

Table 2. Data distribution of experimental datasets

数据集	序列数	序列平均长度	2σ/(%)	3σ/(%)	异常占比/(%)
AWS	17	3985	4.46	1.35	9.32
DY	5	6665	0.098	0.098	0.44
RK	7	5210	0.71	0.22	0.79
ZK	4	8129	0.18	0.16	0.43
注　2σ和3σ分别代表超过平均值2个标准差(2σ)和3个标准差(3σ)的数据点占比.

下载: 导出CSV

表 3 各种检测算法在数据集上的检测结果对比

Table 3. Comparison of detection results of various detection algorithms on datasets

数据集	指标	TiGAN	Z-score	LOF	DBSCAN	Bi-LSTM	TAnoGAN
DY	F1	0.8824	0.7582	0.7018	0.4751	0.6955	0.8842
	Rec	0.7895	0.6106	0.5405	0.3115	0.9997	0.9411
	Pre	1.0000	1.0000	1.0000	1.0000	0.5332	0.8338
	Acc	0.9871	0.9961	0.9974	0.9932	0.5362	0.9913
RK	F1	0.7595	0.5969	0.3435	0.4684	0.7430	0.7239
	Rec	0.7124	0.5418	0.8894	0.3788	0.9987	0.8131
	Pre	0.8133	0.6643	0.2128	0.6135	0.5916	0.6253
	Acc	0.9615	0.9883	0.8887	0.9805	0.6548	0.9566
ZK	F1	0.8861	0.6846	0.2874	0.5495	0.4815	0.5160
	Rec	0.9159	0.7199	0.6445	0.4464	0.9970	0.9837
	Pre	0.8581	0.6526	0.1849	0.7143	0.3174	0.3498
	Acc	0.9823	0.9917	0.8707	0.9930	0.5534	0.8918
AWS	F1	0.5470	0.2156	0.2039	0.1053	0.3755	0.3078
	Rec	0.8733	0.2714	0.7600	0.0563	0.8760	0.7909
	Pre	0.39818	0.1788	0.1177	0.8135	0.2390	0.1911
	Acc	0.6161	0.7861	0.3557	0.8960	0.3611	0.4064
注　加粗字体表示该指标下的最高值.

下载: 导出CSV

表 4 特征学习模块对检测结果影响的对比

Table 4. Comparison of the influence of feature learning modules on detection results

F1 score	有特征学习模块	无特征学习模块
序列1	0.6523	0.6132
序列2	0.6531	0.2203
序列3	0.8659	0.1403

下载: 导出CSV

表 5 插补数量对检测结果影响的对比

Table 5. Comparison of the influence of the number of interpolations on the test results

F1 score	序列1	序列2	序列3
mask=0	0	0	0
mask=1	0.6429	0.4698	0.8552
mask=3	0.6523	0.5152	0.8552
mask=5	0.6523	0.3421	0.8670
mask=10	0.6523	0.6531	0.8659
mask=15	0.6132	0.6452	0.8670
mask=20	0.6523	0.2203	0.8479

下载: 导出CSV

参考文献(20)

[1]	王赤. 大力发展空间科学推动我国空间活动高质量发展[J]. 中国科学基金, 2024, 38(2): 363-364 WANG Chi. Developing space science vigorously to promote high-quality development of space activities in China[J]. Bulletin of National Natural Science Foundation of China, 2024, 38(2): 363-364
[2]	王赤, 邹自明, 胡晓彦, 等. 空间科学大数据与航天强国[J]. 科学通报, 2024, 69(9): 1132-1136 WANG Chi, ZOU Ziming, HU Xiaoyan, et al. Big data in space science and great power of space[J]. Chinese Science Bulletin, 2024, 69(9): 1132-1136
[3]	薛玉雄, 杨生胜, 把得东, 等. 空间辐射环境诱发航天器故障或异常分析[J]. 真空与低温, 2012, 18(2): 63-70 doi: 10.3969/j.issn.1006-7086.2012.02.001 XUE Yuxiong, YANG Shengsheng, BA Dedong, et al. Analyze of spacecraft system failures and anomalies attributed to the natural space radiation environment[J]. Vacuum and Cryogenics, 2012, 18(2): 63-70 doi: 10.3969/j.issn.1006-7086.2012.02.001
[4]	王亚坤, 杨凯飞, 张婕, 等. 卫星在轨故障案例与人工智能故障诊断[J]. 中国空间科学技术, 2022, 42(1): 16-29 WANG Yakun, YANG Kaifei, ZHANG Jie, et al. Case study of in-orbit satellite failures and artificial intelligence based failure detection[J]. Chinese Space Science and Technology, 2022, 42(1): 16-29
[5]	王彧泽, 吕良庆, 何睿. 航天器数据定义方法及其工具设计[J]. 航天器工程, 2024, 33(1): 62-67 WANG Yuze, LYU Liangqing, HE Rui. Spacecraft data definition method and tool design[J]. Spacecraft Engineering, 2024, 33(1): 62-67
[6]	杨甲森. 卫星遥测数据相关性知识发现方法研究[D]. 北京: 中国科学院大学(中国科学院国家空间科学中心), 2019 YANG Jiasen. Research on correlation Knowledge Discovery Method of Satellite Telemetry Data[D]. Beijing: Chinese Academy of Sciences (National Space Science Center, the Chinese Academy of Sciences), 2019
[7]	CHIKODILI N B, ABDULMALIK M D, ABISOYE O A, et al. Outlier detection in multivariate time series data using a fusion of K-medoid, standardized euclidean distance and Z-score[C]//Proceedings of the 3rd International Conference on Information and Communication Technology and Applications. Minna: Springer, 2021: 259-271
[8]	TSIGKRITIS T, GROUMAS G, SCHNEIDER M. On the use of k-NN in anomaly detection[J]. Journal of Information Security, 2018, 9(1): 70-84 doi: 10.4236/jis.2018.91006
[9]	SHERIDAN K, PURANIK T G, MANGORTEY E, et al. An application of DBSCAN clustering for flight anomaly detection during the approach phase[C]//Proceedings of the AIAA Scitech 2020 Forum. Orlando: AIAA, 2020: 1851
[10]	胡智超, 余翔湛, 刘立坤, 等. 基于上下文生成对抗网络的时间序列异常检测方法[J]. 哈尔滨工业大学学报, 2024, 56(5): 1-11 doi: 10.11918/202212029 HU Zhichao, YU Xiangzhan, LIU Likun, et al. A time series anomaly detection method based on contextual generative adversarial network[J]. Journal of Harbin Institute of Technology, 2024, 56(5): 1-11 doi: 10.11918/202212029
[11]	ZHOU T, MA Z Q, WEN Q S, et al. FEDformer: frequency enhanced decomposed transformer for long-term series forecasting[C]//Proceedings of the 39th International Conference on Machine Learning. Baltimore: PMLR, 2022: 27268-27286
[12]	BASHAR A, NAYAK R. TAnoGAN: time series anomaly detection with generative adversarial networks[C]//Proceedings of 2020 IEEE Symposium Series on Computational Intelligence. Canberra: IEEE, 2020: 1778-1785
[13]	LIN S Y, CLARK R, BIRKE R, et al. Anomaly detection for time series using VAE-LSTM hybrid model[C]//Proceedings of the ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing. Barcelona: IEEE, 2020. DOI: 10.1109/ICASSP40776.2020.9053558
[14]	PAPARRIZOS J, KANG Y H, BONIOL P, et al. TSB-UAD: an end-to-end benchmark suite for univariate time-series anomaly detection[J]. Proceedings of the VLDB Endowment, 2022, 15(8): 1697-1711 doi: 10.14778/3529337.3529354
[15]	DARBAN Z Z, WEBB G I, PAN S R, et al. Deep learning for time series anomaly detection: a survey[J]. ACM Computing Surveys, 2025, 57(1): 15
[16]	GOODFELLOW I, POUGET-ABADIE J, MIRZA M, et al. Generative adversarial nets[C]//Proceedings of the 27th International Conference on Neural Information Processing Systems. Montreal: MIT Press, 2014: 2672-2680
[17]	AHMAD S, LAVIN A, PURDY S, et al. Unsupervised real-time anomaly detection for streaming data[J]. Neurocomputing, 2017, 262: 134-147 doi: 10.1016/j.neucom.2017.04.070
[18]	XU S Y, LIU H Y, DUAN L T, et al. An improved LOF outlier detection algorithm[C]//Proceedings of 2021 IEEE International Conference on Artificial Intelligence and Computer Applications. Dalian: IEEE, 2021: 113-117
[19]	MISHRA S, KSHIRSAGAR V, DWIVEDULA R, et al. Attention-based Bi-LSTM for anomaly detection on time-series data[C]//Proceedings of the 30th International Conference on Artificial Neural Networks. Bratislava: Springer, 2021: 129-140
[20]	赵玉炜, 苏举. 基于参数自适应优化聚类的卫星状态异常检测方法[J]. 空间科学学报, 2023, 43(5): 927-937 doi: 10.11728/cjss2023.05.2022-0054 ZHAO Yuwei, SU Ju. Satellite anomaly detection method based on parameter adaptive optimization clustering[J]. Chinese Journal of Space Science, 2023, 43(5): 927-937 doi: 10.11728/cjss2023.05.2022-0054