航空学报 > 2023, Vol. 44 Issue (13): 327931-327931   doi: 10.7527/S1000-6893.2022.27931

联合字典学习与OCSVM的遥测数据异常检测方法

何家辉, 程志君(), 郭波   

  1. 国防科技大学 系统工程学院,长沙 410073
  • 收稿日期:2022-08-18 修回日期:2022-09-05 接受日期:2022-11-28 出版日期:2023-07-15 发布日期:2022-12-06
  • 通讯作者: 程志君 E-mail:chengzhijun@nudt.edu.cn
  • 基金资助:
    国家自然科学基金(72071208);湖南省科技创新团队项目(2020RC4046)

Telemetry anomaly detection method based on joint dictionary learning and OCSVM

Jiahui HE, Zhijun CHENG(), Bo GUO   

  1. College of Systems Engineering,National University of Defense Technology,Changsha 410073,China
  • Received:2022-08-18 Revised:2022-09-05 Accepted:2022-11-28 Online:2023-07-15 Published:2022-12-06
  • Contact: Zhijun CHENG E-mail:chengzhijun@nudt.edu.cn
  • Supported by:
    National Natural Science Foundation of China(72071208);Science and Technology Innovation Program of Hunan Province(2020RC4046)

摘要:

针对航天器遥测数据故障样本少,多参数关联异常缺乏有效检测手段的问题,引入字典学习(DL)改进一类支持向量机(OCSVM)对多参数关联关系异常的识别性能,并提出一种联合优化学习机制,以提升模型的检测效果和泛化能力。首先基于正常遥测数据,构建字典学习与OCSVM的联合学习函数,通过迭代优化的方法获取最优关联关系特征提取字典和异常决策边界;然后测试样本在最优字典下稀疏分解,提取特征输入优化后的OCSVM;最终基于异常决策边界实现异常标记。将所提方法应用于NASA公布的航天器MSL数据集和某卫星天线遥测数据进行验证,结果表明:相比于现有的一类分类异常检测方法,该方法在异常检测率、F1分数和G-mean等性能指标都有所提升,特别是在关联关系异常检测上展现出更优越的性能。

关键词: 典学习, 一类支持向量机, 遥测数据, 异常检测, 联合优化

Abstract:

To solve the problems of small fault sample size in spacecraft telemetry data and lack of effective detection methods for correlation anomalies of multi-parameters, Dictionary Learning (DL) is introduced to improve the recognition performance of One-Class Support Vector Machine (OCSVM). A jointly optimized mechanism is proposed to improve the detection effect and generalization of the model. Firstly, the joint learning function of dictionary learning and OCSVM is given based on telemetry data in the normal state. The optimal dictionary for extracting correlation features and the decision boundary of anomalies are obtained by iterative optimization. Then, test samples are sparsely decomposed via the optimal dictionary, and the extracted features are input into the optimized OCSVM. Finally, anomalies are labeled based on the decision boundary. The proposed method is applied to the MSL dataset published by NASA and telemetry data of a real satellite antenna. The results show that compared with existing one class classification anomaly detection methods, the proposed method shows improved performance in detection precision, F1-score and G-mean, especially in detection of correlation anomalies.

Key words: dictionary learning, one-class support vector machine, telemetry data, anomaly detection, jointly optimized mechanism

中图分类号: