摘要
针对在轨卫星异常检测中现存的异常定义单一、检测流程不规范不灵活的问题,提出一种基于长短期记忆(Long-Short Term Memory,LSTM)网络和多种异常定义的卫星异常检测方法。基于某在轨卫星实测电源遥测数据,首先进行卫星时序数据预处理,随后以LSTM为示例算法对数据的"正常值"进行预测,最后分别以测试数据均值的标准差、预测结果均值的标准差和非参数动态阈值作为异常定义,进行异常的联合投票检测,检测流程可容纳丰富的预测算法和异常定义,且流程模块间耦合度低。仿真结果表明,LSTM模型预测结合多异常定义的联合投票机制能有效提升异常点检测的性能。
In order to solve the problems of single anomaly definition,nonstandard and inflexible detection process in orbiting satellite anomaly detection,a satellite anomaly detection method based on long-short term memory(LSTM)network and various anomaly definitions is proposed.Based on the real power telemetry data of an in-orbit satellite,the time series data is preprocessed firstly,and then the“normal”value of the data is predicted by using the LSTM algorithm.Finally,the standard deviation of the test data mean,the standard deviation of the predicted result mean and the nonparametric dynamic thresholding are used as anomaly definitions respectively,to perform joint voting detection of anomalies.The detection process can contain various prediction algorithms and anomaly definitions,and the coupling between process modules is low.The simulation results show that LSTM model prediction combined with various anomaly definitions of joint voting mechanism can effectively improve the performance of anomaly detection.
作者
范潇杰
陈振安
FAN Xiao-jie;CHEN Zhen-an(School of Computer and Communication Engineering,University of Science and Technology Beijing,Beijing 100083,China;Key&Core Technology Innovation Institute,Guangdong-Hong Kong-Macao Greater Bay Area,Guangzhou 510530,China)
出处
《测控技术》
2021年第11期78-87,95,共11页
Measurement & Control Technology
基金
广东省科技计划项目(科技创新平台类)(2019B090904017)。
关键词
卫星
LSTM
异常检测
时序数据
联合投票
satellite
long-short term memory(LSTM)
anomaly detection
time series data
joint voting
