基于噪声元学习的卫星遥测信号异常检测方法  

Anomaly detection method of satellite telemetry signal based on noise meta learning

作　　者：郭鹏飞[1] 靳锴 陈琪锋 魏才盛 GUO Pengfei;JIN Kai;CHEN Qifeng;WEI Caisheng(School of Automation,Central South University,Changsha 410083,China;The 54th Research Institute of China Electronics Technology Group,Shijiazhuang 050081,China)

机构地区：[1]中南大学自动化学院,湖南长沙410083 [2]中国电子科技集团有限公司第五十四研究所,河北石家庄050081

出　　处：《系统工程与电子技术》2025年第2期352-359,共8页Systems Engineering and Electronics

基　　金：国家级项目(2021YFA0717100);湖南省自然科学基金优青项目(2022JJ20081);中南大学创新驱动计划(2023CXQD066);中南大学研究生自主探索创新项目(2024ZZTS0768)资助课题。

摘　　要：针对卫星遥测数据先验知识稀缺、常规数据驱动的异常检测方法难以准确辨识异常状态的问题,提出一种基于元学习与动态放缩阈值法的卫星遥测信号异常检测算法。首先,通过元学习算法求解一组具备快速适应小样本任务能力的长短期记忆神经网络初始参数,并在训练过程中为网络权重添加噪声,进一步提升模型泛化性能。其次,采用动态放缩阈值法分析预测误差序列,划定动态变化的异常阈值,标记异常点索引以实现卫星遥测数据异常检测。最后,通过两组卫星遥测信号算例验证所提算法的有效性。仿真结果表明,所提方法能够改善预测模型过拟合现象,并降低漏警概率。Due to the scarcity of prior knowledge in satellite telemetry data,conventional data-driven anomaly detection methods are difficult to accurately identify abnormal states,a satellite telemetry signal anomaly detection algorithm based on meta learning and dynamic scaling threshold method is proposed.Firstly,a set of initial parameters of a long short-term memory(LSTM)neural network with the ability to quickly adapt to small sample tasks is solved through meta learning algorithms.The noise is added to the network weights during the training process to further improve generalization performance of the model.Secondly,the dynamic scaling threshold method is used to analyze the prediction error sequence,define the abnormal threshold for dynamic changes,and mark the index of abnormal points to achieve anomaly detection of satellite telemetry data.Finally,the effectiveness of the proposed algorithm is verified through two sets of satellite telemetry signal examples.The simulation results show that the proposed method can improve the overfitting phenomenon of the prediction model and reduce the probability of missed alarms.

关 键 词：卫星遥测信号 异常检测 长短期记忆神经网络 元学习 

分 类 号：V474[航空宇航科学与技术—飞行器设计]