基于遥测数据特征预测的轻量化航天器异常检测  

Lightweight Spacecraft Anomaly Detection Based on Telemetry Data Feature Prediction

作　　者：刘切 李嘉玺 王泽煜 陈小龙 LIU Qie;LI Jiaxi;WANG Zeyu;CHEN Xiaolong(College of Automation,Chongqing University,Chongqing 400044,China;PipeChina Beijing Pipeline Co.,Ltd.,Beijing 100101,China)

机构地区：[1]重庆大学自动化学院,重庆400044 [2]国家管网集团北京管道有限公司,北京100101

出　　处：《宇航学报》2025年第2期232-243,共12页Journal of Astronautics

基　　金：国家重点研发计划(2021YFB1715000);国家自然科学基金(62373068,62203075);中央高校基本科研业务费(2024CDJCGJ-011)。

摘　　要：航天器遥测数据异常检测是保障航天器安全运行的关键技术。航天器异常检测技术面临遥测数据维度高、非线性强、异常类型复杂等挑战。深度学习在异常检测方面取得了巨大成功,但通常面临模型计算过程复杂、规模大、占用内存大、训练时间长等问题。针对此情况,提出了一种基于遥测数据特征预测的轻量化航天器异常检测方法。该方法提出了一种能捕获时序特征的无监督自编码机模型进行特征提取,再利用轻量化的长短期记忆神经网络对所提取的特征进行预测,通过预测值与真实值的误差来判断异常。该方法在特征提取阶段充分考虑了遥测数据的时序特征,在降维的同时保留了指示异常的信息,且不依赖标签数据;在此基础上通过预测方法来检测异常,具有高效、轻量的优势。使用美国航天局公开的遥测数据进行实验验证,结果显示,和其他方法相比,所提方法在检测指标相当的情况下,检测模型更加轻量化,训练时间显著降低,且利于可视化分析。Anomaly detection in spacecraft telemetry data is crucial for ensuring the safe operation of spacecraft.However,the high dimensionality,strong nonlinearity,and diverse anomaly types of such data pose significant challenges.To address these issues,the integration of deep learning technology into anomaly detection for spacecraft telemetry data has gained widespread adoption.Nevertheless,experimental processes often reveal complexities such as intricate calculation procedures,large model scales,substantial memory requirements,and prolonged training times.To tackle these challenges,we propose a novel method for spacecraft anomaly detection based on telemetry data feature prediction,which deviates from traditional time series prediction or feature extraction approaches.This method initially employs a self-encoding machine to extract features from Galway’s telemetry data.Subsequently,Long Short-Term Memory(LSTM)is utilized to predict these extracted features,with anomaly detection achieved by monitoring the discrepancy between predicted and actual values.Experimental verification is conducted using telemetry data published by NASA.The results demonstrate that,compared to other methods,our detection model is more lightweight while maintaining the same detection performance,significantly reducing training time and facilitating visual analysis.

关 键 词：异常检测 航天器遥测数据 特征提取 长短期记忆神经网络 

分 类 号：V557.3[航空宇航科学与技术—人机与环境工程]