Feen-LSTM:一种优化的多遥测参数在线无监督异常检测方法  

Feen-LSTM: An optimized online unsupervised anomaly detectionmethod for multi-telemetry parameters

作　　者：张金垒 庞景月 卢晓伟[2] 宋宇晨[3] Zhang Jinlei;Pang Jingyue;Lu Xiaowei;Song Yuchen(School of Artificial Intelligence,Chongqing Technology and Business University,Chongqing 400067,China;Shanghai Satellite Engineering Research Institute,Shanghai 201100,China;Harbin Institute of Technology,Harbin 150080,China)

机构地区：[1]重庆工商大学人工智能学院,重庆400067 [2]上海卫星工程研究所,上海201100 [3]哈尔滨工业大学,哈尔滨150080

出　　处：《仪器仪表学报》2025年第1期247-257,共11页Chinese Journal of Scientific Instrument

基　　金：国家自然科学基金项目(62001069);重庆市教委科学技术研究项目(KJQN202300841,KJQN202100821);高层次人次科研启动项目(2056009);重庆市博士“直通车”科研项目(CSTB2022BSXM-JSX0008)资助。

摘　　要：随着我国航天事业由航天大国向航天强国迈进,航天器发射数量以及密度屡创新高,保障航天器在轨正常运行成为非常重要的任务。航天器遥测数据是地面长管判断其正常运行的重要依据,增强遥测数据的异常检测能力是目前地面长管提升保障能力的关键。目前工程上遥测数据异常检测主要依赖于专家经验和固定阈值,虽高效可靠,但难以应对复杂多变的在轨运行环境,且检测准确性有待提高。而传统的机器学习方法随着遥测数据量增加,模型的性能与有效性不足。近年来,深度学习方法在异常检测领域展现出巨大潜力,然而现有基于深度学习的航天器遥测数据异常检测仍面临较大挑战:一方面,对异常模式标记的准确性与完整性依赖较强,而实际工程中获取大量准确的异常标记数据较为困难;另一方面,现有方法在线异常检测能力不足,难以满足航天器的在轨监测需求。针对上述问题,提出了一种在线且无监督的异常检测模型Feen-LSTM,其基于Transformer结构提取多维遥测数据的全局时空特征,并结合LSTM来建模局部时间依赖性,从而实现了特征增强的优化结构。通过在NASA公开的两个航天器遥测数据集上的实验,表明Feen-LSTM能够有效地提高异常检测的精度,尤其是在面对复杂数据和未知异常模式时,表现出比其他方法更优的性能。As China′s space industry advances from being a space power to a space strong nation,the number and density of spacecraft launches have reached new heights.Ensuring the normal operation of spacecraft in orbit has become a crucial task.Spacecraft telemetry data is an important basis for ground control to determine normal operation,and enhancing the anomaly detection capability of telemetry data is key to improving ground control′s support capabilities.Currently,anomaly detection of telemetry data mainly relies on expert experience and fixed thresholds.While these methods are efficient and reliable,they struggle to cope with the complex and dynamic operating environment in orbit,and the detection accuracy still needs improvement.Traditional machine learning methods show limited performance and effectiveness as the volume of telemetry data increases.In recent years,deep learning methods have shown great potential in the field of anomaly detection.However,existing deep learning-based anomaly detection methods for spacecraft telemetry data still face significant challenges.On the one hand,they heavily rely on the accuracy and completeness of anomaly labels,while obtaining a large amount of accurate anomaly-labeled data in practical engineering is difficult.On the other hand,existing methods lack the ability for online anomaly detection,which is essential for meeting the real-time monitoring needs of spacecraft in orbit.To address these issues,this paper proposes an online and unsupervised anomaly detection model,Feen-LSTM.This model extracts global spatiotemporal features from multidimensional telemetry data using a Transformer structure and combines LSTM to model local temporal dependencies,thereby achieving an optimized structure for feature enhancement.Experiments on two spacecraft telemetry data sets published by NASA show that Feen-LSTM can effectively improve the accuracy of anomaly detection,especially in the face of complex data and unknown anomaly patterns,and show better performance than other methods.Wi

关 键 词：遥测数据 长短时记忆网络 特征增强 在线异常检测 

分 类 号：TH165.3[机械工程—机械制造及自动化]