基于联邦对比学习的航天器故障诊断

打开文本图片集

中图分类号：TP181 文献标志码：A 文章编号：1001-3695（2025）08-029-2467-07

doi： 10.19734/j. issn.1001-3695.2024.11.0481

Spacecraft fault diagnosis based on federated comparative learning

Yang Xiajie1，Lin Ziqian1†，XuDandan1，YouYanli²，FanChongjun1 （1.BusinessSol，Uesityfghaifoee&Tchlgnghao，hna;.agiJingnDistricAeurU versity（Shanghai Open University Jing'an Branch），Shanghai 2Ooo4O，China）

Abstract：Spacecraftfault diagnosis typicallyinvolves interpreting telemetryparameters.Time-seriesdata colected by key spacecraftdevices exhibitanon-independentidenticallydistributed（Non-ID）patern.Traditional federatedlearning methods tendtocause modeldriftduetolocal modelupdates when handlingsuchdata.Toaddressthis isse，thestudyproposedafederatedoptimizationalgorithmcalledfederatedcontrastivetime-seriesanomalydetection（FclTAD）.Thealgorithmanalyzed spacecraft Non-IDtime-seriesdatatodetectanomaliesatany givenmoment，enabling timelyinterventions.TheFclTADalgorithmconsistedof two stages：local updateand global aggregation.Thelocalupdate phaseused anautoencoder tocapture longtermdependenciesandnonlinearrelationsipsinthtimeseries.Itintroducedajointcontrastiveregularizationlosstoenhance themodel’sdiscriminativeabilityforfaultdiagnosis.Theglobalaggregationphaseaccountedforsensordiferencesandapplieda normalizationandweightedagregationstrategytocorrectinconsistentlocalupdates.Thisadjustmentreducedthe negativeimpactof eroneous updates onthe global model.The studyconductedsimulationsusing the NASA public datasets MSL and SMAP，as wellas the machine prediction serverdataset （PSM）.Resultsshow thatFcITADimproves the fault diagnosis accuracy by approximately 14.37% and the adjusted F1 score by 7.82% compared to other methods.FclTAD protects spacecraftdataprivacy，addressesthe“clientdrift”problem，andefectivelyutilizesrelationshipsbetweenlocalandglobalmodels. The method achieves excellent performance in spacecraft Non-IID time-series fault diagnosis.

Key Words：federated learning；fault diagnosis；time series；comparative learning；aerospace

0 引言

随着科技水平的不断发展，航天器的复杂度也在不断增加，引入新的方法对航天器进行及时的故障诊断能推动航天事业的发展。（剩余17825字）