基于时空图网络和Informer的多元时间序列异常检测

打开文本图片集

中图分类号：TN919-34；TP393.0 文献标识码：A 文章编号：1004-373X（2025）15-0083-08

Multivariate time series anomaly detection based on spatial-temporal graph networkandInformer

YANGChenlong'，SUNYe²，LIUXiaoyue' （1.CollegeofElectricalEnginering，NorthChinaUniversityofScienceandTechnologyTangshanO63210，China; 2.China Academy of Informationand Communications Technology，Beijing1Oo191，China）

Abstract：Efectiveanomalydetection for multivariate timeseriescan ensure thesecurityofIoTsystems.In theexisting methods，localspatial-temporalrepresentationsareusuallylearedfromnearbytimepointsandneighboringnodesinodertoreconstructorpredictsensordata.AmultivariatetimeseriesanomalydetectionmodelSTGINisproposedtoeliminatefalsealarms andanomalyomisioncausedbythedificultyof modelingcomplexnonlinear topologicalrelationshipsanddynamic temporalpatternsinthelocalrepresentations.Firstlytemporalconvolutionalnetwork（TCN）ismbeddedintomultiscaleresidualcovolutionalnetworktoapturetemporalfaturesatteortegaularitylevelandgatingmchansisintroduedtofilter outtheirrelevantinformation.Thentespatialgaphstructureisonstructedandhecomplexspatial-tmporaldependenciesare learedefcientlybythegraphatentionnetwork．Finally，thepredictionandreconstructionmoduleareoptimizedjointlythe variationalself-encoderandInforerareemployedforlongtimeseriesreconstruction，andteextractedglobalandlocalspatialtemporalcorelationsareused todetectanomaliesinnormaldatasamples.ExperimentswereperformedonMSL，SMAPand SWaT public datasets.The obtained F1 -scoresare 0.9623，0.9425and 0.8709，respectively，which are better than thebenchmark models，verifying the effectiveness and feasibility of the proposed method.

Keywords：multivariate time series;anomaly detection;TCN; gated mechanism; graph atention network; Informer

0 引言

随着物联网连接的传感器设备迅速增加，物联网产生了大量传感器数据，其特征是地理和时间依赖性。（剩余12061字）