时间序列的分割及不一致发现分析-analysis of time series segmentation and inconsistency discovery.docx

聚类的静态时间序列不一致发现算法。首先，对给定原始时间序列采用基于PAA方式比特序列化的方法进行分割，该方法不仅提取了原始序列的主要趋势特征，而且能够避免噪声的影响；然后，基于比特表示并利用聚类可以加速的思想，提出了变形的k中心点聚类算法，将具有相似变化模式的子序列归为一类；基于该聚类算法提出了不一致发现的算法，算法中采取了启发式剪枝和根据簇中心距离剪枝两种剪枝策略。实验结果说明，基于比特表示聚类的不一致发现算法有效性较好，效率较高，且具有扩展性。时间序列的不一致发现可以应用于异常检测。为改进已有时间序列流的异常检测的效果，提出了基于分形的时间序列流异常检测算法。由于关联分形维数的改变可以用作数据集中数据趋势改变的指示器，因此，采用含有基本窗口的滑动窗口模型，利用关联分形维数捕获滑动窗口中当前可见数据的模式特征，设计了基于分形的异常检测算法。与基于TSA-tree的方法、基于免疫的方法的对比实验显示，基于分形的异常检测方法可以有效地发现异常。关键词：时间序列，分割，聚类，不一致发现，异常检测AbstractTimeseriesisasequenceofdatawhichchangeswithtimeorder.Timeseriessegmentationanddiscorddiscoveryareveryimportantinvariousdomains,suchasfinancialdatasegmentation,discorddiscoveryinspacetelemetryandmedicaldata,networkmonitoring,trackingandanomalymonitoringovermovingobjecttrajectorystreams.Accordingtothedeficienciesontimeseriessegmentationanddiscorddiscovery,thefollowingworkhavebeenstudied,includingtimeseriessegmentationbasedonsymbolicrepresentation,timeseriessimilaritymeasurebasedonsymbolicrepresentation,segmentationtechniquefortimeseriesstream,bitrepresentationbaseddiscorddiscoveryonstatictimeseries,fractal-basedanomalydetectionovertimeseriesstream.Accordingtotheactualitythatmostofsymbolicrepresentationbasedsegmentationmethodsonlyreflectthemeanvalueofsegmentsbutlosethetrendinformation,proposeTrend-basedSymbolicapproximation(TSX)segmentationmethod.Afterdimensionalityreduction,getthemeanvalueofsegmentsandrefinetheimportanttrendfeature,thendesignmulti-resolutionanglebreakpointintervalsearchtableanddiscretethetrendfeaturetosymbol,thusobtainsymbolicdimensionalityreductionTSX,whichnotonlycollectsthemeanvaluebutalsoreflectsthetrendinformation.ExperimentsindicatecomparedwithSymbolicAggregateapproximation(SAX),TSXhaslowerfalsepositiverateinsimilaritysearch,whichcaneffectivelysupportsimilaritysearch.DuetothemeasureoftimeseriessegmentationbasedonsymbolicrepresentationMINDIST_PAA_iSAXdoesnotholdthesymmetry,putforwardanewmeasureSym_PAA_SAXbasedonSAX.Sym_PAA_SAXconsiderstheequalstatusoftwoevaluatedtimeseriesindistancecalculation,no