Adaptive Anomoly Detection For Data Stream Of Sequence-based Slinding Windows Model

The telemetry data, produced by in-orbit satellites, are the only basis for theground staffs to estimate the health status of satellites. So paying more attention onanomaly detection of the most recent telemetry data is of great significance.However, with the feartures of large, rapid and real-time, the telemetry data aretransmitted to the ground in the form of streams, and the traditional analysis andprocessing methods can not be effectively applied. Reasearch on the slidingwindows model, as the research hotspot, is limited to the data streams in the nearestwindow which meets the anomaly detection applications of the telemetry data.Aiming to achieve anomaly detection for satellite telemetry data, this thesis workson the anomaly detection technologies based on sliding windows model and thedetailed contents are as follows:First, for the contradiciton between a large number of the telemetry streamcontinuing to reach and the computer’s limited memory resources, the thesis workson the generation algorithms of data stream’s summary. To receive thecomprehensive information of the recent window, UBCS is proposed based on thetechnique of basic window. Moreover, the results of experiments verify that thesamples, based on UBCS algorithm, keep the advantage of uniform distributionamong the basic windows. Secondly, for the anormaly detenction of single-pointexception、continuous and aggregated anomaly in univariate-stream arriving inchronological order, the method of anomaly detection based on Gaussian ProcessRgression algorithm(GPR) is developed to achieve the direct output of confidenceintervals. As opposed to Na ve and MLP, the results of experiment show theimprovement of its detection performance. Thirdly, to improve the GPRperformance of continuous anomaly detection, UBCS_GPR method is raised whichcombines GPR with UBCS algorithm. And it reduces the proportion of abnormaldata getting access to training windows effectively, the experiment’s results showthe improved performance compared with GPR method. In addition, for thedetection problem of aggregated anomaly, IUBCS_GPR method is put forward.IUBCS_GPR builts the offline and online model based on UBCS_GPR method.Offline model can be used to model the normal data and output the effectivesampling ratio and initial online model.Then online model updates in real time usingthese inputs to track the changes of data stream. Power and Space Data set verify itsvalidity for the detection of aggregated anomaly. Finally, aiming at anomalydetection for multiple-streams, this thesis proposes the HSWStream (High SlidingWindows) algorithm achieving the effective culster of multiple-streams. As a result, the experiments of anamoly detection achieve good results at different indexs ofdata sets KDD99.This thesis carrys out the reasearch on anomaly detection technology forunivariate-stream and multiple-streams, and the reasearh refers to two differentapplicated levels of anomaly detection. Furthermore, the simulated and real publiclydata sets validate the effectively of different algorithms which can be used for theanomaly detection of telemetry data from satellite afterwards.