Research On Time Series Anomaly Detection Based On Collaborative Learning

Time series data originates from all aspects of human life and industry.Under the background of the sharp increase in data volume today,how to manage and use these time series data more effectively has become the focus of researchers.Although the size and quantity of abnormal data in the time series are relatively small,some of the implied information may have an important impact on scientific research or production and life.Since the problem of time series anomaly detection has been proposed,many researchers at home and abroad have conducted in-depth and extensive research,and also proposed anomaly detection methods applied in many fields,but there are still in the field of time series anomaly detection.Many problems worth studying,discussing and solving.This thesis focuses on time series anomaly detection.In conclusion,this thesis consists of the following achievements.(1)The Method of Multi-Scale Anomaly Detection in Time SeriesAiming at the problem of point detection and linear segmentation in time series,a multi-scale anomaly detection method based on time series is proposed in this paper.In this method,time series is compressed and decomposed on different scales by Haar wavelet transform,and then time series of wavelet transform is segmented into variable length patterns(subsequences)by quadratic regression model.The anomaly value of all patterns are calculated.The original time series is reconstructed by the wavelet reconstruction function,and finally the abnormal pattern in the original time series is detected by comparing with the adaptive threshold.The method was verified and evaluated using the simulation data set Keogh_Data and compared with the three anomaly detection methods IMM,TSA-Tree and Tarzan that tested the Keogh_Data data set.The experimental results show that the three-level wavelet decomposition of experimental data can achieve good anomaly detection results.And compared with the Tarzan method,distance-based method and density-based method in detection accuracy,detection efficiency and detection rate,the method can reach 93.5% and 96.2% in terms of short-term accuracy and efficiency.In terms of long-term accuracy and efficiency,91.7% and 90.5% can still be achieved,and the anomaly detection effect has been significantly improved.(2)Time Series Anomaly Detection Method Based on Collaborative Training and Selective Ensemble LearningAiming at the problem of insufficient detection of time series anomalies in traditional classification methods,a time series anomaly detection method based on collaborative learning was proposed.The traditional classification method is mainly to train the classifier through a given tagged subsequence,so that the classifier predicts and classifies unknown subsequences.However,in actual research,a large number of tagged subsequences take a long time and labor to obtain,and there are a large number of unmarked subsequences in the dataset.Therefore,a cooperative training algorithm combining mark sequences and a large number of unlabeled sequences is proposed to analyze the time series.Collaborative training algorithm will be subject to the diversity of classifiers in the process of training learning.To solve this problem,combined with selective ensemble learning,a time series anomaly detection algorithm RFCL based on collaborative learning is proposed and combined with the multi-scale method of wavelet transform.The assisted collaborative learning algorithm selects unlabeled sequences for classifier training and learning,and is used for abnormal detection of time series patterns.Finally,on the basis of the theoretical research,the simulation data set Keogh_Data,Ma_Data and the real economic data set were used to test the anomaly detection method proposed in this paper.The experimental results show that the time series anomaly detection algorithm RFCL based on collaborative learning can effectively detect the abnormal pattern in the time series,and has high detection accuracy and detection efficiency.