| As the data-generating system becomes larger and larger and its structure becomes more and more complex,it is difficult to grasp the overall state of the system through a single parameter.Most of the data generated by it has the characteristics of time series,and because of the interrelated influence of multiple levels,multiple components,and multiple mechanisms of action,it presents a "diversified" nature.The information contained in these time series data can reflect the overall state of the system.Therefore,how to dig out more valuable information from a large number of multivariate time series has become an important research problem in the field of data mining.Traditional data mining pays more attention to frequent patterns that appear more frequently in time series data,and chooses to ignore or filter out the rare patterns or abnormal patterns.However,the information contained in these rare patterns is often more important,providing important guidance and support for further in-depth data mining,and can even bring us heuristic thinking influence.Based on this,this article mainly focuses on the anomaly detection algorithm of time series,aiming to dig out the abnormal information hidden in the massive multivariate time series through the trend change information and feature representation information of the time series.On this basis,a multivariate time series anomaly detection method based on non-uniform segmented trend information symbols with practical significance and value is proposed.The main work of this paper is as follows:First,it analyzes the difficult problems faced by multivariate time series data in anomaly detection and feature representation based on the research direction and content of this article.In order to explain the problem more comprehensively,on the basis of analysis,several representation methods are explained in detail,mainly to introduce the differences and connections,advantages and disadvantages of piecewise linear representation and symbolic representation methods.Finally,In order to realize the representation of the local trend information of the time series,combined with the characteristics of the trend information,the segment maximum value is introduced for non-uniform segmentation,the trend information is expressed by the fitting slope,and the fitting slope is symbolized and mapped.As a result,a formal representation method based on non-uniform segmentation trend information is proposed,and the distance measurement method is defined and explained,which provides strong support for the subsequent abnormal mining of multivariate time series.The problems of the traditional time series anomaly detection algorithm based on symbolic representation are analyzed,and the algorithm is improved to solve this problem.In the stage of dimensionality reduction and feature representation,segmented trend information is used as important information for the feature representation of multivariate time series.In order to avoid the algorithm from performing more redundant calculations in the process of anomaly detection,clustering algorithm is used to analyze the trend symbols of multivariate time series.Information feature represents the result for clustering optimization,redundant pruning improves the efficiency of the algorithm in the search process and reduces the time complexity of the algorithm.In addition,the representation of trend information can better distinguish between different feature representations and improve the accuracy of the algorithm's detection results.Finally,the algorithm studied in this article is applied to different time series data sets for simulation experiments,and the algorithm shows ideal results.Through comparative analysis,compared with other multivariate time series anomaly detection algorithms based on symbolic representation,the improved algorithm not only can effectively detect the location of anomalies,but also has a greater advantage in detection efficiency. |
PDF Full Text Request