Research On Time Series Anomaly Detection Based On Deep Learning

Anomaly Detection(AD)is a type of problem has attracted the attention of academic and industrial.However,there is no standard and universal definition of what is "abnormal",and it is usually related to specific application scenarios.The currently more recognized expression is given by Hawkins.He believes that samples that deviate from the standard value(or expected value),that is,samples that are significantly different from most of the data,are abnormal,and this type of data accounts for a relatively small proportion of the overall.Due to the extensiveness and common use of time series data,TSAD with AD have the same characteristic,the lack of tags in the historical data,the simulation is usually performed in an unsupervised mode.Besides,people are not only concerned about the ability of the algorithm to detect anomalies,but also interested in the cause of the anomaly.With the widespread application of deep neural networks,its advantages over traditional machine learning algorithms in solving multi-variable timing anomaly detection have become increasingly significant.Using the deep learning networks to deal with the multivariate timing anomaly detection has become a general trend,and the surge in various related documents reflects this.In order to depict the abnormality of abnormal data,the most direct and classic method is to use a variety of distance-based,statistical,and density-based quantitative indicators to describe the degree of alienation between samples.In deep learning,some existing model frameworks are used to build models.However,most models do not pay attention to the relationship between variables in multivariate problems,and just focus on information in temporal pattern when building a network framework.The amount of information brought by multiple features can not only improve the robustness of the model in detecting anomalies,but also provide strong support for anomaly location(that is,anomaly traceability,location of abnormal features)after effective capture and utilization of data information.Based on the above,this article mainly focuses on deep learning to deal with unsupervised MTAD issue.The main achievements are as follows:(1)Two unsupervised anomaly detection models on multivariate time series are proposed,and both algorithms show superior performance on three public data sets.(2)The model combines information from two perspectives of temporal domain with feature domain to capture the relationship of variables in temporal and feature pattern.(3)Carry out abnormal diagnosis experiment.The use of the attention mechanism module in the model can explore the correlation between variables as well as to find the reason of abnormal through the larger fluctuation of the weight between the features at the abnormal moment compared with the normal situation,which has a good explanatory effect on the occurrence of anomalies and can be used for auxiliary abnormal diagnosis function.