Research On Multivariate Time Series Anomaly Detection

Anomaly detection is a crucial research topic in many fields and extensively applied in economies,industry,meteorology,aerospace,etc.And multi-variate time series data is one of the most seen data types in anomaly detection.Compared with anomaly detection on other data types,multi-variate time series anomaly detection has following three unique properties.First,the data is sequential,thus temporal related to each other.Second,types and values of different variables in multi-variate time series may vary widely even in the same system.Last,different variables may affect and rely on each other.There are two major shortcomings of recent researches.Firstly,most works only consider data in multivariate time series as isolated points and neglect the sequential characteristic of the data.Even in some works that already take this sequential feature into consideration,they forget the variable-level relevance among the data or they fail to consider these two characteristic simultaneously.Secondly,recent researches only focus on embedding either normal data or anomalous data into a latent space.To cope with these above shortcomings,this thesis makes following contributions.To explore the characteristic from different aspects in multi-variate time series,a model combining sequence-level semantic and variable-level semantic features simultaneously is proposed,namely,VTG-TAD.VTG-TAD exerts the techniques of graph neural networks to model the sequence-level semantic and variable-level semantic patterns explicitly,then merges the two prediction results and feeds them to a dynamic thresholding module to determine the anomalies.Comparative experiments with several conventional models on two open-source datasets validate the effectiveness of the proposed model.To fully discover the differences between normal and anomalous patterns,an anomaly detection model based on triplet loss and autoencoder is proposed,namely,TAE-TAD.By embedding the normal data and anomalous data in the same latent space,data of the same class can be much closer,while data from different classes are pushed far apart.As the pattern differences get larger,the anomaly detection task becomes easier.Experimental results show that TAE-TAD model does get better results than conventional methods.