Font Size: a A A

Anomaly Detection For Multivariate Time Series Based On Transformer And High-Order Feature Interactions

Posted on:2022-09-14Degree:MasterType:Thesis
Country:ChinaCandidate:L W DengFull Text:PDF
GTID:2518306524989319Subject:Master of Engineering
Abstract/Summary:PDF Full Text Request
Monitoring complex systems results in massive multivariate time series data,and anomaly detection of these data is very important to maintain the normal operation of the systems.Despite the recent emergence of a large number of anomaly detection algorithms for multivariate time series,most of them have the following three defects.Firstly,they ignore the correlations among multivariate,which can often lead to poor anomaly detection results.Moreover,the correlation among multivariate is hidden and hard to predefine,which make the modeling of correlation become a challenge.Secondly,they mostly use recurrent neural networks or convolutional neural networks to model the temporal information,which is difficult to capture the long-term dependencies while these widely exists in time series data.Thirdly,they do not consider the noise of time series and ignore the robustness of the model,which makes the model susceptible to noise in the data and makes it difficult to generate accurate outlier scores.To solve the problems as stated,we propose a novel unsupervised anomaly detec-tion model for multivariate time series based on transformer and high-order feature inter-actions.More specifically,we build multivariate feature interaction graph automatically and use the graph convolutional neural network to achieve high-order feature interactions.Then we adopt Transformer,which is based on attention mechanism,as the backbone of our model.Thanks to the properties of attention,Transformer has better capability for modeling long-term dependencies than recurrent neural networks or convolutional neu-ral networks.Additionally,in order to model the noise in time series and inspired by variational autoencoder,we use variational encoding technique to map the representation of a segment of time series to stochastic space,which can improve the robustness of the model by modeling the randomness of the time series.Finally,we perform extensive experiments,i.e.,overall performance comparison,ablation study,the sensitivity of hy-perparameters and automatic threshold selection,on three publicly available datasets.The results of these experiments show the effectiveness and superiority of the proposed model.
Keywords/Search Tags:Anomaly Detection for Multivariate Time Series, Graph Convolutional Neu-ral Networks, Transformer, Variational Autoencoder
PDF Full Text Request
Related items