Research On Anomaly Detection Of Time Series Based On Multiple Attention Mechanisms

With the rapid development of the Internet,anomaly detection has become more and more widely used in reality,such as fraud detection,abnormal network traffic detection and structural defect identification.Time series can reflect the historical laws of the system and is easy to collect.Therefore,the use of time series for anomaly detection has attracted more and more attention.In practice,a time series usually consists of multiple variables at multiple time points.This characteristic makes time series anomaly detection algorithms often unable to make full use of time series information to reconstruct samples.At the same time,existing algorithms usually judge the abnormal time point according to the difference between the original feature and the reconstructed feature at each time point.The greater the difference between the original feature and the reconstructed feature at a certain point in time,the higher the degree of abnormality.Therefore,how to make the original features at the abnormal time point and the reconstructed features have a large gap is also an important issue.Based on this,this paper explores the solutions to the difficult problems in time series anomaly detection,and carries out the following three tasks:(1)Model design based on dual attention.The current representative deep learning anomaly detection method does not fully consider the relationship between multiple variables and multiple time points in the process of extracting time series features.On the other hand,the self-attention mechanism has superior performance in modeling multiple entity relationships.Therefore,in response to the above problems,this study proposes an anomaly detection model based on a dual attention mechanism,which includes a feature attention module and a temporal attention module,which are used for multi-dimensional feature variables and multiple time points,respectively.The feature extraction between them can fully mine the timing information in the sequence.(2)Training strategy based on consistency loss.Suppressing the reconstruction of outlier samples will amplify the difference between reconstructed features and original features on outliers.The suppression strategy of existing methods relies on the temporal correlation of the time series and cannot be applied to the features extracted from the feature self-attention module.To address this issue,this study proposes a consistency loss.The loss function acts on the model training stage,which will make the hidden layer features of the model at the abnormal time point close to the hidden layer features at the normal time point,and then reconstruct the normal input features,resulting in the gap between the reconstructed features and the abnormal input features.Larger is more conducive to anomaly detection.(3)Anomaly detection system design.To apply the research to practice,this study designs a time series anomaly detection system.The system is mainly divided into server-side and front-end.The server side is divided into an offline processing module and an anomaly detection module.The offline processing is mainly to preprocess the data,and the anomaly detection module has a built-in algorithm proposed in this study.The front-end interface performs alarm and statistics on abnormal time points.This study explores two difficult problems in time series anomaly detection,and proposes a time series anomaly detection algorithm based on dual attention.Based on this,a time series anomaly detection system is designed.