Research On Attention Mechanism Based Multivariate Time-series Anomaly Detection Algorithm

With the advent of the big data era,many kinds of time series data are produced in daily life.Time-series anomaly detection is the process of identifying abnormal sample data from normal time series data.Nowadays,time-series anomaly detection has made great progress in many fields.For example,the monitoring data of spacecraft,servers and sensors are multivariable time series data.Detecting anomalies in time-series data is a great challenge due to its complex characteristics of unevenness,time dependence and randomness,and its complex nonlinear dependence between multiple time steps and variables.With the wide application of deep neural network,it has more and more advantages over traditional methods in dealing with multi-variable time series anomalies,so it is widely used in time series data anomaly detection.In recent years,with the development of deep learning,attention mechanism has become a hot research and application model because of its ability to capture key information.This paper focuses on the multivariate time series anomaly detection based on multi-attention mechanism.(1)In order to solve one of the main problems of time series data analysis,which is that long time series features are difficult to capture,a new multivariate time series anomaly detection framework,MTAD-TCGA,is proposed,which integrates several modules that are good at capturing long time series features.The framework consists of a time convolution network,then two parallel graph attention layers are used to learn the complex dependencies of data from time and feature dimensions,and then data prediction is entered into Gated Recurrent Unit layer based on an improved attention mechanism.The prediction model and reconstruction model are then jointly optimized to find thresholds and detect anomalies.Experimental results show that the MTAD-TCGA algorithm has higher accuracy than other advanced models.(2)Given that the efficiency of the MTAD-TCGA algorithm,despite its high accuracy,is not significantly different from most algorithms,we propose an ACAM-AD time-series anomaly detection framework in this section.This framework combines auto-correlation mechanism,multi-head graph attention mechanism and dot-product attention mechanism,and comprehensively models the dependencies between data in both time and feature dimensions.In order to reduce the complexity of the model and improve the operation efficiency,the algorithm parallels the auto-regressive model with the neural network,and thinned both the autocorrelation mechanism and the graph attention mechanism.Experiments on public data demonstrate the effectiveness of the ACAM-AD algorithm,which is significantly more efficient than the first algorithm.