Research On Multivariate Time Series Anomaly Detection Algorithm Based On Bilateral Sliding Window And Multimode

Anomaly detection aims to find some observations,which have large deviations from other observations.Such deviations may be caused by different reasons or mechanisms.Multivariate time series anomaly detection is an important research content of time series anomaly detection.However,the traditional multivariate time series anomaly detection methods are usually single modal,that is,they only consider the variate correlation in the time domain feature space,and ignore the relationship between different variates of time series information in the multi-modal space.At the same time,the previous methods usually use single sliding window input,the local time dependence of time series is not fully considered,which leads to poor detection efficiency.In order to make use of the variate correlation in the multimodal feature spaces of the time series,this paper designs the Multimodal Variate Correlation Learning Algorithm(MVCL),which maps the time domain features of the time series corresponding to each variate to the frequency domain through the time frequency joint domain analysis method,and then jointly calculates the variate correlation in the time domain and frequency domain feature spaces,realize cross modal interactive learning of variate correlation between time domain features and frequency domain feature space.In order to fully consider the local time dependence of the time series,this paper takes bilateral sliding window sampling of the original time series as the model input to fully capture the change trend of local data.This paper proposes a Multivariate Time Series Anomaly Detection Algorithm Based on Bilateral Sliding Window and Multimode(BSM-MTAD).BSM-MTAD aggregates node features with structural information by using the learned multimodal variate correlation through graph neural network,and introduces attention mechanism,assign different weights to different neighbor nodes,use high weights to focus important information,and use low weights to ignore irrelevant information,so as to more effectively extract features and effectively improve the performance of anomaly detection.Experiments were carried out on four public datasets of two water treatment plant datasets SWa T and WADI,an Internet server dataset SMD,as well as SMAP dataset,and compared with five benchmark methods,which proved that BSM-MTAD can better capture local time dependencies and learn more reasonable graph structures,effectively improving model detection performance.