Research On Multivariate Time Series Anomaly Detection Methods

With the rapid development of the Internet of Things(Io T)technology,various systems have deployed a variety of intelligent sensors and wireless networks to achieve information interconnection,generating a large amount of complex and multidimensional data.Among them,multivariate time series is a common and representative type of data,which plays an important role in anomaly detection of systems.Currently,in the research of multivariate time series anomaly detection,there are two difficulties: one is that the scale of multivariate time series is large,and the noise interference is strong,which easily leads to the situation of missed or false alarms;the other is that multivariate time series has inherent characteristics of high dimensionality and nonlinearity,which makes it difficult to quickly and effectively obtain their temporal and spatial features.Considering the characteristics of large scale,complex structure,and implicit correlation among variables in multivariate time series,this paper mainly studies how to effectively extract the important features of multivariate time series,handle the nonlinearity of data,and improve the performance of anomaly detection algorithms,so as to ensure the robustness and availability of the algorithms.The main research contents and achievements of this paper are as follows.(1)To address the problems of excessive generalization leading to missed anomalous data and decreased precision in traditional autoencoder,and considering there is a temporal dependency in time series,this paper proposes an improved autoencoder model called MSMAE.The model introduces a memory module into the traditional autoencoder architecture,which can better constrain the latent vector and amplify the reconstruction error of anomalies.Meanwhile,as the feature information between different time steps of multivariate time series can better represent the overall operating state of the system,a multi-scale association matrix is constructed.The memory autoencoder is used to reconstruct the multi-scale association matrix,and the residual matrix between the reconstructed association matrix and the original association matrix is used to determine anomalies.Experiments are conducted on three public datasets: SWa T,WADI and SMD.The results show that MSMAE has higher F1 scores compared to other comparison methods,demonstrating the superiority of MSMAE in anomaly detection tasks.(2)Due to the real-time time series generated by various interconnected devices and sensors in networked physical systems,there exists implicit correlation between time and space and highly nonlinear relationships.Therefore,obtaining effective time and space feature information is the key to anomaly detection.This paper proposes a multivariate time series anomaly detection method,BNCNF,based on conditional normalizing flow.The method uses Bayesian networks to model the structural relationships of multivariate time series,and decomposes the joint probability density of graph nodes into conditional density.Meanwhile,a dependency encoder is designed to induce conditional information into a fixed-length representation,and then the representation is introduced as conditional information into the normalizing flow for density estimation.The data corresponding to low density is judged as an anomaly.The experimental results show that the proposed method outperforms the comparison methods on three public datasets,SWa T,WADI,and SMD,which reflects that combining Bayesian networks with conditional normalizing flow can achieve better anomaly detection performance.