A Multivariate Time Series Anomaly Detection Method Based on Generative Model

Abstract

Modern equipment is complex in structure, large in scale and highly integrated, in order to solve the problems of high dimension and a large amount of data collected by equipment, a multivariate time series anomaly detection method based on the deep generative model (DG-MTAD) has been proposed in this paper. Long short-term memory (LSTM) network is used to optimize the model structures of autoencoder (AE) and generative adversarial networks (GAN). While extracting time information, the bidirectional mapping of data between multi-dimensional feature space and low-dimensional latent space is completed. The reconstruction of normal time series in latent space is realized by GAN, and then the combination of generating loss and discriminant loss is calculated as the anomaly score, so as to realize the anomaly detection of multivariate time series. Experiments on the high-dimensional engine degradation monitoring data set published by NASA show that the accuracy of the method is over 90%.