A Multivariate Time Series Anomaly Detection Model Based on Spatio-Temporal Dual Features

Abstract

This paper proposes a novel unsupervised model for multivariate time series anomaly detection (TSAD), targeting the challenges of sparse and unlabeled abnormal data, as well as high dimensionality in IoT applications. The core of our model is to extract spatiotemporal dual features through a coherent architecture that captures both temporal dependencies and spatial correlations among multiple variables. Specifically, a deep autoencoder is employed to capture the spatial features of multivariate time series data, while a multi-scale sparse Transformer network is used to extract the temporal features. An anomaly detection module based on the dynamic threshold POT method is used for detect anomalies, and extensive experiments are conducted on publicly available datasets. The result on the SMAP dataset show that our proposed model improves the precision and F1 score by 11.3% and 5.48% respectively compared with the latest baseline method. On the NAB dataset, the F1 score is increased by 0.47%. On the SWaT dataset, the precision is improved by 0.62%.