MAD-DGTD: Multivariate time series Anomaly Detection based on Dynamic Graph structure learning with Time Delay

Abstract

Anomaly detection of multivariate time series data is extremely important in the industrial operation maintenance of Internet of Things (IoT). Researchers have found that the relationship between multiple sensors can be modeled as graph structure, and most researchers expresses this relationship by learning static graph structures which only contains the information of single modal. However, in actual IoT, the relationship between sensors will change with the changes of operating conditions, and this fixed graph structure cannot capture the relationship between sensors when working mode changes. To compensate the shortage of static graph, we propose a Multivariate time series Anomaly Detection framework based on Dynamic Graph learning with Time Delay (MAD-DGTD). Firstly, time-delay dynamic graph learning module (TDDG) is proposed to learn the changed mutual relationship between sensors over time and model it as a dynamic graph structure. In TDDG, a delay impact learning mechanism was designed to reconfigure the similarity calculation of node embeddings, which is designed to handle the temporal asynchrony of interactions between sensors in IoT. Secondly, we designed a stacked time dimension information extraction module (TDIE) and graph convolution information propagation module (GCIP) to capture information of different fine-grained sizes through multi-scale feature extraction. Finally, experimental research on three real-world datasets shows that our method outperforms the existing 10 competitive baselines in terms of overall performance.