Stationary Multi-scale Hierarchical Dilated Graph Convolution for Multivariate Time Series Anomaly Detection

Anomaly detection within multivariate time series is a pivotal area of study in the field of data mining, playing a crucial role in safeguarding infrastructure integrity. Effective and precise methods for detecting anomalies are essential for users to diagnose issues promptly, thereby preventing significant financial damage. Current methods for anomaly detection often do not adequately account for the multi-scale spatial and temporal relationships present in sequences, and they also overlook the non-stationarity of the sequence. Thus, this paper proposes SMHDG, a novel unsupervised stationary multi-scale hierarchical dilated graph convolution for multivariate time series anomaly detection. The core of this method is to capture multi-scale temporal dependence and feature correlation of multivariable time series by stacking dilated convolution layers and graph convolution. Mean-while, normalization and de-normalization are used to achieve sequence stationarization. Experiments across four authentic datasets have demonstrated that SMHDG outperforms leading baseline in accurately pinpointing anomalies in time series data. Furthermore, ablation experiments confirm the effectiveness of the method's key components.