Dynamic graph contrastive learning for multivariate time series anomaly detection

Multivariate time series (MTS) anomaly detection has significant applications in fields such as industry, transportation, and network monitoring. Although various models have been developed for MTS anomaly detection, most fail to effectively address the intricate dependencies between intra-series and inter-series correlations, limiting their performance. Moreover, there is limited research on how inter-series correlations evolve across different time scales. To address these challenges, we propose DyGCL (dynamic graph contrastive learning anomaly detection framework). Specifically, DyGCL employs frequency-domain analysis to identify prominent periodic patterns in time series and segments the data based on varying time scales. It constructs multi-scale dynamic graphs by measuring variable similarities within segmented periods and uses graph convolutional network to capture the time-varying inter-series correlations. The framework integrates LSTM and self-attention mechanisms to model local and global intra-series correlations, respectively, and incorporates a contrastive structure to detect anomalies without relying on reconstruction errors. Additionally, we introduce a dynamic graph anomaly-denoised strategy, which enhances the model's ability to learn normal data distributions by identifying and removing anomalous node features. Experimental results demonstrate that DyGCL achieves state-of-the-art anomaly detection performance across multiple datasets, offering new perspectives and approaches for MTS anomaly detection.