Dynamic graph neural networks and enhanced transformer for multivariate time series anomaly detection

Multivariate time series anomaly detection is crucial for ensuring the reliability of complex systems in industrial, healthcare, and IoT domains. However, existing methods struggle to effectively capture both dynamic inter-variable relationships and temporal dependencies. To address these challenges, we propose the Dynamic Graph Neural Networks and Enhanced Transformer (DGET) framework, which combines a Dynamic Graph Construction module to model evolving dependencies and an Enhanced Transformer architecture with dynamic temporal encoding, time-decay attention, and local window mechanisms for capturing temporal dependencies. Additionally, self-supervised contrastive learning enhances the model's robustness in label-scarce environments. We evaluate DGET on three benchmark datasets: SMD, SWaT, and WADI. Results show that DGET outperforms baseline methods (SVM, GNN, LSTM) and the state-of-the-art TranAD, achieving superior F1-scores. Ablation studies further confirm the critical roles of the Dynamic Graph Neural Networks and Enhanced Transformer components, demonstrating DGET's effectiveness for multivariate time series anomaly detection.