Time Series Anomaly Detection Model Based on Memory-enhanced Transformer and Graph Network Joint Training

To address the challenges of overgeneralization and inter-variable interactions inherent in multivariate time series data, this study proposes an unsupervised anomaly detection model, METG, which integrates a memory-enhanced Transformer with graph joint training. The memory enhancement module captures and records diverse patterns between normal data and memory items, significantly improving the model's representation capabilities for time series. Simultaneously, a graph convolutional network extracts spatial dependencies and feature interactions by leveraging graph structures. By combining the strengths of temporal and spatial modeling through a joint optimization objective function and anomaly detection criteria, METG achieves robust performance. Experimental evaluations on real-world datasets across four domains demonstrate the model's effectiveness, achieving an average F1 score of 94.5%.