Denoising-Enhanced Dynamic Graph Modeling for Anomaly Detection in Multivariate Time Series

The progress of industrial IoT and sensor technology has made multivariate time series anomaly detection crucial for system safety and reliability. Despite the application of graph neural networks improving detection efficiency, limitations such as static global correlation modeling and robustness to noise restrict the improvement of anomaly detection accuracy. Therefore, this paper proposes a new dynamic graph construction framework for multivariate time series anomaly detection. The framework includes a collaborative denoising feature extraction module, a dual-perspective dynamic graph builder module, and an anomaly detection module. Experiments on four real-world datasets show that our method’s F1 score and AUC value reach or are close to the optimal level, with a 7.34% improvement on the MSDS dataset.