Graph-Based Time-Series Decomposition for Multisource Sensors Anomaly Detection

The detection of anomalies in multisource time-series data is crucial for the efficient operation of the Internet of Things (IoT) devices and large-scale cross-domain systems. Recent methods have primarily focused on using deep learning to model time series directly. However, these approaches often overlook intricate inherent patterns within the data. We have introduced time-series decomposition into anomaly detection, allowing algorithms to explicitly uncover underlying periodicities. Additionally, to address the lead-lag effects among multiple sensors, we integrated the dynamic time warping (DTW) algorithm into the construction of graph neural networks. This integration enables the efficient and accurate construction of adjacency matrices based on sensor similarity. Furthermore, by leveraging graph networks for spatiotemporal modeling of time series and employing a joint loss function, our approach achieves greater accuracy and stability in anomaly detection performance. Comparative analysis with multiple datasets and algorithms demonstrates the consistent superiority and stability of our method.