AGDSF-Mamba: Adaptive metric graph diffusion with sparse filtering for mamba guided anomaly detection via spatiotemporal learning in industrial IoT

Identifying and detecting anomalies in Industrial Internet of Things (IIoT) systems is critical for maintaining industrial safety. In IIoT scenarios, sensor devices collect vast amounts of industrial time-series data. Extracting spatio-temporal features from this time-series data for anomaly detection is a key step in preventing production process accidents and ensuring system security within IIOT. However, the effectiveness of existing industrial time-series anomaly detection methods in capturing the constantly evolving dependency topology and unique temporal characteristics between sensors is limited by various changes in industrial production processes, resulting in low detection accuracy. To address these issues, we propose a novel industrial time-series anomaly detection method, named AGDSF-Mamba. Specifically, we devise an adaptive metric graph diffusion model to learn the evolving sensor dependency topology. At the same time, we partition the time-series data into local periodic part and global trend part. By employing a carefully designed sparse masking transformer network and introducing a multi-scale convolutional model, we effectively focus on locally significant information within production processes and identify global correlation trends across all production processes. Furthermore, a novel spatio-temporal association discrepancy-aware anomaly detection method based on the mamba model is developed to enhance the accuracy of anomaly detection. Experiments on real-world datasets show that AGDSF-Mamba outperforms state-of-the-art methods.