Spatio-Temporal Data Anomaly Detection Using 3G-Net in IoT

Shuo Zhang,Jiayuan Chen,Xiaofei Chen,Qiao Jiang,Hejiao Huang
DOI: https://doi.org/10.1109/icpads56603.2022.00105
2023-01-01
Abstract:In Internet of Things (IoT), smart devices continuously collect various spatio-temporal data via different kinds of sensors, whose anomaly detection is crucial for the smooth operation of the whole system and the decision-making afterwards. Recently, traditional Deep-Learning based methods (i.e., CNN) for anomaly detection gain significant progress in the Euclidean space, but they fail to capture the spatial dependencies of spatio-temporal data from non-Euclidean domain. Besides, there might be few or even no abnormal data available for training when dealing with anomaly detection issue. In this paper, an end-to-end deep learning framework 3G-Net is proposed to address the problem above. The core idea of 3G-Net is to learn the normal pattern of IoT data considering both the long-short-term temporal dependencies and non-Euclidean spatial dependencies, and the more an observation differs from normal pattern, the more likely it is anomalous. Besides, an offline anomaly detection workflow based on Correcting Mechanism and Anomaly Score is also put forward in 3G-Net. Experiments conducted on different real-world and public datasets demonstrate that, major techniques in 3G-Net are effective in IoT scenes and it outperforms other state-of-the-art approaches overall.
What problem does this paper attempt to address?