Multivariate Time Series Anomaly Detection Based on Dynamic Graph Neural Networks and Self-Distillation in Industrial Internet of Things

Time series anomaly detection is critical to securing the Industrial Internet of Things (IIoT). Although numerous deep learning-based methods have been proposed, these methods fail to consider the interdependencies between different dimensions of the data and often neglect the dynamic changes in these dependencies. Moreover, these methods utilize only the global features from the last layer of the network for anomaly detection. However, local features can capture subtle variations in the data, which are crucial for accurately detecting anomalies. To alleviate these problems, this paper proposes a novel framework for detecting time series anomalies, including four parts, namely the graph structure learning module, the dynamic graph module, the anomaly scoring module and the self-distillation. The graph structure learning module generates different graph structures based on the inputs, which will be used in the dynamic graph module. The dynamic graph module employs dynamic graph neural networks to capture the complex relationships within time series from both temporal and spatial dimensions. The anomaly scoring module obtains anomaly scores from predictions and observed values, and the model makes anomaly judgments based on these scores. Additionally, self-distillation enhances model performance by utilizing mutual learning between the teacher and student models, thereby integrating local and global information for better anomaly detection. We carry out a series of experiments on Industrial Internet of Things datasets, which verify the performance of the framework. The experimental results of the proposed method outperform other methods, demonstrating the advantage of our framework.