Knowledge-assisted multi-graph structure learning for multivariate time-series anomaly detection in multi-stage industrial processes

Abstract

Multivariate time-series anomaly detection is important for ensuring the reliable operation of an industrial process. Most real-world industrial processes consist of multiple stages (or sub-processes). In such a multi-stage industrial process, sensors in the same sub-process are highly correlated, and sensors in distinct sub-processes can even exhibit notable correlations when the sub-processes are closely related. Recently, graph neural networks-based methods have achieved remarkable performance in multivariate time-series anomaly detection because they can effectively capture complex inter-sensor dependencies by learning a graph that explicitly represents relationships between sensors. However, there is still room for improvement in their performance by enhancing the ability to capture inter-sensor dependencies when dealing with multistage industrial processes as most of them overlooked pervasively existing two types of partial knowledge regarding the dependencies, \textit{which sub-process a sensor belongs to} and \textit{which sub-processes are closely related,} when learning the latent graph structure. In this paper, we propose a novel graph neural network-based multivariate time series anomaly detection method for multistage industrial processes, which can learn multiple graph structures with the assistance of both types of the partial knowledge about inter-sensor dependencies and has the enhanced ability to capture complex inter-sensor dependencies. We demonstrate the superiority of the proposed method with two sensor datasets from real-world multistage industrial processes.