Physics-Informed Graph Convolutional Recurrent Network for Cyber-Attack Detection in Chemical Process Networks

Abstract

Cyber-attacks pose a significant threat to the safety and operational efficiency of industrial chemical processes. Traditional data-driven detection methods often require sufficient training data to achieve the desired detection accuracy. However, for complex chemical process networks, such data sets are often unavailable or insufficient, limiting the effectiveness of these approaches. To address this challenge, this work presents a physics-informed graph convolutional recurrent network (PIGCRN) that incorporates both spatial and temporal information on chemical process networks and a priori knowledge of attacking patterns to improve the detection of cyber-attacks while reducing the requirement on extensive training data. A chemical process network consisting of two reactors is simulated in Aspen Plus Dynamics to demonstrate its superior performance in detecting cyber-attacks compared to traditional data-driven methods.