Graph Neural Networks for Chemical Process Fault Diagnosis Based on Hybrid Variable Feature Learning

Graph-based data representation allows us to take into account not only the behavior of each sensor signal but also the correlations and hidden relationships between them. Therefore, using graph neural networks can achieve better results in fault diagnosis. However previous fault detection methods tend to focus only on continuous variables, but industrial processes usually contain both continuous and discrete variables. To utilize this information, we propose a method named graph neural networks for fault diagnosis based on hybrid variable feature learning (HV-GNN). The method continuous discrete variables through a self-attention mechanism and then feeds them into the graph neural network as nodes along with continuous variables for fault diagnosis. Simulation experiments were conducted on an industrial coal gasification process platform and satisfactory results were achieved. Comparative experiments have shown that HV-GNN improves trouhlcshooting accuracy.