Graph neural networks for anomaly detection and diagnosis in hydrogen extraction systems

Recent research has been actively conducted on fault diagnosis in hydrogen extraction systems using artificial intelligence. However, existing studies have not considered the characteristics of hydrogen extractors, where multiple processes form a single system and anomalies in one subsystem can impact others. This study proposes a method combining graph autoencoders (GAE) with graph convolutional networks (GCN) to detect and diagnose anomalies in hydrogen extraction systems. The integrated GAE-GCN model generates an adjacency matrix that represents changes in component dynamic relationships based on system topology information and featureaugmented sensor data. Anomalies are detected using reconstruction errors from an autoencoder model trained on the degree centrality of the adjacency matrix in the normal state. The diagnosis of anomalies in a specific heat exchanger is achieved by identifying the associated nodes through graph analysis. This research contributes to effective anomaly detection and diagnosis in hydrogen extraction systems using graph neural networks.