Spatio‐temporal variational autoencoder with Gaussian process latent prior for multimode chemical process monitoring

Abstract In the chemical industry, real‐time process monitoring constitutes a fundamental prerequisite for ensuring operational safety and production stability. With the advancement of information technology, data‐driven process monitoring methods have been extensively adopted in chemical process monitoring. However, the inherent multi‐mode operation characteristics of chemical processes, necessitated by fluctuating product demands, render conventional single‐model monitoring approaches inadequate in maintaining satisfactory accuracy. Moreover, the complex spatio‐temporal coupling characteristics of process data further increase the difficulty in accurately diagnosing fault occurrences. To address these challenges, this study proposes a novel spatio‐temporal variational autoencoder with Gaussian process prior (ST‐GP‐VAE) for process monitoring. The model employs a multi‐scale one‐dimensional convolutional network combined with an improved gated recurrent unit (IGRU) to jointly extract spatio‐temporal features, while incorporating a Gaussian process (GP) prior to explicitly characterize temporal dependencies in latent variables. Operational mode identification is achieved through temporal convolutional networks with sliding windows, while equipment operating status is determined based on statistical metrics of reconstruction errors and predefined thresholds. Comprehensive experimental validation conducted on both the benchmark Tennessee Eastman (TE) process simulation dataset and real‐world continuous catalytic reforming (CCR) unit data demonstrates the effectiveness of the proposed methodology.