Sequential Variational Autoencoder for Industrial Inferential Modeling

Accurate prediction of quality-related variables is crucial for optimizing and controlling chemical processes. Variational Autoencoders (VAEs) excel in deciphering the complexities of nonlinear process data. However, their utility is limited by an assumption of static processes, which contrasts with the dynamic nature of industrial environments that feature longterm sequential dynamics. To address these challenges, this study introduces the Sequential Variational Autoencoder (SVAE). The SVAE model incorporates sophisticated self-attention and crossattention mechanisms, which are highly effective in identifying and interpreting long-term dependencies across diverse time series. Implementing the SVAE in quality prediction tasks for the coal-to-ethylene glycol conversion process showcases its robust capability to address these complex challenges, marking a significant enhancement over conventional methods.