Attention-based Dynamic Latent Variable Modeling Method for Process Monitoring

This study addresses the underexplored challenge of inherent dynamics in industrial processes through an innovative attention-based latent variable modeling method. Utilizing attention mechanisms, the method articulates time-variant dynamical relationships among samples. The framework extends attention-based dynamical inner principal component analysis to extract latent dynamical features, integrating them with static features obtained through static principal component analysis. This results in comprehensive monitoring statistics for online applications. Numerical simulations and real-world application in an industrial ethylene oxychlorination process demonstrate the proposed method's efficacy. Comparative analysis highlights its advantages and superior performance over existing methods. This innovative approach provides more accurate insights into complex industrial processes, promising advancements in data-driven modeling within the field.