Two-stage stacked autoencoder monitoring model based on deep slow feature representation for dynamic processes

Abstract

The slow feature analysis (SFA) method constitutes a robust technique for dynamic process monitoring, capable of extracting slow-varying features to reveal process dynamics. A significant challenge in SFA-based monitoring involves nonlinear relationships within process data. Therefore, this paper introduces a slow feature constraint two-stage stacked autoencoder algorithm for dynamic process analysis. In the first stage, AE units aim to produce decorrelated and normalized signals through nonlinear expansion, with loss term focusing on the related properties. In the second stage, AE units serve to explore deep slow feature representations under constraints on variations of features. By fusing principles of SFA with the representational depth of SAE, the algorithm not only captures nonlinear relationships but also preserves crucial temporal dependencies within data, thereby providing more accurate insights for process monitoring. The proposed algorithm is validated in the vinyl acetate monomer process.