A Novel Temporal-Enhanced Dynamic Latent Variable Regression for Quality Prediction of Dynamic Processes

The dynamic latent variables (DLVs) model, which describes dynamic relations within a reduced-dimension latent space by the autoregressive (AR) technique, is effective for modeling dynamic relations between process variables and quality variables. However, the reliance on linear assumptions and shallow feature extraction significantly limits its available scope of application. To this end, a novel deep dynamic modeling approach called Temporal-Enhanced DLV regression (TE-DLVR) is developed for quality prediction of dynamic processes. The TE-DLVR model can be trained in an end-to-end manner and is mainly composed of a gated recurrent unit (GRU)-based latent-enhanced autoencoder (LEAE) architecture, a GRU-based AR (GAR) module, and a self-attention mechanism-driven regression (SAR) module. First, LEAE extracts spatially enhanced nonlinear DLVs guided by dual constraints of reconstruction error and latent encoding consistency. Meanwhile, the GAR module achieves temporal consistency in these DLVs for the Temporal-Enhanced modeling of LEAE. Then, the SAR module establishes the mapping relations by associating the extracted nonlinear DLVs with quality variables. The collaborative learning of these three components effectively extracts the nonlinear DLVs. Finally, the validity of the proposed TE-DLVR is verified through the sulfur recovery unit and Dow distillation process.