Process-dynamics-guided latent predictability embedding supervised deep networks for soft sensing in industrial processes

Abstract

Soft sensors for complex industrial processes have become a challenging task due to dynamic self-correlation caused by the feedback loop and inertia effects. Although the dynamic latent variable models offer an interpretable solution, the linear latent variables fail to capture the behavioral characteristics of strongly nonlinear industrial processes. Thus, this article proposes a new deep stacked autoencoder with latent predictability embedding for soft sensing, which is called the process-dynamics-guided latent predictability embedding supervised deep network (PDLPSDN). To capture the autocorrelation in the process data, a regularization term based on the point prediction is embedded into the decoding loss. Subsequently, information theory is used to link the contribution from past time steps to the present, which is used to guide the structure of the latent dynamics. Finally, the parameter-guided regularization terms assist in learning the temporal dependencies in the process data and are then trained in an alternating manner. The proposed PDLPSDN decreases the root mean squared error by 16.8% for the debutanizer column and 25.7% for the sulfur-recovery unit, demonstrating the reliable and superior performance of the proposed PDLPSDN-based soft sensing.