Virtual sensing techniques for nonlinear dynamic processes using weighted probability dynamic dual-latent variable model and its industrial applications

摘要

Over the past decades, data-driven virtual sensors have been widely used to predict hard-to-measure key quality variables where process uncertainties, dynamics and nonlinearity have been considered as critical data features in modern industries. As a result, in this paper, a virtual sensing technique is developed based on a probabilistic dynamic dual-latent structure (PDDLS) in which two distinct dynamic latent variables (LVs) are introduced to take care of quality-related and quality-unrelated dynamic information within measurements respectively. By combining the local weighted (LW) strategy, the virtual sensing technique is further extended to nonlinear applications. Finally, the performance of the proposed method is verified by two industrial cases where the superiority is shown compared with previous researches.