Refining data-driven soft sensor modeling framework with variable time reconstruction. (March 2020)

• Record Type: Journal Article
• Authors: Yao, Le; Ge, Zhiqiang
• Is Part Of: Journal of process control. Volume 87(2020)
• Page Start: 91
• Page End: 107
• Publication Date: 2020-03
• Subjects: Data-driven soft sensor -- Variable time-delay -- Variable Time Reconstruction -- Variational Bayesian Regression model -- Integer Differential Evolution algorithm; Process control -- Periodicals; Fabrication -- Contrôle -- Périodiques; Process control; Periodicals; Electronic journals; 660.281
• DOI: 10.1016/j.jprocont.2020.01.009
• Languages: English
• ISSNs: 0959-1524

Abstract

Highlights: VTR-based modeling framework eliminates variable time-delay while building soft sensor. Variable Time-delay and model parameters are cooperatively estimated by IDE and VB algorithms. Three novel soft sensors are developed with reconstructed dataset under the VTR framework. Qualitative process knowledge turns to be quantitative through VTR-based modeling framework. Abstract: Due to the difference of variable positions brought by process structure, time-delay exists between process variables and quality variables. In this paper, this commonly overlooked problem in data-driven soft sensor modeling is illustrated and solved. The main idea in this paper is to take the variable time-delay (VTD) as a model parameter to reconstruct the dataset and then solve it through optimizing the objective function of models. However, the combination of VTD would lead to an intractable high computational complexity, then it is proposed to use an efficient population-based Integer Differential Evolution (IDE) algorithm to select the optimal VTD values and cooperatively learn model parameters. With the help of IDE algorithm, a Variable Time Reconstruction (VTR) modeling framework is then formulated for soft sensor development. As examples, three types of VTR-based soft sensors are developed under this framework to cope with different cases of data features. The presented numerical and industrial cases demonstrate that the proposed VTR-based model can effectively learn the VTD values, which can reconstruct and recover the original data pattern, and thus significantly help increase the generalization performance of soft sensor models.