Just-in-time semi-supervised soft sensor for quality prediction in industrial rubber mixers

Abstract

Increasing data-driven soft sensors have been adopted to online predict the quality indices in polymerization processes to improve the availability of measurements and efficiency. However, in industrial rubber mixing processes, most existing soft sensors for online prediction of the Mooney viscosity only utilized the limited labeled data. By exploring the unlabeled data, a novel soft sensor, namely just-in-time semi-supervised extreme learning machine (JSELM), is proposed to online predict the Mooney viscosity with multiple recipes. It integrates the just-in-time learning, extreme learning machine (ELM), and the graph Laplacian regularization into a unified online modeling framework. When a test sample is inquired online, the useful information in both of similar labeled and unlabeled data is absorbed into its prediction model. Unlike traditional just-in-time learning models only utilizing labeled data (e.g., just-in-time ELM and just-in-time support vector regression), the prediction performance of JSELM can be enhanced by taking advantage of the information in lots of unlabeled data. Moreover, an efficient model selection strategy is formulated for online construction of the JSELM prediction model. Compared with traditional soft sensor methods, the superiority of JSELM is validated via the Mooney viscosity prediction in an industrial rubber mixer.