Soft-sensor design via task transferred just-in-time-learning coupled transductive moving window learner. (May 2021)

• Record Type: Journal Article
• Authors: Alakent, Burak
• Is Part Of: Journal of process control. Volume 101(2021)
• Page Start: 52
• Page End: 67
• Publication Date: 2021-05
• Subjects: Adaptive learning -- Concept drift -- Data-based modeling -- Lasso -- Transfer learning; Process control -- Periodicals; Fabrication -- Contrôle -- Périodiques; Process control; Periodicals; Electronic journals; 660.281
• DOI: 10.1016/j.jprocont.2021.03.006
• Languages: English
• ISSNs: 0959-1524

Abstract

Abstract: Data based approaches have recently gained extensive attention in modern process industries. Accordingly, data based soft sensing technology used for making online predictions of quality variables plays currently an important role in process control and monitoring. Drifts in operating conditions and process characteristics, however, demand novel online learning methods to be developed for maintaining predictive accuracy of soft sensors. While moving window (MW) and just-in-time-learning (JITL) are the most frequently used online learning methods in soft sensor design, they are, indeed, effective against different types of drifts. In the current study, in order to exploit the merits of both methods within a perspective offered by the recent transfer learning paradigm, we propose combining a task transferred JITL model with a MW learner in a transductive learning setting (JITL TT -MW tr ). The proposed method, tuned via a historical training set, is easy to implement and robust due to cooperative/complementary actions of JITL and MW methods. High prediction accuracy of JITL TT -MW tr is demonstrated on one semi-synthetic and five publicly available real datasets, indicating its efficiency and potential for industrial implementation. Highlights: MW and JITL models combined in a single frame against heterogeneous concept drifts. Inductive/transductive learning methods used to adapt different tasks/domains. High predictive accuracy obtained in five different real industrial datasets. Easy to use, robust, and potential for improvement due to it modular structure.