Design of adaptive soft sensor based on Bayesian optimization. (December 2022)

• Record Type: Journal Article
• Authors: Yamakage, Shuto; Kaneko, Hiromasa
• Is Part Of: Case studies in chemical and environmental engineering. Volume 6(2022)
• Publication Date: 2022-12
• Subjects: Soft sensor -- Adaptation mechanism -- Regression -- Hyperparameter -- Bayesian optimization; PLS partial least squares -- EN elastic net -- LASSO least absolute shrinkage and selection operator -- SVR support vector regression -- GPR Gaussian process regression -- LWPLS locally-weighted partial least squares -- MW moving window -- JIT just-in-time -- TD time difference -- BO Bayesian optimization -- DEB debutanizer column -- SRU sulfur recovery unit; Chemical engineering -- Periodicals; Environmental engineering -- Periodicals; Environmental chemistry -- Periodicals; 628.05
• DOI: 10.1016/j.cscee.2022.100237
• Languages: English
• ISSNs: 2666-0164

Abstract

Abstract: When adaptive soft sensors are introduced to industrial plants, an appropriate combination of the type of adaptation mechanism, hyperparameters of the mechanism, regression model, and hyperparameters of the model must be selected for predictive soft sensors. We propose an automatic and efficient selection method for adaptive soft sensors based on Bayesian optimization. A Gaussian process regression model was constructed between the candidates of adaptive soft sensors and their predictive ability to perform Bayesian optimization. The adaptive soft-sensor candidate with the maximum acquisition value function was selected. The effectiveness of the proposed method was confirmed by analyzing two real industrial datasets. Highlights: Adaptive soft sensors are important to maintain the predictive ability. Appropriate adaptive soft sensors differ depending on process and objective variables. We propose to apply Bayesian optimization for adaptation mechanism and modeling methods. The proposed method can select the appropriate adaptive soft sensor for each process dataset. The effectiveness of the proposed method was confirmed through case studies using real datasets.