Adaptive soft sensor based on online support vector regression and Bayesian ensemble learning for various states in chemical plants

• Our goal is to construct high predictive soft sensors for various process states.
• Multiple online support vector regression (OSVR) models are constructed.
• Prediction results of the OSVR models are combined using Bayes' rule.
• The prediction error is estimated from the multiple predicted values.
• The performance is confirmed with a simulated dataset and two industrial datasets.

A soft sensor predicts the values of some process variable y that is difficult to measure. To maintain the predictive ability of a soft sensor model, adaptation mechanisms are applied to soft sensors. However, even these adaptive soft sensors cannot predict the y-values of various process states in chemical plants, and it is difficult to ensure the predictive ability of such models on a long-term basis. Therefore, we propose a method that combines online support vector regression (OSVR) with an ensemble learning system to adapt to nonlinear and time-varying changes in process characteristics and various process states in a plant. Several OSVR models, each of which has an adaptation mechanism and is updated with new data, predict y-values. A final predicted y-value is calculated based on those predicted y-values and Bayes' rule. We analyze a numerical dataset and two real industrial datasets, and demonstrate the superiority of the proposed method.