Development of an adaptive soft sensor method considering prediction confidence of models

Abstract

Soft sensors are widely used to realize highly efficient operation in chemical process because every important variable such as product quality is not measured online. By using soft sensors, such a difficult-to-measure variable y can be estimated with other process variables which are measured online. For estimating y without degradation of a soft sensor model, a time difference (TD) model was developed previously. Though a TD model has high predictive ability, it does not function well when a process is operated under conditions that have never been observed. In order to cope with this problem, a soft sensor model can be updated with newest data. But, updating a model needs time and effort for plant operators. We therefore developed an online monitoring system to judge whether a TD model can predict y accurately or an updating model should be used for both reducing maintenance cost and improving predictive accuracy of soft sensors. The monitoring system is based on a support vector machine or standard deviation of y-values estimated from various intervals of time difference. We confirmed that the proposed system has functioned successfully in a distillation column with real industrial data.