Appropriate Softsensor Models Selection from Process Monitor Information

Abstract

Adaptive models are the techniques of updating the statistical models in a soft sensor, which enable a soft sensor to follow time changes of the relationships between the process data and improve its estimation accuracy. Conventionally, researches about the soft sensor design with adaptive models are developed in the situation that one soft sensor has one adaptive model, though it is, to be considered, difficult to maintain the estimation accuracy in all phases. Then, the previous researches developed the methodical technique to select adaptive models by applying the ensemble learning method to an adaptive model of one side and judging whether to use it or not. In that technique, however, it is a problem that an adaptive model not applied the ensemble learning method could be selected improperly in a scene with inferior estimation accuracy than the other one. Therefore, in this research, the selection technique that evaluates all adaptive models to use based on process state indexes is developed. Through a numerical simulation dataset analysis, the proposed method was verified.