Adaptive weighted relevant sample selection of just-in-time learning soft sensor for chemical processes

Abstract

A new just-in-time learning (JITL) method using adaptive relevant sample selection strategy is proposed for online prediction of product quality in chemical processes. To overcome certain shortcomings in traditional JITL, such as the incomplete usage of primary variable information and inaccurate feature weights assignment, an adaptive sample selection approach is introduced. First, to keep both input and output information, a dual-objective optimization scheme with an adaptive parameter is considered. Then, an adaptive feature weight assignment strategy is present to construct a suitable similarity criterion for JITL. To illustrate the performance of the proposed method, it is applied to online prediction of the crude oil endpoint in an industrial fluidized catalytic cracking unit. The experimental results demonstrate that the proposed method can help improve the prediction performance.