A semi‐supervised JITL paradigm based on manifold regularization for online soft sensor development

Abstract

AbstractJust‐in‐time learning (JITL) is often used for soft measurements of nonlinear time‐varying processes due to its desirable properties. Nevertheless, most existing JITL methods utilize only limited labeled data, ignoring abundant information of plentiful unlabeled data, which prevents the JITL‐based soft sensors from achieving the best performance. To address this issue, this paper proposes a semi‐supervised JITL paradigm based on manifold regularization, denoted as MRSsJITL. The MRSsJITL could effectively use the information of unlabeled data to make up for the deficiency of supervised JITL. Then, a weighted spatial‐temporal similarity metric is developed to calculate the relevance between samples more accurately. The performance of the MRSsJITL is assessed using both numerical example and actual industrial process. The results demonstrate that, compared with the conventional JITL approaches, the MRSsJITL can efficiently improve the predictive accuracy.