Transductive transfer broad learning for cross-domain information exploration and multigrade soft sensor application

Abstract

Without sufficient labeled data, the construction of accurate soft-sensor models for multigrade chemical processes is challenging. To alleviate the dilemma, a transductive transfer broad learning (TTBL) model for cross-domain information exploration is proposed. The features are extracted by the nodes of broad learning system. Then, similar sample information from the current and related domains is captured by the k nearest-neighbor graph and retained by the manifold regularization framework. By exploring the available cross-domain information, unlabeled data in the prediction domain can be utilized for modeling. Finally, a TTBL model is constructed assisted by the fast leave-one-out cross-validation strategy. TTBL can effectively exploit the useful information of cross-domain data to improve prediction performance. Experimental results on two multigrade chemical processes demonstrate its superiorities compared with several traditional methods.