Multi-source unsupervised soft sensor based on joint distribution alignment and mapping structure preservation. (January 2022)

• Record Type: Journal Article
• Authors: Zhang, Zheming; Yan, Gaowei; Qiao, Tiezhu; Fang, Yaling; Pang, Yusong
• Is Part Of: Journal of process control. Volume 109(2022)
• Page Start: 44
• Page End: 59
• Publication Date: 2022-01
• Subjects: Soft sensor -- Hypergraph -- Multi-view clustering -- Dynamic distribution alignment; Process control -- Periodicals; Fabrication -- Contrôle -- Périodiques; Process control; Periodicals; Electronic journals; 660.281
• DOI: 10.1016/j.jprocont.2021.11.009
• Languages: English
• ISSNs: 0959-1524

Abstract

Abstract: Aiming at the problem of mismatch between real-time data distribution and modeling data distribution caused by the change of working conditions in industrial process, which leads to the performance deterioration of the soft sensor model, a multi-source unsupervised soft sensor method based on joint distribution alignment and mapping structure preservation is proposed. Firstly, the method uses the hypergraph to establish the complex structure of feature and label, and clusters the hypergraph matrix in multiple views to completely construct the class pseudo label; then dynamic distribution alignment is used to adapt marginal distribution and conditional distribution between the data of historical working conditions and the current working conditions, and the hypergraph Laplacian operator is introduced for manifold regularization to prevent the mapping relationship between feature and label from being destroyed; finally, similar working conditions are introduced to further enhance the robustness of the model. The experimental results show that compared with the traditional unsupervised soft sensor methods, the method used in this paper can effectively improve the prediction accuracy of the model. Highlights: DASP constructs pseudo label to dynamically distribute adapt the continuous regression data. The new label can retain the mapping relationship between the original data feature and label. Several experimental results verify the effectiveness of the method.