HUANG Yan, LI Haozhi, CHENG Lan, REN Mifeng, YAN Gaowei
Control Engineering of China. 2026, 33(01): 40-48.
Slow change and multi-condition characteristics are common in the process industry. The slow feature analysis only considers the slow change information, and ignores the data distribution difference among different conditions, which leads to poor prediction accuracy for quality variables. To solve this problem, based on the slow feature analysis, combining the transfer learning strategy, and taking into account the interpretability of the slow features to the quality variables and the local geometric structure of the data, a soft sensor model of multi-condition slow feature regression with structure preservation is proposed. Firstly, the correlation between the slow features and the quality variables is maximized to enhance the interpretability of the slow features to the quality variables. Secondly, the domain adaptation strategy is used to reduce the data distribution difference between the historical conditions and the conditions to be predicted. Finally, the neighborhood preserving embedding is introduced to retain local information, and a multi-objective optimization function is designed to predict the quality variables by using the nonlinear iterative partial least squares framework. Three actual industrial datasets are used to test the proposed model in the experiment. The experimental results show that the proposed model can effectively improve the prediction accuracy of the quality variables.
