Soft Sensor Development Based on Large-Scale Pseudo Labeling Optimization

Data-driven soft sensor techniques enable online prediction of key quality variables. However, in real industry, the expensive cost of manual labeling leads to the abundance of unlabeled data and the scarcity of labeled data, which becomes a major obstacle to establishing high-precision data-driven soft sensing. To address this problem, a method based on large-scale pseudo-labeling optimization, LSPLO, is proposed. This method, by introducing confidence assessment, fixes the pseudo labels assessed as high-confidence during optimization, while the low-confidence pseudo labels are used as new decision variables after dimensionality reduction to continue optimization. The original decision variables are reinitialized using superior pseudo labels, and stepwise dimensionality reduction and optimization are performed to obtain the optimal pseudo labels. The superiority of the proposed method LSPLO is demonstrated by a real industrial case of the Chlortetracycline (CTC) fermentation process.