Semi-Supervised Soft Sensor Modeling Based on Ensemble Learning With Pseudolabel Optimization

Key quality variables are critical in the industrial production process. The soft sensor technology, which predicts key quality variables by establishing mathematical models, has gradually become a research hotspot. However, due to the difficulty in obtaining labeled data in industrial fields, a substantial quantity of unlabeled data is not reasonably utilized, which challenges the reliability and accuracy of conventional soft sensor models. Therefore, a semi-supervised model based on the voting ensemble learning is proposed, which combines the outcomes of multiple models’ predictions and utilizes a genetic optimization algorithm to iteratively optimize the generated pseudolabels, improving the accuracy of pseudolabels. By using the channel attention mechanism and multiscale feature fusion method, the feature extraction ability of the model is further improved, thus enhancing the prediction accuracy of the model. Finally, experiments were carried out on industrial debutanizer and industrial steam volume datasets to validate the superior predictive performance of the proposed method.