A soft sensing method for mechanical properties of hot-rolled strips based on improved co-training

Accurately soft sensing of the mechanical properties of hot-rolled strips is essential to ensure product quality, optimize production, and reduce costs. However, it faces the difficulty caused by limited labeled samples, for which co-training based semi-supervised learning offers a potential solution. So in this paper, a novel soft sensing method for mechanical properties based on improved co-training (ICO) is proposed. Compared with the existing co-training framework, the proposed ICO introduces improvements from the aspects of multiple view partition, confidence estimation, and pseudo-label assignment. Specifically, (ⅰ) in the stage of multiple view partition, ICO integrates metallurgical mechanisms of hot rolling processes and statistical mutual information to achieve a balance between view sufficiency and independence, which improves model performance and interpretability; (ⅱ) in the stage of confidence estimation, ICO evaluates the confidence of unlabeled samples at the cluster level rather than at the level of a single sample, which facilitates the exploration of sample distribution and the selection of representative samples; (ⅲ) in the pseudo-label assignment stage, ICO adopts a safe pseudo-label algorithm (which is called SAFER by its author and originally used for each single sample) to assign pseudo-labels for cluster of samples with the highest confidence determined in the previous step stage, to take advantage of the merit of handling unlabeled samples at the cluster level mentioned above on one hand, and the merit of SAFER in enhancing the quality of pseudo-labels on the other hand. The proposed soft sensing method effectively predicts mechanical properties on the real hot rolling dataset, achieving approximately 5% improvement in $R^2$ compared to traditional supervised learning.