A Semi‐Supervised Transfer Approach Based on Pairwise Learning for Tensile Strength Prediction of Hot‐Rolled Cord Steel

The tensile strength prediction of cord steel is critically importance for ensuring product quality and mitigating failure risks; however, the scarcity of labeled data, particularly for customized steel grades, presents a significant challenge. To address this challenge, a multi‐source dataset is introduced, comprising labeled data from source steel grades and unlabeled data from target steel grades, to improve predictive accuracy. Accordingly, a Semi‐Supervised Transfer Learning approach based on Pairwise Deviation Regression Network, termed PDRN‐SSTL, is proposed. First, the Pairwise Deviation Regression Network reformulates regression as learning label deviations between sample pairs, thereby alleviating cross‐domain label shift and capturing relative sample relationships. Second, an unsupervised domain adaptation strategy is employed to construct a pre‐trained model that extracts common features while reducing distributional discrepancies between source and target domains. Subsequently, the shallow‐layer parameters are frozen, and the model is fine‐tuned using limited labeled target‐domain data together with a consistency constraint on unlabeled data, thereby further improving the predictive accuracy. Experiments on two customized steel grades, SWRH72A‐C and SWRH72A‐D, demonstrate the superiority of the proposed method, achieving MAPE values of 1.369% and 1.447%, and 3% Relative Error Hit Rate of 94.23% and 88.24%, respectively. Comparative and ablation studies further confirm its effectiveness.