Research on edge defect prediction of hot-rolled strip steel based on LGBM-LR

Surface quality control of hot-rolled strip remains challenging due to complex defect formation mechanisms and imbalanced datasets from industrial processes. To address this, we propose a predictive framework integrating Light Gradient Boosting Machine (LGBM) and Logistic Regression (LR) to identify defect-related features and enhance defect prediction. The model focuses on edge defects, leveraging historical manufacturing data for early-stage quality guidance. High-dimensional process data are analyzed using LGBM to rank defect-critical features, followed by LR-based interpretation to evaluate high/low-score variables. To address severe sample imbalance caused by scarce defective samples compared to normal samples, the framework integrates cross-stratified undersampling, BorderlineSMOTE oversampling, and adaptive LR threshold optimization. Comparative experiments are conducted using LGBM and hybrid LGBM-LR models with varied sampling techniques and decision thresholds. Results demonstrate that both models achieve 87% accuracy in defect anomaly prediction. The proposed BorderlineSMOTE and adaptive thresholding significantly improve recall rates from 20–60%, effectively addressing class imbalance. Furthermore, 13 of the top 20 features identified in the edge defect importance ranking align with empirical domain knowledge, validating the model's interpretability. This feature-ranking outcome offers actionable insights for prioritizing process parameters in surface quality control. The study highlights the efficacy of combining ensemble learning, resampling techniques, and threshold adjustment to enhance defect prediction in imbalanced industrial datasets, providing a data-driven reference for optimizing hot-rolled strip production.