Physical information-guided ensemble learning model for surface roughness prediction of galvannealed strip

Surface roughness is a critical quality indicator for galvannealed (GA) strip. Despite its importance, a comprehensive predictive model has not yet been established. This study leverages industrial data from a continuous galvannealing line, focusing on the hot-dip and temper rolling stages. Initially, various correlation analysis methods identified critical features of the hot-dip stage. These features were integrated with the theoretical model of the temper rolling stage, resulting in a “hot-dip to temper rolling” (HT) model. Based on features derived from the HT model, an ensemble learning (EL) model was constructed using the stacking method. Specific algorithms were implemented at each layer to ensure both performance and generalization capability. To further integrate domain-specific physical knowledge, this study introduces a physical information-guided loss function, leading to the development of a Physical Information-Guided Ensemble Learning (PGEL) model. The optimal data split ratio of 8:2 was determined through rigorous analysis. The EL model's performance was benchmarked against individual base learners and other efficient models, focusing on predicting the surface roughness of GA strips' upper and lower surfaces. Results indicate that the HT model demonstrates robust performance and high interpretability, serving as a guidance model for on-site production processes. Meanwhile, the EL model outperforms all other models, showcasing significant improvements across multiple evaluation metrics. The PGEL model, while maintaining high prediction accuracy, enhances the physical reasonableness and robustness of the model through physical information guidance, thereby demonstrating its unique advantages in surface roughness prediction.