Predictive modeling of strip width based on incremental learning and adaptive-weight fusion during the hot rolling process

Abstract

The strip width is a critical quality indicator in the hot rolling process. Accurately predicting the finishing width spread is a challenge due to the absence of direct control equipment and the coupling between multiple stands. Most existing forecasting models in the steel industry are offline models, making them unsuitable for managing dynamic changes in real-time production. To address this issue, a novel incremental learning-based adaptive-weight fusion framework is proposed for the online prediction of the finishing width spread. First, an improved mechanism model is developed to integrate the physical information into the fusion framework. To effectively extract insights from production data, a clustering ensemble feature selection method is developed to obtain the optimal feature subset. Subsequently, an incremental learning approach is introduced to construct periodic and real-time models, facilitating continuous online model updating. Finally, an adaptive-weight strategy is constructed to enable accurate and rapid online prediction of the finishing width spread. Experimental results demonstrate that the proposed framework can effectively synthesize mechanism information, historical information, and real-time information, outperforming other models in terms of online prediction accuracy across different datasets and meeting production speed requirements. Through the deployment of the model on a cloud-edge collaborative computing platform, the model demonstrated a strong generalization ability, with 98.49 % of the width prediction errors falling within ±5 mm.