Intelligent diagnosis for hot-rolled strip crown with unbalanced data using a hybrid multi-stage ensemble model

To improve the smart manufacturing capabilities of strip hot rolling, based on digital twin (DT) and cyber-physical system (CPS), this paper proposes a data-driven approach for diagnosing hot-rolled strip crown. Since the hot rolling process features heredity, nonlinearity and strong coupling, the diagnosis of strip crown is an imbalanced problem with ill-defined decision boundaries. Conventional regression methods tend to learn more information from the majority class, which ignore the strip with unqualified crown. To address this challenge, a hybrid multi-stage ensemble model (HMSEN) is presented to classify strip crown. Initially, a novel data-resampling method that combines adaptive synthetic sampling (ADASYN) with repeated edited nearest neighbor (RENN) is proposed to assign more attention to unqualified crown. Subsequently, using the reinforced data, a multi-stage ensemble model is built to enhance the classification performance. Furthermore, the best-performing HMSEN is identified by exploring various combinations of base classifiers. The experimental results demonstrated the proposed novel resampling method outperforms comparison methods on crown dataset. Significantly, the proposed HMSEN outperforms not only the existing regression models but also the mechanism model. Therefore, HMSEN is the most robust and effective model to intelligently diagnose hot-rolled strip crown with unbalanced data.