An interpretable framework for high-precision flatness prediction in strip cold rolling

Abstract

Flatness plays a crucial role in determining the quality of products in strip cold rolling. Data driven methods have shown promise in flatness prediction by effectively capturing the nonlinearities and strong coupling present in cold rolling processing, surpassing the capability of conventional methods. However, existing data driven models remain restricted by a lack of rolling theory guidance, a black-box nature of predictive processes, and gradient conflict of multi-channel flatness. To overcome these limitations, this paper proposes an interpretable mechanism guided multi-channel distributed meta learning framework for flatness prediction. Initially, significant physic-based parameters, such as theoretical rolling force deviation and tension deviation, and controller parameters are calculated to guide data driven modeling. Subsequently, a distributed meta learning framework is modeled for multi-channel flatness to eliminate gradient conflict. Furthermore, eXplainable Artificial Intelligence (XAI) technique is implemented to ensure the transparent predictive processes of multi-channel flatness. The analysis results present that theoretical parameters and controller parameters effectively improve the performance of flatness prediction. In addition, the comparative results demonstrate that the proposed framework outperforms the existing flatness prediction methods and other state-of-the-art machine learning methods by 4.24%. Importantly, the XAI-based explanation of the proposed framework effectively enhances the credibility of data driven flatness prediction.