A physics guided data-driven prediction method for dynamic and static feature fusion modeling of rolling force in steel strip production

The accuracy of rolling force prediction is key to improving the precision of strip thickness control. The compressive load required for the strip in the rolling process is not only related to the size of the billet and process parameters such as deformation speed, temperature, and reduction, but also to the deformation boundary conditions of the billet between the rolls, such as the wear state of the rolls, lubrication conditions, etc. These influencing factors are interrelated and constantly changing, which is particularly prominent in small-batch and multi-specification intermittent production modes. The existing rolling force prediction models are constructed based on the rolling deformation mechanism through numerous simplifications. Due to challenges in fully and accurately characterizing various complex rolling deformation processes, their mapping relationships with process parameters, and constantly changing boundary conditions, the accuracy of the simplified rolling force prediction model is difficult to meet the control requirements of actual production. This paper proposes a physics-guided data-driven (PGDD) rolling force modeling method. It separates rolling condition features into static and dynamic parts using mechanistic and empirical knowledge and introduces a machine learning framework that integrates these parts for modeling. In this framework, the static feature fitting part can establish the influence of process parameters such as billet chemical composition, size, rolling speed, temperature, etc. on the rolling force. Meanwhile, the dynamic feature fitting part is responsible for the collaborative modeling of influencing factors reflecting the evolution rules of roll state, learning the cumulative effects of various complex processing states from a large amount of time-series data formed by different combinations of rolling conditions. Experiments with real production condition data show that the proposed physics-guided data-driven modeling method can accurately predict the rolling force under complex and variable conditions, and its adaptability and accuracy are superior to the online original model and traditional data-driven model.