A temperature curve generation method for predicting the mechanical properties of hot-rolled strip steel

In response to the problems of sparse temperature measurement point distribution and poor cross-interval continuity in the hot continuous rolling process, a full-process temperature field prediction and reconstruction method based on the combination of a physics-informed neural network (PINN) and a Bayesian-XGBoost surrogate model is proposed in this paper. First, a PINN model across five process areas (from the reheating furnace to the coiler) is established, taking time nodes as input and directly outputting discrete points of the temperature-time curve along the strip in each process area. Subsequently, a Bayesian-XGBoost surrogate model is used; at this time, hierarchical surrogate decision-making determines the number of prediction points automatically by the maximum error principle and thereby improves the prediction accuracy. Finally, a piecewise cubic spline interpolation algorithm is designed, based on curvature detection to achieve a precise high-precision temperature curve reconstruction of discrete prediction results. The experimental results show that this method has high accuracy in temperature curve reconstruction and good reliability in mechanical property prediction; it verifies the effect and practicability of the proposed framework in real hot continuous rolling processes.