Dynamic recrystallization during strip rolling of HSLA steels and prediction of roll forces using artificial neural networks

Abstract

The softening mechanisms taking place during hot rolling of steels influence both the mechanical properties of the final product and the steel flow stress during deformation. The knowledge of the material's constitutive behavior is an essential requirement for the design and control of rolling processes. Steel manufacturers are looking for more accurate models being able to predict the material deformation resistance, micro-structural evolution of steel, and roll forces in order to produce strips with a more consistent output gauge and mechanical properties. In this regard; the occurrence of dynamic recrystallization during strip rolling of HSLA steels and its effects on the flow stress, roll forces, and final properties are of importance. The occurrence of dynamic recrystallization during hot strip rolling still remains controversial.

In this research, the experimental techniques were used to simulate the whole rolling process to study the occurrence of dynamic recrystallization. Axisymmetric compression tests were used to study the kinetics of static recrystallization. Torsion simulations were performed to verify the occurrence of dynamic recrystallization. An industrial mill log was analyzed which further confirmed the occurrence of the dynamic recrystallization and torsion test results.

In spite of drawbacks in terms of ease of development, adaptability, accuracy and speed, empirical stress-strain relationships and traditional roll force equations, along with look-up tables, are being commonly used. In this research, a Neural Network simulator code, based on the gradient descent learning rule, was developed. This code was used to predict the steel and aluminum flow stresses at high temperatures and strain rates, experimental rolling forces during cold and hot rolling of aluminum strips, and rolling forces during industrial strip rolling of a high Nb HSLA steel. The model predictions were compared to those of the statistical models and existing on-line industrial models. The approach based on Neural Networks is shown to be superior in terms of accuracy, speed, and ease of development.

Principal Component Analysis was used as a data pre-processor to remedy data deficiencies when there is an excessive linear correlation between input variables of a database. This analysis was integrated to the Neural Network simulator code in order to decouple linearly correlated input data. The code was applied to an industrial hot rolling database to develop a model to predict the occurrence and the effects of dynamic recrystallization on the rolling forces. This model not only predicts the occurrence of dynamic recrystallization, but also predicts the extent of the resulting softening. The results also proved the beneficial effects of the integration of Principal Component Analysis with Neural Network modelling.