Tensile Strength Prediction for Hot Rolled Steels by Bayesian Neural Network Model

Abstract A neural network model under the Bayesian framework (referred to as Bayesian neural network hereafter) is developed to predict the tensile strength of hot rolled products. The motivation of using a data-driven model is that in a modern steel mill, there exist huge online measurements and offline data, and can be exploited to extract the underlying relationships. The main advantage of using Bayesian neural network (BNN) is the robustness against over-fitting, thanks to the probabilistic reasoning behind the BNN. Preliminary modelling results are encouraging, and the properly trained BNN model can be used for online control and optimisation of the rolling mill to achieve desired mechanical properties.