Development of a Self-Updating System for the Prediction of Steel Mechanical Properties in a Steel Company by Machine Learning Procedures

This study is focused on the implementation of statistical learning methods for the prediction of the mechanical properties of steel products from the chemical profile of the raw material and the process parameters. The integration of this model into the production process allows a large-scale steel industry to predict steel properties with heightened accuracy, optimizing the manufacturing process for minimal waste and improved consistency. A workflow for process data analysis has been developed, based on the use of machine learning algorithms to build an interface for data treatment to be directly used online. The proposed approach has a comprehensive connotation, starting from data pre-treatment and cleaning, to model building and prediction. Different machine learning algorithms are compared (Polynomial Regression, LASSO, Random Forests and Gradient Boosting, ANN, SVM, and k-NN), to provide the best predictive ability, also exploiting human reinforcement. The results proved to be very promising for all the types of steel investigated, with very good RMSE and R2 values both in fitting and in prediction. The application here presented is being integrated into Total Quality Tutor (TQT) software, developed in-house in C# language, for predicting the mechanical properties of steel.