Prediction of the Iron Ore Pellet Abrasion Index using Random Forest

Abstract

The current industry depends on technologies for the acquisition, communication and processing of massive volumes of data. Sometimes, however, the expertise needed to extract intelligence from this data is not available. In the steel sector, vast resources are invested in instrumentation, but this does not always translate into results. This work describes a machine learning application for predicting abrasion rates in iron ore pellets. The entire pelletizing process is instrumentalized, generating a large volume of data. The working hypothesis is that it would be possible to create a predictive model for the abrasion index. Through a thorough analysis of the raw data sets, the most promising predictor variables were selected and a predictive model was generated from them. This model was successful in estimating the pellet abrasion rates from the production process variables, allowing to act on these variables and improve the abrasion rates and the overall quality of the final product. The gains from using machine learning techniques have proven to be beneficial and led to insights that even experts had not previously identified.