Uso de machine learning na mineração: revisão de literatura e aplicação do algoritmo random forest para otimização da recuperação mássica durante o beneficiamento de ferro

Abstract

Mining is undergoing important changes made possible by the exponential technological development of the last decades. New environmental challenges and regulatory changes encourage the search for tools that enable more efficient and sustainable mining. In this context, the evolution of statistical and computational methods of artificial intelligence (AI) and machine learning (ML) are important tools that allow the massive analysis of high-resolution and large-scale data, allowing process improvement and increased productivity. An underexplored application is the use of AI to reduce losses during the beneficiation process. In iron mining, for example, it is estimated that losses in Brazilian plants can reach 20% and in Brazil 20 to 40% of the weight of the total ore mined is destined for tailings dams. Several factors are related to a lower-than-expected mass recovery, from variables associated with processing at the plant to those associated with the mineralogical composition of the site, however, establishing the contribution of each of these variables remains a challenge. Thus, the objective of this work is to review the literature on the concepts and applications of ML methods in mining and use data from a public bank to train and test a supervised ML algorithm in the discovery of possible factors involved in mass losses and recovery of iron. Iron was chosen due to its importance for Brazilian mining, especially in the state of Minas Gerais, and due to the availability of mineral databases containing this element. The Random Forest (RF) algorithm was selected and the analyzes were performed using codes implemented in Python language. The trained model had an overall accuracy of 74% in predicting variables associated with high or low iron recovery from 13 prediction variables. This approach proved to be a fast, cost-effective, and efficient which can provide several important information in the elaboration of hypotheses related to iron mining. Adjustments to the model can provide the highest expected accuracy of the RF, and the adjustment of high and low mass recovery parameters according to field or industry demands allow the application and adjustments of the model generated here in multiple research and industrial contexts.