Assessing Machine learning-based approaches for Silica concentration estimation in Iron Froth flotation

In the mining industry, specifically in the flotation process, there is a challenge associated to noninvasive, real-time contaminant and impurities estimation. Achieving predictions on contaminant levels has a high impact on quality insurance and it can help technicians and engineers to make adjustments in advance to improve the quality of the final product, and thus profits. In this paper, exploratory research is performed on the problem of silica concentrate estimation for iron ore froth flotation using machine learning techniques, with the goal to identify algorithms that may be suitable for industry soft sensor development. For this purpose, a public, real process dataset is used and three different machine learning techniques are implemented: Random Forest (RF), Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU). The techniques were implemented, tested and compared in terms of their error percentage, mean absolute error, mean square error, and root mean square error. Obtained results show acceptable performance for LTSM and GRU implementations, being LSTM network the out-performer with errors below 9%.