The flotation process emerged to improve the sorting in fine particles, needed in areas such as mining.This process consists of the capture and flotation of hydrophobic particles by air bubbles, separatingthese particles from the hydrophilic ones, which are decanted. In order to allow the separation ofminerals that naturally do not exhibit differences in its hydrophobic properties, reagents are addedto induce the desired characteristics. Given that the process' primary objectives are concentrate gradeand recovery, and the measurement of the former is realized through laboratory analysis, in this workone proposes the development of a soft sensor to infer the recovery in real-time in order to improvethe response time in setpoints and reagents dosages to process fluctuation. Soft sensors are inferentialmodels that are fed with measured variables to estimate process variables that either is hard tomeasure directly or require very specialized equipment and technical staff. A month's worth of aflotation line data was prepared and then the variables were classified with the Regressional ReliefF(RReliefF) algorithm to determine the most relevant ones, which were used in the modeling. AnArtificial Neural Network (ANN) was used to create the model and establish a correlation betweenthe data available in the process, inputs to the ANN, and the laboratory results, output of the ANN.The ANN created was a Multi-Layer Perceptron (MLP) model with an error backpropagationtraining algorithm. The resulting model exceeded the desired accuracy, obtaining a preliminarylinear correlation coefficient above 0.9, thus allowing for a possible proposition of on-site trial. Theproposed soft sensor demonstrates robustness and consistency in the predictions, providing a wayto decrease response time in the process which, in turn, improves its performance and reduces wasteof reagents.
展开▼
