Grade prediction of zinc tailings using an encoder-decoder model in froth flotation

Abstract Accurate grade prediction is conducive to proper flotation operation or control. Different from grade monitoring, grade prediction needs to obtain the target grade in advance. However, there is usually a time delay between the flotation cell and the predicted grade in froth flotation. This time delay makes it difficult to match the features of the observed flotation cell with the predicted grade. To solve this problem, this article studies the method of zinc tailings grade prediction using encoder-decoder models. The proposed model considers the feature time series of the first rougher and the previously measured tailings grades. First, according to the sample ratio between froth video and X-ray fluorescence (XRF) analyser, the feature time series of the first rougher can be automatically extracted by finding the nearest available feature vectors. Next, the feature time series of the first rougher is fed into the encoder to generate a context vector, and then the context vector and previously measured grades are sent into the decoder to predict the current tailings grade. The proposed model effectively captures the dynamic consistency between the feature time series and previously measured grades. The effectiveness of the proposed model in the froth flotation has been verified by experiments. Compared with the traditional recurrent neural network (RNN)-based models, the root mean squared error (RMSE) and mean absolute percentage error (MAPE) of the proposed model decrease by about 17.8% and 1.9%. respectively, and the R-squared (R2) score of the proposed model increases by about 13.8%.