Concentrate grade prediction using improved Transformer in zinc froth flotation process

Early and accurate prediction of concentrate grade is important for flotation process. Existing methods are usually insufficient to extract the time series information of flotation system, which cannot guarantee the accuracy and validity of zinc concentrate grade prediction. Therefore, we propose an improved Transformer model. First, we extract feature vector sequences from the froth video as input features to obtain the dynamic time information of the input data. Meanwhile, the position encoding module and decoder module are improved according to the characteristics of the flotation data. Then, we use the improved Transformer model to learn the internal rules of the sequences and the long-range correlation between the sequences to improve the accuracy and efficiency of the prediction. The method is validated on real flotation data, and the results show that the improved Transformer method outperforms other algorithms in terms of accuracy, while ensuring real-time performance.