Abstract
Accurate grade monitoring is an important task in froth flotation. Current grade monitoring methods focus on the visual froth features from an individual flotation cell. As the froth flotation is a long process industry, the multisource features from different flotation cells can further improve the accuracy of grade monitoring. However, the multisource features are from different measurement sources and locations, and they have varied dynamic characteristics for grade monitoring. Therefore, we propose a grouped time series network (GTSN) with multisource features for grade monitoring of zinc tailings in this article. First, we group the multisource features by their additional measurement properties with self-clustering algorithm. Next, we send the grouped features into different time series subnetworks with varied dynamic characteristics and extract the feature vector for each grouped feature. After that, we integrate these feature vectors to represent the tailings grade. Especially, to minimize the grade prediction errors by each individual grouped feature and the integrated feature, we consider multitask learning in the model training. Finally, we implement a case study in a real zinc froth flotation process. Compared with the recurrent neural network (RNN)-based models using the multisource features, the averages of mean absolute percentage error (MAPE) and root-mean-squared error (RMSE) of the proposed GTSN-based models decrease by about 2.93% and 6.21%, respectively, and the average of R-squared (R2) increases by about 4.73%.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have