Grade Prediction of Froth Flotation Based on Multistep Fusion Transformer Model
Cheng Peng, Yikun Liu, Yuyao Ouyang, Zhaohui Tang, Luo Liang, Weihua Gui
Abstract
Accurate and timely foam grade prediction plays an important role in the flotation foam industry process. However, the information between foam characteristic series and foam grade series at different sampling times often does not match, making the prediction result lagging behind. A multistep fusion transformer (MSFT) model is designed in this article. First, we extract multiple froth time series as input to correlate feature information and grade information under multiple time series, then, a self-attention structure is designed to fuse at multiple scales, which enhances the degree of information correlation under different time series, finally, the information matrix is passed through the fully connected layer to obtain the final prediction result. Compared with the existing froth grade network recurrent neural networks (RNN), long short-term memory (LSTM), gated recurrent unit, Transformer, Enc–Dec (RNN), feature reconstruction–regression, Siamese time series and difference (LSTM), and FlotationNet models, the MSFT model has reduced the baseline by 30.3%, 30.3%, 30%, 66.9%, 30%, 45.8%, 55.2%, and 52.5%, respectively, among all indicators.