Concentrate grade prediction of industrial zinc flotation process based on Cross-Temporal Feature Fusion Transformer

Abstract

Flotation industrial process data usually have temporal characteristics and feature nonlinearities. Aiming at the problem that the existing Transformer-based prediction model only considers the temporal information of time series data and ignores the importance of different feature variables, a Cross-Temporal Feature Fusion Transformer (CTFF-Transformer) is proposed for the prediction of concentrate grade of industrial zinc flotation process. The feature multivariate correlation and temporal dependence of the industrial data are captured by the feature attention module and the temporal attention module, respectively, and post-fusion is performed to enhance the model prediction performance. Due to the unsynchronized sampling time of froth video data and concentrate grade data in the flotation process, a fusion feature vector extraction strategy based on the froth video temporal segmentation is proposed, which improves the characterization ability of the data by constructing multi-segment froth video feature vectors and fusing the related grades. The proposed method is validated by using zinc rougher flotation froth video data, and comparative experiments show the merits in predicting the concentrate grade.