Two-Stream Deep Feature-Based Froth Flotation Monitoring Using Visual Attention Clues

In froth flotation monitoring, machine-vision-based soft sensors provide stable and reliable online estimations for the concentrate grade, which is difficult to be measured online owing to technical or economic limitations. It is known that the froth surface appearance and movement characteristics of the froth layer are closely related to the froth grade and usually used as visual indicators for the concentrate grade. However, it is hard for the most used handcrafted features to fully explore patterns of froth surface behaviors from appearance and movement perspectives. Furthermore, the relevance between image features extracted from different time intervals and the concentrate grade can be different. To estimate the concentrate grade appropriately, soft sensors need to explore and exploit the different importance in image features extracted at different intervals. In the context of these issues, this study developed a deep learning-based two-stream feature extraction model to extract the froth appearance and movement features. Also, a hybrid prediction model, which contains a time-series analysis module and an attention mechanism, is proposed to build the prediction relationship between the image features and the concentrate grade. Comparison experiments using historical industry data verified the advantage of the proposed monitoring method. In addition, industrial experiments conducted in a realworld flotation plant show that the coefficient of determination achieved by this method is 0.9256, which has a 7.9% improvement compared to an existing expert system.

Automated summary

This study developed a deep learning-based two-stream feature extraction model to extract froth appearance and movement features, and proposed a hybrid prediction model that includes a time-series analysis module and an attention mechanism to establish the prediction relationship between image features and concentrate grade. Comparison experiments using historical industry data confirmed the superiority of the proposed monitoring method, with industrial experiments achieving a coefficient of determination of 0.9256, representing a 7.9% improvement over an existing expert system.