Timed key-value memory network for flotation reagent control

Abstract

Computer vision-based flotation reagent control (FRC) is a nonintrusive, cost-effective, and reliable technique for changing a deviated flotation status to an optimal status, which produces a product at the target concentrate grade and recovery. It is known that deep froth image features can depict the complex behavior of the froth layer comprehensively and accurately. However, it is still a major challenge to use high-dimensional deep features to construct FRC controllers. In this study, a novel FRC method based on memory networks is proposed. First, a convolutional neural network-based image encoder is constructed to extract deep froth image features. Then, a key–value structured memory network is designed to learn heuristic rules relating deep froth image features to the corresponding corrective operation actions for the reagents. The proposed method can be incrementally updated without catastrophic forgetting and can explore time-series information and changing patterns of froth surface appearance. Additionally, it can handle input images whose size is variable. Experiments using real-world production data verified the effectiveness of the proposed FRC method. In addition, industrial experiments in a real lead–zinc flotation plant in China demonstrated that the new method could obtain reliable flotation process control.