Convolutional memory network-based flotation performance monitoring

Abstract

In flotation process monitoring, visual soft sensors provide stable and reliable online estimations for a concentrate grade, which is difficult to be measured online owing to technical or economic limitations. It is known that hand-crafted features are at times inefficient in extracting features consistent with the experiences of operators. Furthermore, the froth layer always contains some undesired impurities. To grade the product appropriately, soft sensors need to be capable of differentiating patterns produced by different substances. In the context of these issues, this study developed a convolutional memory network-based visual soft-sensor, in which GoogLeNet, trained in a Siamese paradigm, is used to learn the representations for froth images and a semantic key-value memory network is used to recall similar historical records, helping to achieve accurate grade estimation. Simulations using real-world production data verified the effectiveness of the proposed monitoring method. Further, industrial experiments conducted in a lead-zinc flotation plant in China corroborated the fact that our method can provide reliable concentrate grade estimation.