Long short-term memory-based grade monitoring in froth flotation using a froth video sequence. (1st January 2021)

• Record Type: Journal Article
• Authors: Zhang, Hu; Tang, Zhaohui; Xie, Yongfang; Gao, Xiaoliang; Chen, Qing; Gui, Weihua
• Is Part Of: Minerals engineering. Volume 160(2021)
• Publication Date: 2021-01-01
• Subjects: Grade monitoring -- Machine vision -- Froth video -- Froth video sequence -- Long short-term memory; Mines and mineral resources -- Periodicals; Ressources minérales -- Périodiques; Mines and mineral resources; Periodicals; Electronic journals; 622
• DOI: 10.1016/j.mineng.2020.106677
• Languages: English
• ISSNs: 0892-6875

Abstract

Highlights: A LSTM-based network estimates the tailings grade from Zn flotation circuit. The froth video sequence improves the grade monitoring accuracy. The rougher feed grades are integrated into the features by up sampling. The video sequence considers varied sample rates between vision and XRF systems. The video sequence uses the unlabeled froth videos effectively. Abstract: With the rapid development of machine vision technology, machine vision systems are being widely used in flotation plants for online grade monitoring. In an industrial flotation plant, the sample rate of a froth video is considerably higher than that of the labelled grade data obtained using special detection devices; this results in several unlabelled froth videos. However, these unlabelled videos are valuable and should not be discarded, as they can provide temporal information regarding the video sequence, which can reflect the variation trend of the grade. Therefore, a long short-term memory (LSTM)-based network is proposed herein to estimate the tailing grade of the first rougher from a zinc flotation circuit. First, the froth video representation is obtained by calculating the average of the visual features and integrating the rougher feed grades. Then, the representations of the froth video sequence are introduced to the LSTM network to match the labelled grade. Through network training with the historical data, a grade monitoring model can be constructed, in which the froth video information is utilised adequately, and the problem of different sample rates is also solved. Experimental results show the effectiveness of the proposed grade monitoring model. Compared with those of the traditional neural network without unlabelled froth videos, the root mean squared error of the proposed model decreases by 8.48% and the R-squared score of the proposed model increases by 9.32%.