Prediction Of Concentrate Grade In Bauxite Flotation Process Based On Multi-Cameras

Abstract:Concentrate grade is an important production target which measures the effect of bauxite flotation. Currently, online prediction of the concentrate grade based on machine vision mainly focuses on a single flotation cell or single flotation stage. It can only reflect the local froth state of flotation process but cannot fully characterize the entire flotation process, which leads to a poor accuracy of the prediction model. Therefore, based on the analysis of bauxite flotation process and froth image features, a concentrate grade prediction method based on multi-cameras of bauxite flotation process is proposed in this paper, which can provide operational guidance for the production process. The main research and innovation thesis are as follows:(1) The large correlation of froth image features between each flotation stage results in high model input dimension and a large amount of computation. Therefore, a rough set attribute reduction algorithm was proposed to reduce the froth image features. Based on the degree of general importance, the key features of froth images from different flotation stages were obtained. Considering the correlation between froth features and concentrate grade and the dynamic relationship among image features from different flotation stages, a multiple data correlation analysis method based on experimental analysis was proposed, which includes correlation analysis between concentrate grade and the key features of a single flotation cell, as well as correlation analysis between key features of different flotation cells. Experiments show that the froth image trends of different flotation stages are consistent, but the fluctuation of image features in each cell is obviously different.(2) Considering the poor accuracy of concentrate grade prediction using froth features from a single cell, an integrated concentrate grade prediction model based on multi-cameras was proposed. Key froth features of the rougher cell were used as input variables, and a least squares support vector machine (LS-SVM) prediction model of concentrate grade based on improved particle swarm optimization was established. Considering the model error caused by long process, key froth image features of the cleaner cell synchronized with rougher process were used as input, then the error compensation model based on relevance vector machine (RVM) was established. Finally, the integrated model was used to improve the grade prediction accuracy.(3) In this paper, the proposed method was simulated and comparatively analyzed with the existing grade prediction method based on a single cell. The results show that the proposed method has a higher accuracy. The grade prediction method based on multi-cameras was simulated in a domestic monitoring system of bauxite flotation. The verification results show that the proposed method can get a more satisfactory accuracy, which can provide operational guidance for flotation process. Figure34, table8, reference75.