Abstract
In the bauxite flotation process, concentrate grade and tailings grade are key production indicators; however, they are difficult to measure online. It is also difficult to develop an effective mathematical model for the process because of the complex non-linear and uncertain relationship among the feed parameters (feed grade, pulp density, slurry particle size, etc.), froth features and production indicators. Therefore, an online hybrid modeling method is proposed by analyzing the multiple parameters that affect the production indicators. First, according to the correlation and redundancy in the feed and froth feature parameters, the kernel principle component analysis (KPCA) is used to reduce the number of the parameters. Then, a neutral network model of the regular extreme learning machine (RELM), which is based on wavelet function, is presented to predict these two indicators. To improve generalization capability and prediction accuracy, information entropy is used to distribute the weight of the two models based on their predicting error. At last, an on-line updating strategy of the hybrid model is constructed in order to investigate the influence of the working conditions. The proposed method is tested on the diasporic-bauxite flotation process and shows high predictive accuracy and generalization capability. It lays the foundation for optimal control of the operation parameters based on mineral grade in the flotation process.
Keywords:
froth flotation; image features; online predictive model; extreme learning machine (ELM)
