Concentrate Production Output Prediction In The Mineral Process Based On Online Learning Decorrelated Neural Network Ensembles With Random Weights

Beneficiation process is the process that makes useful minerals in the raw ore mined enrich by physical and chemical changes.The overall concentrate output is significant index in the hematite beneficiation process.How to set and adjust operational indices under the concentrate output is the key to improve the beneficiation production efficiency.Therefore,the relationship model between the concentrate output and operational indices need to be established.Use the set indices as the input of the model,then predict the corresponding concentrate output to adjust the operational indices.In addition,due to the varied working conditions in beneficiation process,we need the new indices train the model to make it adapt to the changes.Thus,this paper is focused on finding a way to model concentrate output prediction model based on online learning algorithm.This paper is supported by the National Natural Science Foundation project "closed-loop optimal decision-making approach of technology index for complex industrial processes under dynamic environment".The method "online learning decorrelated neural network ensembles with random weights" is proposed and applied into concentrate output prediction.The detailed work has been summarized as follows:1)The problem description of concentrate output prediction is presented.In this description,it gives the introduction of beneficiation process,gives the production indexes classification including technic index(the magnetic tube recovery rate,particle size of high-and low-intensity grinding unit,concentrate and tailings grade of high-and low-intensity magnetic separation unit),the nature of raw ore and working condition(such as raw ore grade,capacity per hour and run time of high-and low-intensity grinding unit,and grade of waste ore,etc.)and the overall production index(the overall concentrate grade and output),gives the analysis of the relationship between the production indexes,and gives the difficulty and necessity of the prediction of concentrate output.2)Online learning decorrelated neural network ensembles with random weights is proposed.In this part,the paper shows that how online learning algorithm works and how to choose the individual networks.To test the performance of the algorithm we proposed,some standard test functions were used.Generate four datasets based on the variable domain of each test function and show the results of the comparison of the algorithm we propose and the latest ensemble online learning algorithm for regression on each dataset.The results show that the algorithm in this paper has stronger ability to learn the new data.3)The way to model concentrate output prediction model is proposed based on "online learning decorrelated neural network ensembles with random weights".In this part,optimize select the number of individual networks and the input weighting matrix in random pool and give the best penalty coefficient of ensemble model and suitable number of hidden layer nodes.Make the comparison between our method and current methods.The results shows the concentrate output forecast precision of the algorithm in this paper is higher.