Prediction of Quality Indicators of Iron Ore Processing Operations Using Deep Neural Networks

Taking into consideration the development of computing power and algorithms to process big data, machine learning technologies are more widely used to solve production tasks. In this paper, our aim is to develop a prediction system that allows to show the possible trend of the processing of iron ore concentrate in real time. Currently, the operator controls the quality of final product on the basis of laboratory analysis results. They are made in a discrete way every 2 hours. But such period of time is much longer than the time, which is needed by ore to get through the process line from start to finish. Therefore, setpoints values of the process state variables are chosen by the operator conservatively, having a negative scenario in mind. But that results in decrease of the energy efficiency of the technological process.Complexity of development of such system lies in the fact that the given stage of production is a discretely continuous distributed process. It includes many technological objects, the functioning of which cannot be approximated by linear dependencies. Also there is an unsteadiness of properties of incoming raw materials, which causes changes in characteristics of the working equipment. For approximation of complex dependencies, which are curves of technological indexes, the authors have used the multilayer artificial neural networks. It allowed to obtain high quality of modeling. However, it has been revealed that such method of modeling of functioning dynamics of objects leads to negative effect - occurrence of autocorrelation, negatively affecting the predictive properties of the model. In order to avoid such a phenomenon, we studied the application of deep neural networks. It allowed to obtain the prediction results, which meet the requirements of technology and without the autocorrelation effect. It leaded to increase of the stability of the models of technological units.