Modelling of Complex Technological Processes and Material Processing Systems by Use of Neural Networks

"The present work is an attempt of neural networks application to the description of complex technological processes of useful minerals enrichment. Neural networks are among the most convenient tools that have been added recently to the professional toolbox of modelling specialists, and they are particularly useful in realisation of multiple practical tasks, characterised by great complexity and relatively low amount of available a priori knowledge about the nature of the modelled processes. They are successfully applied in a wide variety of problems related to data processing and analysis, their prediction, classification or process control. The object of the study was the technological site for copper ore enrichment by floatation, being a part of technological line in the Ore Enrichment Plant ZG „Polkowice”, KGHM „Polska Miedz” S.A.The main floatation is described by parameters measured by the supervision and control system, which can be grouped with respect to the functions performed for the object:a) input function: aFG - copper contents in the feed material for the main floatation [% Cu]b) control: LFG I - level in the first section of the floatation unit [%], LFG II - level in the second section of the flotation unit [%], LFG III - level in the third section of the floatation unit [%], LFG IV - level in the fourth section of the floatation machine [%]. These quantities compose the five elements long input vector.c) output function: ? - copper contents in the waste from main flotation (final waste) [% Cu]The measurement data, containing the above-mentioned parameters, have been collected during the normal operation of the plant. The data contain momentary values of the parameters, registered every 10 minutes, covering the period of 196 hours of the plant’s operation (1176 measurements for each quantity).For construction of the model used in the study classical multi-layer perceptron-type neural network has been used, with one hidden neuron layer, trained by the error back propagation method."