An adaptive artificial neural network to model a Cu/Pb/Zn flotation circuit

In this paper, we describe the planning and development of an artificial neural network model of line 3 of the copper/lead flotation circuit at Brunswick Mining´s concentrator at Bathurst, New Brunswick. The prototype model predicts the copper and lead assays of the concentrate streams of this rougher flotation circuit. In the model, the actual values and rates of change in the main process variables such as head grades, reagent addition, mass flow, density, pH, temperature, cell level and grind size are treated as inputs. The global error in both training and testing of the model is used to indicate the accuracy of the model. The model is fully adaptable, i.e., it can be updated when required to account for ore and/or processing changes that are not currently included in the ANN because of lack of instrumentation or reliability of measurements. The adaptation algorithm is used to select current data to replace records in the existing training and testing datafile. Retraining is conducted whenever the model accuracy declines to a pre-defined target value. The algorithm determines the frequency of retraining. The final system will be expanded to calculate a total of 12 assays using a separate ANN model for each. All models are independently updated. This approach to artificial neural networks provides plant engineers with a process model that is always current and reasonably accurate. Model access provides flexibility in adjusting set-points to achieve increased efficiency in the control of process variables