Modeling Electrostatic Separation Process Using Artificial Neural Network (ANN)

Abstract

In this paper, the characteristics of an electrostatic separator were modeled using artificial neural network (ANN). The model was constructed by considering the misclassified middling product during separation, where system parameters (voltage level, rotation speed, electrode position, etc) were varied. The ANN architecture was optimized through the variation in the neuron number, percentage of testing data and percentage of validation data. Performance of the network was assessed by the error indicators, namely mean square error (MSE) and coefficient of determination (R-square). It is found that, lesser number of neurons and lower percentage of both training and validation dataset contributes to better network performance. Additionally, network architecture thus derived was selected for a detailed study on the various combinations performance corresponding to the input and output variables. The results consequently suggest a simplified network structure with reduced number of input variables for modeling of this nonlinear process.