Abstract

This paper presents a comparison of two Artificial Neural Network architectures; i.e. a Standard Feedforward network and an Autoassociative network; both as particle size distribution estimators in a wet grinding circuit. The relevant variables to be considered as the network's inputs are selected, using an index, based on the Lipschitz quotient. Their parameters are calculated by a Levenberg-Marquardt algorithm. The results demonstrate that both architectures can provide reasonable estimates of the size distribution. However, the autoassociative one has some attractive properties concerning noise reduction at the expense of a larger number of parameters.