Artificial neural network for predicting the physicochemical composition of milk from physical analyses

Abstract

The aim of the research was to predict, using Artificial Neural Network (ANN) the physicochemical composition of milk from physical analyses. Forty-three milk samples were analyzed for fat, solids-non-fat, protein, lactose, and mineral salts. In addition to temperature, freezing point, and density by means of an ultrasound milk analyzer. The network architecture used was feed-forward multilayer, with data divided into 70% for training and 30% for ANN testing. The artificial neural network transfer function was Relu, Adam as the training algorithm for weight change with a constant learning rate. The number of neurons in the hidden layer was determined using the mean square error and coefficient of determination for training and test data. Twenty neurons in the hidden layer were analyzed and considered appropriate. The ANN model was able to predict the physicochemical composition of milk, the result obtained was a robust coefficient of determination, with values above 0.88.