Prediction of nutritional parameters in milk from its physicochemical properties using deep learning

Abstract

The objective of this study was to determine the optimal structure of a Neural Network that allows for the prediction of physicochemical quality parameters of milk, such as fat, protein, lactose, non-fat solids, total solids, and minerals, based on easily determinable variables like methylene blue reduction time, density, and pH at the Nestle - Cajamarca company. A Feedforward Artificial Neural Network (ANN) was employed, using the Backpropagation training algorithm and the Levenberg-Marquardt weight adjustment algorithm, with the following topology: meta error of 10-2, learning rate of 0.01, momentum coefficient of 0.5, 3 input neurons, 6 output neurons, and 50 training epochs. It was found that the mean absolute deviation (MAD) was minimized to 0.00715952 in a Neural Network with 2 hidden layers, consisting of 18 and 19 neurons, respectively. The activation functions used were Hyperbolic Tangent Sigmoid (Tansig) and Hyperbolic Logarithmic Sigmoid (Logsig), resulting in a regression coefficient of 0.99837. Predictions were compared with a nonlinear multivariable regression model, and no statistical differences (p > 0.95) were observed for all output variables, except for protein.