Prediction of fresh milk quality by using Artificial Neural Network and Multivariate Regression

Resumen

The objective of this research was to compare the best structure of a Neural Network (ANN) with a multivariate nonlinear regression model (MNLR) to predict the physicochemical quality parameters of milk. To create a predictor model for the livestock sector, 3 input and 6 output variables were used. To achieve this, a Feedforward ANN with Backpropagation training algorithms was applied. For the models, the Matlab 2020a software was used. The lowest mean absolute deviation (MAD) was found to be 0.00715952, corresponding to a Neural Network with 2 hidden layers (18 and 19), with Tansig and log sig type function, respectively. MNLR models had R2 values greater than 0.9. Cross-Validation with 10 interactions was used for this purpose. For comparison, a Duncan test was used where it was found that there are no statistically significant differences between the real sample, the MNLR, and the ANN, with a 95.0% confidence level.