Algoritmos de aprendizagem de máquina para predição e classificação de variáveis no drip test da indústria de frango

Abstract

One of the main problems in the chicken meat industry is controlling the amount of water absorbed by the carcasses during the production process. Machine learning algorithms have been applied to solve problems in several areas. In the present work, 907 tests were performed using three machine learning algorithms, the RBF (Radial Basis Function) network, the multilayer perceptron network and SMOReg (Sequential Minimal Optimization), were used to predict the result of the drip test and the most influential variables in the process. The RBF network obtained the best result, reaching 63.68% of accuracy, the multilayer perceptron reached 55.79% and the SMOReg 52.44%. Seeking to optimize the results, the input variables were reduced from eleven to five. At this stage the multilayer perceptron obtained 58.48% of accuracy and the SMOReg 53.36%. With the reduction of the input variables, there was an improvement in the prediction capacity of the two algorithms, approaching the best result obtained by the RBF network with eleven variables. The use of fewer variables makes the industrial application more affordable. The algorithms pointed out the absorption, the temperature of the lairage facility, ambient temperature, the temperature at the beginning of the pre-chiller and the average mass as the most influential variables. Such variables agree with the others found in the literature.