Rede neural artificial para detecção de adulteração em leite por adição de água

Abstract

To control the quality of the milk, analyzes are carried out by sampling, using laboratory methods that demand time, are sometimes costly and, in certain situations, use dangerous/toxic reagents. Official routine analytical and fraud detection methods may not be fully effective in diagnosing milk irregularities, requiring additional analysis. In this sense, the main objective of the work was to develop an artificial neural network (ANN) to detect adulteration in raw milk by adding water and as a secondary objective, to compare official methodologies and ultrasound spectroscopy for the evaluation of the physicochemical characteristics of milk. For this, 125 samples of raw milk were used. Of these, 25 were collected in borosilicate glass bottles with a capacity of one liter to carry out the official methods (OM) and compare them with ultrasound spectroscopy (US) for the determination of fat and protein contents, density, cryoscopic index,

solids total (ST) and non-fat solids (SNG). The remaining 100 samples were placed in Falcon- type tubes with a capacity of 200 mL and used to prepare the RNA, with adulteration of 20%

(n=20) of these samples by adding increasing percentages of distilled water (1%, 5%, 10% and 20%), which resulted in 80 adulterated samples and 80 unadulterated samples, evaluated solely by ultrasound spectroscopy. For the development of the ANN architecture, the 160 milk samples were randomly distributed into three subgroups: (i) training (60%); (ii) validation (20%); and, (iii) test (20%). The values of the variables were normalized between 0 and 1. To determine the ANN configuration, the SNN® software was used and the Radial-Basis Function (RBF) and Multilayer perceptron (MLP) ANNs were tested. As a result of the study, the average densities and average protein contents determined by US and MO did not differ from each other (P > 0.05). The mean fat contents and the mean values of the cryoscopic index, ST and SNG, obtained by US and MO were different (P = 0.026, P = 0.040, P < 0.001, P = 0.014, respectively), the correlations obtained in the analysis of these parameters were positive (R = 0.470, R = 0.118, R = 0.087 and R = 0.315, respectively) and the accuracies of the methods were 0.180, 0.058, 0.155 and 0.075, respectively. For the classification of milk samples, the best network was the one that presented the lowest number of classification errors, the lowest difference between the classification errors in the subgroups (training, validation and testing), and the lowest number of neurons in the hidden layer. The best classification network was the RBF neural network that presented 10 neurons in the input layer, 40 neurons in the hidden layer and two in the output layer, resulting in 95.37% of adulterated samples classified correctly. It is concluded that US is a fast technique and the results obtained are correlated with those of the official analyses. The ANNs have great potential to be used as a tool for evaluating adulterations

of milk with the addition of water. Therefore, the development of automatic, non-invasive, low- cost and intelligent systems could play a decisive role in the diagnosis of food fraud and in

decision-making regarding the judgment and destination of adulterated products. But, for the reliability of the results, it is necessary balanced and sufficient data for training the model to be tested.