Application of artificial neural network to predict benzo[a]pyrene based on multiple quality of smoked sausage

This research aimed to establish the benzo[a]pyrene (BaP) prediction model based on multiple quality of smoked sausages including color, peroxide value (POV) and thiobarbituric acid (TBA), while the back propagation-artificial neural networks (BP-ANN) were applied to model the relationship between the BaP content and multiple quality. In this study, multiple quality parameters were used as input variables, and BaP was used as output layer parameters. The Levenberg = Marquardt back-propagation training algorithm with 13 hidden layer neurons and 0.4 learning rate was the best predictive performance, which the correlation coefficients (R) of validation and test were 0.9510 and 0.9264, mean square error (MSE) was 0.01108. Furthermore, we also conduct sensitivity analysis to analyze the relative contribution of color and lipid oxidation to determine the key factor of influencing the content of BaP. In terms of relative contribution, the color, lipid oxidation were the important parameters with the most discriminative power, specifically the b*, POV and TBA values, which have a critical effect on the prediction of BaP contents. Results indicated that the BP-ANN has great potential in predicting the BaP of smoked sausages based on multiple qualities.