Prediction of sensory quality of UHT milk – A comparison of kinetic and neural network approaches

Data on deteriorative processes that progressed during storage (9, 15, 25, 35 and 45 °C) of ultra-high-temperature (UHT) treated milk were modeled using chemical kinetics and artificial neural network approach to predict sensory quality of the product. Parameters that were considered in the study represented changes associated with proteolytic, lipolytic, oxidative and Maillard reactions whereas sensory quality was evaluated in terms of flavour score and total sensory score as dependent variables. Kinetic models were developed by integrating multiple regression equations for the five physico-chemical parameters as independent variables with their Arrhenius parameters. Artificial neural network (ANN) was developed with the same five variables taken as input data and flavour and total sensory scores as the output quality criteria. Of the different ANN approaches examined, the Bayesian regularization algorithm provided the most consistent results and was therefore used for developing the ANN models. The prediction performance, judged on the basis of percent root mean square error, of the ANN-based models were better than the kinetic models.