Prediction of moisture in industrial cheese using artificial neural networks

Cheese yield dictates the profitability of a cheesemaking operation and cheese moisture is the composition factor that most influences cheese yield. Moisture in stirred-curd Cheddar cheese from commercial production, made from standardized and non-standardized milk, was modeled using three-layer artificial neural networks (NN). Data counted over 10 000 cheese vats from four years of production in a large-scale Quebec cheese plant. Each cheese vat was described by milk composition and quality parameters, starter production and activity measures, ingredient quantities, and processing conditions and measures from the vat and drainage steps. Data were validated to remove erroneous or non-representative values through the use of several statistical outlier detection techniques, including Fourier series modeling of seasonal variables. It was found that outlier detection by single and multivariable approaches were able to identify different outliers and that the different methods were complementary. The verified data were then used to generate NN models and study the effect of several modeling parameters on cheese moisture prediction error. It was found that the optimum number of hidden neurons was determined by the fraction of data used for model validation (FV), and that this fraction had a minimum value below which the random sample used for validation lost its representativity. An initial model with 41 input process variables was reduced to one of 21 input variables using 16 hidden neurons and 15% FV. This model yielded a mean of the validation mean absolute prediction error of 0.53% cheese moisture, within a cheese moisture range of 13.2%. The groups of variables found to be most influential in moisture prediction were: (1) cutting and stirring of newly formed curd, (2) starter quantity, activity and strain, (3) rinse water temperature during draining and (4) seasonal variations in milk composition. The modeled effects of these variables on predicted cheese moisture, as well as the interactions between them were analyzed using the best NN models.