Prediction of moisture in cheese of commercial production using neural networks

Control of cheese moisture is paramount to maximizing yield and profitability of a cheesemaking operation. Modeling and prediction of cheese moisture prior to pressing from a large industrial database for stirred-curd Cheddar cheese made with non-standardized and standardized milk was carried out using neural networks (NN). The number of model input variables was reduced by removing or combining some of them, based on cheesemaking knowledge and on the results of two tests estimating the impact of each model input. Input removal was carried out until the validation mean absolute prediction error (MAPE) increased. An initial NN cheese moisture model with 38 input process variables, coded as 57 NN inputs, was reduced to one with 21 input process variables, coded as 34 NN inputs. For the latter, the validation MAPE was 0.53% cheese moisture in a range of cheese moisture of 13.2%, and 0.51% for the best 25% of models (out of 100). For the range of operating conditions of the process in this study, four main groups of variables were found to be the most influential on the prediction of cheese moisture: cutting and subsequent stirring of the curd, curd rinsing temperature, starter quantity, activity and strain, and seasonal variation of milk composition. The NN model with the selected input variables and optimized number of hidden neurons was then used to predict cheese moisture for ranges of these variables. This study showed that NN models can successfully extract input-output variable relationships from industrial production data in spite of the inherent error in these data. The resulting NN models can be used both for research to develop the base of knowledge on production variables and their complex interactions, as well as for the prediction of cheese moisture.