Impact of modeling parameters on the prediction of cheese moisture using neural networks

Abstract

In the cheese manufacturing process, yield is of great economic importance, and cheese moisture is the single most important factor influencing yield. A large industrial database of 41 process variables was used to develop a comprehensive model for cheese moisture using three-layer feedforward neural networks (NNs). The scope of this research was to analyze NN modeling parameters related to the process and data, in view of obtaining the best predictive model for cheese moisture. An analysis on the number of hidden neurons and the fraction of data used for on-line validation (FV) showed that the latter influences the choice of the number of hidden neurons and determines how the selected model is representative. For the cheese moisture model, six hidden neurons and 20% FV were optimal. Numerous trials were performed with the objective of reducing the number of NN inputs while aiming for better or equivalent model accuracy. These included the development of specialized models associated to specific values of some qualitative variables, various encodings for qualitative variables, removal of variables to increase the number of modeling data, and combinations of some of these strategies. It was found that NN could be used to derive predictive models for cheese moisture from industrial data, capable of attaining a validation mean absolute prediction error of 0.56% cheese moisture, for data with a range of 13.2% cheese moisture. Trials indicated that gains in accuracy were achieved by using several specialized NN models for each value of a qualitative variable instead of an overall model, as well as by encoding qualitative inputs as quantitative inputs directly related to the output. The validation prediction error exhibited a high correlation with the number of training data per NN coefficient, but was highly influenced by the actual choice of model inputs.