Abstract

The final filtration and bottling of wine represents the last and most important stage within the entire process of its production and commercialization. In this context, wine is filtered to retain particles and ensure its microbiological and organoleptic properties. The measurement of the Laurenty index during this process helps to prevent problems in flow and pressure, which can increase production costs and times. However, this index can only be measured when the wine is ready to be bottled, and in some cases, only a few hours before bottling, which limits its usefulness for proper production planning.

To overcome these limitations, the use of Machine Learning is proposed to estimate the Laurenty index from 27 physicochemical parameters through an analysis of data over 30 weeks. For this purpose, missing data are imputed, dimensionality is reduced using principal component analysis (PCA) and factor analysis (FA), and models are trained and evaluated using the factors during the first 24 weeks. Finally, the model with the highest predictive power is selected and used to evaluate the subsequent 6 weeks. This is done by comparing the direct application of the model on a weekly basis to its retraining, considering the data generated in the immediately preceding week.

             Finally, it is concluded that retraining the model with the new data generated in the week prior to prediction ensures a high level of confidence, allowing its practical use at the industrial level to plan production.