Abstract

Data mining techniques have already shown useful to classify wine fermentations as problematic. Then, these techniques are a good option for winemakers who currently lack the tools to identify early signs of undesirable fermentation behavior and, therefore, are unable to take possible mitigating actions. In this study we assessed how much the performance of a clustering K-means fermentation classification procedure is affected by the number of principal components (PCs), when principal component analysis (PCA) is previously applied to reduce the dimensionality of the available data. It was observed that three PCs were enough to preserve the overall information of a dataset containing reliable measurements only. In this case, a 40% detection ability of problematic fermentations was achieved. In turn, using a more complete dataset, but containing unreliable measurements, the number of PCs yielded different classifications. Here, 33%f the problematic fermentations were detected. © 2009 Springer Berlin Heidelberg.