Data preprocessing enhances the classification of different brands of Espresso coffee with an electronic nose

Abstract

Two different ways of preprocessing chemical sensor data are presented as a means to improve the interpretation and the classification ability of an electronic nose (EN). The practical problem at hand is the distinction between four commercial coffee blends — containing up to 12 types of coffees — all of which are to be consumed as Espresso. Coffee was sampled in three successive preparation phases: as beans, ground (powder) or liquid (the actual Espresso). In the case of beans, stress is put on the improved clusters visualization after the preprocessing and before the actual classification is performed. Different catalysed sensors and successive extractions were used to differentiate the response pattern towards the various coffees. The features which permitted the best samples' classification as judged from Principal Component Analysis (PCA) score plots were selected. To this end, an empirical search strategy inside the feature space is presented. Scores from PCA were subsequently utilized as inputs for a feed forward multilayer perceptron (MLP) with cross-validation resulting in 100% correct classification with just two sensors. In the case of ground coffee, a (supervised) drift compensation algorithm was developed. It essentially consists of removing the first principal component (PC) for every cluster since this is seen to be given by the drift. An 87.5% classification performance was achieved. Liquid coffee, on the other hand, was not successfully classified, probably due to the difficulty in assuring reproducible sampling conditions.