Classifier comparison and sensor selection for e-noses

Abstract

The authors analyze 4 datasets produced by an e-nose based on 20 compositionally distinct polymer-carbon black sensors exposed to laboratory designed gas mixtures and one dataset produced by a second e-nose based on five thin films metal oxide sensors exposed to the headspace of seven different coffee blends. We aim at selecting the best 5 sensors (for the e-nose which has 20) using as FS criterion the test performance of Fisher linear discriminant analysis (LDA) and multilayer perceptrons (MLP). The datasets were chosen to give classification problems of varying hardness: from the discrimination of two almost Gaussian distributed analytes to the discrimination of two analytes in the presence of interferents at different concentration levels. The performance of the best 5-sensors subset selected with MLP was found to be better - but not significantly better - than the performance of the LDA-selected subsets. This is true also for classes with a strong multimodal distribution. In only one case the test set performance distribution over all 5-sensors subsets was found to be clearly better with MLP than with LDA.