Discriminative power of independent component analysis applied to an electronic nose

Electronic noses have been developed to analyze gas composition in many applications. The potential analytical power of an electronic nose needs to be investigated with regard to detection in a complicated gas environment. To date, principal component analysis (PCA) has been a popular analytical tool for use with electronic noses; however, PCA reaches its limits when the measurement data exhibit non-Gaussian behavior. Components extracted by independent component analysis (ICA) are expected to have a high discriminative power, conducive to improving the performance of an electronic nose. To verify this hypothesis, the performance of principal components (PCs) and independent components (ICs) in classification and regression tasks is compared. The data were obtained from a chemical gas sensor array for volatile organic compounds (VOCs). In the classification task, silhouette coefficients and K-means are employed to measure the performance of ICs and PCs. ICs have evident classification advantages over PCs because of their higher accuracy. In the regression task, ICs and PCs are applied via a neural network to predict the concentration of four VOCs. The errors in the prediction are calculated and analyzed, and it is found that ICs continue to outperform PCs during the regression task. The results of this study demonstrate the potential and ability of ICA to extract features when used with an electronic nose in a complicated gas environment.