Improving the classification accuracy in electronic noses using multi-dimensional combining (MDC)

Traditional pattern recognition (PARC) methods, used in electronic noses (e-noses) are either parametric (such as k-nearest neighbors, KNN, and linear discriminant analysis, LDA) or non-parametric (such as artificial neural network and fuzzy logic). Multi-dimensional combining (MDC) is proposed to combine the classification outputs of individual classifiers into a more robust and accurate one. Two implementations are proposed to find the individual classifiers, one is based on various feature extraction methods and the other is based on various dimension reduction methods, with three means of combining. Six household fragrances were sampled using the Cyranose 320 e-nose device. The acquired data (600 measurements) was split into two sets, training and testing. Experiments were conducted at various concentrations of the sample smell, various sample numbers and various training numbers. Results show the advantage of MDC over the individual classifiers, and over the other traditional PARC methods under all conditions.