An alternate method of hierarchical classification for E-nose: Combined Fisher discriminant analysis and modified Sammon mapping

摘要

An alternate method of hierarchical classification combining Fisher discriminant analysis (FDA) and modified Sammon mapping (MSM) is presented in this paper. The FDA-MSM method could project most of the samples information for classification onto a new two or three dimensions object space and show the samples distribution directly. Meanwhile, the other part of the method for new sample classification within the object space is also provided, where a parameter P of rationality is defined to represent the degree of confidence that the new sample belongs to an assumed affiliated class, and the class which gets the highest rationality is the class in which the new sample belongs to. A dataset with seven classes to be discriminated was used to validate the proposed method. The methods used in the data analysis were the k-nearest neighbour (k-NN), back-propagation artificial neural network (BP-ANN), FDA, Sammon mapping and the FDA-MSM. The correct classification rates of all samples by k-NN, BP-ANN and FDA-MSM were 73.8, 97.6 and 98.8%, respectively.