A Minimum Distance Inliers Probablity (MDIP) Feature Selection Method To Enhance Gas Classification For An Electronic Nose System

摘要

To extract as much information as possible from the gas sensor responses of an electronic nose (E-Nose) system, feature extraction methods are adopted to obtain meaningful information of data. However, among a large quantity of the extracted features, only a few are actually informative for gas classification. Even worse, some of the features may degrade the accuracy of classification. To solve this problem, we propose a minimum distance inliers probability (MDIP) feature selection (FS) method that eliminates unnecessary features by considering their degree of clustering and degree of separation. The performance was validated using an open-access dataset. After applying the MDIP method, the number of features was reduced from 48 to 18 in average, while the average classification accuracy was improved from 46.45% to 80.6%, validating the efficiency of the MDIP method.