Selective detection of target volatile organic compounds in contaminated air using sensor array with machine learning: Aging notes and mold smells in simulated automobile interior contaminant gases

Abstract

We investigated the selective detection of target volatile organic compounds (VOCs) which are age-related body odors (namely, 2-nonenal, pelargonic acid, and diacetyl) and a fungal odor (namely, acetic acid) in the presence of interference VOCs from car interiors (namely, n-decane, and butyl acetate). We used eight semiconductive gas sensors as a sensor array; analyzing their signals using machine learning; principal-component analysis (PCA), and linear-discriminant analysis (LDA) as dimensionality-reduction methods; k-nearest-neighbor (kNN) classification to evaluate the accuracy of target-gas determination; and random forest and ReliefF feature selections to choose appropriate sensors from our sensor array. PCA and LDA scores from the sensor responses to each target gas with contaminant gases were generally within the area of each target gas; hence; discrimination between each target gas was nearly achieved. Random forest and ReliefF efficiently reduced the required number of sensors, and kNN verified the quality of target-gas discrimination by each sensor set.