Detection of Volatile Organic Compounds in Contaminated Air Using Gas Sensors with Machine Learning

Abstract

We investigate selective detection of target volatile organic compounds (VOCs) as unpleasant odor substances in the presence of interference VOCs from car interiors. We selected four unpleasant odor substances as target gases, such as age-related body odors and fungi, and two interference VOC substances, such as the main olfactory components from car interiors, in this study. We used eight semiconductive gas sensors including six general semiconductive gas sensors and two bulk-response type semiconductive gas sensors. Discrimination of each target gases was carried out using machine learnings, such as principal component analysis (PCA) and linear discriminant analysis (LDA), which are dimensionality reduction methods. PCA is an unsupervised learning method, and finds the axis that maximizes the data’s variance. LDA is a supervised learning method, and finds maximizes the separation between multiple classes. The machine learnings used sensor responses (r=Ra/Rg; r: sensor responses, Ra: resistance in humid pure air or in contaminant, Rg: resistance in the target gases) from all sensors. The plots of PCA scores tended to move radially outward as concentrations of target gases increased. PCA scores of target gases with contaminants showed almost the same as those of lower concentration target gases without contaminants. However, the PCA scores in the presence of contaminants were within the area of each target gas so that the discrimination between the target gases was almost achieved. The LDA scores were classified by target gases without concentrations of target gas and contamination, so the LDA scores tended to aggregate with target gases. The plots of LDA scores seemed to be more gathered in each class (target gas) area than those of PCA scores.Moreover, we investigated to reduce the number of sensors using another machine learning, i.e. feature selection technique, which used the sensor responses for obtaining sensor weights against all data. Based on the results of another machine learning, we selected three sensors which possess high scores in another machine learning. The PCA and LDA scores from selected three sensors showed almost the same qualities as those from eight sensors, whereas those from non-selected sensors cannot discriminate target gases, so the amount of information of the target gases has been reduced by excluding the selected three sensors for separating each target gas. From this result, it is possible to reduce the number of sensors to keep the discrimination between target gas by selecting sensors obtained using feature selection techniques.Fig. PCA scores and eigenvectors from eight sensors to four target gases. Figure 1