Sensor response analysis, feature extraction and classification of volatile organic compounds

Abstract

This project focuses on analyzing sensor response from different types of functionalized reduced Graphene Oxide (rGO) based VOC sensors on three selected VOC gases (acetone, toluene and isoprene), performing feature extraction and employing supervised learning for VOC classification. The rGO as sensing material was functionalized with nanoparticles (NPs); such as gold (Au), silver (Ag) and platinum (Pt) and plasma treatment; such as ammonia (NH3), hydrogen (H2) and Octafluorocyclobutane (C4F8). The sputtering duration and relative frequency power (WRF) are varied for the functionalization of the nanoparticles while the temperature is varied for the plasma treatment functionalization. The sensor response then was measured from the change of resistance signal during the presence of the VOC gas at low concentration, from 1 to 6 parts per million (ppm). Sensors with thin-film from rGO/Au NPs, rGO/Ag NPs, rGO/Pt NPs, rGO/H2 (RT) and rGO/ C4F8 had shown a good response toward the VOC gas while sensor with rGO/NH3 and rGO/H2 (except for the RT recipe) showed poor responses. Sensors that have a good response then proceed with the analysis, feature extraction and machine learning part. Average Resistance value at the presence of clean dry air (CDA) only, Rair and in the presence of the VOC gas, Rgas were extracted from the original sensor signal and manipulated into 10 new features. Then, five supervised learning models such as k-Nearest Neighbors (kNN), Random Forest (RF), Logistic Regression (LR), Support Vector Machine (SVM) and Artificial Neural Network (ANN) were benchmarked for the VOC classification task. The model performances were evaluated using k-Fold Cross- Validation and the prediction of the classification are visualized by using a Confusion Matrix. The results showed that RF and kNN have good performances with a mean of accuracy and standard deviation, 0.813 ± 0.035 and 0.803 ± 0.033, respectively. However, ANN, LR and SVM (Polynomial kernel) showed poor performance with 0.447 ± 0.035, 0.403 ± 0.041 and 0.419 ± 0.035 respectively. Based on the reported performance, it shows that 2 out of the 5 models could deal with the feature selected in the VOC dataset and it is feasible to analyze the classification of VOC gases based on single sensor arrays. It is therefore interesting to explore the analysis of combined sensor arrays for such tasks in future research.