Recognizing the disease-specific odor in human breath is a feasible non-invasive disease detection strategy. The most critical issue of olfactory diagnosis lies in recognizing the target gas among complex respiration compounds accurately. This paper demonstrated a novel approach to binary gas mixture recognition, exploring nature-related features of the target “odor object” (acetone gas) regardless of its “noisy background” (ethanol or methanol gases). Those features were extracted from time-series resistance fluctuations collected with one conventional SnO2 thin film gas sensor (TF) or our sensitivity-controllable lines gas sensor (TL). 6 features (variance, root mean square, band power, relative band power, entropy, and the number of values crossing mean value) were selected and analyzed with k-Nearest Neighbor regressor (KNN) and Support Vector Regressor (SVR). 18 types of acetone/ethanol mixtures and 3 acetone/methanol mixtures were tested with the proposed models and further explained with SHAP (SHapley Additive exPlanations). The model of TL and KNN achieved the optimal performance of regressing acetone concentrations with the coefficient of determination (R2) of 0.97 and 0.95 in individual gas and binary gas mixtures, respectively. Entropy was found to be the key feature that was linked to sensor properties. Finally, acetone gas data mixed with simulated gas and random noise data were experimented to evaluate our models’ potential for acetone/random gas mixture recognition and robustness.
Read full abstract