Presentation Information
[23p-22B-3]Study on correlation between GCMS mapping and QCM sensing signals for ternary gas mixtures
〇(B)THANISORN BEST OONPITIPONGSA1, Chaiyanut Jirayupat2, Wataru Tanaka1, Takuro Hosomi1, Tsunaki Takahashi1, Jiangyang Liu1, Takeshi Yanagida1,3 (1.University of Tokyo, 2.Mi6 Corporation, 3.IMCE, Kyushu Univ.)
Keywords:
integration of sensors and AI,AI processing in sensor,GCMS
In the biomedical field, especially cancer prediction, researchers have identified multiple Volatile Organic Compounds (VOCs) as biomarkers for diseases using Gas Chromatography-Mass Spectrometry (GCMS), while gas sensor arrays are claimed to differentiate these diseases with notable accuracy. This highlights a key question in gas sensor applications: is there a link between gas sensing signals and GCMS analysis profiles? In this study, we introduce a novel method that establishes a strong correlation between GCMS mapping and sensor signals, a long-standing gap in gas analysis, with machine learning techniques. We applied this method to ternary gas mixtures of ethanol, toluene, and dichloromethane with QCM gas sensors covered with 4 types of metal oxides. Utilizing Convolutional Neural Networks (CNN) and Principal Component Analysis (PCA), we successfully predict GCMS mapping from gas sensors signal with a remarkable average R2 = 0.99 in stratified cross-validation and average R2 = 0.92 in leave-one-out cross-validation. Furthermore, we employed GradCAM (Gradient-weighted Class Activation Mapping) for feature importance visualization of sensor signals, providing insights into the interaction between various gas species and sensor materials. This advancement not only proves the capability of gas sensors in distinguishing between different gas mixtures but also paves the way for enhanced gas mixtures’ fingerprint studies in diverse applications such as healthcare, personal authentication, and food quality determination.