New Dynamic Fingerprint in Derivative-Based Phase Space: Rapid Gas Sensing in Seconds

Abstract

Many studies have focused on smart electronic noses combining machine learning and gas sensor arrays, but feature extraction for training has generally relied on dimensionality reduction techniques based on raw time-series data. These methods do not reflect the principles of sensor responses, limiting their applicability in diverse gas environments. In this study, we propose a new phase space, expressed through the first and second derivatives of dynamic response signals, to effectively characterize the nonlinear responses between gas sensors and gases. Sensing data transformed into a phase space showed unique patterns depending on the type and concentration of gases, and these were investigated for alkanes with various chain lengths (CH₄, C₃H₈, C₄H₁₀). By applying these patterns as a preprocessing method, CNN-based gas identification machine learning achieved a high classification accuracy of 99.1% and a low concentration prediction error of 2.23 ppm using only a single sensor. Additionally, the algorithm was trained and validated across various regions of the phase space, identifying the minimum time and region required for simultaneous gas classification and concentration prediction. This study presents a novel strategy for the fast and accurate identification of gases within seconds and is expected to have significant scalability in diverse gas environments.