Development of Machine Learning Models for Gas Identification Based on Transfer Functions

Potential applications using gas sensors are expected in various fields by combining MEMS sensors with advanced Information and Communication Technology. In a conventional gas identification approach, gas injection patterns must be strictly controlled to obtain a comparable sensing features which are used for gas identification. Toward a simpler gas sensing system than existing ones, a gas identification protocol which does not depend on gas injection patterns is required. In this study, we have developed a data processing method for gas identification by integrating system identification and machine learning. By using transfer functions as a feature descriptor, we have developed gas identification models which can identify gas species with an arbitrary gas input pattern. For solvent vapors, we have successfully identified gas species with an accuracy of 0.98±0.03 even with a random gas injection pattern. We have also identified complex samples, that is odors of herbs and spices, with an accuracy of 0.94±0.04, showing the high feasibility of the present approach.