Several ML Algorithms and Their Feature Vector Design for Gas Discrimination and Concentration Measurement with an Ultrasonically Catalyzed MOX Sensor

Although gas-borne ultrasound catalysis has been developed as a new method to discriminate gas species and measure the concentration, applications of machine learning methods in gas analyses with a single metal oxide (MOX) gas sensor catalyzed by gas-borne ultrasound are still scarce. In this work, with an ultrasonically catalyzed MOX gas sensor, we explored the effectiveness of K-nearest neighbors (KNN), support vector machine (SVM), and single-hidden-layer BP-ANN (SHBP) in gas discrimination and the application of the SHBP in concentration measurement. The target gases in this work are ethanol, acetone, methanol, hydrogen, and n-butane in clean air, respectively, and the discrimination and concentration regression are implemented by two different ML models. With the properly designed feature vectors, the SHBP method has an acceptable capability of both of species discrimination and concentration regression (success rate of gas discrimination = 99.5%, relative error of concentration regression = 6.406%). The KNN and SVM methods have similar capabilities of gas discrimination as the SHBP. This work also demonstrates a method to design the feature vectors for the ultrasonically catalyzed MOX gas sensor and to choose the feature parameters.