From microwave gas sensor conditioning to ammonia concentration prediction by machine learning

This paper proposes an innovative approach to understand the conditioning process of a microwave gas sensor operating at room temperature based on the combination of its response and the interpretation of the mass spectrometer data. A large variation of the dielectric parameters and thus of the microwave response is due to water departure from the sensor surface. Consequently, the first step of the conditioning process is a carrier gas sweep of the sensor surface. The second step consists in pre-saturating the microwave gas sensor surface with a high concentration of the polluting gas which will be detected (here ammonia). This process results in a very good quality microwave response on the qualitative and quantitative aspects for the detection of ammonia in the air and is helping the work carried out in this article in artificial intelligence on microwave responses. A regressor machine learning model is applied on time samples of this sensor response to predict the ammonia concentration. Several machine learning algorithms are tested and compared. Principal Component Analysis is also tested to reduce the input data dimension, but results are not conclusive. The concentration profile is revised to reduce the bias induce by the presence of too much measurement data when no pollutant is present in the air. And the Mutlti-Layer Perceptron regressor give the best results with a mean absolute error of 32.13 ppm (8 %) over a range of 0–400 ppm and R-squared score of 0.87.