2024 , Volume 29, № 4, p.4-23

The study examines the application of machine learning methods for determining the concentration of carbon monoxide (CO) in the air based on data from metal-oxide (MOX) gas sensors. High levels of concentration are hazardous for human and animal health, making air quality control critically important. The output data from the sensors were investigated, and new features were created to account for the daily temporal variation of gas concentration’s. Multiple linear and polynomial regression models, as well as neural networks, were developed to predict CO concentration. The impact of regularization on the accuracy of gas sensor data interpretation was also explored. The analysis revealed that the primary source of error in CO concentration recovery was the data with low concentration values. Creating new features through daily averaging of resistance sensor values and temperature, as well as deviations from the mean values for the day, improved the results of the MAPE and GRE metrics. It was found that the best loss function for training neural networks is the absolute error (MAE), and the best activation function for a neuron is the hyperbolic tangent function (tanh(𝑥)). The study demonstrates the potential use of machine learning methods for air quality control.