A comparative study of machine learning models for identifying noxious gases through thermal fingerprint measurements and MOS sensors

Abstract

Recently, there has been a notable surge of interest in developing swift and precise gas detection and categorization tools through artificial intelligence. This work specifically focuses on designing a simple sensor array configuration and empirically compares various machine-learning models for identifying multiple gases. The study employs eight metal oxide semiconductor sensors, each operating independently based on the thermal fingerprint principle. These sensors are utilized to gather data on responses to different toxic gases, and a range of traditional machine learning classifiers are trained and assessed using this dataset. The classifiers undergo investigation across various train/test dataset split ratios and different principal component analysis dimensionalities. Average accuracy metrics are employed over the entire testing dataset to evaluate model performance. The results demonstrated the challenge of selectively detecting each gas type solely through conventional analyses with the gas response data of the proposed sensor array. Nevertheless, the utilization of machine learning models has exhibited promising outcomes in augmenting gas identification and classification. The findings reveal a consistent improvement in testing accuracy for the Gradient Boosting model, which achieves an outstanding accuracy rate, with instances of reaching 100% accuracy in the training dataset when three or more principal components are considered. This study highlights the potential of the proposed straightforward sensor array and presents effective recognition algorithms applicable to creating electronic noses for gas-sensing applications.