The identification of gas mixtures is critical in chemical engineering, food safety, environmental science, and medical applications. This study employs a four-sensor array to classify ethylene-methane and ethylene-carbon monoxide (CO) mixtures, with concentrations ranging from 0 to 20 ppm for ethylene, 0–600 ppm for CO, and 0–300 ppm for methane. To enhance robustness, mean sensor response over time is used to mitigate noise, ensuring high classification accuracy. Each sample is evaluated using sixteen features, including temporal dynamics (e.g., rise time) and statistical metrics (e.g., baseline variance). Tree-based machine learning models—Decision Tree (DT), Random Forest (RF), and Extra Trees (ET)—are developed for gas classification, significantly reducing training set size (to 60%) and prediction time. The proposed ET model achieves superior classification accuracy (99.15%) compared to RF (95.86%) and DT (93.89%) on the UCI dataset, offering an efficient and accurate method for experimental gas identification.
In controlled dynamic laboratory conditions representative of real-world variability, the gases are usually mixtures rather than individual components. In chemometrics and electronic-nose (E-nose) systems, accurate concentration is a key challenge. Standard algorithms like support vector machine (SVM), k-nearest neighbors (KNN), and shallow multi-layer perceptron (SMLP) have limited feature extraction and poor generalization ability when faced with overlapping responses from sensors, as well as strong cross-gas interactions. While deep learning methods offer enhanced performance, they generally rely on large labeled datasets and cannot inherently maintain robustness in mixed gas scenarios. In order to overcome this limitation, the paper presents a hybrid approach of 3D convolutional neural networks (3DCNNs) with semi-supervised generative adversarial network (SGAN). The 3DCNN component captures spatiotemporal dynamics of sensor array responses, while the SGAN improves generalization under limited labeled data by generating realistic synthetic samples. Experimental results demonstrated that the model obtains a classification accuracy of 99.10%, which is higher than SVM (93.80%), KNN (92.60%), and SMLP (95.30%). These results indicate that the model can be well used for high-precision gas mixture analysis. This work supports SDG 9 (Industry, Innovation and Infrastructure) through advanced AI-based gas sensing, SDG 11 (Sustainable Cities and Communities) by enabling accurate air-quality monitoring, and SDG 3 (Good Health and Well-Being) by improving detection of hazardous gas mixtures that impact public health.