Gated Transformer Networks-Assisted Intelligent Sensor Array for Concentration-Generalized Identification of VOCs

Conventional gas sensors are limited to concentration detection due to non-specific adsorption, which impedes accurate gas identification. The integration of artificial intelligence algorithms with gas sensors has improved recognition capabilities; however, discrepancies between the training set concentrations and real-world environments remain a barrier for commercialization. This study presents a deep learning-based gas recognition system using a sensor array comprising six commercial metal oxide semi-conductor sensors, which generated a comprehensive dataset of 28 gas categories across six concentration levels (20–100 ppm). The system employs a Gated Transformer Network (GTN) deep learning model to process multivariate sensor data, achieving a recognition accuracy of 99.35±0.04%, which surpasses traditional methods and other deep models. Notably, the system demonstrates strong generalization to unseen concentrations, with an average accuracy of 90.66% in concentration generalization experiments. t-distributed Stochastic Neighbor Embedding (t-SNE) analysis further confirms the extraction of highly discriminative features, which reduces dependence on full-range concentration data and offers a reliable solution for industrial volatile organic compounds (VOCs) monitoring.