Physics-Aware Spatiotemporal Graph Attention Network for Variable Frequency Temperature Modulation of MOS Gas Sensors

Abstract

Metal oxide semiconductor (MOS) gas sensors are widely used in industrial mixed-gas detection. However, conventional temperature modulation fails to fully activate their dynamic response characteristics, limiting the performance of electronic nose systems. To address this issue, a variable frequency temperature modulation (VFTM) mechanism is proposed, which jointly regulates heating frequency and duty cycle to enhance sensor–gas physicochemical interactions and obtain richer gas fingerprint information. In light of this, a multitask physics-aware spatiotemporal graph attention network (PST-GAN) is developed for simultaneous gas type recognition and concentration estimation. A physics-aware temporal encoder (PATE) incorporates heating excitation into feature encoding, enabling the network to learn the physical correlations between sensor responses and heating states. Spatiotemporal graph modeling is then performed to jointly characterize temporal dependencies within each sensor and cooperative responses across sensors. These features are hierarchically aggregated through layered graph convolution, while a gating mechanism adaptively fuses temporal and spatial graph features. Finally, a compact electronic nose system is designed and fabricated for mixed-gas detection and verified through experiments. The results show that PST-GAN achieves a recognition accuracy of 99.68% and a mean absolute error of 1.9881 ppm in concentration estimation, demonstrating strong potential for industrial gas detection applications.