Topology-Aware Residual Graph Attention Networks for MOS Gas Sensor Calibration

In the current era, air quality monitoring is vital owing to the increasing environmental hazards. Yet, high-precision sensors are not feasible for widespread edge deployment. However, low-cost Metal Oxide Semiconductor (MOS) sensors, such as the MQ series, offer a scalable solution but suffer from cros-ssensitivity and environmental drift. Traditional calibration methods rely on linear regression or MLPs and often fail to capture the complex and non-linear electrochemical interactions governed by surface adsorption dynamics (Wolkenstein isotherm). To address these limitations, this paper proposes a Residual Graph Attention Network (ResGAT) based approach to model the sensor array. Unlike standard “black box” approaches, the ResGAT architecture models the sensor array as a graph architecture. The edges represent the physical and environmental influence pathways. This integration of mathematical feature transformation with graph deep learning allows the model to disentangle the target gas from interferents and thermal drift. We validated the approach using a hybrid dataset containing raw signals from MQ-7, MQ-135 and DHT-11 sensors. Experimental results demonstrate that the ResGAT achieves a Coefficient of Determination ($R^{2}$) of 0.982 and a Root Mean Square Error (RMSE) of 2.45 PPM. This performance significantly outperforms baseline methods, including Random Forest ($R^{2}=0.979$) and MLP ($R^{2}=0.964$), confirming the efficacy of topology-aware calibration.