H₂ and CO identification and concentration prediction using a pulse-driven MEMS gas sensor with multi-feature, multi-task learning

Metal-oxide-semiconductor (MOS) gas sensors are attractive for compact monitoring but remain constrained by cross-reactivity and reliance on steady-state measurements. Here, a suspended Cu–SnO2 MEMS chemiresistor was integrated with a multi-feature, multi-task temporal network for hydrogen (H2)/carbon monoxide (CO) identification and quantification. The heater cycled between approximately 90 °C for 0.8 s and 350 °C for 1.6 s at an average power of 19.8 mW. Currents recorded at 10 Hz during 200–900 ppm exposures were transformed into 4.8 s moving-average, Hilbert-envelope, and temporal-derivative channels. Sliding time windows were analyzed using a multi-scale convolutional neural network (CNN)–squeeze-and-excitation (SE)–bidirectional long short-term memory (BiLSTM) architecture with attention-based classification and gas-specific regression heads. For separately acquired test sequences at 200 and 250 ppm, the three-model ensemble achieved classification accuracy of 99.95% and a gas-only mean absolute error (MAE) of 25.79 ppm using a 12 s window with a 0.1 s stride in offline replay. These results demonstrate a compact, low-power approach to multi-task gas identification and concentration prediction using a single microsensor.