Design of Enhanced Multicomponent Gas Sensor Based on Neural Network Cross-Architecture Knowledge Distillation Algorithm

To satisfy the demand for high-precision online monitoring of early signature gases, carbon dioxide (CO2) and carbon monoxide (CO), in lithium-ion battery thermal runaway, we propose a multi-component gas sensing system based on a Convolutional Neural Network (CNN) via Cross-Architecture Knowledge Distillation (CAKD). Prior knowledge learned by the teacher model from noisy second harmonic (2ƒ) signals is distilled into a structurally simpler student model, reducing inference latency and computational burden. Under laboratory conditions, the system achieved concentration fitting accuracies of 0.9997 for CO2 and 0.9990 for CO. The mean absolute errors (MAE) for CO2 and CO were 62.615 and 30.110 ppm, respectively, and the mean relative errors (MRE) were 0.500% and 1.759%. Allan deviation analysis further yields detection limits of 3.1171 ppm for CO2 at an averaging time of 146 s and 1.6604 ppm for CO at 122 s. Deployed on a Raspberry Pi, the system achieved an average per-sample inference time of 0.2037 ms, satisfying real-time online monitoring requirements. The model parameter count and computational cost were reduced by 88.47% and 94.10%, respectively, compared with the teacher model. The sensor was further applied to real-time monitoring of CO2 and CO during lithium-ion battery thermal runaway, successfully identifying characteristic gas signals 874 s before thermal runaway onset. These results demonstrate that CAKD significantly reduces model complexity and inference latency while maintaining accurate trace gas detection, demonstrating its application value in embedded safety monitoring systems.