Metric Network for E-Nose Drift Compensation: Few-Shot Learning for Robust Gas Sensing

Abstract

This study introduces the metric drift compensation network (MDCN) to address the issue of sensor drift in electronic noses (E-noses). E-noses mimic the olfactory sense of mammals to detect odors. Sensor drift, which refers to the change in sensor outputs over time, poses a significant challenge to the reliability of E-noses. MDCN utilizes metric learning and few-shot learning (FSL) within a metric learning framework to enhance stability against drift. Its advantage lies in maintaining good classification performance even when there are few samples in the target domain or when new categories emerge in the target domain. We evaluated the performance of MDCN in scenarios of category symmetry (where the source and target domains share the same categories) and category asymmetry (where there are fewer categories in the source domain) on two datasets: the pure gas dataset and the mixed gas dataset. In category symmetry scenarios, MDCN outperformed traditional and advanced methods, demonstrating high accuracy with a minimal number of reference samples. In category asymmetry scenarios, it also showed strong generalization capabilities and high accuracy. Comprehensive ablation experiments were also conducted to prove the rationality of the model architecture and its nondependence on a large number of target domain samples. In addition, tests have shown that the model has good device transferability. Source code can be found at https://github.com/TYaDream/Metric-Drift-Compensation-Network.