摘要
Long-term drift in metal oxide gas sensor arrays remains a major obstacle for reliable electronic-nose deployment. Recent work has shown that meta-learning can mitigate drift on legacy feature-engineered datasets, but the behavior of modelagnostic meta-learning (MAML) on raw time-series responses from modern electronic noses is still unclear. In this work we present an initial empirical study of MAML on the "Long-term drift behavior in metal oxide gas sensor arrays" one-year dataset, which provides 700 recordings from a 62-sensor commercial electronic nose with both raw signals and pre-extracted features. We first establish a supervised 1D residual network baseline trained on early days and tested on later days, confirming strong accuracy degradation due to drift. We then define day-wise and batched-day meta-learning tasks and apply MAML directly to the raw multi-channel sensor responses under few-shot adaptation. Our results show that naïvely applying MAML on this dataset does not automatically outperform a well-tuned supervised baseline; instead, performance is sensitive to task construction and optimization details. We situate these findings with respect to prior feature-based meta-learning work and discuss implications for future electronic-nose algorithms on long-term drift datasets.