A Wise Multi-Scale Encoder Based on Additive Angular Margin Loss and Kullback-Leibler (KL) Divergence for Gas Sensor Drift

Sensor drift compensation is a common issue in the field of gas sensors. Due to unavoidable chemical properties such as sensor aging and poisoning, there often exists a significant difference in probability distribution between collected data and target data. Recently, due to the excellent performance of attention mechanisms (transformers) on time series data, many studies have applied them to gas recognition tasks. This paper proposes a multi-scale encoder based on Additive Angular Margin Loss (AAML) for gas classification tasks, along with using KL divergence for domain adaptation to address drift compensation issues. The proposed model, WMSE-AAML, achieves a classification accuracy of 97.75% on a public gas sensor array dataset and WMSE-DA achieves nearly 95% prediction accuracy in the drift compensation task of predicting the target domain using the source domain.