摘要
Aging in electronic-nose sensors induces batch-wise distribution drift, severely degrading cross-batch classification accuracy. To address this problem, we propose a drift-compensation framework that combines 1D CycleGAN-based pseudo-domain generation [1], [3], Domain-Adversarial Neural Networks (DANN) [2], and a support vector machine (SVM) ensemble. Three pair-specific CycleGAN mappings are first learned from the shared classes (2-4) and then applied to source-domain unseen gases (5-6) to generate pseudo target representations. These pseudo samples are used for downstream classification and adversarial alignment, while real target samples of gases 5-6 remain unlabeled during training. In this framework, CycleGAN captures transferable batch-shift geometry from shared classes, whereas DANN further aligns pseudo and real target feature distributions to improve robustness. Experiments on three representative drift scenarios-severe drift with sparse shared data (1→6), moderate and coherent drift (\mathbf{6} \rightarrow \mathbf{1 0}) , and highly nonlinear drift with the lowest baseline accuracy (\mathbf{9} \rightarrow \mathbf{1 0}) -show that the proposed method consistently outperforms conventional baselines. In the most challenging \mathbf{9} \rightarrow \mathbf{1 0} case, baseline SVM achieves only 20.33% accuracy, whereas the proposed method reaches \mathbf{8 5}-\mathbf{8 8} \boldsymbol{\%} . These results indicate that drift geometry extracted from shared classes provides transferable structure for compensating unseen-gas drift and substantially improves cross-batch classification robustness in MOX-based electronicnose systems [4]-[6].