Unsupervised Attention-Based Multisource Domain Adaptation Framework for Drift Compensation in Electronic Nose Systems

Reliable gas identification in industrial environments is critical for safety monitoring. However, electronic nose (E-nose) systems suffer from reduced accuracy over time due to gas sensor drift. Most existing studies focus on single-source domain adaptation, with limited exploration of multi-source scenarios involving unlabeled target data, especially when multiple source domains are available at the initial stage. To address this challenge, we propose an unsupervised attention-based multi-source domain shared-private feature fusion adaptation (AMDS-PFFA) framework for drift-resilient gas identification. AMDS-PFFA leverages labeled data from multiple source domains collected in the early stage to accurately identify gases in unlabeled target domain drift signals. Extensive experiments on the UCI drift dataset (36 months) and a self-developed E-nose dataset (30 months) demonstrate that AMDS-PFFA achieves superior performance, with average accuracies of 83.20% and 93.96%, respectively, significantly outperforming recent drift compensation methods.