A semi-supervised domain adaptation for cross-gas concentration prediction with multi-target and bidirectional transfer for E-nose system

Although current unsupervised domain adaptation (UDA) methods enable knowledge transfer without labeled data, their performance in regression tasks is often hindered by noisy pseudo-labels in the target domain. This challenge is particularly pronounced in gas concentration prediction, which demands higher label accuracy than classification tasks, yet lacks systematic research. To address this issue, this study proposes a novel semi-supervised cross-domain regression framework named Online-Teacher Domain-Adaptive Regression Network (OT-DARN) for accurate multi-gas concentration estimation. OT-DARN integrates an online teacher-student architecture that incorporates domain adversarial learning, along with a pseudo-label selection mechanism, to enhance both transferability and generalization. Extensive experiments on 15 commercially available gas sensor datasets demonstrate that OT-DARN achieves state-of-the-art performance with a minimum mean absolute error (MAE) of 0.7769 parts per million (ppm) under n-propanol concentrations of 10–100 ppm at (50 ± 10) % relative humidity and a temperature of (25 ± 2) °C, with a maximum R² of 0.9986 while reducing the labeling cost by 70 %. Even under fully unlabeled settings, the model achieves an MAE of 6.1206 and R² of 0.9317. Furthermore, bidirectional transfer experiments reveal no statistically significant differences between migration directions, validating its robustness. OT-DARN also consistently outperforms four representative UDA methods across various source–target combinations. Ablation studies further confirm the individual and joint contributions of key modules. This work pioneers semi-supervised cross-domain regression for gas concentration prediction and offers a cost-efficient, high-precision solution applicable to real-world sensing scenarios.