A Domain Generalization Network With Inverse Gaussian Weighted Contrastive Learning for Gas Identification Under Sensor Drift
IEEE Sensors Journal · 2026-04 · 26 (8) · 12319-12330
DOI: 10.1109/jsen.2026.3671730 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Sensors Journal, 2026-04. Verfasst von Yang Liu und 2 weiteren Autor:innen. Die Arbeit wurde bisher 2-mal zitiert (Crossref) und nennt 34 Literaturangaben. Gefördert von: National Natural Science Foundation of China.
Abstract (Originalsprache)
With the increasing demand for industrial gas detection, methods based on deep learning (DL) are becoming increasingly important for gas recognition. However, sensor drift is inevitable, making recognition models trained on the source domain difficult to adapt to the target domain, thereby reducing gas recognition performance. Therefore, this article proposes a domain generalization (DG) network with inverse Gaussian weighted contrastive learning (WCDG) to achieve gas recognition under sensor drift. The network decouples multidomain features to separate recognition-related and domain-specific features, employs saliency-guided feature fusion to extract domain-invariant features, and applies weighted contrastive constraint to reduce intraclass divergence and enlarge interclass margins, thereby enabling effective transfer of source domain knowledge derived solely from multiple batches of early stage data to a target domain that contains mid-stage and late-stage data. Specifically, a multiscale feature extraction (MSFE) module is first proposed to capture the latent discriminative features of the input samples. Second, adversarial training is incorporated to achieve decoupling between recognition-related and domain-specific features. Subsequently, a saliency-guided feature fusion (SGFF) module is proposed to further extract domain-invariant features without drift. Finally, an inverse Gaussian weighted contrastive learning (IGWC) strategy is proposed to optimize the feature space and enhance discriminability. Experimental results demonstrate that the proposed method achieves average recognition accuracies of 91.80% and 94.00% on a public dataset and a laboratory-collected dataset, respectively, significantly outperforming the compared methods, thereby confirming its superiority under sensor drift and further indicating strong potential for practical applications.
Förderung
- National Natural Science Foundation of China
Zitieren
APA
Liu, Y., Feng, Q., & Shi, Y. (2026). A Domain Generalization Network With Inverse Gaussian Weighted Contrastive Learning for Gas Identification Under Sensor Drift. IEEE Sensors Journal, 26(8), 12319-12330. https://doi.org/10.1109/jsen.2026.3671730
Vancouver
Liu Y, Feng Q, Shi Y. A Domain Generalization Network With Inverse Gaussian Weighted Contrastive Learning for Gas Identification Under Sensor Drift. IEEE Sensors Journal. 2026;26(8):12319-12330. doi:10.1109/jsen.2026.3671730
BibTeX
@article{Liu2026,
author = {Liu, Yang and Feng, Qiaohua and Shi, Yunbo},
title = {A Domain Generalization Network With Inverse Gaussian Weighted Contrastive Learning for Gas Identification Under Sensor Drift},
journal = {IEEE Sensors Journal},
year = {2026},
volume = {26},
number = {8},
pages = {12319--12330},
doi = {10.1109/jsen.2026.3671730}
}
EUVIMED