EUVIMEDEUVIMED entdecken
Crossref

Discrete wavelet transform and multi-scale parallel convolutional network for gas recognition

Jingxiao Li, Lingwei Yu, Hai Yang, Zhe Wang

Measurement Science and Technology · 13.11.2026 · 37 (45) · 456105

DOI: 10.1088/1361-6501/aeb565 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Measurement Science and Technology, online erschienen am 10.11.2026. Verfasst von Jingxiao Li und 3 weiteren Autor:innen. Die Arbeit nennt 31 Literaturangaben. Gefördert von: Natural Science Foundation of China, National Key Research and Development Program of China, Shanghai Municipal Education Commission’s Initiative on Artificial Intelligence-Driven Reform of Scientific Research Paradigms and Empowerment of Discipline Le….

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Electronic nose (e-nose) systems based on metal oxide sensor arrays are widely deployed in food safety, environmental monitoring, and medical diagnosis. However, accurate gas identification remains challenging due to two factors: (1) most deep learning methods insufficiently exploit the transient dynamics of sensor response signals, discarding critical time-frequency information; (2) the limited receptive fields of standard convolutional neural networks hinder the modelling of long-range dependencies across sensor channels and temporal evolution. To address these issues, we propose a discrete wavelet transform and multi-scale parallel convolutional network (DWT-MPCN) that formulates gas recognition as a time-frequency image classification problem. Specifically, a fixed five-level db5 discrete wavelet transform (DWT) module converts multi-channel raw time-domain signals into two-dimensional wavelet coefficient maps without manual feature engineering, preserving both rapid transient details and slow recovery trends. A Multi-scale Parallel Convolution (MPC) module then employs parallel convolutional branches with different kernel sizes to simultaneously capture local transient features and broader contextual patterns while maintaining a lightweight architecture (only 1.75 M parameters). Finally, a convolutional Transformer encoder adaptively weights cross-sensor and cross-scale features to produce discriminative representations. Evaluated on three public e-nose datasets with distinct sensor configurations and gas types, DWT-MPCN achieves mean accuracies of 99.69 ± 0.10 % on the OPEN SAMPLING dataset, 99.11 ± 0.16 % on the TWIN dataset, and 99.82 ± 0.11 % on the large-scale array sensor dataset across five independently seeded runs. Component ablations, capacity-aware comparisons, hyperparameter analyses, and wavelet-parameter sensitivity studies validate the contribution of each component and clarify the design choices. The proposed method provides a compact and interpretable framework for edge-intelligent olfaction systems. Our supporting code is publicly available at https://github.com/ECUSTLiJingxiao/DWT-MPCN-support .

Förderung

  • Natural Science Foundation of China
  • National Key Research and Development Program of China
  • Shanghai Municipal Education Commission’s Initiative on Artificial Intelligence-Driven Reform of Scientific Research Paradigms and Empowerment of Discipline Le…

Zitieren

APA

Li, J., Yu, L., Yang, H., & Wang, Z. (2026). Discrete wavelet transform and multi-scale parallel convolutional network for gas recognition. Measurement Science and Technology, 37(45), 456105. https://doi.org/10.1088/1361-6501/aeb565

Vancouver

Li J, Yu L, Yang H, Wang Z. Discrete wavelet transform and multi-scale parallel convolutional network for gas recognition. Measurement Science and Technology. 2026;37(45):456105. doi:10.1088/1361-6501/aeb565

BibTeX
@article{Li2026,
  author = {Li, Jingxiao and Yu, Lingwei and Yang, Hai and Wang, Zhe},
  title = {Discrete wavelet transform and multi-scale parallel convolutional network for gas recognition},
  journal = {Measurement Science and Technology},
  year = {2026},
  volume = {37},
  number = {45},
  pages = {456105},
  doi = {10.1088/1361-6501/aeb565}
}
Kostenlos starten →