EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Time-Aware Transformer With Wavelet Transform for Missing-Resilient GNSS Denoising

Fang Xinjian, Guo Yuxiang, Yang Xu

IEEE Access · 2026 · 14 · 169823-169834

DOI: 10.1109/access.2026.3744083 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Fang Xinjian und 2 weiteren Autor:innen. Die Arbeit nennt 22 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Open Research Fund of Anhui Province Engineering Laboratory for Big Data Analysis and Early Warning Technology of Coal Mine Safety, National Natural Science Foundation of China, Engineering Research Center of Mining Area Environmental and Disaster Cooperative Monitoring.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

GNSS coordinate time series denoising under missing data conditions remains a persistent challenge. Conventional methods require strictly continuous, equally spaced observations and resort to interpolation prior to denoising, which amplifies errors through Runge’s phenomenon during prolonged gaps. Deep learning methods tolerate missingness but lack the fine-grained noise suppression capability of traditional approaches. To address these limitations, we propose a cascaded missing-resilient denoising framework—TATRWT (Time-Aware Transformer Regressor with Wavelet Transform). In the first stage, a time-aware Transformer regressor extracts a continuous macroscopic trend directly from sparse, gap-containing series by leveraging a masked loss function and explicit injection of annual and semi-annual periodic features, thereby circumventing any interpolation pre-processing. In the second stage, wavelet soft-thresholding strips residual high-frequency noise from the trend, combining the complementary strengths of neural networks and classical signal processing. We validate the framework using real observations from five GNSS reference stations in North China, introducing holdout testing as a generalization audit. Comparative analysis with conventional spline-interpolation denoising and a pure neural-network baseline reveals that Spline+Wavelet ranks best on fidelity yet worst on holdout generalization—an evaluation trap. The pure TATR baseline matches TATRWT in holdout error yet exhibits a Smoothness of 0.85 versus 0.18 for TATRWT, producing physically implausible high-frequency artifacts. TATRWT preserves the baseline’s masked prediction fidelity while reducing Smoothness by nearly a factor of five. This confirms that wavelet post-processing removes high-frequency components lacking generalization capacity. The results demonstrate that fidelity alone is dangerously misleading for assessing denoising quality under missing data, and that TATRWT achieves the best overall performance under a three-dimensional evaluation framework encompassing accuracy, smoothness, and generalization.

Förderung

  • Open Research Fund of Anhui Province Engineering Laboratory for Big Data Analysis and Early Warning Technology of Coal Mine Safety
  • National Natural Science Foundation of China
  • Engineering Research Center of Mining Area Environmental and Disaster Cooperative Monitoring

Zitieren

APA

Xinjian, F., Yuxiang, G., & Xu, Y. (2026). Time-Aware Transformer With Wavelet Transform for Missing-Resilient GNSS Denoising. IEEE Access, 14, 169823-169834. https://doi.org/10.1109/access.2026.3744083

Vancouver

Xinjian F, Yuxiang G, Xu Y. Time-Aware Transformer With Wavelet Transform for Missing-Resilient GNSS Denoising. IEEE Access. 2026;14:169823-169834. doi:10.1109/access.2026.3744083

BibTeX
@article{Xinjian2026,
  author = {Xinjian, Fang and Yuxiang, Guo and Xu, Yang},
  title = {Time-Aware Transformer With Wavelet Transform for Missing-Resilient GNSS Denoising},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {169823--169834},
  doi = {10.1109/access.2026.3744083}
}
Kostenlos starten →