Fast Radio Burst Classification on Dispersed Dynamic Spectra Using a Wavelet-based Convolutional Network
7 affiliations
- 1Southwest Petroleum University
- 2Qilu Normal University
- 3Shandong Key Laboratory of Space Environment and Exploration Technology
- 4State Key Laboratory of Radio Astronomy and Technology
- 5Swinburne University of Technology
- 6University of Nottingham Ningbo China
- 7Shanghai Dianji University
Abstract
Fast radio bursts (FRBs) are millisecond-duration extragalactic radio transients of great scientific value for studying the intergalactic medium and extreme astrophysical environments. Large-scale radio surveys such as the Five-hundred-meter Aperture Spherical radio Telescope (FAST) produce petabytes of observational data annually, making efficient automated FRB identification an urgent priority. Traditional FRB search pipelines rely on computationally expensive dedispersion preprocessing, and existing deep learning methods struggle to extract weak dispersed burst features from heavy background and radio frequency interference (RFI) contamination. In this work, we address image-level FRB/non-FRB classification on dispersed dynamic spectra without dedispersion preprocessing. We propose FRBWaveletNet, a wavelet-enhanced convolutional framework that embeds explicit multiscale frequency decomposition into the network backbone. This design explicitly targets the multiscale morphological characteristics of dispersed FRBs, improving the representation of weak burst signals masked by interference. On a real FAST test set constructed from observations of FRB 20201124A, FRBWaveletNet attains 99.09% accuracy and an F1-score of 98.94%, outperforming matched convolutional neural network (CNN) baselines. Applied to four FAST repeating sources using sliding-window inference, it recovers 599 of 679 cataloged bursts. The results support the use of wavelet-based multiscale representation for FRB classification and characterize cataloged-burst recovery and false-positive candidate output in burst-associated FAST observations.
Related
- Bayesian classification of astronomical spectra with class uncertainties
- Seeing Through the RFI Mask: DT-FISTA Inpainting of Pulsar Dynamic Spectra via Fourier Sparsity
- HierArchical Wavelet Coefficients on the Sphere (HAWCS): the Scattering Transform on Spherical Maps
- A Wavelet-Integrated Search Pipeline for Narrowband Technosignatures in FAST Observations of 33 Exoplanet Systems
- The Wavelet Detection Filter: a real time unmodeled pipeline for gravitational wave transients, ranking coincidences with a graph neural network
BibTeX
@article{gu2026fast,
title = {{Fast Radio Burst Classification on Dispersed Dynamic Spectra Using a Wavelet-based Convolutional Network}},
author = {Xinyu Gu and Wangxuan Zhu and Xianfeng Yang and Zerui Wang and Lei Zhang and Zheng Lu and Jie Zhang},
journal = {The Astrophysical Journal},
year = {2026},
month = oct,
doi = {10.3847/1538-4357/aea21a},
url = {https://iopscience.iop.org/article/10.3847/1538-4357/aea21a},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Fast Radio Burst Classification on Dispersed Dynamic Spectra Using a Wavelet-based Convolutional Network
AU - Xinyu Gu
AU - Wangxuan Zhu
AU - Xianfeng Yang
AU - Zerui Wang
AU - Lei Zhang
AU - Zheng Lu
AU - Jie Zhang
PY - 2026
DA - 2026/10/01
JO - The Astrophysical Journal
DO - 10.3847/1538-4357/aea21a
UR - https://iopscience.iop.org/article/10.3847/1538-4357/aea21a
AB - Fast radio bursts (FRBs) are millisecond-duration extragalactic radio transients of great scientific value for studying the intergalactic medium and extreme astrophysical environments. Large-scale radio surveys such as the Five-hundred-meter Aperture Spherical radio Telescope (FAST) produce petabytes of observational data annually, making efficient automated FRB identification an urgent priority. Traditional FRB search pipelines rely on computationally expensive dedispersion preprocessing, and existing deep learning methods struggle to extract weak dispersed burst features from heavy background and radio frequency interference (RFI) contamination. In this work, we address image-level FRB/non-FRB classification on dispersed dynamic spectra without dedispersion preprocessing. We propose FRBWaveletNet, a wavelet-enhanced convolutional framework that embeds explicit multiscale frequency decomposition into the network backbone. This design explicitly targets the multiscale morphological characteristics of dispersed FRBs, improving the representation of weak burst signals masked by interference. On a real FAST test set constructed from observations of FRB 20201124A, FRBWaveletNet attains 99.09% accuracy and an F1-score of 98.94%, outperforming matched convolutional neural network (CNN) baselines. Applied to four FAST repeating sources using sliding-window inference, it recovers 599 of 679 cataloged bursts. The results support the use of wavelet-based multiscale representation for FRB classification and characterize cataloged-burst recovery and false-positive candidate output in burst-associated FAST observations.
SN - 0004-637X
SN - 1538-4357
ER -