Spectrum Prediction With Deep 3D Pyramid Vision Transformer Learning
This paper uses three spectrum datasets collected in the real world, including frequency-modulated (FM) spectrum dataset, long-term evolution (LTE) spectrum dataset, and cross-validation spectrum dataset, to prove the effectiveness of the proposed methods. The datasets are collected at the Jiangjun Road Campus of Nanjing University of Aeronautics and Astronautics (NUAA) in Nanjing, China. You can obtain all datasets from Google Drive. Specific details about these three datasets are as follows:
Figure 1. Spectrum sensor map (left) and actual sensor (right).
The FM dataset is collected by a spectrum measurement sensor located at [118.7905 (east longitude), 31.9378 (northern latitude), 12.10 (altitude)] on the Jiangjun Road Campus of the NUAA in Nanjing, China (see Fig. 1 (left) node 1, and the actual sensor is shown in Fig. 1 (right)). The collected bandwidth is 90 MHz-110 MHz. The collected data types are the in-phase (I) and quadrature (Q) signals. The range of collection time is from 17:20 on Sep. 23rd, 2022, to 20:20 on Sep. 23rd, 2022, with a sampling interval of 1 second.
We convert the FM I/Q data collected every second to a spectrogram via the short-time Fourier transform (STFT). The configurations are: the sampling frequency is 125 MHz, the descending sampling coefficient is 4, the STFT number is 32508, the center frequency of the FM is 99 MHz, and the length-window is 256. These spectrograms constitute the FM dataset. We split the FM dataset into the training set (7200 samples with 17:20-19:20), validation set (1800 samples with 19:20-19:50), and test set (1800 samples with 19:50-20:20) with a 4:1:1 ratio in chronological order.
This dataset is used to demonstrate the superiority of the proposed methods over the baseline methods.
The LTE dataset is collected by a spectrum measurement node located at [118.7905 (east longitude), 31.9378 (northern latitude), 12.10 (altitude)] on the Jiangjun Road Campus of the NUAA in Nanjing, China (see Fig. 1 node 1). The collected bandwidth is 690 MHz-710 MHz. The data type is the same as the FM spectrum dataset. The range of collection time is from 17:52 on May 2nd, 2023, to 18:32 on May 2nd, 2023, with a sampling interval of 1 second.
The STFT is also used to convert LTE I/Q data into spectrogram series. The parameters are set the same as for the FM dataset except that the STFT number is 16254 and the center frequency is 700 MHz.
This dataset is used to demonstrate the transfer learning performance of the proposed methods.
The cross-validation dataset is collected by a spectrum measurement node located at [118.7907 (east longitude), 31.9386 (northern latitude), 36.80 (altitude)] on the Jiangjun Road Campus of the NUAA in Nanjing, China (see Fig. 1 node 2). The bandwidth and data type collected are the same as the FM dataset. The range of collection time is from 17:52 on May 2nd, 2023, to 18:32 on May 2nd, 2023, with a sampling interval of 1 second. The collection time was from 22:14 on Jul. 10, 2024, to 00:04 on Jul. 11, 2024, with a sampling interval of 1 second.
As with the first two datasets, the STFT is used to convert the I/Q data into spectrogram series. The parameter configuration is the same as that of the FM dataset.
This dataset is used to cross-validate the superiority of the proposed methods over the baseline methods.
If you find this repository useful, please cite our paper.
@article{pgl2024,
title={Spectrum prediction with deep {3D} pyramid {Vision Transformer} learning},
author={Guangliang Pan and Qihui Wu and Bo Zhou and Jie Li and Wei Wang and Guoru Ding and David K.Y. Yau},
journal={IEEE Transactions on Wireless Communications},
year={2025},
volume={24},
number={1},
pages={509-525}}