This is a automatic modulation recognition project based on deep learning (DL). In this project, we generated four different types of datasets and provided two neural networks including CNN and Transformer as feature extractors to classify the modulated signals.
- Dataset
We provide 4 datasets including the popular modulation methods in 5G and 6G. The datasets can be obtained by linking https://mailnwpueducn-my.sharepoint.com/:u:/g/personal/houdongbin_mail_nwpu_edu_cn/Ef8WIcCdVwFGhXZyQhUk-w0BZOb0MluwHo-rDzm8jFTR3A?e=qlxUet
The "Gauss" dataset contains the OFDM dataset under Gaussian channel conditions. The dataset contains 6 types of OFDM modulations: BPSK, QPSK, 8PSK, 16QAM, 64QAM and 256QAM, with signal-to-noise ratios ranging from -10dB to 20dB at intervals of 2 dB. There are 2000 pieces of data generated for each signal type under a specific SNR and the dataset has a total of 192,000 data. Before the signal is transmitted, it undergoes signal processing steps such as channel coding, modulation, serial-to-parallel conversion, spread spectrum and pulse forming, respectively.
The "Rayleigh" dataset contains the OFDM signals in Rayleigh channel conditions, and its basic information is same as the OFDM dataset in Gaussian channel conditions. The difference is that the signal-to-noise ratio covers -20dB to 10dB at intervals of 2 dB.
The "OTFS" dataset contains the OTFS modulation signals. OTFS modulation [1] is a modulation technology for high-speed environment (strong Doppler effect, such as UAV communication, high-speed railway communication, etc.) and is considered as a key technology for 6G. In this project, we generate OTFS signals under high speed channel, and the dataset contains 6 OTFS signals of BPSK, QPSK, 8PSK, 16QAM, 64QAM and 256QAM with signal-to-noise ratios ranging from -20dB to 10dB at intervals of 2 dB.
The "real_channel_OFDM" dataset contains the OFDM signal dataset generated by our HackRF hardware in a real channel, which contains 12 modulation classes, each containing 30,000 signal data. The hardware consists of two HackRFs in an indoor environment with complex electromagnetic interference (computers, cell phones and other devices) at a distance of between 5 and 10m. We generate 12 modulated OFDM baseband signals with header modulation and payload modulation as BPSK+BPSK, BPSK+QPSK, BPSK+8PSK, BPSK+16QAM, BPSK+64QAM, BPSK+DQPSK, QPSK+BPSK, QPSK+QPSK, QPSK+8PSK, QPSK+16QAM, QPSK+64QAM, QPSK+DQPSK, respectively.
- Feature extractor
We designed two feature extractors to extract modulated signal features to achieve accurate recognition of modulated signals.
OFDMmodel is a multilayer convolutional neural network based on CNN.
GPT is a Transformer model which is similar as GTP-2/GTP-3 (at a much smaller scale than them) [2], which includes MASK layers, position encoding, etc.
We created this project to contribute OFDM and OTFS modulation signals to the study of automatic modulation identification (which few researchers have mentioned before), as well as to help beginners in signal modulation identification to learn the basics knowledge quickly. If this is helpful to your research, please cite the paper:
@ARTICLE{10102600, author={Lin, Wensheng and Hou, Dongbin and Huang, Junsheng and Li, Lixin and Han, Zhu}, journal={IEEE Transactions on Vehicular Technology}, title={Transfer Learning for Automatic Modulation Recognition Using a Few Modulated Signal Samples}, year={2023}, volume={}, number={}, pages={1-5}, doi={10.1109/TVT.2023.3267270}}
Reference
[1] R. Hadani et al., "Orthogonal Time Frequency Space Modulation," 2017 IEEE Wireless Communications and Networking Conference (WCNC), San Francisco, CA, USA, 2017, pp. 1-6.
[2] Brown T, Mann B, Ryder N, et al, "Language models are few-shot learners," in Advances in neural information processing systems, vol. 33, no. 1, pp. 1877-1901, Jun. 2020.