A project on RF modulation classification using different neural architectures and RF signal representations.
I explored various neural architectures (CNN and LSTM variants), different data representations (IQ, amplitude-phase, constellation diagrams) and different RF datasets (radioML, Matlab comms toolbox) and tried to answer the following questions:
- What is the best model architecture?
- What is the best feature?
- How well do models adapt to different channel conditions?
Initially, I expected CNN models trained on RF data represented as coloured constellation images to be the best performing, as it made sense that CNNs would perform well on a standard image classification task.
However, it was surprising that my 1D CNNs working on time-series outperformed the constellation CNN models, particularly in low SNR conditions. In fact, I found that a CNN + LSTM architecture with skip connections looks promising for exploiting patterns in time-series features, and that amplitude-phase time-series improves classification accuracy at high SNRs.
Finally, I also found that the models tried did not adapt well to datasets exposed to diferent channel conditions. Surprisingly, even when a model was trained on a dataset with more difficult channel conditions, it was not able to perform that well when tested on a dataset with less harsh channel conditions.
These models were chosen because they represent a wide range of popular architectures now and there was enough information in the papers for me to reproduce them.
Open the notebooks in Colab which has a moving content page to make it more navigable. The first part of every notebook is a lot of helper functions and setting up, can be ignored.
This notebook explores the radioML datasets, in different data representations: constellation diagram, IQ time-series, amplitude-phase time-series and power density spectra. It also visualises the dataets with Rician Fading that I created following this Matlab tutorial.
This notebook builds neural architectures from literature that performed quite well for the task of RF modulation classification and trains them on the 2016.10A radioML dataset. The architectures tried are basic CNN, Inception, ResNet, CLDNN and LSTM. I trained the models with IQ and amplitude-phase (AP) time-series and compared the results. CLDNN-AP was found to be the best performing.
This notebook explores constellation image representations with different resolutions and colour schemes. It was found that increasing resolution improved performance while colour didn't have much of an effect. Be careful of memory usage.
This notebook compares the best performing time-series and constellation models on the 2016.10A radioML dataset. CLDNN-AP outperforms the constellation models, and a model combining both ResNet-IQ and CLDNN-AP performs well in both low and high SNR conditions.
This notebook compares the best performing models on a realistic but more difficut dataset I generated from Matlab (see report for details on data). Here, the models performed a lot worse especially for higher order modulations like QAMs and PSKs, and the constellation model performed significantly better than IQ models for these high order constellations. This suggests the potential of having a model that have both time-series and constellation images as input. Also, I found that models trained on one Matlab dataset didn't perform well on another, which isn't a good sign for modulation recognition in real-life.
[1] T. O’Shea and N. West, “Radio machine learning dataset generation with gnu radio,” Proceedings of the GNU Radio Conference, vol. 1, no. 1, 2016. [Online]. Available: https://pubs.gnuradio.org/index.php/grcon/article/view/11
[2] T. J. O’Shea and J. Corgan, “Convolutional radio modulation recognition networks,” CoRR, vol. abs/1602.04105, 2016. [Online]. Available: http://arxiv.org/abs/1602.04105
[3] N. E. West and T. J. O’Shea, “Deep architectures for modulation recognition,” CoRR, vol. abs/1703.09197, 2017. [Online]. Available: http://arxiv.org/abs/1703.09197
[4] M. Kulin, T. Kazaz, I. Moerman, and E. D. Poorter, “End-to-end learning from spectrum data: A deep learning approach for wireless signal identification in spectrum monitoring applications,” CoRR, vol. abs/1712.03987, 2017. [Online]. Available: http://arxiv.org/abs/1712.03987 23[5] T. J. O’Shea, T. Roy, and T. C. Clancy, “Over the air deep learning based radio signal classification,” CoRR, vol. abs/1712.04578, 2017. [Online]. Available: http://arxiv.org/abs/1712.04578
[6] S. Rajendran, W. Meert, D. Giustiniano, V. Lenders, and S. Pollin, “Distributed deep learning models for wireless signal classification with low-cost spectrum sensors,” CoRR, vol. abs/1707.08908, 2017. [Online]. Available: http://arxiv.org/abs/1707.08908
[7] S. Peng, H. Jiang, H. Wang, H. Alwageed, and Y. Yao, “Modulation classification using convolutional neural network based deep learning model,” in 2017 26th Wireless and Optical Communication Conference (WOCC), 2017, pp. 1–5.
[8] Y. Wang, M. Liu, J. Yang, and G. Gui, “Data-driven deep learning for automatic modulation recognition in cognitive radios,” IEEE Transactions on Vehicular Technology, vol. 68, no. 4, pp. 4074–4077, 2019.
[9] B. Tang, Y. Tu, Z. Zhang, and Y. Lin, “Digital signal modulation classification with data augmentation using generative adversarial nets in cognitive radio networks,” IEEE Access, vol. 6, pp. 15 713–15 722, 2018.