A neural network-based signal modulation recognition system implemented on a FPGA

DOI: 10.34759/trd-2021-121-13

Аuthors

Bakhtin A. A.^*, Volkov A. S.^**, Solodkov A. V.^***, Sviridov I. A.^****

National Research University of Electronic Technology, Bld. 1, Shokin Square, Zelenograd, Moscow, Russia, 124498

*e-mail: bah@miee.ru
**e-mail: leshvol@mail.ru
***e-mail: solodkov_aw@mail.ru
****e-mail: igor@mekatto.com

Abstract

In cognitive radio systems including software-defined radio, an important task is to recognize modulation type of received signals under various signal-to-noise ratios in the communication channel. The detection of the modulation type in a received data packet can be used in ad hoc networks, as well as to provide dynamic spectrum access.

To solve this problem, there are few common approaches, including the shape detection of the signal constellation, the study of the statistical characteristics of the signal, the use of deep neural networks and others. The use of deep convolutional neural networks leads to higher accuracy for large sets of different types of modulation. In addition, implementing the neural network on an FPGA allows not only changing the weights of the neural network, but also configuring the types and arrangement of layers without replacing the hardware component.

The proposed system consists of hardware and software parts. The hardware part includes a Digilent Zedboard and an AD-FMCOMMS3-EBZ development board (based on an AD9361 configurable radio transceiver connected via an FMC connector). The software part of the system consists of the Petalinux distribution kit version 2019.1, the Linux industrial I/O driver, developed software for pre-processing the received signal and a trained neural network model.

The developed software part performs the pre-processing of the signals and controls the configurable coprocessor located on the FPGA. The pre-processing consists in normalizing the received signal. Moreover, as the coprocessor does not support the softmax layer of the neural network, the necessary calculations are performed at the post-processing stage.

To train the neural network, a set of samples of radio signals with different types of modulation (dataset) was generated in the Matlab environment. The generated signal samples were transmitted at a 1 GHz carrier frequency over a low noise wireless channel. The carrier frequency was chosen arbitrarily as it does not affect the processing algorithm. Before transmitting the signal, the transceivers were manually calibrated in order to reduce the frequency desynchronization, however, during the transmission; the frequency deviation value was being changed arbitrarily. The received signals were pre-processed and divided into bursts of 1024 samples in the in-phase and quadrature channels. Then AWGN with different power was added to the samples. Thus, samples with various SNR in range −5...15 dB with step 1 dB were formed in the dataset.

The use of the DPU v3.2 coprocessor allows it to perform necessary computations for the neural network in the FPGA. We decided to use a compact neural network with a small number of Inception modules and fast connections.

The most significant impact on errors in determining modulation at high SNR is the incorrect classification of the 8-PSK and 16-PSK modulations — instead of the first type, the second is often predicted and vice versa. At low SNR QAM modulations of different orders are erroneously classified as higher order QAMs due to the influence of noise.

Despite this fact, the developed system shows an average accuracy of 90% of successful recognitions for SNR values above 12 dB and 70% and higher for SNR values exceeding 2 dB. The proposed system has great flexibility and a large possibilities for further improvement of performance.

Keywords:

signal receiving, digital modulation, cognitive radio, neural networks, automatic modulation recognising, FPGA, Digilent Zedboard, AD9361

References

1. O’Shea T.J., Roy T., Clancy T.C. Over-the-air deep learning based radio signal classification, IEEE Journal of Selected Topics in Signal Processing, 2018, vol. 12, no. 1, pp. 168-179. DOI:10.1109/JSTSP.2018.2797022
2. Amjad M., Rehmani M.H., Mao S. Wireless multimedia cognitive radio networks: A comprehensive survey, IEEE Communications Surveys & Tutorials, 2018, vol. 20, no. 2, pp. 1056-1103. DOI:10.1109/COMST.2018.2794358
3. Bychkov E.D., Kovalenko O.N., Belenkov D.V. Dinamika sistem, mekhanizmov i mashin. 2019, vol. 7, no. 4. DOI: 10.25206/2310-9793-7-4-192-199
4. Zuev A.V. Sistemy upravleniya, svyazi i bezopasnosti, 2019, no. 3, DOI: 10.24411/2410-9916-2019-10302
5. Shevtsov V.A., Borodin V.V., Krylov M.A. Trudy MAI, 2016, no. 85. URL: http://trudymai.ru/eng/published.php?ID=66417
6. Wang Y. et al. Data-driven deep learning for automatic modulation recognition in cognitive radios, IEEE Transactions on Vehicular Technology, 2019, vol. 68, no. 4, pp. 4074-4077. DOI:10.1109/TVT.2019.2900460
7. Ibnkahla M. Adaptation in wireless communications Taylor & Francis Group, 2009, 485 p.
8. Borodin V.V., Petrakov A.M., Shevtsov V.A. Trudy MAI, 2015, no. 81. URL: http://trudymai.ru/eng/published.php?ID=57894
9. Borodin V.V., Petrakov A.M., Shevtsov V.A. Elektrosvyaz’, 2016, no. 11, pp. 41-45.
10. Rembovskii A.M. Radiomonitoring. Zadachi, metody, sredstva (Radiomonitoring. Objectives, methods, means), Moscow, Goryachaya liniya-Telekom, 2010, 623 p.
11. Adzhemov S.S., Klenov N.V., Tereshonok M.V., Chirov D.S. Vestnik Moskovskogo universiteta. Seriya 3. Fizika. Astronomiya, 2015, no. 6, pp. 19-27.
12. O’Shea T.J., Corgan J., Clancy T.C. Convolutional radio modulation recognition networks, International conference on engineering applications of neural networks, Springer, Cham, 2016, pp. 213-226. DOI:10.1007/978-3-319-44188-7_16
13. Hazza A. et al. An overview of feature-based methods for digital modulation classification, 2013 1st international conference on communications, signal processing, and their applications (ICCSPA), IEEE, 2013, pp. 1-6. DOI:10.1109/ICCSPA.2013.6487244
14. Pu D., Cozma A., Hill T. Four quick steps to production: Using model-based design for software-defined radio, Analog Dialogue, 2015, vol. 49, pp. 1-5.
15. Tridgell S. Low Latency Machine Learning on FPGAs, Ph.D, Thesis, University of Sydney, Sydney, Australia, 2020.
16. Xilinx DPU Product Guide. URL: https://www.xilinx.com/support/documentation/ip_documentation/dpu/v3_2/pg338-dpu.pdf
17. Kedia R., Goel S., Balakrishnan M., Paul K., Sen R. Design Space Exploration of FPGA-Based System With Multiple DNN Accelerators, IEEE Embedded Systems Letters, 2021, vol. 13, no. 3, pp. 114-117.
18. Simonyan K., Zisserman A. Very deep convolutional networks for large-scale image recognition, arXiv preprint: arXiv:1409.1556, 2014.
19. He K., Zhang X., Ren S., Sun J. Deep residual learning for image recognition, Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770-778. DOI:10.1109/CVPR.2016.90
20. Szegedy C. et al. Going deeper with convolutions, Proceedings of the IEEE conference on computer vision and pattern recognition, 2015, pp. 1-9. DOI:10.1109/CVPR.2015.7298594
21. Peng S. et al. Modulation classification based on signal constellation diagrams and deep learning // IEEE transactions on neural networks and learning systems, 2018, vol. 30, no. 3, pp. 718-727. DOI:10.1109/TNNLS.2018.2850703

Download