Deep Learning-based Modulation Detection for NOMA Systems

Wenwu Xie, Jian Xiao, Jinxia Yang, Ji Wang, Xin Peng, Chao YuarXiv 2020 · DOI · arXiv

Abstract

Since the signal with strong power should be demodulated first for successive interference cancellation (SIC) demodulation in non-orthogonal multiple access (NOMA) systems, the base station (BS) should inform the near user terminal (UT), which has allocated higher power, of modulation mode of the far user terminal. To avoid unnecessary signaling overhead in this process, a blind detection algorithm of NOMA signal modulation mode is designed in this paper. Taking the joint constellation density diagrams of NOMA signal as the detection features, deep residual network is built for classification, so as to detect the modulation mode of NOMA signal. In view of the fact that the joint constellation diagrams are easily polluted by high intensity noise and lose their real distribution pattern, the wavelet denoising method is adopted to improve the quality of constellations. The simulation results represent that the proposed algorithm can achieve satisfactory detection accuracy in NOMA systems. In addition, the factors affecting the recognition performance are also verified and analyzed.

The NOMA Signaling Challenge

In the push for 5G and massive IoT, NOMA (Non-Orthogonal Multiple Access) has emerged as a key technology to improve spectral efficiency. Unlike traditional methods that give each user their own frequency or time slot, NOMA stacks multiple users on the same resource by varying their power levels.

NOMA System Model and Power Allocation

To recover their own data, a "near user" (with a better channel) must first demodulate and subtract the "far user's" signal using Successive Interference Cancellation (SIC). Traditionally, the Base Station (BS) must send extra signaling data to tell the near user exactly which modulation format (e.g., QPSK vs. QAM16) the far user is using. In a massive IoT network with millions of devices, this extra signaling creates massive overhead and latency. This research proposes a method to enable the near user to "blindly" detect the modulation format on its own.

Proposed Methodology

The authors propose a three-step pipeline to move from raw radio waves to a classification result:

  1. Wavelet Denoising: NOMA signals are often buried in noise. The authors use Wavelet Transforms (WT) with "soft thresholding" to scrub noise while preserving the underlying constellation pattern.
  2. Joint Constellation Density Extraction: Instead of feeding raw I/Q samples into a neural network, the authors plot the signal on a 2D plane. They then count the number of points in specific regions to create a grayscale density image, transforming a telecommunications problem into a computer vision task.
  3. Deep Residual Network (ResNet): A ResNet architecture takes the $100 \times 100$ density matrix as input and outputs the detected modulation mode. ResNets are particularly effective here due to their ability to handle deep architectures without vanishing gradient issues.
Comparison of Superposition Constellations before and after Power Allocation

Key Insights and Results

  • The "Image" Advantage: By converting signals into density diagrams, the system becomes robust to the number of sample points. As long as the "shape" of the constellation is visible in the density map, the CNN can recognize it.
  • Power Allocation Impact: Detection accuracy is positively correlated with the power difference between users. When the "far user" is allocated significantly more power, its constellation pattern dominates the joint diagram, making it much easier for the ResNet to identify.
  • Performance: For 2-user NOMA, the system achieves near 100% accuracy once the SNR reaches approximately 6–10 dB. Lower-order modulations (like $\pi/2$-BPSK) are detected much more accurately at low SNRs than higher-order ones (like QAM64).
  • Denoising Efficacy: The study demonstrates that the proposed denoising and density extraction significantly outperforms a "Raw ResNet" that attempts to learn from unprocessed, noisy constellations.

Limitations and Future Work

While effective for standard 5G formats, the authors acknowledge that higher-order modulations like 256QAM present a significant challenge due to their extreme density. Furthermore, as more users share a sub-channel, the joint constellation becomes increasingly cluttered, causing accuracy to drop. Future work will focus on high-density NOMA scenarios and testing on real-world software-defined radio (SDR) hardware to account for hardware impairments like phase noise.

Summary updated

Abstract

Since the signal with strong power should be demodulated first for successive interference cancellation (SIC) demodulation in non-orthogonal multiple access (NOMA) systems, the base station (BS) should inform the near user terminal (UT), which has allocated higher power, of modulation mode of the far user terminal. To avoid unnecessary signaling overhead in this process, a blind detection algorithm of NOMA signal modulation mode is designed in this paper. Taking the joint constellation density diagrams of NOMA signal as the detection features, deep residual network is built for classification, so as to detect the modulation mode of NOMA signal. In view of the fact that the joint constellation diagrams are easily polluted by high intensity noise and lose their real distribution pattern, the wavelet denoising method is adopted to improve the quality of constellations. The simulation results represent that the proposed algorithm can achieve satisfactory detection accuracy in NOMA systems. In addition, the factors affecting the recognition performance are also verified and analyzed.
BibTeX
@misc{xie2020deep,
  author = {Wenwu Xie and Jian Xiao and Jinxia Yang and Ji Wang and Xin Peng and Chao Yu},
  title = {{Deep Learning-based Modulation Detection for NOMA Systems}},
  year = {2020},
  eprint = {2005.11684},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2005.11684},
}
Figure