Sub-Nyquist spectrum sensing and learning challenge

Yue Gao, Zihang Song, Han Zhang, Sean Fuller, Andrew Lambert, Zhinong Ying, Petri Mähönen, Yonina C. Eldar, Shuguang Cui, Mark D. Plumbley, Clive Parini, Arumugam Nallanathan

Introduction

As wireless communication transitions toward 5G-Advanced and 6G, the industry is increasingly leveraging millimeter-wave (mmWave) and Terahertz (THz) bands. While these frequencies offer massive bandwidth, spectrum utilization remains inefficient. Dynamic Spectrum Access (DSA) provides a pathway for "secondary users" to utilize idle frequencies, but implementing this requires continuous, high-speed spectrum sensing across wide bands.

The primary technical bottleneck is the Nyquist-Shannon sampling theorem. To sense a 2GHz wide band using traditional methods, hardware must sample at a minimum of 4GHz. This requirement leads to prohibitive power consumption, high hardware costs, and a data deluge that often exceeds the capabilities of standard Analog-to-Digital Converters (ADCs). Sub-Nyquist sampling, or compressive sensing, offers a theoretical solution by reconstructing signals from fewer samples, provided the signal is sparse. However, transitioning this from theory to practice involves overcoming challenges such as unknown signal sparsity, rapidly changing dynamic environments, and high computational overhead for real-time signal recovery.

Methodology: The Challenge Framework

To address these hurdles, a global challenge was established to move compressive sensing from theoretical elegance to practical deployment. Rather than proposing a single algorithm, the challenge invites researchers to develop end-to-end solutions comprising innovative sampling architectures (e.g., Modulated Wideband Converters or multicoset samplers) and robust recovery algorithms.

A defining feature of this challenge is its hardware-in-the-loop evaluation. Unlike traditional competitions that rely solely on simulations, submissions are validated on a physical testbed using National Instruments (NI) mmWave SDR systems operating at 28.5 GHz. This ensures that proposed solutions can handle the non-idealities and noise inherent in real radio hardware.

Hardware-in-the-loop evaluation setup featuring the Tx and Rx systems used for real-world validation.

Dataset and Benchmarking

Participants are provided with a specialized dataset to benchmark their algorithms against real-world conditions:

  • Signal Characteristics: Real-world wideband signals, specifically Verizon 5G OFDM with 64-QAM modulation.
  • Bandwidth: A 2GHz complex baseband spanning from -1GHz to 1GHz.
  • Activity Patterns: 1 to 3 active channels (100MHz each) at unknown positions to test sparsity detection.
  • Ground Truth: Data is provided at a Nyquist-rate sampling frequency of 3.072 GHz for direct performance comparison.

Evaluation Metrics

Success in the challenge is determined by a multi-dimensional scoring matrix designed to balance theoretical performance with practical viability:

  1. Sensing Performance (25%): Accuracy in identifying active channels.
  2. Code Quality (25%): Readability, documentation, and reproducibility.
  3. Approach Ingenuity (15%): Novelty in the sampling or recovery architecture.
  4. Cost (20%): A split between sampling rate efficiency and computational overhead.

Conclusion and Future Work

The challenge highlights a significant shift toward machine-learning-based spectrum sensing, which may adapt more effectively to complex radio environments than traditional "greedy" algorithms. While the competition establishes a reference for sub-Nyquist processing, several open problems remain, including detecting secondary user coexistence, optimizing sampling window selection for dynamic speed, and further reducing the latency of recovery algorithms. The goal is to eventually make sub-Nyquist sensing viable for consumer-grade cognitive radios in the 6G era.

Summary updated

BibTeX
@article{gao2021sub,
  author = {Yue Gao and Zihang Song and Han Zhang and Sean Fuller and Andrew Lambert and Zhinong Ying and Petri Mähönen and Yonina C. Eldar and Shuguang Cui and Mark D. Plumbley and Clive Parini and Arumugam Nallanathan},
  title = {{Sub-Nyquist spectrum sensing and learning challenge}},
  journal = {Frontiers of Computer Science},
  year = {2021},
  doi = {10.1007/s11704-021-1275-y},
}
Figure