Deep Spiking Delayed Feedback Reservoirs and Its Application in Spectrum Sensing of MIMO-OFDM Dynamic Spectrum Sharing

Kian Hamedani, Lingjia Liu, Shiya Liu, Haibo He, Yang YiAAAI 2020 · DOI

Introduction

As 5G and future wireless standards evolve, the radio frequency spectrum is becoming an increasingly scarce resource. Dynamic Spectrum Sharing (DSS) has emerged as a key solution, allowing "Secondary Users" to utilize frequency bands when "Primary Users" are inactive. However, implementing DSS effectively requires highly accurate and energy-efficient prediction of subcarrier occupancy. While Recurrent Neural Networks (RNNs) excel at such time-series tasks, their high training complexity and power demands often preclude edge deployment.

The DSDFR Architecture

To address these challenges, this research introduces the Deep Spiking Delayed Feedback Reservoir (DSDFR). This architecture leverages Reservoir Computing (RC), which simplifies training by using a fixed recurrent layer, and Spiking Neural Networks (SNNs), which mimic biological efficiency through pulse-based communication.

Figure 1: The DSDFR architecture, featuring an encoding layer, a reservoir layer with a delay loop and virtual nodes, and a trained output layer.

The system operates in three stages:

  1. Encoding: Analog signal energy is converted into discrete spikes. The authors found that Interspike Interval (ISI) encoding outperforms Time-to-First-Spike (TTFS) because it does not require a global temporal reference, making it more robust for real-time analysis.
  2. Reservoir Layer: The core innovation is a single Leaky Integrate-and-Fire (LIF) nonlinear node coupled with a delay loop. This loop is partitioned into "virtual nodes" that capture complex temporal correlations without the hardware overhead of a massive physical network.
  3. Readout: A Multi-Layer Perceptron (MLP) is trained to classify the reservoir's state as either "busy" or "idle."

Performance and Insights

A critical finding of the study is the sensitivity of the reservoir to the delay parameter ($\tau$). The system achieves its highest predictive power at the "edge of chaos," where its short-term memory and high-dimensional mapping capabilities are maximized.

Figure 6: Area Under Curve (AUC) vs. Delay (ms), showing peak performance at approximately 30ms.

In experimental evaluations using real-world occupancy data and simulated MIMO-OFDM signals, the DSDFR-ISI model achieved an impressive AUC of 0.99 at -20dB SNR, surpassing both SVM and traditional energy detection methods.

Hardware Implementation

The model was implemented on a Xilinx Zynq-7000 FPGA, demonstrating extreme efficiency. By utilizing a single physical nonlinear node to simulate many virtual ones, the design consumed only 0.199 Watts and used a tiny fraction of the FPGA's resources (1% of BRAM and 5% of flip-flops). This low power footprint and high processing speed (over 5,000 samples in 0.3 seconds) make it an ideal candidate for real-time spectrum sensing at the network edge.

Conclusion

The DSDFR provides a powerful framework for next-generation wireless sensing. By combining the temporal processing of reservoirs with the energy efficiency of spikes, it achieves state-of-the-art accuracy within a hardware-parsimonious envelope. Future work will explore stacking multiple DFR layers to further enhance performance for even more complex signal environments.

Summary updated

BibTeX
@inproceedings{hamedani2020deep,
  author = {Kian Hamedani and Lingjia Liu and Shiya Liu and Haibo He and Yang Yi},
  title = {{Deep Spiking Delayed Feedback Reservoirs and Its Application in Spectrum Sensing of MIMO-OFDM Dynamic Spectrum Sharing}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i02.5484},
}
Figure