Pushing the Boundaries in CBRS Band: Robust Radar Detection within High 5G Interference

TL;DR

This work tackles radar coexistence in the CBRS band under strong 5G interference by introducing two ML pipelines that operate on IQ data and spectrograms to detect radar signals and classify radar waveforms. The spectrogram-based pipeline (Pipeline 2) achieves robust radar detection at SINR as low as $-5$ dB, exceeding the FCC requirement of $20$ dB and enabling more aggressive spectrum sharing, while also supporting multi-class waveform identification for six radar types. Synthetic and real-over-the-air SDR datasets demonstrate superior performance and generalization compared to IQ-based methods, albeit with higher inference latency for time-frequency processing. The authors provide public code and data to facilitate replication and future work on radar parameter estimation and distributed deployment in ESC sensing for CBRS.”

Abstract

Spectrum sharing is a critical strategy for meeting escalating user demands via commercial wireless services, yet its effective regulation and technological enablement, particularly concerning coexistence with incumbent systems, remain significant challenges. Federal organizations have established regulatory frameworks to manage shared commercial use alongside mission-critical operations, such as military communications. This paper investigates the potential of machine learning (ML)-based approaches to enhance spectrum sharing capabilities within the Citizens Broadband Radio Service (CBRS) band, specifically focusing on the coexistence of commercial signals (e.g., 5G) and military radar systems. We demonstrate that ML techniques can potentially extend the Federal Communications Commission (FCC)-recommended signal-to-interference-plus-noise ratio (SINR) boundaries by improving radar detection and waveform identification in high-interference environments. Through rigorous evaluation using both synthetic and real-world signals, our findings indicate that proposed ML models, utilizing In-phase/Quadrature (IQ) data and spectrograms, can achieve the FCC-recommended $99\%$ radar detection accuracy even when subjected to high interference from 5G signals upto -5dB SINR, exceeding the required limits of $20$ SINR. Our experimental studies distinguish this work from the state-of-the-art by significantly extending the SINR limit for $99\%$ radar detection accuracy from approximately $12$ dB down to $-5$ dB. Subsequent to detection, we further apply ML to analyze and identify radar waveforms. The proposed models also demonstrate the capability to classify six distinct radar waveform types with $93\%$ accuracy.

Figures (6)

• Figure 1: Proposed machine learning pipelines for radar detection in the presence of $5$G interference and subsequent radar waveform identification.
• Figure 2: Laboratory testbed emulating a real‐world spectrum‐coexistence scenario with an incumbent radar transmitter, ESC receiver, and interfering $5$G transmitter.
• Figure 3: Different neural network architectures used for the two parts of hierarchical classifier using: (a) IQ samples for radar detection (Pipeline 1) and waveform identification (Pipeline 2), (b) Spectrograms for radar detection (Pipeline 2), (c) Spectrograms for radar waveform identification (Pipeline 2).
• Figure 4: Trade-off between the inference time and the radar detection accuracy at SINR= $-5$dB.
• Figure 5: Comparison of feature separability via $3$D t-SNE for IQ-based Pipeline 1 and spectrogram-based Pipeline 2. Spectrogram features exhibit better class distinction.