Radio spectrum is finite and mostly licensed, yet much of it sits idle at any moment. Cognitive radio aims to opportunistically use these gaps, which requires fast, accurate spectrum sensing — deciding whether a band is occupied. Classical detectors (energy detection, matched filtering) struggle at low SNR or with unknown signals; deep learning now outperforms them.
Working principle
Raw IQ samples or spectrogram images are fed to a neural network — a CNN over the time-frequency representation, or a model over raw IQ — trained to output occupancy and even the modulation type. Because the network learns signal features directly from data, it generalises to noise and interference far better than fixed thresholds, enabling robust detection at low SNR.
| Method | Needs prior knowledge | Low-SNR performance |
|---|---|---|
| Energy detection | No | Poor (threshold-sensitive) |
| Matched filter | Yes (signal model) | Good but inflexible |
| Cyclostationary | Partial | Good, high compute |
| Deep learning | Training data | Strong, data-driven |
Trade-offThe trade-off shifts from hand-designing detectors to curating training data; models can fail on signal types absent from training, so robustness and domain shift are active concerns.
Applications
- Dynamic spectrum sharing in 5G/6G and CBRS bands
- Spectrum monitoring and interference / jammer detection
- Automatic modulation classification for SDR and defence
References & further reading
- O'Shea et al., “Over-the-Air Deep Learning Based Radio Signal Classification,” IEEE JSTSP, 2018.
- Mitola & Maguire, “Cognitive radio: making software radios more personal,” IEEE Pers. Comms, 1999.
- Arjoune & Kaabouch, “A Comprehensive Survey on Spectrum Sensing in Cognitive Radio Networks,” Sensors, 2019.