Spectrum Sensing in Interference and Noise Using Deep Learning

Wireless devices are ubiquitous and consequently the spectrum is congested. Dynamic spectrum access is becoming more widespread in unlicensed bands and as a means to allow secondary users white space in licensed bands. Spectrum sensing capabilities are the cornerstone of dynamic spectrum access. In this paper, the spectrum-sensing problem is transformed into an image recognition problem and machine learning is employed to distinguish between noise and the presence of a signal. An existing Convolutional Neural Network (CNN), AlexNet, is repurposed to sense the spectrum for energy using only a small training set of a few hundred samples. The performance of the CNN detector is then compared to the performance of more traditional energy detection as well as other published results of machine learning used for signal detection. The CNN detector presented here surpasses the other machine learning methods for signal detection. The CNN detector does not require a measurement of the noise floor, which offers a significant improvement over the classic energy detector. The CNN detector also surpasses the performance of the energy detector in the presence of narrowband interference.