Experimental evaluation, spectrum sensing, unsupervised detection, software defined radio (SDR
Abstract
Spectrum sensing plays a key role in cognitive radio (CR) networks in order to determine the availability of unused frequency bands. So far, presumed models have been employed to conceive statistical tests such as eigenvalue-based detectors. Nevertheless, their detection performances are degraded as the accuracy of these models depart from real-world measurements. In this paper, we assess the performance of an unsupervised learning spectrum sensing (ULSS) detection through experimental evaluations. In this approach, model assumptions are no longer required, while avoiding labeled data often not available in practical CR scenarios. The ULSS consists of a two-stage training, where an unsupervised Gaussian mixture model (GMM) is employed to provide training data for a deep neural network (DNN). The experimental results shows that it outperforms model-based detectors by learning from real measurements.
