Limited Data Spectrum Sensing Based on Semi-Supervised Deep Neural Network
Y. Zhang, Zhijin Zhao
Abstract
Spectrum sensing methods based on deep learning require massive amounts of labeled samples. To address the scarcity of labeled samples in a real radio environment, this paper presents a spectrum sensing method based on semi-supervised deep neural network (SSDNN). Firstly, a deep neural network is established to extract the features of signals by using small amounts of labeled samples; Then, plenty of unlabeled samples are used for self-training process, and the ones with high confidence are marked with pseudo-label to expand the labeled dataset. Finally, the extended dataset is used to retrain the network. Plentiful experiments are carried out on a dataset of 124,800 samples. The results demonstrate that the proposed algorithm has good detection performance over multi-path fading channel and additive white Gaussian noise channel due to the utilization of a great deal of unlabeled dataset. When the labeled samples account for only 5% of the traditional fully supervised deep learning model and the SNR is higher than -13 dB, the detection probability of SSDNN is higher than 90%.