SEMI-SUPERVISED DEEP LEARNING USING PSEUDO LABELS FOR SPECTRUM SENSING

As the first operation in cognitive radio systems, spectrum sensing is the basis of cognitive radio and therefore plays an important role. Spectrum sensing needs to find spectrum holes to improve spectrum utilization. Many papers propose deep learning based methods to improve probability of detection, but most of the existing methods are supervised learning, which requires a large number of training sets with labeled samples. However, it is difficult and time-consuming to obtain a great many labeled samples in real physical scenes. Thus, in order to overcome the drawback of some existing methods, we present a spectrum sensing method based on pseudo labels semi-supervised deep learning (PL-SSDL). Firstly, the semi-supervised integrated decision tree module SDTEM is pre-trained by using a small number of labeled samples. Then, the pre-training SDTEM is used to label a large number of high confidence unlabeled samples, expand the training set. Finally, the expanded samples are used to train the CNN model. Extensive studies results show that suggested PL-SSDL is superior to conventional methods, such as energy detection and entropy-based method, and second only to the supervised CNN spectrum sensing method.