High-Performance Blind Spectrum Sensing: An Iterative Unsupervised Learning Approach

In the research of blind spectrum sensing, unsupervised learning has been regarded as a promising technology to detect spectrum status without any prior knowledge or labeled signals. However, the detection performance of the existing unsupervised sensing methods is usually far below that of supervised sensing methods. In order to overcome this problem, we propose an iterative unsupervised learning based cooperative spectrum sensing (CSS) algorithm, where the concept of iteration is introduced to improve detection performance. Specifically, in each iteration, i) The CSS technology is first used to make global labels by combining the local labels of the individual detectors constructed in the last iteration; ii) And then, individual detectors are constructed by utilizing supervised learning to fit the mapping from local sample set to the current global labels. During above iterative process, the global labels contribute to improving individual detectors, while these improved individual detectors can cooperatively make more accurate global labels, which will be utilized in next iteration. Finally, with the increase of the number of iterations, these individual detectors can be used to online cooperatively detect spectrum status when there are only slight changes in global labels. Simulation results show that the proposed algorithm can achieve comparable individual and cooperative detection performance with the existing supervised sensing methods over a wide signal-to-noise ratio (SNR) region.