Research On Machine Learning Based Spectrum Sensing In Cognitive Radio

With the fast development of wireless communication, spectrum resource is now being scarcer than before. The existing scheme for spectrum allocation can not meet the increasing demand for wireless communication. As an effective solution to alleviate spectrum scarity in wireless communications, cognitive radio technology plays an important role in the development of wireless communication in which spectrum sensing has become a hotspot in this area. This dissertation focuses on the research of machine learning based spectrum sensing.The main innovations and working points are as follows.1. In this dissertation, the current technology development of machine learning and spectrum sensing technology is introduced. The feasibility of combining machine learning with spectrum sensing is discussed, and the advantages and disadvantages of existing research results are discussed.2. In this dissertation, a spectrum sensing algorithm based on kernel principal component analysis (KPCA) and K-means Clustering is proposed,which is divided into the off-line training stage and the on-line detection stage. The feature of the received signal is extracted by KPCA. In the offline training process, the clustering algorithm is exploited to cluster the training sample in order to obtain the threshold. The online classification step presents the results of spectrum sensing by similarity calculation between the received signal and the clustering centers. With the help of offline training stage, the dependence on the primary signal information can avoid without significantly increasing the complexity of the algorithm.Simulation results based on low SNR scenario show that the algorithm can improve the performance of spectrum sensing and reduce the influence of noise.3. In this dissertation, a Multi Scale Kernel Principal Component Analysis (MSKPCA) and K-means Clustering based spectrum sensing is proposed for multi-node collaborative spectrum sensing scenarios. With the help of the multi-scale kernel function and the multi scale sample sheme,the feature of the received signal can be extracted in a more efficient way.In order to balance the effect of the feature extraction algorithm in terms of detection time, the local sensing results of different scales are fused by sequential detection to obtain the results of spectrum sensing. Simulation results show that the performance of the proposed algorithm is superior to that of the traditional spectrum sensing scheme especially in low SNR scenarios. Besides, the proposed schem can improve the detection probability of the system without significantly increasing the time delay.