Research On Spectrum Sensing Method Based On Extreme Learning Machine

Cognitive Radio technology is an intelligent communication technology which can improve the spectrum utilization under the condition of more and more crowded spectrum.Spectrum sensing is the core of Cognitive Radio technology,which affects the performance of the whole system.The core idea is to make Cognitive devices have the ability to discover the free spectrum without affecting the Primary Users' communication,so that Second Users can use the free spectrum communication.In order to improve the accuracy of spectrum sensing of Second Users and reduce the interference to Primary Users,a variety of spectrum sensing methods came into being.This paper studies the existing problems in two scenarios: single user spectrum sensing and multi-user cooperative spectrum sensing.The main research work and innovations are as follows:For the single user spectrum sensing scenario,the performance of the existing traditional algorithm in the low SNR scenario is still not ideal.Machine learning algorithm can achieve higher detection performance under low SNR,but the existing Artificial neural network(ANN)algorithm is easy to fall into local extremum during training,and Support vector machine algorithm(SVM)has over fitting problem in low SNR spectrum sensing.In this paper,a spectrum sensing method of Cognitive Radio Network based on QPSO-ELM algorithm with structural risk is proposed.According to the characteristics of the extreme learning machine algorithm,Quantum particle swarm optimization(QPSO)is used to optimize the parameters of the Extreme learning machine(ELM),and the QPSO-ELM model with the idea of structural risk is constructed to balance the empirical risk and structural risk of the algorithm and improve the spectrum sensing performance of the algorithm.Through the simulation,compared with the spectrum sensing performance of three algorithms: ANN,SVM and ELM,the spectrum sensing performance is improved by 16%,28% and 9% respectively at-15 d B,which can be effective in the scene of low SNR.In the centralized multi-user cooperative spectrum sensing scenario,the existing algorithm improves the accuracy of spectrum sensing by increasing the number of Second Users,but a large number of Second Users contain a large number of repetitive features,resulting in a waste of Second User resources,and a large number of Second Users may contain more fading or malicious users,resulting in a larger number of Second User needs.To solve this problem,a packet cooperative spectrum sensing algorithm based on kernel extreme learning machine(KELM)is proposed.(1)Firstly,the KELM cooperative spectrum sensing model with suitable kernel function is determined,and the sensing users are divided into normal users and abnormal users(severely fading users and malicious users)according to the training condition of the characteristics.(2)Then,the correlation degree of the characteristics is calculated respectively to eliminate redundant users.(3)Finally,the optimal grouping strategy is solved by optimizing the algorithm,and the spectrum sensing accuracy is satisfied as many users as possible are used for spectrum sensing in other frequency bands,so that Second Users can be more fully utilized and reduce the cost of spectrum sensing resources.Simulation experiments show that compared with direct cooperative spectrum sensing,more Second Users can be liberated for spectrum sensing in other frequency bands,and malicious users can be separated to a certain extent,and the use of Second User resources is more optimized.