Research On Spectrum Sensing Technology Based On Support Vector Machine Under Impulsive Noise

In recent years,the demand for spectrum resources is growing larger and larger with the rapid development of wireless communication.As a limited and scarce public resource,the utilization effeciency of radio spectrum resources is very low due to its relatively fixed allocation strategy.Cognitive radio technology provides a feasible scheme to improve spectrum efficiency.The primary task of a cognitive radio system is to perform real-time sensing of the spectrum state in the wireless environment,so as to identify whether there exists an idle spectrum,which is called spectrum sensing.Spectrum sensing technology can make full use of idle spectrum resources,leading to its great important research value.At present,there are a plenty of spectrum sensing methods,but most of them are addresseed by assuming that the background noise obeys the Gaussian distribution.However,in practical cognitive radio systems,the background noise often exhibits impulsive manners.The sensing performance of traditional spectrum sensing methods is severely degraded,and may even fail in the presence of impulsive noise.Because of their fast development,machine learning algorithm and multiantenna technology have been widely utilized by the scholars for the problem of spectrum sensing.As a supervised learning algorithm,Support Vector Machine(SVM)has a solid mathematical theoretical foundation and strong generalization ability.It has shown good performance in many fields.Therefore,this thesis combines SVM with spectrum sensing technology to conduct in-depth research on bi-antenna spectrum sensing under the background of impulsive noise.The main works are as follows:Firstly,under the background of impulsive noise interference,this thesis proposes a biantenna single-point spectrum sensing method based on SVM.For the transmitted signal contaminated by impulsive noise,we use Kendall's Tau(KT),Polarity Coincidence Correlator(PCC),Fractional Lower Order Moments(FLOM)and Logarithm Moment(LM)as the feature extraction method.After performing feature extraction of the sensing signal providing that the pectrum state is known,the SVM algorithm is employed for offline training,and the classification decision function is then obtained.During model training,Fisher's criterion is adopted to optimize the kernel function parameters.Finally,online sensing is to extract the features of the signal perceived by the cognitive user,and then substitute it into the decision function for calculation,thereby obtaining the spectral state.Simulation experiments show that the proposed SVM-based bi-antenna single-point spectrum sensing method has good sensing performance under impulsive noise.Further,in view of the limited data processing capability of single-point spectrum sensing,and problems such as path fading,shadow effect,and hidden terminals,the centralized cooperative spectrum sensing technology is adopted.Specifically,multiple cognitive users send their sensed data to the fusion center.Then,the fusion center determine the spectrum state after the feature extraction,and finally the results are sent back to the sensing users.We adopt the Fuzzy Support Vector Machine(FSVM)rather than the originalSVM because of the sensitivity of SVM to outliers embedded in the training set.As an improved version of SVM,FSVM is able to resolve the problem that SVM is very sensitive to outliers by utilizing different membership degrees.This thesis proposes to use the k-Nearest Neighbor(k-NN)algorithm to calculate the membership degree of the sample points in each training set,so that different sample points have different contributions to the classification hyperplane,thus alleviating the problem of outlier sensitivity.The experimental results show that in the case of multi-user cooperative spectrum sensing,the proposed spectrum sensing method effectively improves the performance of spectrum sensing.