Research On Spectrum Sensing Based On Random Forest In Cognitive Networks

Cognitive networks are characteristic of relatively good flexibility,autonomy as well as adaptability to environmental changes.They are able to effectively improve the spectrum efficiency,enhance communication performance and to alleviate the spectrum shortage problem during wireless communication.Thus,they are widely applied to public security,military training,disaster relief,spectrum trading and so on.Spectrum sensing,as an important basis and key technique of cognitive networks,can be exploited to search the primary user's idle frequency band in the time domain and frequency domain by detecting,analyzing and judging spectrum resources.The current spectrum shortage problems during wireless communication can be largely alleviated via this method.Thus it has recently become a hot topic.In this dissertation,the recent progress on cognitive networks is given.Then the main ideas and the fundamental principles of spectrum sensing are introduced in detail.This dissertation aims at improving the performance of spectrum sensing and recognizing modulation type of primary user signals in low SNR environments.Considering the fact that random forest algorithm has good classification performance,robust to noise,suffers no overfitting and offers possibilities for explanation,we novelly apply it to solve the spectrum sensing problems in low SNR.Firstly,the spectrum sensing algorithm based on random forest model is built.Then three solutions are proposed based on the sensing model.and are introduced as follows:(1)Spectrum sensing algorithm based on cyclic spectrum and random forest;(2)Spectrum sensing algorithm based on principal component analysis and random forest;(3)Spectrum sensing algorithm based on manifold learning and random forest.The dissertation's main contents and innovations are given as follows:(1)The spectrum sensing algorithm model based on random forest is built.Considering the fact that the sensing performance of primary user signal is mainly affected by noise and other interference factors in low SNR environment and the fact that random forest is a strong classifier which is robust to noise,suffers no overfitting and offers possibilities for explanation,the mechanism of random forest algorithm is integrated into the model,on the basis of the further study on the binary model of spectrum sensing in cognitive networks.The random forest is generated by extracting the features of the signal cyclic spectrum over and over,and then the spectrum sensing model based on random forest is built.(2)On the basis of the above spectrum sensing model,a novel spectrum sensing algorithm based on cyclic spectrum and random forest is proposed.Under low SNR environment,a novel spectrum sensing algorithm based on cyclic spectrum and random forest is proposed.By taking advantage of received signal spectrum analysis,the algorithm fully combines the multiple weak classifiers to enhance the classification performance.The mean and variance of the maximum received signal cyclic spectrum energy are extracted to construct the training and testing samples in low SNR environments.The algorithm has a lower computational complexity and better real-time performance.It is capable of improving the performance of spectrum sensing as a whole.It mainly applies to solve the spectrum sensing issues with lower computational complexity and higher real-time requirements in low SNR conditions.(3)In order to improve the accuracy of spectrum sensing,a novel principal component analysis and random forest based spectrum sensing algorithm is proposed.Given the fact that the the limited number of features effects the sensing accuracy,a spectrum sensing algorithm based on principal component analysis and random forest is proposed.This approach is on the basis of the aforementioned cyclic spectrum sensing random forest.Signal cyclic spectrum parameters are extracted and the principal components of these parameters are obtained by using principal component analysis algorithm.Random forest is then built to achieve the primary use signal detection and perception.Principal component analysis algorithm could reflect the overall characteristics and the strong classification performance dramatically.Then,a spectrum sensing algorithm based on kernel principal component analysis and random forest is proposed.The data from the input space is mapped to higher dimensional space.In the higher dimensional space,the features are extracted by nonlinear transform.Therefore,the samples are extracted more comprehensively and then the performance of spectrum sensing is improved.It mainly is applied to solve the problem with higher accuracy requirement in lower SNR conditions.(4)In order to obtain the overall features of the received signal and improve the accuracy,a new spectrum sensing algorithm based on manifold learning and random forest is given.Considering the impact of the nonlinear components of the received signal to spectrum sensing,a new spectrum sensing algorithm based on manifold learning and random forest is proposed.The prevailing feature extraction method of nonlinear dimensionality reduction"manifold learning" has been intensively studied and been used to obtain the overall characteristics of the received signal.The presented algorithm is based on local linear embedding algorithm in manifold learning methods.Dimensionality reduction parameters of signal cyclic spectrum data are extracted through the linear reconstruction between sample neighborhood points.Then random forest is built and used for classification and perception to primary user signal.The overall characteristics of the information component of the received signal can be extracted more completely by the proposed algorithm,and thereby more excellent performance of spectrum sensing is achieved.The proposed algorithm is mainly applied to spectrum sensing scenarios with higher accuracy,looser real-time requirements,complex harsh wireless channel environment and lower SNR.In addition,the applicable range of these proposed algorithms are also introduced by the comparison of their computational complexity.