| With the rapid development of mobile telecommunication technology,its business types and application fileds also continue to expand.The types,quantity and density of mobile terminals increase,leading to the nonlinear increasing demand for spectrum resources.Therefore,it is important to promote the scientific research on spectrum resourse utilization and spectrum sensing technology.This thesis mainly studies automatic modulation classification(AMC)technology of cognitivie radio in dynamic spectrum sensing technology.In order to improve the accuracy and robustness in automatic modulation classification,machine learning methods are introduced to solve the problem of high-order modulation signal recognition in complex spectrum environment.The main contributions of this paper include two parts:(1)In order to sovle the problem of redundancy in machine learning method based on human designed features and the problem of optimization goal is single,a novel AMC method based on multi-objective genetic programming(MOGP)named MOMC is proposed.To reduce the redundant features,the original mutli-dimensional features are recombined into a single feature by MOGP method.Two fitness functions named classification error rate and robustness variance are designed as two quantization targets,improving the performance of this method.The simulation results show that under the same channel fading condition,the proposed MOMC increases more than 2%compared with contrast methods,illustrating the good robustness of MOMC.(2)In order to solve the problem of incompleteness artificial designed features and the ignore of signal time correlation,a novel method based on gated recurrent residual neural network is proposed.The convolutional neural network(CNN)with residual structure(ResNet)is designed to extract highly representative features from the received signal.In addition,the gated recurrent unit(GRU)can deal with the characteristics of signal time correlation.It is verified that compared with the existing algorithm,when classification accuracy reaches at 90%,the proposed method improves more than 2.5 dB.At the same time,the proposed method is closer to the ideal maximum likelihood classifier when SNR is high.It is proved that GRR has excellent classification accuracy performance. |
