Modulation Recognitionalgorithm Based On SVM

With the great development of technology in communications, modulation types have become more and more complex. How to efficiently monitor and recognize the modulation types is an important problem. Support Vector Machine has powerful ability in pattern recognition, and it is widely used in modulation recognition with its better stability and mistake tolerance.This thesis mainly researches on the modulation recognition algorithm based on SVM. At first, the algorithm with instantaneous features via AWGN channel is introduced, and feature extraction methods are analyzed. Owing to present C-SVM algorithm giving no consideration on the location of the hyper-plane and the inference of abnormal samples, the model is improved by introducing the parameter representing the location and fuzzy factor. Moreover, the solving process of the improved model is given. Follow that, a SVM classifier based on binary tree is designed for automatic modulation recognition. Simulation results show that the recognition rates of improved C-SVM are higher than that of present C-SVM.In this thesis the method of recognizing signals modulated by QPSK, 16QAM, 64QAM and OFDM is presented based on SVM algorithm with the characteristic parameter of forth order cumulants. SVM maps the feature values of classification into high dimension space, in which the optimal separating hyperplane is constructsed to realize the separation of the signals with targeted modulation method. The influence of AWGN channel, Rayleigh and Nakagami fading channels on the characteristic parameter is analyzed respectively, and the explicit expression of the characteristic parameter is further derived. The simulation results show that the average recognition rate of AWGN channel reaches 80%, while those of Rayleigh and Nakagami fading channels are higher than 70% when SNR is -5dB. Moreover, the performance of the SVM-based method is better than that of the decision tree method.At last, a new method of cooperative modulation recognition (CMR) is proposed to improve the performance of modulation recognition at low SNRs. CMR algorithms based on feature fusion and decision fusion are analyzed according to different fusion methods. In CRM algorithm based on feature fusion, each node sends feature samples to the fusion center, where recognition results will be got using the SVM-based method. But in CRM algorithm based on decision fusion, each node distinguishes the received signals first, and then sends the decision bits to the fusion center, which determines the final modulation types with voting fusion rule or maximum posterior probability rule. Simulation results show that the cooperative modulation recognition method has better performance than the non-cooperative method at low SNRs.