Research On Digital Modulation Recognition Of Cognitive Radio Signals

Cognitive radio is widely used in civilian and military fields because it can improve spectrum utilization.In a cognitive radio system,how to effectively identify the modulation types of signals sent by authorized users is of great significance to the improvement of spectrum sensing performance.Modulation recognition,as an intermediate step of signals detection and signals demodulation,is one of the key technologies to ensure reliable communication.This paper mainly studies the modulation recognition of multiple single-carrier digital modulation signals in the cognitive radio system.First,by analyzing the theoretical values of the high-order cumulants of different digital modulation signals,it can be seen that under less prior informations,the second,fourth,sixth,and eighth-order cumulants of the extracted signal sequences are the most effective in identifying the eight modulation types of 2ASK,4ASK,QPSK,8PSK,16 PSK,8QAM,16 QAM,and 32 QAM.Based on the high-order cumulants theoretical value of the modulation signals and traditional decision theory,we artificially design decision features and decision thresholds,and build a hierarchical classification structure to achieve modulation recognition.The simulation results show that for signal sequences that use symbol-spaced sampling in the additive white gaussian noise channel,the traditional decision theory can achieve a recognition rate of more than 95% with a10 d B signal-to-noise ratio.Secondly,because artificially designed features and thresholds are easily interfered by noise and other factors which cause large errors,this paper proposes a modulation recognition method that combines high-order cumulants and Discriminative Restricted Boltzmann Machine,and training Discriminative Restricted Boltzmann Machine based on the discriminantive objective function to realize modulation recognition.Compared with traditional decision theory method,Discriminative Restricted Boltzmann Machine has better recognition performance under different interferences.And compared with the commonly used Decision Tree,Random Forest and Support Vector Machine,it has a simpler model parameter selection process.The simulation results show that Discriminative Restricted Boltzmann Machine has obvious advantages over Decision Tree,Random Forest or Support Vector Machine in recognition performance when the signal contains clock offset.Finally,although the method of using the extracted high-order cumulants to achieve modulation recognition has the advantages of strong pertinence and low feature expression dimension,it also loses other modulation characteristics contained in the original signal sequences.In response to this problem,this paper proposes a modulation recognition method based on the original signal sequence and Convolutional Neural Network,and designs a Convolutional Neural Network that can adaptively perform feature extraction and modulation type recognition.In the network design process,the Global Average Pooling layer is introduced to replace the commonly used fully connected layer,which effectively reduces network training parameters and training complexity.The simulation results show that when the other structures of the network are the same,the Global Average Pooling layer has better recognition performance than the fully connected layer.Moreover,the proposed Convolutional Neural Network has higher recognition accuracy than the previous algorithms under different interferences.