Studies On Digital Modulation Signal Recognition Based On Convolutional Neural Network

Signal modulation recognition,as one of the related research contents in the field of signal processing,is of great significance for accurately distinguishing the modulation type of various signals from complex communication environments.Therefore,aiming at the inadequacies of the existing modulation recognition methods,this thesis combines convolutional neural network(CNN)to conduct the further studies.The specific contents are as follows:1.Aiming at the problem that the existing modulation recognition methods seldom take the relationship among different features into consideration,a method of modulation recognition based on CNN and features fusion is studied.Specifically,the Smooth Pseudo Wigner-Ville Distribution(SPWVD)and the Born-Jordan Distribution(BJD)are applied to represent the signal into images.Then,a CNN model is employed to automatically learn image features under different representations,and the transfer learning is introduced to speed up and optimize the learning efficiency of the model.Meanwhile,the several handcrafted features are extracted to improve overall performance.On this basic,a multi-modality fusion(MMF)mechanism is introduced to integrate different images features and handcrafted features to yield further improvement.The simulation results show that the proposed method has the superior performance at low signal-to-noise ratio(SNR)compared with those methods without considering the correlation among features,and the MMF method is more effective than traditional feature fusion method.2.Aiming at the problems that the existing deep learning(DL)based modulation recognition methods neglect the complementarity among different models,and the CNN can only extract the local spatial features of the signal,a dual-stream structure modulation recognition model based on CNN and LSTM is studied.In particular,the signal is represented into in-phase/quadrature(I/Q)and amplitude/phase(A/P)format,and the CNN and long short term memory(LSTM)network are combined to explore the temporal and spatial correlation of the signal under two representations.On this basis,the signal features under different representations are interacted in pairs to increase the diversity of features,and these features are fed into softmax classifier to complete the modulation recognition task.The experimental results show that the proposed model outperforms the modulation recognition method without considering the complementarity between the models,and the two-stream structures based on CNN and LSTM provides a better recognition effect than single stream structure.