Research On Modulation Recognition Of Wireless Communication Signal Based On Deep Neural Network

Automatic modulation recognition technology is the basis of signal demodulation,spectrum management and parameter estimation in wireless communication.It is widely used in military and civilian fields such as information interception,electronic countermeasures,spectrum monitoring,etc.,and plays an important role in non-cooperative communications.The modulation mode of wireless communication signals is becoming increasingly complex with the rapid development of wireless communication technology,which increases the difficulty of identifying modulation signals.Therefore,the efficient modulation signal identification methods become the key of research.As a hot technology in the field of artificial intelligence,deep learning is widely used in machine translation,image recognition and other fields,and has achieved good results.Therefore,using the self-learning and powerful feature extraction capabilities of deep neural networks to improve the modulation recognition rate in multiple types and complex environments,it is the focus and difficulty of research in the field of automatic modulation recognition.With the wide application of neural network models,the training and optimization of neural networks is still a difficult problem.Therefore,the optimization of neural network hyperparameters has become a hot research topic in the field of deep learning.Based on the deep learning method,this thesis studies the automatic recognition of various modulation types,and for the training optimization of neural network,this thesis also studies the effect of learning rate methods on network model training.For the above two problems,the main research contents of this thesis are as follows:For the limited recognition of modulated signals by the current network structure and the optimization of neural network training.In order to reduce the time cost of model training,this thesis proposes an exponential decay adaptive cyclic learning rate method based on the triangular cyclic learning rate method.And in order to improve the recognition performance of the modulated signal,this thesis proposes a dual-channel hybrid model termed as CLDR(convolutional long short-term deep neural and residual network).The CLDR consists of the convolutional long shortterm deep neural network(CLDNN)and the residual network(Res Net),where CLDNN is to reduce the variations in the spectrum and time,Res Net is to avoid gradient vanishing or exploding.Simulation results show that the training time cost of CLDR using exponential decay adaptive cyclic learning rate is reduced by 14.6% and 32.1% compared with triangular cyclic learning rate and fixed learning rate.Therefore,exponential decay adaptive cyclic learning rate method effectively reduces the training time cost of the model in the same classification accuracy.The CLDR model yields a recognition accuracy of 93.1% when the signal-to-noise ratio is 0?18d B,which effectively reduces the influence of channel environment such as noise and fading on recognition accuracy.In order to solve the problem of neural network model training time cost,this thesis proposes a piecewise adaptive learning rate adjustment method based on exponential decay adaptive cyclic learning rate.In addition,to verify the reliability of the piecewise adaptive learning rate method,this thesis proposes a modulation recognition model termed as CSSGNet(Convolutional Channel and Spatial Attention Bi GRU Network),which combines the advantages of CNN to reduce signal spectrum and Bi GRU to reduce signal temporal variation,and introduces channel and spatial attention to extract local and global features of modulated signals.The experimental results show that the piecewise adaptive learning rate shows better robustness on different models compared with the fixed,polynomial decay,exponential decay and cosine decay learning rate adjustment methods,the CSSGNet model achieves convergence after only 25 epochs of training with a piecewise adaptive learning rate,which significantly reduces the training time overhead of the model.The CSSGNet model for the 16 QAM and 64 QAM modulated signals reaches 91% and98%,which effectively reduces the noise and fading interference on the recognition of the similar16 QAM and 64 QAM modulated signals of the constellation diagrams.