Research On Automatic Modulation Recognition Based On Deep Learning

In a modern society where science and technology are developing rapidly,modulation recognition is an intermediate step between signal detection and signal demodulation,and the further development of automatic modulation recognition technology becomes more and more important.At present,modulation recognition technology generally has the characteristics of manual selection,and it is difficult to unify the selection of features in different communication systems,which brings about problems such as low processing efficiency and poor robustness of the recognition model.With the driving of big data and the continuous development of deep learning technology,the ability of feature extraction and fitting of deep learning models can promote the advancement of modulation recognition.Therefore,this thesis studies the appropriate model in the field of modulation recognition based on deep learning,and automatically extracts the characteristics of the modulated signal to realize the modulation recognition of the signal.In this thesis,the characteristics of IQ(In-phase & Quadrature)signals are converted into amplitude and phase characteristics in the time domain and the characteristics of the fast Fourier transform in the frequency domain,and the three signal information is unified into the data dimension to obtain a data set of three different characteristic information of the signal.In view of the particularity of the modulation signal data set,this thesis designs three different convolution models based on convolutional neural networks and based on different convolutional layer structure methods.Experimental comparisons are made on three signal data sets respectively.The convolutional network with densely connected structure of feature multiplexing has better recognition effect on amplitude and phase feature data sets than other convolutional networks.At the same time,the timing of the existence of the signal is considered,and the recurrent neural network can make use of the correlation of the timing information.In this thesis,a longshort-term memory network model suitable for training signal data is designed.Through experiments on the three signal data sets,the recognition accuracy of the model in the amplitude and phase characteristic data sets is further improved.Finally,by combining the characteristics of convolution and recurrent neural networks in modulation recognition,a deep learning model based on the combination of one-dimensional convolutional network and long-short-term memory network is proposed.The model also removes the fully connected layer structure of parameter redundancy,and uses a variety of optimization methods to solve the training problem of the model,which further improves performance and recognition accuracy of the model.The experimental results show that based on the amplitude and phase feature data set,the best modulation recognition accuracy is achieved in the model studied in this thesis,and compared with the existing modulation recognition method,its recognition accuracy has also been steadily improved.