Research On Modulation Recognition Based On Deep Learning

Modulation recognition in complex electromagnetic environments is an important method for optimizing spectrum utilization efficiency,identifying and minimizing interference,and is of great significance in military and civilian fields.Applying deep learning to modulation recognition can increase the upper limit of performance of traditional modulation recognition algorithms,enhance people's cognitive ability of wireless signals,reduce the construction and operation and maintenance costs of wireless devices,and promote the development of deep learning in the field of communications.Based on the perspective of deep learning,this paper studies modulation recognition algorithms,feature fusion framework,and factors that affect the performance of the model.The main contributions are as follows:Firstly,a modulation recognition network framework and corresponding training algorithm are proposed.Both the supervised method and the unsupervised method can obtain the characteristic that the modulated signal is separable.This paper constructs the overall loss based on reconstruction loss and cross-entropy loss by combining the convolutional auto-encoder and the convolutional neural network.CAE-CNN algorithm framework and corresponding training algorithm are performed.The simulation results show that the proposed algorithm has higher recognition accuracy(95% or more)under high SNR(Signal to Noise Ratio),and has stronger robustness;it has a higher recognition rate than traditional methods under low SNR conditions.,and the accuracy is still about 85% when the SNR is-4d B.Sencondly,we propose a fusion framework of traditional features and depth features.The characteristics of the traditional modulation recognition and the depth features obtained by CNN are adaptively batched,and then feature fusion models are constructed using algorithms such as Softmax,random forest,and deep neural network.The simulation results show that the performance of the fusion algorithm based on random forest and Softmax is better than the benchmark convolutional neural network,and the fusion algorithm based on random forest has relatively optimal classification accuracy and robustness.Affected by the characteristics of the fusion algorithm itself,the misclassification signals under different fusion frameworks are different.Lastly,performance of modulation recognition was studied from the bottom layer of the network,and the simulation results were explained from the perspectives of under-fitting and over-fitting as well as deviation and variance.The simulation results show that the performance of CNN modulation recognition is almost not limited by the depth of the network.When the number of convolution layers is 3,the optimal performance can be achieved.At this time,the depth of the convolution layer is increased,and the system performance may even be reduced;using a CNN network with two convolutional layers,when the number of first-level convolutional cores is small,the performance of the network decreases almost as the number of convolutional cores in the second layer increases.With the number of first-level convolution kernels increases,the system classification performance is relatively stable when changing the number of second-level convolution kernels.Overall,the network with large convolution kernel widths is superior,but the performance is more stable and less fluctuation when the convolution width is increased to 7.