Research On Application Of Deep Learning In Modulation Recognition

Automatic modulation identification is a key step in signal demodulation and parameter estimation.It is usually applied to interference identification,electronic countermeasures,and spectrum detection in civil and military wireless communications.It is the basis of research in cognitive radio and non-cooperative communication.With the continuous change of communication environment and modulation style,the extraction of feature parameters from modulated signals is not enough to accurately distinguish the signal modulation method.Deep learning technology can extract implicit features from large-scale data.Compared with machine learning methods,the calculation speed is faster.The high degree of feature abstraction,and the use of deep neural network's powerful pattern recognition capabilities,to improve the automatic identification of modulated signals at low signal-to-noise ratios is an inevitable trend in the development of automatic modulation recognition technology.Aiming at the problems of traditional modulation recognition algorithms,this paper studies the application of deep learning technology in modulation recognition from the aspects of features and networks.The specific content can be summarized as:(1)For the problem of low recognition rate of six signal types such as MPSK and MQAM under low signal-to-noise ratio,this paper first proposes a modulation recognition method based on cyclic spectral feature combination and deep learning.By analyzing the characteristics of the three-dimensional cyclic spectrum of the modulation signal from different angles,three features of the contour map,the cyclic frequency axis ? projection map,and the frequency axis f projection map of the cyclic spectrum three-dimensional image are extracted to convert the original image into a grayscale image.Then,the three feature images are combined in the third dimension through matrix transformation,and the feature data obtained by the combination can reflect the differences between signals in a deeper level.Aiming at the problems of the traditional convolutional neural network structure,this paper makes improvements to the structure of the convolutional layer,pooling layer,and fully connected layer to reduce the amount of model calculation and increase the speed of network convergence.Experimental results show that: for MPSK signals,when the SNR is-5d B,the average recognition rate can reach more than 90%;for MQAM signals,when the SNR is-4d B,the average recognition rate can reach more than 95%;a mixture of MPSK and MQAM Modulation type,when the SNR is 0d B,the average recognition rate is as high as 100%.Compared with the traditional decision tree model and SVM model,this method has better recognition effect under low signal-to-noise ratio.(2)From the perspective of feature preprocessing,this paper proposes to use the original image of the modulation signal self-fuzzy function as the feature data,use the feature combination method to obtain multi-channel fusion data,and then send the data to the convolutional neural network for training.The experimental results show that the low signalto-noise ratio In the following,MPSK and MQAM intra-and inter-class signals have higher recognition accuracy,which verifies the validity of the feature combination.From the perspective of image processing,this paper extracts the pixel information of each channel in the image,calculates the image histogram data on each channel,and stitches in the third dimension to obtain the 3D histogram data as the input of the network model.(3)In terms of network structure optimization,this paper uses the advantages of CNN network to extract spatial features and LSTM network to extract timing information,and proposes CNN-LSTM and LSTM-CNN models.For CNN networks,this article embeds the two steps of Squeeze and Excitation in SENet networks into CNN.In the network,the feature weights are learned according to the network loss value,so that effective feature channel weights are large,invalid or small effect feature channel weights are small,and the network model training efficiency is improved.For the three shortcomings of the traditional LSTM network gating structure,Adjust the structure of forget gate,input gate and output gate,increase the contribution rate of LSTM network long-term memory to data output,and strengthen the structural connection between forget gate and output gate.Experimental results show that compared with other machine learning and deep learning models,the network model proposed in this paper has significantly improved the accuracy of modulation recognition at low signal-to-noise ratio.