Research On Modulation Signal Recognition Algorithm Based On Deep Neural Networ

The modulation mode of communication signals changes with the change of communication environment,so automatic modulation classification is necessary.Automatic modulation identification technology is widely used in the field of communication and plays a vital role.Modulated signal classification algorithm based on deep learning have achieved better performance than traditional modulated signal classification algorithms.And Modulated signal classification algorithm based on deep learning have attracted a lot of attention from researchers.The thesis is based on deep neural networks for the modulation classification algorithm,and the main research includes the following three aspects.1.A modulation classification algorithm based on MSRNN is proposed.This algorithm focuses on the problem of low accuracy due to overfitting of deep neural networks.The algorithm adopts an optimized loss function,selects the SNR of the signal as the weight,and uses the modulation factor to pay reasonable attention to the signal sample.A new residual unit is designed to increase the network depth and alleviate the overfitting problem by reasonably increasing the number of skip connections.A new Res Net is built to simultaneously learn and extract shallow features and deep features of modulated signals,and then realize modulated signal recognition.The results demonstrate that the algorithm has a classification accuracy of90.31%,which is better than comparison networks.2.A modulation classification algorithm based on DRCNet and RCTLNet is proposed.This algorithm focuses on the problem of low accuracy of automatic modulation classification algorithms due to noise interference.STLNet is designed to give appropriate thresholds to signal samples with different SNR and eliminate noise-related features.DRCNet is constructed,which uses the convolutional layer to transform the signal in the domain and inverse transformation,and uses STLNet to remove the noise features,so as to realize the denoising of the signal.A new network structure consisting of Res Net and LSTM in series is proposed,which can extract the spatial features and temporal features of signal samples,and then achieve modulated signal recognition.The results demonstrate that the algorithm has a classification accuracy of 92.05%,which is better than comparison networks.3.A modulation classification algorithm based on SCBLNet is proposed.This algorithm focuses on the problem of poor performance of neural networks due to low feature weight assignment capabilities.The algorithm uses the amplitude components and phase components to represent the input signal samples,which changes the traditional representation of the input signal samples.A deep network architecture consisting of Res Net and Bi LSTM is constructed.The spatial features of the input signal are extracted by Res Net.The temporal characteristics of signals are extracted and memorized by Bi LSTM.CSANet is introduced into this network.CSANet can assign feature weights to signals,perform adaptive refinement of signal features,and enhance signal features.Thus,the classification of modulated signals can be realized.The results demonstrate that the algorithm has a classification accuracy of 93.52%,which is better than other comparison networks,and the confusion of signals is improved to some extent.