Research On Modulation Recognition Algorithm Based On Neural Network

In modern communication systems,signal modulation is an important technique.Modulation not only improves the transmission rate of information,but also shifts the spectrum of signals.Hence the communication system can adapt to different channel characteristics and improve its reliability.In many non-cooperative communication scenarios,the modulation recognition of the received signal is one of the most critical and fundamental steps.Such as in military electronic countermeasures,only when we recognizing the modulation mode of enemy's transmitting signal correctly,can we intercept enemy's information successfully or transmit high-power interference signals in the same modulation mode to achieve the purpose of interference enemy's normal communication.The early modulation recognition method is artificial recognition,which is dependent on lots of experiences and inefficient.Modern recognition methods mainly include maximum likelihood hypothesis algorithm based on Bayesian decision theory,pattern recognition algorithm based on feature extraction and recognition algorithm based on deep learning.However,the above algorithms have problems such as excessive computational complexity or additional data preprocessing to extract signal features.In order to solve the above problems,this thesis proposes a modulation recognition method based on graph neural network.Meanwhile,in order to improve the recognition accuracy under low SNR,a predenoising modulation recognition algorithm is proposed.This thesis firstly proposes a modulation recognition method based on graph neural network.Compared with traditional deep neural networks,the graph neural network is good at processing non-Euclidean spatial data such as graph.This algorithm includes two important network structures,namely feature embedding network which is used to extract key features and graph neural network.The graph neural network includes two cores: adjacency matrix computing block and graph convolution block.The former block can map the feature embedding matrix of samples into a graph,and the latter block can make predictions by merging all information in the graph.The simulation results show that compared with convolutional neural networks and other machine learning algorithms,the graph neural network modulation recognition method's accuracy is higher.The advantage of this algorithm is more obviously especially when SNR is low.The existing modulation recognition algorithms have higher accuracies in the case of high SNR,but the accuracies decreased significantly when SNR is low.For this problem,this thesis proposes a pre-denoising modulation recognition algorithm based on neural networks.This algorithm first denoises the signal using a denoising network,then input the denoised signal into a modulation recognition network to recognize its modulation mode.This thesis constructs two different denoising networks whose core structures are convolutional neural networks and convolutional self-encoders.The latter one uses auto-encoder structure,in order to prevent the loss of data's important information in the encoding process,jump connections are added between the corresponding coding layers and decoding layers,so that the information of the coding layers are directly transmitted to deep layers.Residual learning and batch normalization are used in denoising neural networks to accelerate model training.The simulation results show that the above two denoising networks can achieve the same denoising effect.The denoising network based on convolution auto-encoder has higher training efficiency and shorter computing time.The pre-denoising modulation recognition algorithm can improve the accuracy under low SNR effectively.