| Modulated signal recognition is a technique for classifying signal modulation types in the case of unknown received signal modulation types,which is a key prerequisite for signal demodulation and has important applications in both military and civilian fields.With the development of wireless communication technology,the modulation mode of signals is gradually diversified,and the electromagnetic environment is becoming increasingly complex,which puts forward higher requirements for modulation recognition technology.In addition,deep learning technology has been recognized by academia and industry as the most promising solution to solve electromagnetic problems because of its powerful feature extraction ability and low a priori requirements.Therefore,the work in this thesis will use deep learning algorithms to improve the performance of current modulated signal recognition algorithms from both accuracy and complexity,and the main research content is as follows:Aiming at the shortcomings of the Adaptive Visibility Graph(AVG)algorithm that is too complex and the accuracy improvement is not obvious,an automatic modulation recognition framework based on the Single-channel Multi-scale Graph Neural Network(SMGNN)is proposed,and the interpretability of each part of the framework is studied.Firstly,the multilayer perceptron and one-dimensional convolution are used to adaptively realize the mapping between single-channel signal sequence and graph,which effectively reduces the complexity of AVG algorithm.Secondly,a multi-scale graph neural network is designed to fuse features of different resolutions and improve the accuracy of model recognition.Experiments show that the SMGNN algorithm proposed in this thesis saves nearly half of the parameters compared with the AVG algorithm,and the recognition accuracy is greatly improved.Although the SMGNN model reduces a certain amount of computation,it has certain limitations in improving the speed of the model due to its inherent graph transformation and graph update steps,and a self-calibration network algorithm based on graph knowledge distillation is designed to solve this problem.Firstly,the convolutional self-coding network,long short-term memory network and fully connected network with adaptive threshold algorithm are formed into a self-calibration network model,which reduces the impact of lowquality data on the model.Secondly,in order to further improve the network performance,this thesis combines graph neural networks and recurrent convolutional neural networks for the first time through knowledge distillation,and adopts adaptive loss value weights to reduce the cumbersome parameter tuning steps in the training process.Experiments show that the use of graph neural network as the teacher model for knowledge distillation can improve the performance of cyclic convolutional network,compared with SMGNN,the inference time used in this method is reduced by about 1/2,the model parameters are reduced by more than 5 times,and the recognition accuracy is similar to it. |
