Specific Emitter Identification Methods Based On Signal Capusle Network

As a typical pattern recognition problem,specific emitter identification(SEI)is a crucial step to achieve spectrum sensing.The process of traditional SEI methods is quite complex,and it is difficult to quickly obtain the required information from the complicated electromagnetic environment in real time.Current deep learning-based SEI methods also have many shortcomings such as single feature description,poor discriminative ability and limitation of the input format.This thesis,in view of the above issues,based on the framework of the capsule network in deep learning,designs three signal capsule network models for SEI.Specific work and results are as follows:1.In order to solve the problem that current deep learning-based SEI methods using single character,a specific emitter identification method based on multi-view signal capsule network is proposed.First,sampling signal,the time-frequency spectrogram and Wigner-Ville transformation map are used as the model inputs from multiple perspectives.Then fusing multi-vision features via attention mechanism to enhance the feature representation ability.Finally,the capsule network provides a higher degree of robustness.Compared with four current deep learning-based methods,extensive analysis and experiments on the Tianyan Cup competition dataset show that the prediction accuracy of our method can increase by 1.7%,which verifies the effectiveness and robustness of the proposed method.2.In order to improve the discriminative ability of the signal feature extracted by the previous method,a specific emitter identification method based on discriminative feature signal capsule network is proposed.First,the original sampling signal is used as the model input,then extract character feature by one-dimensional convolution.Secondly,employ a margin constraint in the original softmax loss to enhance stronger discriminant power of extracted features.Finally the capsule network promotes model's ability of anti-jamming.We do validity tests on the Tianyan Cup competition dataset.The results show that the discriminative loss function designed by the proposed method can effectively improve the model performance.3.In order to overcome the limitation of the input format in the previous section,a specific emitter identification method based on graph signal capsule network is proposed.First,the signal is transformed into the undirected graph according to the Euclidean distance from the sampling point,and then take the undirected graph as the input of the network.Second,extract topological structural characteristics by graph convolution operation.Finally,this thesis introduces capsule network to improve generatization ability and enhance the robustness of the model.Extensive analysis and experiments on the Tianyan Cup competition dataset show that the method can increase the accuracy of specific emitter identification to 85.0%,which proves the feasibility and validity of the signal capsule network.