Research On Emitter Signal Recognition Technology Based On Deep Learning

With the rapid development of electronic information technology,electronic countermeasure has gradually become the core of modern military war.The identification of emitter signal is an important part of electronic countermeasure,and plays an important role in electronic countermeasure.The recognition of emitter signal is to intercept the signals emitted by the emitter,extract the individual characteristics of different emitters by signal processing technology,and identify different emitters by classification algorithm,which provides reliable basis for communication intelligence investigation and military action decision.The traditional feature matching method has some limitations and low reliability.In recent years,many scholars have applied deep learning to the recognition technology of emitter signals.Deep learning can classify and recognize the signals according to the time-frequency characteristics of signals.Due to the difference of hardware equipment,the signals emitted by different individuals of the same type are different,how to accurately classify and identify the emitters when each emitter has multiple signals is the focus of current research.In this thesis,the deep learning method is used to identify the emitter signals of different individuals in the same type and working environment.The main work is as follows:(1)Aiming at the signal processing and feature extraction of the emitter,this thesis uses a windowed energy detection method to intercept the useful signals,and Then the short-time Fourier transform is selected to extract the features of each emitter signal,and the two-dimensional time-frequency matrix obtained is converted into a threedimensional matrix,so as to better express the signal features and provide stable and reliable classification features for emitter signal recognition.(2)Aiming at the signal recognition of emitter,this thesis uses five classification algorithms,namely full-connection neural network,VGG11 network,Resnet18 network,SVM and KNN,to identify the signals respectively.The recognition accuracy of fully connected neural network is less than 90%,and the recognition accuracy of Resnet18 network is more than 98%.However,SVM and KNN are not as accurate as VGG11 and Resnet18.Then,in view of the insufficient training samples in the complex electromagnetic environment,this thesis constructs a data enhancement method based on generative adversarial network,and improves the overall recognition performance of the model in the scenario of insufficient training samples.(3)Aiming at the recognition of unknown signal of emitter,the KNN algorithm is firstly used to identify the unknown signals in the case of a large number of training samples,and the recognition accuracy reaches 84% when the value of K is 100.Then,the prototypical network model based on deep learning and metric learning was built.The known signal samples were used to build their respective "prototypes",and the distance between unknown signal samples and each "prototype" was calculated to complete the classification.The distance measurement algorithm and other network parameters that are most suitable for the prototypical network are obtained through experiments.The recognition accuracy of this model can reach more than 90% in the case of fewer training samples.Compared with traditional machine learning methods,it requires fewer training samples and has higher recognition accuracy.In addition,this paper uses the data enhancement method to balance the data in the unbalanced training sample scenario,which improves the recognition performance of unknown signals.The effectiveness of the data enhancement method constructed in this thesis is demonstrated again.