Individual Indentification Of Communication Emitter In Complex Environment Based On Deep Learning

The individual identification of communication emitter has been used widely in military reconnaissance,wireless security and other fields.Traditional method normally extracts expert features before taking identification with classifier.However,it can no longer meet the requirements in the complex electromagnetic environment.In recent years,preliminary attempt have been taken in the field of emitter identification with the help of deep learning technique,while much of the existing studies are limited to apply the successful models in other fields directly to emitter identification,cared less to the problems of emitter identification in practical application.Therefore,based on various models of deep learning,this paper studies the identification problems in three complex environments.1.In order to solve the problem of identification difficulty caused by the appearance of unknown types of signal,a semi-supervised individual identification method of unknown emitter is proposed.This method can automatically learn more essential and comprehensive fingerprint features through supervised training feature extraction network,and then identify unknown types of emitters based on clustering algorithm,which simplifies the traditional method of complex locking operation and reduces the dependence on prior knowledge.By analyzing the accuracy of emitter number identification and ARI index,it is verified that the performance of this method is better than that of traditional artificial design features,which shows that this method can extract more representative features without prior knowledge.2.In order to solve the problem that the model is prone to over fitting and unable to learn the local characteristics due to the small number of labeled signal in non-cooperative scenarios,a communication emitter signal enhancement model based on GAN is designed.The network consists of a generator and a discriminator.The network makes the generator learn the distribution of the real signal,and to generate fake signal enough to deceive the discriminator,so as to complete the expansion of signal samples.According to the characteristics of emitter fingerprint signal,the network structure is designed,and the structure of generator and discriminator is optimized.By analyzing the fingerprint feature distribution of the generated signal and the real signal,the usefulness of the expanded signal is proved,and the sample enhancement network is combined with the emitter identification network to verify the effectiveness of the sample enhancement network by comparing the identification accuracy before and after the expansion.3.In order to solve the problem that the distribution of the signal to be identified and the training signal is inconsistent in the dynamic channel,the emitter recognition network is constructed based on the idea of transfer learning,which can realize the knowledge transfer from the source domain to the target domain and improve the recognition accuracy in the target domain.In this paper,the method of getting better effect in target domain based on source domain signal training is studied in theory.Combined with the characteristics of emitter signal,an emitter recognition model based on transfer learning is designed.The main improvement is to increase the domain discriminator and gradient inversion layer.By maximizing the error of the domain discriminator,the feature extraction network can learn the features with both discrimination and domain invariance,and achieve the alignment of the source domain and the target domain in the feature space.