Research On Application Of Attention Mechanism In Communication Emitter Identification

Radiation source individual identification is a potential physical layer authentication technology that plays an increasingly important role in both military and civilian scenarios,and has gradually become a focus of research.Currently,various emitter individual recognition methods based on deep learning have been proposed,there are few issues related to feature optimization during intelligent feature extraction,which limits the feature learning ability of the model.In addition,existing deep learning methods are data driven and require a large amount of supervised data to effectively complete the emitter individual identification task,greatly limiting their recognition performance in limited training sample scenarios.In response to the above issues,this thesis focuses on improving the emitter feature learning ability of network models,reducing feature information loss,and emitter individual recognition in limited sample scenarios.The specific work is as follows.(1)An algorithm based on attention mechanism and multi-level feature aggregation is proposed to improve the performance of emitter recognition.The algorithm consists of an emitter feature acquisition module and a feature fusion module.For the feature extraction part,attention mechanism is introduced into the fingerprint feature extraction of convolutional neural networks.Channel attention technology is used to filter useful information from the perspective of feature channels,guiding convolutional neural networks to focus on more important channels.At the same time,time-frequency spatial attention technology is used to globally filter regional information from the perspective of feature space,strengthening the network’s feature extraction ability,in order to obtain more radiation source discriminant features.For the feature fusion part,in order to reduce the information loss of features during the convolution process,the algorithm is based on the hierarchical structure of convolutional neural networks,and a multi-scale feature fusion module is embedded in the emitter feature extraction network.This module adds and fuses multi-layer feature information with a certain weight,maximizing the use of shallow and deep information in the network,and the weight is automatically obtained by the network,avoiding manual setting.This thesis proves the effectiveness of the proposed method through experiments such as feature visualization and ablation,and finds that this method can effectively improve the accuracy of individual recognition of radiation sources and enhance the feature expression ability of the model.(2)Aiming at the problem of poor robustness of existing deep learning algorithms with limited training samples,this thesis introduces metric learning and proposes a small sample emitter identification method based on attention mechanism and metric learning.This method includes two parts: feature extraction and classification.For the feature extraction part,attention mechanisms are used to selectively receive and process information in the neural network,improving the network’s ability to analyze and understand data.In order to improve the generalization ability of the network and adapt to feature learning with small samples,the algorithm introduces both center loss and triple loss functions to assist the model in obtaining distinguishable features,and enhances the intra class aggregation and inter class differentiation of features.For the classification part,Euclidean distance measurement technology is used instead of traditional recognition classifiers to maximize the use of the feature space information of the constructed small samples of radiation sources.This thesis establishes a small sample data set of radiation sources,and through simulation results,it is found that this algorithm can complete the individual recognition technology of radiation sources under small samples,and is superior to several existing small sample recognition technologies...