Intelligent Recognition Of Radar Emitter Based On Deep Learning

Electronic reconnaissance is an important part of modern war,an important methods of living electronic intelligence,and a prerequisite for the implementation of electronic countermeasures.Radar emitter signal recognition is an important component of electronic reconnaissance.It is not only the purpose of signal processing of reconnaissance system,but also an important basis for judging the threat status of enemy weapons.Traditional recognition methods can not meet the requirements of automatic and intelligent recognition of radar emitter signals in modern battlefield.In this thesis,we use deep learning theory to carry out intelligent recognition of radar emitter signals,and mainly carry out the following work:(1)For the radar emitter signal recognition based on time-frequency image features,the short-time fourier transform of radar emitter signal is used to obtain the time-frequency diagram,and the Rayleigh entropy feature and the center distance feature are extracted from the time-frequency diagram.Canonical correlation analysis is used to carry out feature fusion of the two features to obtain stable and reliable classification features.(2)For radar emitter signal recognition based on convolutional neural network,we use Dense Net121 model to identify radar emitter signal.And for the problem of low recognition rate under low SNR,two solutions are proposed: one is a combined model based on on-line feature stitching,which combines a 10-dimensional features fused by canonical correlation analysis of time-frequency image and center-distance with features extracted by convolution network,the other one is to achieve knowledge distillation of Res Net152 model by using Dense Net121 model,the model size can be controlled and the model effect is also improved.(3)For an open set radar emitter signal recognition,we propose an open set radar emitter signal recognition model based on one-class support vector machine.Firstly,the test samples are input into the convolutional neural network trained by closed set to obtain softmax probability output,the classification and the confidence score can be obtained by the probability output.Then,the confidence score is input into the one-class of support vector machine to determine whether the sample is ”abnormal”,if yes,mark it as a new category signal,if not,output its category.At the same time,a new confidence score calculation method is proposed.Compared with the maximum output probability of softmax,the confidence score can identify new category signal better.