The Study On Specific Aerial Emitter Accurate Identification Method

Specific emitter identification refers to the technology of identifying aerial emitter by extracting one or more unintentional modulation features of the signal emitted by the transmitter.Specific emitter identification technology has broad military and civilian application prospects.With the substantial increase in aerial emitter equipment,it is necessary to judge whether the aerial emitter is working normally and the enemy or friend attributes of the abnormal aerial emitter,and these judgments need to be based on the aerial specific emitter identification.This thesis mainly studies the identification method based multi-modal and multi-scale permutation entropy,and specific emitter identification based on deformable convolutional neural network in a distributed federated learning system.Aiming at the problem of poor identification performance of traditional specific emitter identification methods in a low signal-to-noise ratio environment,a specific emitter identification method based on multi-modal and multi-scale permutation entropy is proposed.First,use empirical mode decomposition and variational mode decomposition to obtain the intrinsic mode functions of the signal at different frequencies.Remove the noise component,and then calculate the multi-scale permutation entropy of each intrinsic mode function.All multi-scale permutation entropy components are used as features.Use the support vector machine classifier to identify the emitters.The simulation results show that the identification performance of this method on subtle features such as phase noise and harmonic distortion is significantly better than some traditional methods,and it has good noise immunity.Aiming at the problem that the traditional deep learning network model is not highly adaptable to the contour feature maps extracted from signals and insufficient training sample data,a distributed specific emitter identification method based on deformable convolutional networks and federated learning is proposed.On the basis of the traditional convolutional neural network,part of the convolutional layer is replaced with a deformable convolutional layer.The deformable convolutional layer can calculate the offset of the convolution kernel according to the input feature map,change the shape of the convolution kernel,and make the convolution operation more biased towards the useful information content of the feature map with higher energy,ignoring part of the background noise information.In the convolution operation,a depth-separable structure is adopted,and the convolution kernel offset is calculated separately for different data in each channel.Through simulation experiments,the effectiveness of the deformable convolutional network is verified.Under the two conditions of insufficient sample size and insufficient samples under certain signalto-noise ratios,it is verified that the distributed federated learning system realizes the multiple local models joint training without exchanging the original sample data,and has a good recognition effect.