Individual Feature Extraction And Identification Technology Of Specific Emitter Based On Deep Learning

Specific Emitter Identification(SEI)has promising applications in identifying electromagnetic interference,managing radio signals,and detecting and identifying enemy signal types.However,the key challenge in SEI is extracting essential features from signals with strong differentiation power.The use of network models,such as Convolutional NeuralNetworks(CNN),can be continuously trained to extract individual signal features and improve SEI performance.In this thesis,we focus on SEI based on the CNN network model using the current UAV public dataset,DroneRF,for SEI.We investigate three identification tasks: two-category UAV detection,four-category UAV type identification,and ten-category UAV flight pattern identification.The main work of this thesis is as follows:(1)For the two identification tasks of UAV detection and UAV type identification,a simpler network model is considered to require fewer computational resources to reduce the cost of building the model.A network model with only 10 layers is proposed,which contains a OneDimensional(1D)CNN and a Long Short Term Memory(LSTM)network.In order to facilitate the extraction of features from the data samples by the 1D CNN,the original Radio Frequency(RF)signals are pre-processed with data,such as signal splicing and conversion to time-domain signals for modal values.Then a CNN-LSTM network model is constructed,which includes a 1D CNN layer,a Max pooling layer,a LSTM layer,a Dropout layer,a Flatten layer and a Fully Connected layer.The convolutional layer is used to extract features from the input signal,and the Max pooling layer is used to reduce the dimensionality of the feature map to reduce the number of model parameters,the LSTM layer is added to better capture the temporal feature information output by the CNN layer.The simulation results show that the model has good identification effect for both UAV detection task and UAV type identification task,with99.80% and 98.25% identification accuracy,respectively.However,the model is not effective in identification various UAV flight patterns.(2)In response to the low effectiveness of CNN-LSTM in(1)for the identification of various UAV flight patterns,propose the Inception-ResNet-v2-ECANet network model.Firstly,the input of the model is preprocessed.The process involves converting 1D time-domain data into Gramian Angular Field(GAF)image showing the characteristics between the two data,and extract the spatial association information of the input data samples.Additionally,due to the unbalanced samples of each signal class in the dataset,Gaussian Noise or Salt And Pepper Noise is added to the generated GAF images for data augmentation.This ensures that the same number of samples from each class is used for model training and testing,thereby improving model generalization.Afterwards,three Efficient Channel Attention(ECA)modules are added to the original Inception-ResNet-v2 model.Inception-ResNet-v2 inherently has a large number of trainable parameters to ensure that the model learns subtle features.The addition of the ECA module helps Inception-ResNet-v2 to dynamically adjust to important details,capture feature maps Interdependencies between channels and spatial location relationships improve model performance.The Inception-ResNet-v2 network model has high identification efficiency.Simulation results show that,The proposed Inception-ResNet-v2-ECANet identification performance is improved by about 50% compared with the Deep NeuralNetworks(DNN)model.(3)Considering that the Inception-ResNet-v2-ECANet network model has a high cost of spatial complexity under the condition of limited computing resources,A lightweight D-A XceptionNetwork model based on Dilated Convolution and Dual Attention(DA)module is proposed to solve the practical problem of network model in reality.Firstly,data preprocessing was carried out for the input data samples of the network model.The process is as follows:Savitzky-Golay filter is performed for each RF signal in the frequency-domain as the frequency changes,then normalization is performed to fix the signal relative intensity scale,and finally the corresponding signal relative intensity image is generated,and two types of noise are added for data enhancement in the same way.After that,on top of the original Xception network model,DA is used to extract different subtle features present in the feature map,and Dilated Convolution is introduced to expand the perceptual field to better capture the contextual information in the feature map,so the D-A Xception network model has better results.The simulation results show that the D-A Xception effectively reduces the network complexity when the accuracy is unchanged.Compared with the Inception-ResNet-v2-ECANet network model,the space complexity is reduced by half.