Research Of Specific Emitter Identification With Small Samples

Individual identification of communication emitter is gradually playing an important role in the fields of spectrum management,electronic countermeasures and access identification.So it gets more and more attention from researchers.With radiation source models become more and more abundant,there comes more and more complex signals.Extracting expert features is time-consuming and laborious and may not be able to extract subtle features,especially for signals of different individuals of the same type.Different from past algorithmes,deep learning can automatically extract high-level features,and fit more complex functions.So deep learning has become the mainstream research direction in the field of specific emitter identification.But deep learning models are data-driven methods,and each category requires a large number of labeled samples to support training.However in practical application scenarios,it is difficult to obtain labeled data that satisfies deep learning models.When dealing with small sample data sets,it is very difficult to reach the satisfactory identification accuracy,which severely limits the applicable scenarios of radiation source identification.Aiming at the problem of individual identification of communication emitters under the condition of small samples,this thesis combines genetic algorithm with siamese network,and merges a varity of information,and classifies sample pairs to realize the effective identification of communication emitters.The main work of this thesis includes the following two aspects:(1)In this thesis,the small sample data set includes multiple different individuals of the same type,and different individuals have the same bandwidth,waveform/modulation method.Premeditating the strong uncertainty of artificial neural network construction,this thesis builds an algorithm based on genetic algorithm to construct adaptive network and siamese network,which can effectively identify communication emitters with small samples.In the thesis,convolutional neural network is considered as a structure consisting of multiple convolutional blocks and one output block.Then the parameters such as the number of convolutional layers in the convolutional block,the number of convolutional kernels in each convolutional layer and the kernel size are optimized by genetic algorithm,which can discover a better model structure and be called adaptive network.Further,the adaptive network is regarded as a feature extractor and the output of the middle layer is seen as the network feature.Next the network feature is input into the siamese network to get the identification result.The proposed method can improve the recognition accuracy and enhance the stability of the model by combining the adaptive network and the siamese network.(2)In the small samples condifion,there are little information that the model can use to learn.In order to make full use of the information,this thesis constructs a small-sample radiation source identification method based on multi-layer information fusion.The proposed method extracts the output of each convolution block,then reduces the dimension of the outputs of adjacent convolution blocks to obtain multiple new features,and inputs the new features and the output of the last convolution block into the output block to obtain multiple decisions.Finally multiple decisions are merged to get the final result.Meanwhile,the Siamese network is trained by randomly generated sample pairs.However the randomly generated sample pairs may include sample pairs that are not helpful or even unfavorable for network training.The situation is classified and discussed,then the chosen sample pairs will be recognized or deleted.So the network is trained by the optimized training set.The proposed algorithm merges information of multi-layer,and optimizes the sample pairs for training the Siamese network.The results shows that designed method can accelerate the model convergence,further improve the model accuracy,and reduce the fluctuation of the recognition rate.