Reasearch On Key Technologies Of Communication Emitter Identification Based On Deep Learning

With the vigorous development of mobile communication technology,the number of communication devices has increased rapidly,and the types have become more abundant,which has brought broader development prospects for communication emitter identification technology.Deep neural network(DNN)is the most representative algorithm of deep learning and has powerful feature extraction capabilities which plays an increasingly important role in the field of communication emitter identification and is gradually becoming a research hotspot.The existing deep learning-based communication emitter identification algorithms are mainly based on supervised training methods,which realize high-precision communication emitter identification under the condition of a large number of labeled signal samples.However,in actual application scenarios,problems such as insufficient signal samples and only a small number of labeled signal samples will cause the deep learning model to fail to be fully trained,and it is difficult to obtain a satisfactory identification accuracy.Aiming at the three scenarios where there are only a few training signal samples in practical applications,enough training signal samples but only a few labeled training signal samples,and lack of labeled signal samples in specific scenarios,this thesis carries out the following research:(1)The problem of communication emitter identification under the conditions of small sample size is studied.Aiming at the small sample size problem for network training in practical application scenarios,a communication emitter identification algorithm based on siamese convolutional neural network(SCNN)is proposed.The algorithm is based on a two-way convolutional neural network(CNN)with shared parameters.It uses a contrastive loss function to learn similarity metric,and achieves the dimensionality reduction of signal samples under the condition of maintaining the similarity between pairs of samples,and then designs a simple twolayer fully connected network is used as a classifier to classify the features output by SCNN.The experimental results on the simulated datasets and the actual measured datasets verify that SCNN has a higher identification performance under the small sample size conditions,and the identification accuracy is better than that of CNN.(2)The problem of communication emitter identification in a semi-supervised scenario is studied.In a non-cooperative communication scenario,only a small number of labeled signal samples can be directly obtained,and a large number of signal samples are unlabeled.Aiming at the problem that the performance of the existing deep learning algorithm for communication emitter identification relies heavily on a large number of labeled signal samples for supervised training,a semi-supervised learning algorithm for communication emitter identification based on deep clustering is proposed.This method first uses a convolutional autoencoder(CAE)to map all signal samples to a low-dimensional latent space,and then constructs pairwise constraints between labeled signals with less than 10% of the total data volume and a large number of unlabeled signals to drive hidden spaces to form clusters.The experimental results on the simulated datasets and the actual measured datasets verify the effectiveness of the method,and the identification accuracy is close to the results of the CNN trained with a large number of labeled signal samples.(3)The problem of communication emitter identification under dynamic noise scenarios is studied.Aiming at the problem that the identification accuracy of the existing identification algorithm is seriously degraded due to the inconsistency of the channel environment noise between the unlabeled signal samples to be identified and the labeled training signal samples,a domain adaptation based communication emitter identification algorithm is proposed.This method combines the idea of transfer learning,uses the clustering constraints within the domain and the tightness constraint between classes,and aligns the features of the signals under different signal to noise ratios(SNRs),and extracts the “unchanged” fingerprint features under different SNRs,thereby achieving the identification of communication emitters in the scene of tagged signal samples under the condition of a specific SNR is lacking.The experimental results on the simulated datasets and the actual measured datasets verify that this method can improve the identification accuracy of deep learning models under dynamic noise conditions.