Research On Individual Communication Transmitter Identification Under Small Sample Condition

Communication radiation source individual identification technology refers to the technology that the receiver determines whether it belongs to a certain transmitter by receiving the radiation source signal,which belongs to an important research topic in the field of electronic countermeasures and reconnaissance.With the development of classification technology based on deep learning in recent years,the application of deep learning algorithms in the field of communication radiation source classification has become mainstream,but deep learning algorithms require a large amount of training data for network extraction of class features.The effect of individual identification of radiation sources is poor.Therefore,this thesis studies the individual identification technology of communication radiation sources under the condition of small samples when using deep learning algorithms.The main results are as follows:1.Prototype network and Siamese network method based on metric learning are implemented.Both of them design the neural network and loss function,so that the distance between the feature vectors of the input samples of the same category is small,and the characteristics of samples of different categories are smaller.The distance between vectors is large.By measuring the distance,the recognition accuracy of small samples is improved,and an improved Siamese network method is further proposed to optimize the loss function,so that it can take into account the distance of samples and feature classification.2.A few-shot communication radiation source identification method based on transfer learning is realized.Starting from the structure of the neural network,the method of pre-training and fine-tuning is adopted,so that the neural network shallowly shares the feature classification of communication radiation sources.Therefore,only training is performed in the fine-tuning stage.The network parameters of the latter layer do not require too many samples to participate in training.3.In order to solve the problem of individual identification of unbalanced communication radiation sources,a method based on data synthesis is implemented from the data level,and the SMOTE algorithm is used to expand the sample size of the small sample data,which solves the problem of inconsistent sample size between categories.4.In order to solve the problem of individual identification of unbalanced communication radiation sources,the method based on Focal Loss is implemented from the level of network structure,and the modulation factor and balance factor are introduced into the loss function to improve the neural network's ability to deal with samples in categories with a small sample size.The feature mining of samples that are difficult to identify,and on this basis,the weight optimization of the Siamese network is carried out to solve the problem of training emphasis on categories with a large sample size.Experiments are carried out to verify the effectiveness of the algorithm.5.For the multi-task few-shot communication radiation source individual identification problem,the non-forgetting incremental learning algorithm is used to realize the neural network's recognition effect for both the old task with a large sample size and the new task with a small sample size.The same neural network is used to improve the classification effect of old and new data.The above work has been verified and analyzed by experiments.Through the above results,the existing deep learning algorithms can solve the problems of overfitting,inability to extract enough category features,and poor recognition effect under the condition of small samples.Application scenarios of individual identification of communication radiation sources using deep learning algorithms.