Research And Application Of Transfer Learning In Individual Identification Of Few-Shot Radiation Sources

With the rapid development of the Internet of Things,more and more radiation source devices are widely deployed,and radiation source identification has important application value in the field of safety supervision.The radio frequency fingerprint identification technology of the radiation source comes from the signal characteristics of the circuit difference,which is a safer and more reliable identity authentication technology than traditional cryptography.The research on radio frequency fingerprint identification technology for large sample data sets is relatively comprehensive,but in the situation of increasing number of new radiation source devices,the acquisition and labeling of signal data are more difficult.For few-shot radiation source scenarios,the existing radio frequency fingerprint identification technology and There are still gaps.In order to solve the problem of individual identification of radiation sources with few-shot,this study adopts the application of radio frequency fingerprints to transfer learning to realize the identification method of radiation sources with few-shot.In this study,through the research on the feature extraction method of radio frequency fingerprints,the method of using the original I/Q data as the input of the model under non-cooperative conditions is obtained,which is also applicable to the individual identification of radiation sources.In order to solve the problem of performance degradation caused by overfitting of the few-shot training model,transfer learning can be used to learn the knowledge of related tasks and then migrate to the target few-shot data set.Therefore,this study studies the transfer learning based on metric learning,and improves and designs two models.First of all,this study proposes an improved relation network model,using LSTM as a supplementary feature extraction on a single feature extractor,and the embedding module in the form of two branches can mine more feature information of I/Q signals in time series,so as to obtain more good classification condition.Then,according to the characteristics of I/Q signal,this study proposes a prototypical network model based on complex-valued convolutional neural network,which is composed of complex-valued convolutional neural network as feature embedding and classifier based on ensemble learning,which can better Adapt to the radiation source identification problem and improve the identification accuracy of the model.The experimental results show that the improved relation network model improves the recognition accuracy of the system and has more accurate recognition results.The experimental results also show that the improved prototypical network model not only performs well in recognition effect,but also performs better in terms of training time and total number of model parameters.Both improved models have their own advantages in few shot radiation source individual recognition.Finally,the radiation source detection and analysis system designed based on the above two models also has good practicality.By invoking the model,the situation to which the radiation source belongs can be quickly analyzed to help the regulatory authorities to detect and exclude abnormal radiation sources in a timely manner.