Research On Identification Of Specific Radiation Sources Based On Deep Learning

Radio frequency fingerprinting is a process of identifying radiation source equipment,which is divided into the identification of trained radiation sources and the identification of newly added untrained radiation sources.The trained radiation sources have been trained in advance.The radiation source categories in the training library,The newly added untrained radiation source is the radiation source category that appears for the first time in the identification process and is not in the training library.The primary problem of current research is to achieve accurate and efficient identification of radiation source equipment.However,with the increase of the types of radiation sources,The trained radiation source cannot completely cover the device to be tested,and the problem of detection and identification of newly added radiation source equipment becomes increasingly prominent.this essay,the neural network is applied to the field of radiation source recognition.On the one hand,the lightweight network improves the recognition rate and reduces the amount of network computation.On the other hand,for the detection and recognition of new radiation sources,two schemes of clustering and generative adversarial networks are investigated.The main works of this paper are like follows:1.A new radiation source detection scheme based on unsupervised SK-Gand neural network is proposed.The anomaly detection of traditional GAN(generative adversation network)is applied to the new radiation source detection.On the basis of the traditional GAN network,SK(Selective Kernel)module is added,and the traditional convolution layer is changed to the multi-channel convolution layer,and the original pooling layer and full connection layer is removed.Two thresholds are set to determine the boundary between the identified radiation source and the new radiation source,and a multi-discriminator model is proposed.Experimental results show that the improved SK-GAND network has 90% discriminant rate within 5radiation sources and 80% discriminant rate within 12 radiation sources,which verifies the feasibility of binary detection in the field of newly added radiation sources detection,and proves that SK-GAND is suitable for radiation source detection stage.2.A new data processing method is to be proposed,which uses lightweight neural network for trained radiation sources identification.The original I/Q data is normalized and then segmented into fixed image size graphics to form images data setting.Aiming at the problems that traditional network has too many layers and a large amount of computation,the lightweight network is applied to the field of radiation source recognition.The radiation source recognition effects and network model training time of VGG16 traditional convolutional neural network and Res Net18 and Shuffle Netv1 lightweight neural network based on I/Q data set and image data set are built and compared.The feasibility of image mapping in neural network is verified,and this essay selects the Shuffle Netv1 residual network which is more suitable for the field of radiation source identification.3.Design a newly added untrained radiation source recognition scheme based on unsupervised learning clustering algorithm.Firstly,the fingerprint features of signal data were obtained through VMD(variation modal decomposition),box dimension calculation and PCA dimensionality reduction,and k-means and DBSCAN clustering algorithm models were established.By comparing the systematic clustering effects of the two algorithms,the DBSCAN algorithm with the best clustering effect was selected.The results verify the feasibility of the clustering algorithm in the field of new radiation source identification.When the number of equipment is small,the new radiation source can be accurately identified.The more the number of equipment,the worse recognition result,and the clustering algorithm is more suitable for the new radiation source classification after the detection of radiation source.