Research On Radiation Source Identification Technology Based On Joint Data And Model

Radiation source identification is an important part of judging the enemy's combat platform,strategic intent and threat level.From the previous methods based on templates and artificial features,to the recent methods of machine learning and artificial intelligence,the existing radiation source identification technologies have achieved excellent identification performance under the condition that the radiation source mode is fixed and the data is sufficient.However,with the introduction and development of software radio technology,the working mode of the radiation source has become more and more flexible.In addition,in order to combat interference,radiation sources have a large difference between wartime mode and peacetime mode.Existing data-driven recognition methods have poor cross-mode recognition ability for unknown modes.Aiming at the problem of cross-mode recognition of radiation sources,this thesis studies from three aspects: signal fingerprint extraction,multimodal information fusion and robust identification.The main contents and contributions are as follows:First,an improved short-time Fourier transform preprocessing algorithm based on hyperspherical distribution is proposed for the problem of Fat-tailed distribution in the cross-mode characteristic distribution of radiation sources.By improving the distribution of sample features,the cost of class judgment is reduced and the convergence of the network is accelerated.Experiments on cross-mode measured data of radiation sources show that the improved preprocessing algorithm makes the central distribution of radiation sources in the same mode and different types samples more dispersed,and the central distribution of same type samples across modes is more convergent.Second,for the problem of cross-mode steady-state feature extraction of radiation sources,a cross-mode feature extraction method of neural network based on partial modulation feature guidance is proposed.The network is guided to extract mode-insensitive features by embedding the modulated features in the signal-generating model into the training sample labels.Under the condition of using linear discrimination,the cross-mode recognition accuracy is improved by 31.99% compared with the traditional methods.Third,aiming at the multimodal fusion problem of radiation source identification,a multimodal information fusion method based on the heterogeneous graph of sampleplatform information is proposed.A graph convolutional neural network is used to fit the information transfer process between radiation sources.Under the condition of multimodal samples,93.66% accuracy of same-mode recognition and 82.47% accuracy of cross-mode recognition were obtained.Fourth,for the problem of robust multimodal fusion radiation source identification,a radiation source cross-mode identification network based on the relationship between modalities is proposed.By utilizing graph convolution and graph shrinkage techniques,robust identification in the absence of sample modalities is achieved.Under the condition of multimodal information missing samples accounting for 50%,the same-mode recognition accuracy rate of 95.35% and the cross-mode accuracy rate of 86.30% were obtained.