Specific Emitter Identification Based On Multi-source Data

Specific emitter identification is a crucial technology in cognitive electronic warfare and has experienced unprecedented development in recent years.This technology relies on fine fingerprint features formed by transmitter components for identification.Since the process of receiving signals is the inverse of transmitting them,the receiver introduces receiver fine features during signal reception.However,current research mainly focuses on recognition accuracy and algorithm spatio-temporal efficiency and doesn’t consider the influence of receiver features on radiation source recognition.In actual application scenarios of specific emitter identification,multi-platform multi-receiver identification and old receiver damage often occur,requiring the use of data from different receivers for identification.To address this,this paper utilizes multi-source radiation source signals from various application scenarios and data sources,combined with machine learning algorithms,to perform the following:(i)Design and analyze a single-receiver data specific emitter identification method based on residual networks.We use the residual network as a feature extractor to identify the radiation source with single-receiver data,further enhancing the feature extraction capability of the network model by adding convolutional attention.Experiments show that the residual neural network with added convolutional attention has the best recognition effect,with recognition accuracy of 99.9% and 93.77% on different datasets,respectively.The recognition performance of the residual recognition network decreases when the training and test data come from different receivers,with an average recognition accuracy decrease of 46.4% and 46.59% under different datasets,respectively.(ii)Design a receiver device-independent recognition method based on information entropy.In the absence of new receiver data,receiver-independent features are extracted for specific emitter identification by feature separation,information entropy maximization,and step-by-step training to achieve receiver device-independent identification.Experimental results show that the receiver device-independent recognition network improves recognition accuracy by 8.61% and 21.65% on different datasets,respectively,compared to the residual recognition network when using data from two different receivers for training and the third receiver data for testing.(iii)Design a self-supervised receiver data-based adaptive recognition method.When unlabeled data from a new receiver is available,and only the trained network model of the original receiver can be used,we propose an adaptive specific emitter identification method for receiver data.Firstly,the network model trained on the original receiver is loaded,and the network is adjusted by information entropy minimization,pseudo-label self-supervision,and contrast learning to achieve adaptive new receiver data.Experiments show that compared to the residual recognition network,the receiver data adaptive recognition network improves recognition accuracy by 18.78% and 23.10% on different datasets,respectively.