Itelligent Open Set Recognition Technology For Specific Emitter Identification

Specific emitter identification technology of radiation source is a kind of technique to extract external characteristics from the signal to accurately classify and identify the radio transmitter.By applying the individual identification technology of radiation source,it can detect illegal radiation source or illegal frequency equipment in electromagnetic environment and improve the ability of perception and control of electromagnetic target in electromagnetic environment.The current method of individual identification of radiation sources generally uses pattern recognition or deep learning methods to extract features,and then uses classifiers to make reasoning predictions.Because of the advantages of deep learning method in generalization ability and robustness,the traditional pattern recognition method is gradually replaced,and the deep learning method applied to individual identification of radiation source is generally only better on closed data,and the performance in open set data needs to be improved.However,the number of individuals with radiation sources in electromagnetic environment is increasing,and the individual identification of radiation sources must be faced with the identification of new target individuals.In this paper,the intelligent open set recognition method based on deep learning is introduced in the field of radiation source individual identification,and three methods of individual open set recognition of radiation source are proposed for different situations:1.Since the SoftMax function is used for label prediction,the typical neural network recognition model cannot perform open set recognition.Without retraining the neural network,it is necessary to optimize the inherent closed set properties of the SoftMax function.It is proposed to use the activation vector of the last layer of the neural network to construct the probability of unknown individuals through the extreme value theory to expand the original neural network model.Output probability to realize open set recognition.Experiments show that by using a network model trained based on the closed set hypothesis,the method achieves a 96% open set recognition rate on simulated data,and an open set recognition rate close to 90% on measured data.2.Aiming at the difficulty that the high similarity of individual transmitters leads to small distances between classes,this paper proposes to use prototype learning technology to build individual feature prototypes of the radiation source,and use the two loss functions of discriminant loss and prototype loss for model training to reduce the number of features in the feature space.The inter-class distance and the intra-class distance fully shrink the area occupied by a single radiation source individual category in the feature space to achieve a better open set recognition effect.Experimental results show that the method can achieve the current best recognition performance.The open set recognition rate on the simulated data reaches 97%,and the open set recognition rate on the measured data reaches 96%.3.Considering that the actual data situation of individual radiation source recognition is big data with few labels,this paper proposes to use semi-supervised learning method to improve the utilization of unlabeled data.In order to reduce the reconstruction error and simplify the network structure,this paper optimizes and improves the existing semi-supervised confrontational autoencoder model in image recognition,removes the identification network that controls the generation style in the network,and adds the identification network formation in the sample reconstruction process In the confrontation process,the reconstruction error between the input sample and the reconstructed sample is finally used to realize the open set recognition.Experiments show that through the use of unlabeled data,only 30% of the labeled data is needed on the actual data set to obtain an open set recognition rate that exceeds the comparison algorithm.