Research On Technology Of Fingerprint Feature Extraction And Classifier Design Of Emitter

It is of great significance to identify specific emitter,whether it is the identification of friend or foe and identification of identity in military applications,or the spectrum management,equipment fault diagnosis and identification of network access in civil applications.Due to the hardware difference of the emitter,the transmitted signal may be unintentionally modulated.Through the analysis of the received signal,the signal characteristics that can reflect the transmitter's identity can be extracted.These characteristics are unique,stable,subtle and undeceiving.Based on the labels of train set,this paper studies the key technologies of specific emitter identification from three aspects: single class feature extraction and joint feature extraction under supervised conditions;neural network classifier design under unsupervised conditions,respectively considering the correct identification rate,robustness,universality of fingerprints and new scenes.The main work is as follows:1.In order to improve the accuracy and universality of fingerprint recognition,a feature extraction algorithm based on texture of bisectrum image was proposed.Bacause dimension of bisectra feature is high,while traditional dimensionality reduction methods,such as seletive bispectra,can only retain partial information of bispectrum and ignore the whole information.First,the bisectrum is calculated to obtain its image.Secondly,the image texture features are extracted by the gray level co-occurrence matrix,and the 16-dimensional feature vectors are formed.Finally,the individual recognition is completed by using the support vector machine.The simulation results show that the algorithm has a higher recognition accuracy than seletive bispectra or the texture features of Hilbert spectrum.At the same time,when the number of targets and modulation mode change,the algorithm has better adaptability.2.Considering the robustness and universality of fingerprint,a joint feature extraction algorithm based on signal complexity and entropy is proposed to solve the problem that a single class of feature is easy to drift.First,the instantaneous amplitude,frequency and phase of the signal are calculated.Secondly,the box dimension,information dimension and information entropy are calculated respectively.Thirdly,the entropy of envelope spectrum is extracted by this algorithm.Finally,10 features are taken as joint features to identify specific emitters.Simulation results show that,compared with single feature,the joint feature has larger variance,that is,higher dispersion.And the accuracy is improved at a small time cost.When the signal to noise ratio(SNR)and modulation mode change,the joint feature performance is better.3.Based on the transfer idea of domain adaptive,this paper proposes a classifier design method based on the maximum classifier discrepancy.Because distribution alignment between the target and the source is difficult and class boundary often generate ambiguous features which lead to misclassification.The algorithm uses the existing database as the source domain,and the unmarked samples to be detected as the target domain.First,the bispectrum is input into two independent CNN classifiers as the original feature.Secondly,according to the output discrepancy of the two classifiers to the target domain,the minimum-maximum optimization method is adopted to realize the adversarial learning between the feature generator and the classifier.Finally,in the process of adversarial learning,domain adaptive is realized,that is,the target domain samples away from the distribution boundary of the source domain are detected,and generate the characteristics of the target domain near the distribution boundary of the source domain repeatedly.The measured results show that it has a good transfer effect for the identification of newly emerging equipment and user signals collected in different time periods.