Research On Individual Identification Method Of Communication Emitter Under Small Sample Condition

In the field of military combat,individual identification of communication emitter is of great significance to improve the combat capability of communication countermeasures.By extracting the subtle features of signals from the enemy's communication emitter and using the prior information to determine which communication emitter the signal comes from,it provides an important basis for accurately predicting the enemy's strategic and tactical intentions.Moreover,in military-to-civilian applications,this technology also has important application prospects and value in wireless communication network security,cognitive radio and mechanical fault diagnosis.Based on the field of communication countermeasure and the in-depth analysis of the current status of individual identification methods for communication emitter,this dissertation proposes an individual identification method of communication emitter under small sample condition.This method focuses on the key technologies such as feature extraction,feature dimensionality reduction,and classifier design in individual identification of communication emitter under small sample condition.The effectiveness of the proposed method is verified by actual collected data.The main work of this dissertation is summarized as follows:(1)According to the different states of transmitted signals,the feature extraction method of communication emitter is studied from two aspects: transient features and steady-state features.In the aspect of transient feature extraction,to solve the problem of accurate detection of the starting point and ending point of transient signals,the improved phase starting point detection method and the end point detection method based on HHT energy trajectory adaptive threshold were studied.On this basis,a feature extraction method based on HVG-NTE was proposed to improve the anti-noise performance of features.In terms of steady-state feature extraction,the SIB-based feature extraction method is studied after the comparison and analysis of the four integral bispectral properties.The steady-state features obtained by this method can suppress the Gaussian noise and better characterize the differences among communication emitters.(2)To solve the high dimensionality and small sample size problem of the extracted features,two kinds of nonlinear feature dimensionality reduction method for communication emitter are studied according to whether the tag information is used.When tag information is not used,an unsupervised feature dimensionality reduction method based on MKPCA is proposed,which can effectively reduce the dimensionality of high-dimensional bispectral features and improve the operation efficiency.When partial tag information is used,a semi-supervised feature dimensionality reduction method based on ESDA is proposed to map high-dimensional bispectral feature data to low-dimensional subspace,so as to improve the separability of features.(3)In the aspect of classifier design,to illustrate the advantages of the scheme of dimensionality reduction and then recognition,firstly,a classifier is designed to recognize high-dimensional features directly.Based on the correlation entropy model,the classifier combines the reconstruction residues and discrimination information in the sparse representation coefficient,which can alleviate small sample size problem to some extent.Then,for the dimensionality reduction features,an improved SVM-KNN classifier is proposed based on the research of SVM-KNN classifier.The classifer defines the distance from each class of sample to be detected to the support vector belonging to the class as manifold geodetic distance,which is insensitive to the selection of kernel function parameters.While greatly improving the real-time performance,it has better identification performance than the CSRC,and is suitable for the case of small sample size.In this dissertation,comparative experiments are carried out in terms of feature extraction ability,feature dimensionality reduction effect and classification performance.The experimental results show that in terms of feature extraction,the HVG-NTE-based feature extraction method improves the average recognition rate by 5.7% compared with the NPE based method.Compared with methods based on RIB,CIB and AIB,the SIB-based feature extraction method improves the average recognition rate by 4.5%,6% and 8.9%,respectively.In the aspect of feature dimensionality reduction,the average recognition rate of the dimensionality reduction method based on MKPCA is 16.9% and 5% higher than that based on PCA and KPCA respectively.Compared with the RDRA-based and SDA-based methods,the average recognition rate of the ESDA-based method is 19.3% and 4.2% higher,respectively.In the aspect of classifier design,the recognition rate of CSRC is 5.2% higher than that of SRC.The average recognition rate of the improved SVM-KNN classifier is over 2% higher than that of the SVM-KNN classifier,5.2% higher than that of the CSRC,and the computation time was reduced by 165 times.In order to verify the effectiveness of the proposed method,a verification environment composed of signal receiving and collecting subsystem and individual identification software subsystem is constructed.According to the state of signal samples,whether the tag information is used,three identification modes are established: an individual identification experiment was carried out for the similar communication emitter with the same plant,batch and model.The recognition rate was between 85% and 98%,which can meet the individual recognition requirements of communication emitter under small sample conditon.