| Electromagnetic space is a national strategic space,and accurate identification of specific emitters is one of the important links to achieve the security protection and management control of electromagnetic space in extreme conditions and non-cooperative scenarios,there are certain data limitations in the identification of specific emitters,such as few shot label samples,incomplete signal,and uneven distribution of samples.Transfer learning technology,which transfers empirical knowledge from similar domains to the target domain,has received extensive attention in the research field of specific emitter identification under data constraints.Currently,the core problem in the research of transfer learning technology for specific emitter identification under data constraints is how to extract empirical knowledge suitable for transfer from limited specific radiation source data and how to transfer empirical knowledge after obtaining it can ensure accurate identification.Aiming at the above core issues,from the perspective of data distribution domain adaptation,this thesis focuses on the three key sciences of "data distribution adaptability,high-dimensional feature subspace potential sharing,and data structure stability in feature subspace" for specific emitter identification.The objective is to explore the theory and method of transfer learning for specific emitter identification under data constraints.In terms of theoretical modeling,this study emphasizes key issues like high-order feature sharing component analysis,and employs mathematical tools and learn the shared space of specific emitter datasets with distribution differences with mathematical tools such as tensor networks and kernel learning.In terms of method innovation,a new approach for dynamic domain adaptation of high-dimensional specific emitter feature distribution is proposed to solve the problem including specific emitter identification under data constraints,category alignment,and open sets.The main content and innovative results of this thesis are as follows.1)For specific emitter identification under data constraints,there is redundancy in high-dimensional features and the difficulty of constructing feature subspaces leads to low recognition accuracy.Based on the distribution adaptability of radio frequency data and the potential sharing of high-dimensional feature subspaces,a specific emitter identification method based on tensor embedding transfer,named the tensor embedding radio frequency domain adaptation(TERFDA)model,is proposed in this thesis.Initially,the model extracts high-dimensional features of specific emitter by considering the temporal stability of signals,which expands one-dimensional signals into two-dimensional matrices to retain the time-domain variation of signals and compensate for the incomplete information loss of signals.Then,tensor shared subspace learning is introduced to construct high-dimensional feature subspaces for both auxiliary data(source domain)and recognition task data(target domain).Further,dynamic domain adaptation is performed on the distribution of high-dimensional feature data of source and target domains applied in the shared subspace.Finally,the prediction of target domain emitter labels is realized based on structural risk minimization theory.TERFDA is a basic framework for identifying specific emitters under data constraints.The algorithm module in the basic model can be improved to achieve the target task if the environment changes and the difference in data distribution increases.2)In order to achieve specific emitter identification under data constraints and address category misalignment,a specific emitter identification method based on substructure-preserving migration is proposed,namely the tensor embedding substructure preserving domain adaptation(TESPDA),which can solve the problem of increased inter-class confusion in the tensor shared subspace caused by the increase of target types in the source domain,and the decrease in the prediction accuracy of pseudo-labels.First,the substructure preservation constraint is added to the tensor shared subspace learning.The constraint matrix operates on the sample dimension of the tensor,ensuring the inter-class structure of the source domain remains unchanged during the feature space learning process,and can optimize the data structures in the target domain.Secondly,this thesis proposes a double-cluster fusion(DCCSC)pseudo-label prediction algorithm,which combines the unsupervised clustering results of the source domain clustering cluster centers and the target domain structure prediction cluster centers.Unlike the pseudo-label prediction method that solely relies on source domain class prototype clustering,the algorithm in this thesis enhances the data structure of the target domain and further performs complementary fusion of the two clustering results through the D-S fusion algorithm.Consequently,the accuracy of pseudo-label prediction is effectively improved when the categories between the two domains are not aligned.3)Facing the specific emitter identification task with limited data and open set,for the closed set dynamic domain adaptation method,the unknown class is misidentified as known,which leads to the problem of increasing the error of the data distribution adaptive algorithm.This thesis proposes a specific emitter identification method based on unsupervised open-set domain adaptation,namely tensor embedded substructure preserving open set domain adaptation(TESPOSDA)to address the problem of misidentification of unknown classes as known classes in the closed-set dynamic domain adaptation method leading to increased errors in data distribution adaptation.Firstly,the model designs the open set rejection domain module to divide the known classes and unknown classes in the target domain.Within this module,an open set rejection domain learning method based on metric-based meta-learning is proposed,in which the extreme value distribution function parameters are used as known class constraint parameters in open set.To prevent the problem of inflexibility of known class constraint parameters caused by fixed extreme value thresholds,a metric-based meta-learning is used to optimize the optimal extreme value thresholds.Furthermore,building upon the initial division of known and unknown classes in the target domain,an open set pseudo-labeling algorithm based on double-cluster clustering fusion under open set(DCOSC)is proposed.It takes the unknown cluster centers in the target domain as the prototype of the N(10)1th class in the source domain,so as to obtain the clustering result with the unknown class,and fuse it with the N(10)1 class clustering result based on the structured prediction in the target domain.Finally,through unsupervised open-set domain adaptation,the data distribution of known and unknown classes is respectively adaptive.The approach in this thesis successfully increases the number of unknown targets in the open-set scenario without compromising the recognition accuracy of known targets. |
