Communication Emitter Identification Based On Combinations Of Data Driven And Knowledge Informed Methods

Communication emitter identification refers to the process of identifying emitter individuals based on the hardware characteristics contained in received communication signals,which has wide application prospect in both military and civilian fields.Current communication emitter identification methods are divided into two categories: methods based on expert knowledge and methods based on deep learning.Methods based on expert knowledge require small amounts of data,but they are unable to model emitter characteristics precisely,and are only effective under certain conditions.Methods based on deep learning overcome the weakness of expert knowledge based methods,but they require completeness of training data,which are not desirably practical.To address the issue of data incompleteness in practice for deep learning based methods,this thesis combines data-driven and knowledge-guided methods for communication emitter identification,by introducing expert knowledge to deep learning based methods.More specifically,this thesis aims at three kinds of data incompleteness: sample incompleteness,individual incompleteness and scenario incompleteness,which lead to three issues: identification with limited samples,open-set identification and cross-domain identification.Three ways of combining data-driven and knowledge-guided methods are considered: observational bias,inductive bias and learning bias.For each kind of data incompleteness,feasible ways are selected to combine data and knowledge and corresponding approaches are proposed.Finally,all approaches are combined to build a communication emitter identification system for incomplete data,which achieves desirable performance with coexistence of the three data incompleteness conditions.The main contributions of the thesis are summarized as follows:1.For emitter identification with limited samples,the research is carried out from observation bias and learning bias,respectively.From the perspective of observational bias,an communication emitter identification approach based on data augmentation is proposed.Effectiveness of existing data augmentation methods for communication emitter identification is analyzed,and a new data augmentation method,namely AmplitudePhase Mix,is proposed to overcome the weakness of current data augmentation methods.From the perspective of inductive bias,an communication emitter identification approach based on unsupervised pretraining is proposed.The approach regularizes neural networks by learning generation information of signal samples unsupervisedly and improves generalization ability of learning models.Experimental results show that the two proposed approaches improve performance of communication emitter identification with limited samples effectively,and combination of the two approaches can further increase accuracy.2.For open-set communication emitter identification,the feature extraction network and classifier are improved by introducing learning bias and inductive bias,respectively.For the feature extraction network,from the perspective of learning bias,an approach based on additive angular margin loss is proposed.The approach enhances discriminability of learned features by mapping features to angular space with additive margin.For the classifier,from the perspective of inductive bias,an approach based on adaptive Open Max is proposed.The approach distinguishes unknown emitters by modeling distributions of known emitters using extreme value theory,and the original Open Max classifier is modified to match feature extractor trained by additive angular margin loss and to select the hyperparameter adaptively.Experimental results show that the proposed feature extraction network enhances discriminability of learned features,and the proposed classifier increases accuracy of distinguishing unknown emitters with cost of slightly decreasing accuracy of known emitters.Combining the two approaches can improve performance of open-set emitter identification,especially when the number of training emitters is limited.3.For cross-domain communication emitter identification,problems of cross-channel identification and cross-frequency identification are addressed by introducing inductive bias and learning bias respectively.For cross-channel identification problem,an approach based on self-attention network is proposed,from the perspective of inductive bias.The approach can overcome weakness of convolutional neural networks,which are hard to capture long-range information and contain limited learning capability.For cross-frequency identification problem,an approach based on unsupervised domain adaptation is proposed,from the perspective of learning bias.The approach assumes emitter characteristics of different frequencies are similar and utilizes adversarial training to align features of different frequencies.Based on the frequency-invariant features,accurate cross-frequency communication emitter identification can be achieved.Experimental results show that the proposed approaches can improve performance cross-channel and cross-frequency communication emitter identification respectively.4.A communication emitter identification system for incomplete data is constructed by combining proposed approaches,and is evaluated by real data with coexistence of the three kinds of data incompleteness.Experimental results show that the constructed system achieves desirable performance with coexistence of the three kinds of data incompleteness,and each approach proposed by this thesis is critical for the identification performance of the system.