Method Research Of Specific Emitter Identification Based Machine Learning

Specific emitter identification refers to using the signals transmitted by the emitters to identify the emitters,their modulation modes and so on.Specific emitter identification technology has become one of the core technologies in electronic reconnaissance system because of its important role in battlefield situation awareness and decision-making.The existing algorithms for specific emitter identification have some problems,such as poor feature completeness,poor identification effect of unknown emitters,and poor stability of cross-session specific emitter identification.Therefore,it is very challenging to achieve accurate and robust specific emitter identification.To this end,this thesis study the feature fusion,unknown specific emitter identification and cross-session specific emitter identification.The main work of this thesis is as follows:1.Aiming at the problem that the existing specific emitter identification algorithm based on artificial features has low accuracy due to poor feature completeness,this thesis proposes a Feature Fusion based on Self-Attention mechanism(FFSA).The algorithm combines the advantages of spectrum,amplitude and phase features to represent information,and mines the synergy between them based on self-attention.The simulation experiment results on three datasets show that the proposed FFSA algorithm improves the identification accuracy by 2.74%,4.73% and 1.82% respectively compared with the existing methods.2.Aiming at the problem that the existing unknown emitter discrimination algorithm based on metric learning is limited by the selection of positive and negative sample pairs,and have poor clustering effect on feature spaces,which leads to the low precision of unknown emitter discrimination,this thesis proposes a new unknown emitter discrimination algorithm based on Generative Adversarial Metric Learning(GAML).This algorithm improves the clustering effect of feature space by generating difficult negative sample features.The simulation experiment results on ADS-B dataset show that the proposed GAML algorithm is superior to existing methods in identifying known emitters and identifying unknown emitters.;especially,compared with the conventional deep metric learning algorithm,the precision of unknown emitters discrimination is improved by4.33% on average;in the case of different numbers of known emitters,the precision of unknown emitters discrimination remains above 90%.3.Most of the existing specific emitter identification algorithms identify emitters at the same session.However,there are differences in the distribution of emitter signal data in cross-session scenario,resulting in a decline in the accuracy of existing specific emitter identification algorithms.Therefore,this thesis proposes the Adversarial Domain Adaptation with Wasserstein distance(ADAW)algorithm for cross-session emitter identification.The algorithm aligns the data distribution in different sessions and extracts common features to realize cross-session specific emitter identification.The simulation experiment results on ADS-B dataset show that the proposed ADAW algorithm improves the crosssession specific identification accuracy by 4.82% and 6.52% respectively compared with the existing methods in the 5-month and 1-year time span.The simulation experiment results on Wi Sig dataset show that the cross-session specific identification accuracy of the proposed ADAW algorithm is improved by 5.81% compared with the existing methods in a one week time span.