Research On Specific Emitter Identification Method Of Open Environment Based On Deep Learning

Specific Emitter Identification(SEI)is a technology that utilizes the radio frequency fingerprints of signals to distinguish different emitters.This technology has received a lot of attention in recent decades and has met the needs of various applications such as military,electronic intelligence,wireless network security,and the Internet of Things.However,in the real-world open environment,the complex electromagnetic environment caused by channel fading makes it difficult to extract individual fingerprint features of radiation emitters.Additionally,an unknown radiation emitter can cause the entire identi-fication system to lose credibility,and the system needs to maintain autonomous learning from the received information.To address these challenges,this thesis proposes an open-set SEI algorithm and an incremental SEI algorithm,as well as a SEI framework based on edge-cloud collaborative reasoning using the two algorithm models.The main work of this thesis is as follows:1.For the open-set SEI scenario,this thesis proposes an open-set SEI algorithm based on the Maximum Interval Open Boundary(MIOB),which addresses the problem of de-creased closed-set recognition accuracy caused by channel changes,while existing algo-rithms only consider the rejection effect on unknown categories.The MIOB algorithm first introduces attention mechanism to alleviate the feature confusion between categories by aligning the feature space,and finally constructs a classification boundary between known and unknown categories through a one-vs-all classifier.On the publicly available Wi Sig dataset and the self-collected ADS-B dataset,the harmonic mean accuracy of the MIOB algorithm improved by 3.19%and 3.95%compared to existing algorithms.2.In the continuous learning of new category knowledge scenario for the SEI model,existing research mainly addresses the catastrophic forgetting problem by replaying histor-ical data,but ignores the limitations on algorithm performance imposed by device storage and data privacy.Therefore,this thesis proposes an algorithm based on Self-Supervised Synergistic Augmentation(S~3A)for incremental learning.The S~3A algorithm is based on a signal data augmentation method that uses category prototypes to replace old cate-gory samples,reducing storage costs and protecting data privacy and security.Compared with existing algorithms,the multi-stage recognition accuracy of the S~3A algorithm in-creased by 2.11%and 2.27%on the publicly available Wi Sig and ORACLE datasets,and by 10.65%on the ADS-B dataset.3.For the deployment scenario of SEI systems in open spaces,this thesis addresses the problem that existing SEI systems are difficult to deploy for a long time in uncertain,unstructured,and dynamic real environments.Therefore,this thesis proposes an evolvable SEI framework based on the collaborative inference of the MIOB algorithm model and the S~3A algorithm model.The framework includes multiple modules such as edge-side rapid inference,cloud-side asynchronous verification,edge-cloud collaborative conflict resolution,and model evaluation and evolution.The implemented system can be deployed on a large scale in real open environments and can to some extent meet the needs of military SEI and wireless network security.