Intelligent Individual Identification Technology Of Electromagnetic Targets In Open Environments

The core task of conventional electromagnetic(EM)target recognition is to effectively distinguish between different types of electromagnetic signals,systems,and working status by mining differences in EM signal parameter features,such as frequencys,bandwidth,and modulation types.The core task of individual EM target identification is to accurately identify different individuals with the same signal type.Since there are no differences in frequency parameters between individuals,it is necessary to extract and express distinguishable subtle "radio frequency fingerprint features" from the radiated signals.This enables the identification of individual wireless communication devices,wireless network devices,radar systems,spectrum monitoring equipment,and other EM targets.It has significant application value in military and civilian fields such as electromagnetic spectrum operations and electromagnetic spectrum management.Through the supervised training with the abundant and high-quality annotated datasets,the deep learning-based intelligent electromagnetic target recognition methods develop the extraction capability of underlying and subtle RF fingerprint features,allowing accurate and efficient identification of target individuals in the closed-static electromagnetic environment.However,the real electromagnetic environments exhibit the characteristic of being open and dynamic,where the electromagnetic targets are complex,varying,dense,and adversarial,leading to the negative factors such as the sparse annotated data(few-shot),the diverse operating modes(cross-mode),and the emergence of unknown categories(new-class).These factors seriously impact the robustness,adaptability,and generalization of deep learning-based electromagnetic target recognition methods.To address these challenges,this dissertation conducts the research on electromagnetic target recognition technologies in open scenarios,focusing on improving the quality of RF fingerprint feature expression and innovating the models of feature extraction and recognition.Specifically,the methods of enhanceing,the robustness of small sample,the adaptability of cross-mode,and the generalization of new category in electromagnetic target recognition are studied comprehensively,and an intelligent recognition system for electromagnetic target is built for validation,assessment,and visualization demonstration.The main contributions of this dissertation are listed as follows:1.For enhancing the quality of RF fingerprint feature expression,the proposed approach focuses on two aspects,namely improving the quality of electromagnetic target signals and highlighting subtle features of RF fingerprints.Firstly,a robust Minimum Variance Distortionless Response(MVDR)beamforming method is proposed by using the Alternating Direction Multiplier Method(ADMM)optimization framework.This method efficiently and robustly controls the sidelobes to physically isolate the multipath clutters,interferences and noise from the sidelobe directions,thereby enhancing the reception of electromagnetic target signals in the main lobe.Additionally,a feature expression method for RF fingerprints based on Variational Mode Decomposition(VMD)and Synchronous Sparse Transform(SST)is designed,where the entangeled and overlapping RF fingerprints caused by different modules are separated by using VMD and then the SST with high time-frequency resolution is employed to transform the subtle fingerprint features from the one-dimensional time domain to the two-dimensional timefrequency joint domain.As a result,the RF fingerprint features are highlighted to provide high-quality data foundation for the subsequent recognition models.2.With regard to RF fingerprint feature extraction and recognition model design,in order to address the robust recognition task for electromagnetic targets under few-shot condition,a recognition model based on the pre-training and composite Siamese network is proposed.Specifically,this model utilizes Vision Transformer(Vi T)and Masked Autoencoder(MAE)self-supervised pre-training to enhance the ability of data representation.Subsequently,multiple simple binary classification Siamese sub-networks are combined to simplify the task of multi-class recognition as multiple binary classification taskes.In this case,the special structure of Siamese network is leveraged effectively to implement the data augmentation.Under conditions of limited training samples(with only 100 samples per class),the proposed method achieves a 10% higher recognition accuracy compared to CNN-based methods,enhancing the robustness of small-sample target recognition.3.Based on the triplet network,a recognition network is designed to recognize the unbalanced cross-mode electromagnetic targets.Under the triplet framework,the network structure and loss function are designed to focus on the common fingerprint features across operation modes.By means of the decision fusion feedback mechanism,the adaptation capability of recognition network is improved and the self-growing capability is formed.The recognition performance of the designed network is increased by 20%compared to conventional CNN networks,significantly enhancing the adaptability of cross-mode individual identification for complex multi-functional targets.4.A class-incremental learning method based on feature distribution learning is proposed to deal with the task of recognizing new-class electromagnetic targets in open environments.By introducing a simple structure of distribution learner between the feature extractor and the classifier,the extremely small amount of distribution information of RF fingerprint features is utilized to represent the electromagnetic targets.In the process of incremental learning for new categories of electromagnetic targets,the feature distribution information of electromagnetic targets is relearned,and the discriminative ability of new categories is enhanced by the feature distribution information of the old categories.As a result,the proposed method,while retaining the recognition capability for old categories of electromagnetic targets,achieves a maximum enhancement of 22%in the identification accuracy for new categories of targets,ensuring the generalizability of recognition for new categories.5.In order to verify the performance of few-shot,cross-mode and new class electromagnetic target recognition in open scenarios,an intelligent recognition system for electromagnetic targets is constructed based on the Spark distributed memory computation framework and the HDFS distributed file system.In this recognition system,the performance validation and the visualized interaction of recognition process are implemented by utilizing the collected real data.The research work in this dissertation comprehensively enhances the robustness of few-shot recognition,the adaptability of cross-mode recognition,and the generalization of new-class recognition,which contributes to improve the performance of electromagnetic target recognition in practical open environments and promote the practical application of electromagnetic target recognition technology in military and civilian fields.