| Electromagnetic spectrum resource is one of the most critical national strategic resources,which has high economic,social and military value.With the rapid development of mobile communication,satellite communication,Internet of Things and other radio technologies,the process of information society continues to deepen,but it also makes the electromagnetic spectrum resources increasingly scarce and the electromagnetic spectrum space increasingly “crowded”.It has been a great challenge that how to intelligently and efficiently monitor and control the electromagnetic spectrum in the development of radio technology.Electromagnetic signal identification technology is one of the most significant part of electromagnetic spectrum monitoring,which plays an important role in the fields of unknown signal modulation and parameter analysis,illegal radiation source identification,spectrum usage measurement and so on.In view of the outstanding performance of deep learning(DL)in many complex and nonlinear practical problems and its capability to mine the inherent laws and deep features from massive and high-dimensional samples,many researchers have introduced DL into the electromagnetic signal identification technologies,such as automatic modulation classification(AMC)and specific emitter identification(SEI),and achieved much better performance and efficiency than traditional algorithms.However,there are a series of problems in the existing DL-based electromagnetic signal identification methods,such as “weak generalization,poor robustness”,“redundant parameters,high complexity” and “huge cost,poor privacy”.Based on the theories of universal approximation,vicinal risk minimization(VRM),few-shot learning ensemble learning,model compression and acceleration,sparse representation and distributed learning,this paper proposes limited/few sample-based,lightweight,and distributed electromagnetic signal identification methods,and verifies the effectiveness of the methods proposed in this paper on the open-source dataset.The main research content of this thesis is given as follows.For “weak generalization,poor robustness”,a mixed training,data augmentation and consistency regularization-based robust AMC method is proposed under dynamic environment and limited samples.Firstly,the receiver signal-to-noise ratio(SNR)has a large dynamic range and changes rapidly,bacasue of the complexity and dynamics of the electromagnetic environment.The traditional AMC method based on high-precision SNR estimation algorithm is lack of robustness in dynamic SNR environments.Inspired by the universal approximation theorem,this paper proposes a robust AMC method based on convolutional neural network(CNN)and mixed training.The proposed method uses multiple SNR signal samples to train CNN,and these features extracted by CNN are generalized and robust in dynamic SNR environments;Secondly,considering that there are too few signal samples available in non-cooperative scenarios,this paper proposes a robust AMC method based on VRM under limited samples.This method can reduce the risk of overfitting by using rotation data augmentation,random scale cyclic shift transformation and consistency regularization.The experimental results show that the AMC method based on mixing training can adapt to SNR with a larger dynamic range,and the training sample size required by the AMC method based on VRM is only 2.5% of that of the traditional data-driven AMC method.For “weak generalization,poor robustness”,a few-shot specific emitter identification(FS-SEI)method based on metric learning and ensemble learning is proposed.This paper focuses on the metric learning-based discriminative feature embedding method for generalized RFF extraction and a fast identification method based on few-parameter machine learning classifier in few-shot scenarios.Inspired by the FSL theory,FS-SEI can be composed of a good feature embedding and asimple classifiers,where the former is applied to extract generalized radio frequency fingerprinting(RFF)features,while the latter is used to construct a feature-to-category mapping relationship in few-shot scenarios.Thus,this paper proposes a discriminative feature embedding method based on CVCNN and hybrid loss function consisting of Softmax loss function,center loss function and triplet loss,and few-shot identification method based on logistic regression(LR)and multi-model ensemble method.The experimental results show that compared with the existing feature embedding methods,the RFF features extracted by discriminative feature embedding have higher intra-class consistency and greater inter-class difference,and the multi-model ensemble method can improve 1.18%～ 6.10%identification performance,when compared with single LR classifier.For “redundant parameters,high complexity”,lightweight electromagnetic signal identification methods based on model compression and acceleration technology are proposed,respectively,and the design methods of electromagnetic signal identification model with few parameters and low computational complexity is studied from different perspectives.For AMC,this paper applies complex-valued convolution,separable convolution,channel shuffle,channel attention and other methods to design an ultra-light convolution neural network(ULCNN)for AMC.The experimental results show that the classification performance of the lightweight AMC method based on ULCNN is similar with that of the state-of-the-art method,and there are only 9,751 parameters and 0.2M computation,which is far lower than the existing AMC method.For SEI,this paper uses sparse structure selection algorithm to extract lightweight structure from the original CVCNN,and uses knowledge distillation to compensate for the performance loss of SEI based on lightweight structure.The experimental results show that the identification performance of SEI based on lightweight structure is similar with that of SEI based on original CVCNN,but the parameters of lightweight structure for SEI is reduced by at least 80%,and the computational complexity is also reduced by more than 90%.For “huge cost,poor privacy”,the gradient averaging(GA)and model averaging(MA)-based the distributed electromagnetic signal identification methods are proposed,respectively,and the distributed electromagnetic signal identification methods under non-independent identically distributed condition are evaluated from performance,convergence and communication cost.For AMC,the distributed AMC method based on balanced cross entropy is studied;For SEI,the distributed SEI method is based on metric learning.The experimental results show that the performance for the GA-based method is closer to that of the centralized method,and compared with the MA-based method,the GA-based has more than 2% improvement;the convergence speed of the GA-based method is also slightly faster than that of the latter;the communication overhead of the GA-based method is several times or even dozens of times that of the latter,because the communication frequency of the former is higher than that of the latter.Finally,the main work of this paper is summarized and the future work is prospected. |
PDF Full Text Request