Research On Emitter Modulation Recognition Technology In Non-Gaussian Electromagnetic Environments

Emitter recognition is a useful technology to analyze and obtain information,e.g.modulations or the specific transmitter identities,of wireless signals in non-cooperative communication scenarios.As an important branch,modulation recognition and classification,a.k.a.automatic modulation classification(AMC),is an essential part.AMC under the assumption of Gaussian electromagnetic environments has been researched a lot,but the assumption of non-Gaussian electromagnetic environments matches non-cooperative communication scenarios.A constellation diagram is a common type of signal representation,whose formation relies on blind symbol timing estimation(STE)to obtain the optimal sampling timing of a signal.However,in non-Gaussian electromagnetic environments,the signal distortion leads to the incorrect estimation of optimal symbol sampling timing and the failed formation of the constellation diagram that is good for AMC,which puts bad effects on AMC to a large extent.Therefore,the main approaches to improving AMC performance include the optimization of blind STE and classifiers,etc.Firstly,a novel constellation diagram representation for AMC called multitiming constellation diagrams(MCD)is proposed to solve that constellation diagrams of optimal symbol timings are difficult to obtained via conventional signal-level STE in non-Gaussian electromagnetic environments.Specifically,MCD enumerate the constellation diagrams of all symbol timings to both inherit the advantage of discriminative AMC features and reserve all the sampling points.Based on MCD,the following AMC methods with feature-aided STE optimization can select optimal symbol timings accurately at the feature level,which benefits AMC.Secondly,an AMC architecture with manual feature-aided STE optimization is proposed to realize STE at the manual feature level and format constellation diagrams of optimal symbol timings for deep learning based AMC.Observably,the constellation diagram with the highest inner-cluster stellar aggregation is the optimal one within MCD for AMC,which is similar with that constellation diagrams of higher signal-to-noise ratio(SNR)in the additive white Gaussian noise(AWGN)channel have higher inner-cluster stellar aggregation for better AMC as well.So,the SNR of the constellation diagram in the AWGN channel is proposed as the manual feature,and two convolutional neural networks(CNN)are utilized for the first-stage manual feature estimation and the second-stage constellation diagram based AMC,respectively.Afterwards,an AMC architecture with deep feature-aided STE optimization is proposed and realized by attentive Siamese networks(ASN),which fuses STE and AMC into an end-toend neural network.In term of attentive weights,ASN can guide its modulation classification module to the deep feature vector of the constellation diagram of the optimal symbol timing for feature-level STE.Then,ASN can classify the modulation.For further interpretability,the modules of ASN are designed to function like the functional modules in practical communication reconnaissance receivers.Meanwhile,the deep features for STE and AMC in the end-to-end ASN can adjust to non-Gaussian electromagnetic environments.Finally,stellar cluster counting(SCC)knowledge driven AMC method is proposed to solve that deep learning based AMC methods are easy to overfit and cannot be solved by adjusting hyper-parameters simply due to the lack of diversity of the electromagnetic environments of the training set.Specifically,a neural network is pretrained firstly by the SCC task that is related with stellar clusters directly,which can guide the network to stellar clusters in constellation diagrams.Then,this SCC knowledge is transferred to the AMC network,which is driven to extract intrinsic AMC features related with stellar clusters,whose generalization ability is better.Above all,experiments based on two non-Gaussian electromagnetic environments,i.e.the public Radio ML 2018.01 A dataset and the simulated additive general Gaussian noise dataset,verify the effectiveness of the above-mentioned two-stage CNN and ASN from the perspectives of visualized analysis of symbol timing estimation and objective correct classification probabilities contrasts.At the same time,from the perspectives of subjective network attention thermodynamic diagrams of class activation maps and objective correct classification probabilities contrasts,experimental results show the SCC knowledge driven AMC method can focus on stellar clusters effectively and improve AMC accuracies and generalization ability,without increasing temporal and spatial complexities of conventional constellation diagram based CNN methods.