Research On Communication Signal Recognition Based On Deep Learning

With the rapid development of the Internet of Things(IoT),the electromagnetic environment is becoming increasingly complex.The number of IoT devices has grown significantly,and the security of devices has been seriously threatened.Especially in the environment of low signal-to-noise ratio and large-scale scenarios,wireless signals are seriously interfered,which undoubtedly increases the difficulty of verifying the individual process of the device(called Authentication).Physical layer-based authentication has the advantage of being naturally not easy to be tampered with,since it uses rich,unique,and fine-grained fingerprint features to verify.Recently,thanks to the great breakthrough in deep learning,deep learning methods with strong feature extraction abilities have made it possible to extract discriminative fingerprint features directly from the raw signal.A deep learning method to process the signal of the emitter requires solving the following three problems:(1)fingerprint modeling and dataset construction;(2)effective identification at low signal-to-noise ratios;(3)identification in large-scale scenarios.1.Create a unified model of multiple fingerprints and generate fingerprint datasets.To solve the difficulty of unifying a wide variety of emitters and build a comprehensive fingerprint model,a unified model of multiple fingerprints is created.Based on this model,datasets with different degrees of fingerprints are generated for training deep learning models.2.Propose an identification method based on a nonlinear statistical convolutional neural network(NSCNN).Aiming at addressing the small difference in fingerprint features and the difficulty of identifying the raw signals under low signal-to-noise ratios,a NSCNN method is proposed.A normalized constellation feature map is created to enhance the nonlinear statistical properties of different symbols in a signal.Simulation experiments demonstrate that the proposed method can achieve better convergence performance and robustness of noise.The recognition accuracy of the proposed method is closer to the maximum likelihood method than the traditional methods.When the signal-to-noise ratio is 5dB,the recognition accuracy increases by 24.79%.3.Propose an identification method based on a multi-task residual network(MTResNet).To solve the problem of the consistent carrier frequency offset(CFO)interference in large-scale scenarios,an end-to-end MTResNet is created,in which the frequency offset estimation network and the signal compensation unit are designed.They are integrated into the input end of the identification network.Then,a dual loss fusion is used for training MTResNet.With large simulation experiments are conducted,the effect of multiple consistent CFOs is mainly studied.The proposed MTResNet exhibits good recognition performance and robustness to multiple consistent CFOs.It can also achieve good recovery of phase information in fingerprint features.When the signalto-noise ratio is 10dB,the recognition accuracy of 14 emitters is as high as 86.09%.We also verify that the inconsistent CFOs can be used as a fingerprint feature,which results in an accuracy improvement of 6.58%in 5dB.In addition,a higher estimated loss weight(λ2=0.75)can recover the phase information and help improve recognition accuracy further.