Waveform Recognition Research Based On Deep Transfer Learning

Waveform recognition technology plays a vital role in cognitive passive radar,electronic warfare,and wireless network security.Traditional waveform recognition is based on feature extraction.Typical features include time-frequency distribution features,fractal features and high-order cumulant features.With the deep maturity of Deep Learning technology,neural network methods are widely used in waveform recognition,and the performance greatly exceeds the traditional waveform recognition method.However,during the actual measurement process,it is found that the generalization ability of the trained model is not ideal for the data collected by the receivers in different channel scenarios and different sampling rates.Considering the characteristics of the channel conditions in the online waveform recognition,it is necessary to introduce Transfer Learning methods for knowledge transfer.This paper first builds a hardware platform to collect the protocol dataset of the real channel scene,including different channel conditions and different sampling rates.Secondly,for the same-domain isomorphic waveform recognition problem,a convolutional neural network model is designed,which has achieved good performance in both modulation signal recognition and protocol signal recognition.A two-channel convolutional neural network model is proposed,which combines bidirectional Long Short-Term Memory network,thus it can extract features at different scales and make full use of the time series characteristics of signals to achieve better results.Finally,combined with convolutional neural network and software design radio platform,an intelligent communication signal online recognition system is designed and implemented.For the waveform recognition problem of different channel scenes and different sampling rates,two cross-domain waveform recognition methods based on Deep Transfer Learning are designed.For the case where the target domain has some labeled data,the method of parameter transfer is designed to complete the knowledge transfer.This method is based on the source domain model,and only a small amount of target domain labeled data can quickly realize the accurate identification of the target domain.For the case where the target domain does not have any labeled data,a general unsupervised domain adaptation method is designed,and adversarial learning is introduced to learn domain invariant feature.An encoder network is used to transform the data in different domains to a high dimensional feature space,and the features similarity between source domain and target domain is realized by a discriminator.In addition,for the individual identification problem of radio stations,considering the similarity of different individuals,this paper proposes a method based on short-time Fourier transform and convolutional neural network to achieve individual identification and service identification.Firstly,the time-frequency diagram is obtained by short-time Fourier transform of individual signals.Then the feature extraction process and classification of time-frequency diagrams are performed using convolutional neural networks.The experimental results demonstrate the effectiveness of the method.