Deep Learning Based Automatic Modulation Recognition Algorithm For Communication Signals

Automatic modulation recognition(AMR)plays an important role in a wide range of applications,e.g.,cognitive radio,spectrum sharing,and dynamic spectrum access,which has been widely applied in military and civilian applications.In recent years,due to the breakthroughs in deep learning(DL),it has been utilized in almost all fields.It is worth noting that DL-based AMR methods are becoming extremely popular,which have greatly improved the performance of modulation recognition.However,most existing DL-based methods still suffer from some difficulties:(i)the information of the modulated signal is not fully used;(ii)the advanced recognition models have high computational complexity and they are difficult to be deployed on computing resourceconstrained devices;(iii)it may result in excessive resource overhead to labeled signal samples for supervised learning;(iv)the well-trained models generally suffer from performance degradation when the distribution characteristics of test data change.To address these issues,this thesis explores the techniques in DL and employs them in modulation recognition algorithm,which can be summarised as follows.An intelligent algorithm based on phase constellation and trajectory clustering is proposed to classify the types of amplitude-phase modulation.This thesis recovers an ideal baseband signal by preprocessing the signal for timing synchronization.By exploiting the DL-based techniques,the signal modulation recognition problem is transformed into a featured image classification problem.The extracted features are input into a residual structure network with two parallel inputs.Then,hierarchical learning and feature fusion training are applied to achieve the target modulation recognition.Simulation experiments show that the recognition rates achieved by the feature fusion method outperform those achieved by the methods based on constellation diagrams,signal waveform and high-order statistics.The recognition rate for seven types of modulation can reach above95.14% when the signal-to-noise ratio(SNR)is larger than 2 d B.Moreover,the ablation experiment shows that fusing phase vector trajectory features can effectively enhance the characteristic properties of the modulated signal and improve the identification performance of the modulated signal.However,this method could suffer from performance degradation when the carriers are not perfectly synchronized.To address the problem that it is difficult to recognize multi-type modulation signals in harsh environments,this thesis considers a recognition algorithm based on adaptive feature extraction from IQ(Inphase and Quadrature)data.Existing DL-based methods are unable to deal with complex-format data,and need multiple complex networks to mix the information from the time series or its transformed representation.To overcome these difficulties,we present a complexvalued convolution and frequency global filter unit(CGFU)and propose a hybrid neural network,namely CGF-HNN,which can efficiently exploit features from time and frequency domains.The recognition performance of the proposed model is evaluated on two well-known datasets.Simulation results show that the proposed model outperforms the existing state-of-the-art models and validate the effectiveness of the proposed CGFU.Moreover,to reduce the excessive computational complexity of the CGF-HNN network,we design a lightweight recognition network by exploiting the self-attention mechanism(SAM),namely CGF-SAN.Simulation results show that the performance of CGF-SAN still matches the existing advanced networks with significantly reduced complexity,and the proposed model achieves an efficient balance between recognition performance and complexity.Furthermore,this thesis proposes a modulation recognition network based on transfer learning(TL)methods to address the mismatches problem that the pre-trained DL-based AMR network processes different data distributions.The network transfers the knowledge obtained from supervised training in the source domain to unsupervised learning in the target domain.Simulation results demonstrate that great gain can be obtained about 24.63% from the complex distribution data to the Gaussian distribution data.It can provide an unsupervised learning method in that the DL-based model decides the modulation type across domains.