Research And Implementation Of Automatic Modulation Classification Algorithm Based On Deep Learning

Automatic modulation classification(AMC)refers to the technology that determines the modulation schemes according to the received signals when the modulation schemes of received signals are unknown to the receivers.It is an important task between signal detection and demodulation in non-cooperative communication systems.Due to the growth number of wireless network users and the diversification of user requirements,the interference and noise in wireless channels are becoming more and more serious,and the types of modulation schemes are gradually increasing.It may increase uncertain factors in received signals,which makes AMC algorithms unable to achieve the classification task effectively.In order to improve the effectiveness and robustness of AMC algorithms for multi-candidate modulation classification in non-cooperative communication systems,this thesis studies the design and verification of AMC algorithms based on deep learning.To overcoming the difficulty of modulation classification of diverse candidate modulations in wireless non-cooperative communication systems,this thesis proposes an AMC algorithm based on deep learning with constellation and cyclic-spectrum fusion(DLCCF).In the proposed DL-CCF algorithm,the pre-extracted constellation and cyclic spectrum features are fused and then put into a deep learning model for processing and classification.Meanwhile,the Bayesian optimization method is introduced to optimize model hyper-parameters.The two pre-extracted features represent amplitude-phase information and frequency information respectively,which guide the further adaptive feature extraction in deep learning.Therefore,the problem of inferior classification performance caused by insufficient signal feature discrimination is solved.The simulation results show that the proposed DL-CCF algorithm has fewer model parameters compared with other algorithms and improves the classification accuracy effectively under medium and high signal-to-noise ratios(SNRs).However,the classification performance under low SNRs is poor,and classification accuracy of high-order quadrature amplitude modulation(QAM)is not excellent.In order to classify low SNR signals and QAM signals effectively,this thesis proposes an AMC algorithm based on adaptive feature extraction and fusion(AFEF).On the basis of the DL-CCF algorithm,the proposed AFEF algorithm adds a branch of convolutional neural network that uses the samples before feature pre-extraction as input.Then parallel training and transfer learning are introduced to improve the training and optimization process of the model,which reduces the model training time.Because of the additional information compensation provided by the new branch,the algorithm avoids the information loss in the feature pre-extraction process and makes up for the lack of reliability of amplitude and phase features caused by the weak noise immunity of constellation features.The simulation results show that the proposed AFEF algorithm improves the classification accuracy of low SNR samples effectively and ensures unobvious changes of the model complexity,training time and testing time.The classification performance of QAM signals under medium and high SNRs is also improved.In order to verify the practicability of the AFEF modulation classification algorithm,universal software radio peripherals and GNU Radio are used to build an experimental system which sends and collects different types of modulation signals to analyze the classification performance of the algorithm.The experimental results show that the proposed AFEF algorithm has high classification accuracy by fine-tuning the model parameters with transfer learning.The two algorithms and verification methods proposed in the thesis improve the classification accuracy of the AMC algorithm effectively in actual communication scenarios,and have bright application prospects in software defined radio,military countermeasures and other fields.