Research On Deep Learning Based Modulation Recognition Technologies

Modulation recognition technology plays critical roles in both commercial and military fields.The modulation recognition technology can be categorized into maximum likelihood-based methods and feature-based methods.Due to the high computational complexity of maximum likelihood-based modulation recognition,feature-based modulation recognition technology has been more widely adopted.The traditional feature-based methods need to extract handcrafted features from raw signal data and then design classification rules to determine the modulation type.Manually designed signal features and classification rules usually have limited recognition accuracy in complex channel environments.Therefore,the deep learning-based modulation recognition technology is proposed to automatically learn the classification rules from training data to improve classification accuracy.With the development of deep learning in recent years,deep learning algorithms have achieved great success in computer vision and natural language processing.The state-of-the-art deep learning algorithms can automatically learn the signal features as well as classification rules from raw data simultaneously,further improve the modulation recognition performance.First,this thesis studied the modulation recognition using handcrafted signal features and fully connected neural network classifiers for decision rules.The theoretical analysis and the simulation results show that the fully connected neural network classifier can achieve better classification accuracy compared to the handcrafted classification rules.Then,this thesis studied the modulation recognition system based on convolutional neural network.The theoretical analysis and simulation results show that the well-designed and trained convolutional neural network can simultaneously achieve the target of feature extraction and modulation classification.The features extracted by the convolutional neural network can lead to better modulation recognition performance.This thesis further studies the impact of training sets on convolutional neural network and finds that training sets with diversity can improve the robustness of convolutional neural network.Finally,this thesis studied the performance degradation of modulation recognition caused by the channel variation in the convolutional neural network-based modulation recognition.To overcome the problem,this thesis proposed three different transfer learning methods.The theoretical analysis and simulation results show that the first transfer learning method demonstrated the optimal modulation recognition performance in the new channel environment.The second transfer learning method demonstrated the optimal modulation recognition performance and converges faster at the same time when the training data is sufficient in the new channel environment.The third transfer learning method demonstrated the suboptimal modulation recognition performance and converges faster at the same time when the training data is moderate in the new channel environment.