| With the development of wireless communication technology and the increasing complexity of electromagnetic environment,various communication methods are changing with each passing day.Among them,non-cooperative communication is widely used in the field of electromagnetic signal monitoring,and signal modulation classification is of great significance as a necessary step before signal demodulation.In military applications,electronic countermeasures and electronic reconnaissance under non-cooperative communication mode also require real-time and portable modulation classification system.The continuous development of artificial intelligence technology has led to a new research direction for modulation classification,especially the wide application of deep learning,which realizes the classification of various modulation methods by virtue of its powerful fitting ability,and it can be deployed on embedded edge devices to meet real-time and portable requirements with the help of neural network model lightweight techniques.In this thesis,the classification of various modulation methods will be realized based on neural network,and then the neural network lightweight technology is used to improve the execution efficiency of the algorithm,finally,the modulation classification system with application value is built based on edge devices.In view of the classification task of various modulation methods,this thesis conducts model training based on inphase and orthogonal components and MobileNet,after theoretical analysis and experimental verification,the conclusion is drawn:the obtained model has poor generalization ability to modulated signals generated by different source sequences.Therefore,a method of using instantaneous feature joint distribution matrix to represent signals is proposed,and a network model insensitive to source sequences is obtained based on MobileNet,that is,MobileNet trained by instantaneous feature joint distribution matrix performs well on the modulated signals generated by different source sequences.Then in order to improve the accuracy of the algorithm,the idea of grouping convolution in MobileNet is used to improve the feature pre-extraction convolutional layer of ResNet34.Finally,the experimental results show that the improved ResNet34 has higher accuracy than MobileNet and the original ResNet34.To meet the real-time requirements of algorithms on edge devices,this thesis proposes an algorithm called Unilateral Knowledge Distillation with Adaptive Temperature Coefficient Teacher Network,taking the improved ResNet34 as the teacher network and MobileNet as the student network to conduct unilateral knowledge distillation,compared with the original knowledge distillation,the algorithm can significantly improve the accuracy of the student network,at the same time,through experimental verification,when the teacher network after pruning is used as the student network,since it has more similar parameters and structure with the original teacher network,it can obtain accuracy performance closer to the teacher network.Finally,the improved ResNet34 after pruning and unilateral knowledge distillation was used to replace the original teacher network to improve the efficiency of algorithm execution.This thesis finally builds a modulation classification system based on edge devices,and develops embedded software to deploy the trained neural network model.The final system testing shows that the classification accuracy of the algorithm is basically consistent with the simulation experiment,and it meets the real-time requirements.The theory and results of this thesis are finally applied to the National Major Instrument Research and Development ProjectâDevelopment and Application of Illegal Electromagnetic Signal Monitoring and Classification Technology,meeting relevant requirements. |
