Research On Automatic Modulation Classification Technology Based On Neural Architecture Search

Automatic Modulation Classification(AMC)technology is an indispensable and important technology in the development of cognitive radio.It can automatically determine its Modulation mode according to the signals collected by the receiver.In the past,traditional modulation classification techniques have been based on probability and features,so it requires efforts to find more efficient features and classifiers.In recent years,Deep Learning(DL)replaces the traditional modulated signal automatic recognition with the method based on decision theory and pattern recognition,and has achieved good results.However,the development of neural network classification model is mainly concentrated in the field of construction engineering,discovering the most advanced neural network structure requires a lot of prior knowledge and human experts’ efforts.Neural Architecture Search(NAS)can be regarded as a sub-field of Automatic Machine Learning(Auto ML),which can use Neural networks to adjust structure and parameters automatically.At the same time,the search strategy is followed to maximize the performance as the goal to obtain the network model structure required by researchers.Firstly,the classical Deep Convolutional Neural Network(DCNN)is introduced into the AMC task,and the data sample and model structure are slightly adjusted.The experimental results show that the classical DCNN can predict and classify the modulated signals at high SNR,but the cost of computing resources used by these networks is relatively high.At the same time,modular design,skip connect and deeply separable convolution in classical network connection operation also play a good reference for the subsequent NAS algorithm search space design.Then,this thesis introduces the NAS approach to AMC tasks.According to the results of our simulation experiment,the network model designed based on NAS algorithm has the same predictive classification ability as common DCNN in the modulation signal classification scenario presented in this thesis.However,in terms of parameter number and Floating Point Operations(FLOPs),NAS can save a lot of computing resources and avoid very complicated manual design.Finally,this thesis aims at the further research of DARTS-based algorithm,and designs the NAS from the two perspectives of model compression and improving the practicability of the algorithm.In this thesis,the model structure of the search space is rearranged to design the final model with fewer stacked cells in a more reasonable way.At the same time,the search process of the network architecture is accelerated through partial channel clipping at the search stage,and regularization operation is introduced at the edge of each connection.Theoretical analysis and final experimental results show that the network designed by the above ideas is more reasonable,has higher adaptability to the AMC task in this thesis,and can achieve the expected classification effect.