Research On Automatic Modulation Classification Via Progressive Neural Architecture Search

Automatic Modulation Classification(AMC)technology is a vital research direction in the field of signal processing.It is a key step between signal detection and demodulation,and plays an important role in various civilian and military applications.In the past,modulation classification technology was mainly based on likelihood ratio and feature-based methods.The former usually requires large computational resources,so the feature-based methods have been widely used.However,traditional machine learning(ML)methods need manual design of classification rules and manual feature extraction,and their recognition performance is very limited in the complex channel environments.In the past few years,the method of deep learning(DL)has been extensively used in the field of AMC.This method automatically extracts features end-to-end by constructing neural networks with a large number of parameters,greatly improving classification performance.However,building a DL network is still an error-prone and time-consuming process.Neural architecture search(NAS)is a sub-field of auto machine learning(Auto ML)that automatically adjusts the structure and parameters of neural networks to obtain the optimal network model structure by constructing search spaces,optimization algorithms,and model evaluations.Firstly,this paper studies the AMC method using Convolutional neural network(CNN).By making slight modifications to the model structure of some classic CNN networks,the simulation results of modulation signal classification under different signal-to-noise ratios show that welldesigned and trained CNN networks can complete modulation signal classification tasks well,but these networks also require huge computational resources.At the same time,various connection operations in classic CNN networks provide good inspiration for the design of NAS algorithms.Secondly,this paper studies the AMC method based on NAS.By using this method to search for the final network model structure,the simulation results of modulation signal classification under different signal-to-noise ratios show that the predictive classification ability of this method is greatly improved compared to the CNN-based AMC method.Furthermore,this method does not require manual design of the neural network structure.Finally,this paper studies the further optimization of the NAS-based AMC method.By dividing the search process into three stages using the progressive search method,the model is further compressed and the search efficiency of the NAS algorithm is improved.The simulation results of modulation signal classification under different signal-to-noise ratios show that the AMC method based on progressive NAS further reduces the computational resources compared to the NAS-based AMC method,and improves the predictive classification accuracy,thus achieving the expected classification performance.