Research On Automatic Modulation Classification Algorithm Based On Deep Learning

Automatic Modulation Classification(AMC),as the core technology of software radio and cognitive radio,is widely used in intelligent communication systems and has important military and civilian value.With the development of Deep Learning(DL)technology,the combination of neural network and AMC algorithm is becoming more and more abundant.This thesis aims to study the design and verification of AMC algorithm based on DL,mainly including the following three parts.(1)Aiming at the problem that the diversity of signal modulation methods in wireless communication systems leads to low classification accuracy,this thesis proposes a Feature Fusion and Attention Cooperation Frameworks(FF-ACF)AMC algorithm.The proposed algorithm fuses the pre-extracted signal constellation diagram features and cyclic spectrum features as the input of the DL model.A variety of feature extraction modules are designed,including recurrent connection convolutional neural network,bidirectional recurrent neural network,attention mechanism and depth feature fusion.The pre-extracted two features represent the amplitude,phase and frequency information of the signal respectively,which effectively solves the problem of low discrimination of signal features.The simulation results show that the average recognition rate of FF-ACF algorithm can reach 92.3% under the medium and high signal-to-noise ratio(SNR ?0 d B).Compared with the method based on convolution long-term and short-term deep neural networks,the average recognition rate is improved by about 8%,and the recognition accuracy rate is 94.15% when the SNR is 18 d B.Experimental results show that the proposed method has good classification performance.(2)In order to solve the problem of large number of parameters and low efficiency of AMC algorithm based on DL,this thesis proposes an AMC algorithm based on Lightweight Convolutional Neural Networks(LCNN).The algorithm uses a variety of compact designs,including complex convolution,separable convolution,channel shuffling,attention mechanism,feature reconstruction,and cross-level fusion.In addition,the rotation expansion method is used in the data expansion part,which verifies the importance of training sample number to the model.The simulation results show that the parameter quantity of LCNN algorithm model is 49.751,which is 30%-32% lower than the adaptive method based on parameter estimation and transformation and the lightweight AMRnet method,and is also far less than other models.Compared with complex convolution neural network,the average recognition accuracy is improved by about 5%.The experimental results show that the proposed method can effectively reduce the parameters of the model while improving the classification accuracy.(3)In order to verify the practicability of the proposed AMC classification algorithm,an experimental verification platform is built by using USRP 1 and GNU Radio.By sending and receiving different modulation signals,the classification performance of the algorithm is analyzed.The experimental results show that the algorithm has better classification performance after finetuning the parameters by transfer learning.