Research On Modulation Signal Recognition Algorithm Based On Convolutional Neural Network

Automatic Modulation Classification(AMC)is a complex and important technology before demodulation.It has been widely used in military and civilian fields.In recent years,AMC technology has attracted much attention due to the introduction of machine learning algorithms.Compared with traditional algorithms,the AMC algorithm based on machine learning has higher recognition accuracy and stronger robustness.This dissertation has conducted in-depth research and done the following work:1.To solve the problem of low classification accuracy due to the noise interference of modulated signal in complex channel,a modulation classification algorithm based on Wavelet Threshold Denoising(WTD)and Convolution Long Short-term Memory Neural Network(CLNN)is proposed.This algorithm improves the traditional wavelet threshold denoising function and adds variable factors which enable the function to make adaptive adjustments with the change of wavelet decomposition scale.Compared with traditional threshold functions,it has better continuity and higher-order derivability.In addition,the spatial feature extraction of convolutional neural network and the time features extraction of long and short term memory network are utilized to fully learn the global feature of denoised signal.The experimental results show that the classification accuracy of this algorithm is higher than the comparison algorithm in the environment which closes to the real channels.2.In order to solve the problem of low recognition accuracy caused by ignoring the correlation between current features and global features,a modulation classification algorithm based on the attention mechanism and bidirectional long short-term memory network is proposed.Firstly,the amplitude and phase components are used as input of signal by changing the traditional in-phase quadrature components.Then the signal is fed into a Bidirectional Long Short-Term Memory(Bi LSTM)neural network to extract and remember the features in both directions.The attention mechanism is employed to learn the correlation of each feature,and then different weights are allocated to the features.Finally,the network can automatically learn the importance of different weight features so as to realize the recalibration of feature learning.Experimental results show that the algorithm can significantly improve the classification accuracy.3.Aiming at the problems of traditional deep neural networks with complex structure,huge parameters and long training time,a modulation classification algorithm based on the Group Lasso lightweight neural network is proposed.In this algorithm,a Group Lasso regularization function is designed to induce network sparsity.The algorithm can effectively reduce the overfitting phenomenon by accelerating the operation speed of the convolutional layers,and promote the deep neural network to prune itself automatically to obtain a highly compact lightweight network.In addition,in order to solve the problem of signal types confusion such as QAM16,QAM64,WBFM,AM-DSB,etc.,a Double-trained Lightweight Neural Network(DLNN)is designed to improve classification accuracy.Experimental results show that the proposed algorithm can significantly lighten the depth neural network and shorten the computation time.