Research On Signal Modulation Classification Based On Deep Auto-encoders

Modulation classification is a key technology in the fields of intelligent communication,non-cooperative communication and radio spectrum management,which aims to identify the modulation type of received signals.Deep auto-encoders is an artificial neural network applied in semi-supervised and unsupervised learning in the field of deep learning,which has a powerful data feature extraction ability.In the present study,the classification way and performance of modulation signals applied by deep auto-encoder are investigated from three perspectives i.e.,signal noise reduction,feature extraction and signal identification.The main investigated contents are as follows:(1)A signal denoising method based on stacked convolutional denoising auto-encoders(SCDAE)is investigated and implemented.The denoising network is stacked by three denoising auto-encoders(DAEs)in a convolutional way,and each of DAE adopts the least square estimation(LSE)as targeted function.The constructed LSE function based on first-layer input and last-layer output is taken as the loss function of the entire denoising network.During the training process,the denoised signal is input into the network with the output goal of the free-noise signal.After trained and mapped layer by layer,the noise distribution is learned by the noise reduction network and the free-noise signal is additionally reconstructed in the output layer,which consequently realize the ability of noise reduction.Experimental results show that:(1)When the signal-to-noise ratio(SNR)of the original signal is-20 d B?4 d B,the SNR of the output noise reduction signal can be increased by 7 d B;(2)The recognition performance can be effectively improved based on the LSTM or CNN modulation classification algorithms by means of the present method.(2)A signal feature extraction method based on wavelet convolutional auto-encoder(WCAE)is investigated and implemented.The fully connected layers of the onvolutional autoencoder(CAE)is replaced by the1-dimension convolutional layers,which is able to reduce the network parameters.Firstly,the input samples in the encoding network are convoluted by three convolutional layers,aiming to reduce the dimension and learn the potential features of the samples.Then,the signal reconstruction is carried out using the extracted feature data through the up-sampling and symmetrical transposed convolutional operation in decoding network,and the difference between the input and output layer is controlled by the targeted function consisted of the mean square error function and the Jacobian penalty term.The optimized parameters can be ascertained by minimizing the reconstructed error.When the network becomes convergent,the output of the encoding network is the extracted features.Experimental results show that:(1)when the SNR is less than-10 d B,the identification accuracy of extracted signal feature from the input samples is relatively high;(2)the identification accuracy for the extracted features of amplitude/phase representation is higher than that of the I/Q representation;(3)when the SNR is higher than 4 d B,the identifucation results of the 8 modulated signals are above 95%.(3)A modulated signal classification method based on the combined network of SCADE and WCAE is proposed.The SCADE,WCAE and Softmax classifier are connected to construct a modulated signal classification network.Among them,the SCADE is used to complete the input signal pre-processing task to realize the denoising function,and the WCAE is used to extract features from the SCDAE output.Then the signal identification is accomplished by the Softmax classifier.Experimental results show that:(1)when the SNR is less than 0 d B,the identification accuracy of the present proposed method is better than that of other methods;(2)when the SNR is higher than 2 d B,the overall identification accuracy of the present proposed method can achieve 91.7%.In summary,the present work provides the implementation method and the results of performance analysis for the application of signal noise reduction and feature extraction using deep-autoencoders.In addition,the proposed classification method,which is combined deep auto-encoders signal,can effectively improve the signal identification accuracy under the situation of low SNR.