Research On Blind Source Separation Of Communication Signals Based On Deep Learning

With the rapid development of communication technology,the electromagnetic environment in the air is becoming increasingly complex,and the spectrum resources are increasingly strained.In addition to noise,received signals often contain other communication signals that are aliased,which not only increases the difficulty of signal reception,but also poses serious challenges to subsequent signal processing work.Therefore,studying how to effectively achieve signal noise reduction and blind source separation is of great significance for improving communication signal processing capabilities.Deep learning is one of the most popular research methods nowadays,with powerful data learning capabilities,and has been widely used in fields such as image,voice,and so on.Due to the mature application of deep learning algorithms in these fields,thesis applies it to signal processing.According to the different interference sources in communication mixed signals,thesis introduces deep learning algorithms to conduct research from two aspects: signal noise reduction and signal blind source separation.The main contents include:(1)An improved signal denoising algorithm based on convolutional noise reduction auto-encoder is proposed for mixed signals with random noise as the interference source.Residual learning and attention mechanisms are integrated into the stacked denoising auto-encoder to enhance the ability of feature selection and expression.The mean square error is used as the loss function.In the training stage,the pure signal is taken as the label,and the normalized noisy signal is taken as the input.After layer by layer learning and weight updating of the network,the noise reduction signal is reconstructed in the output layer.In the test stage,the noise reduction signal can be obtained by inputting the normalized noise signals.This thesis conducts simulation experiments on MPSK,MQAM,MFSK,and MAPSK modulated signals.The experimental results show that:(1)The noise reduction algorithm is applicable to different modulation types of signals.(2)When the signal-to-noise ratio of noisy signals is between 2 and 20 dB,the signal-to-noise after noise reduction can be improved by more than 3.7 dB,and the noise reduction performance is relatively stable;(3)The demodulation error rate of the noise reduction signal is lower than that of the noisy signal,which is beneficial for subsequent signal demodulation and other work.(2)A blind source separation algorithm based on deep convolutional neural network is proposed for mixed signals where the interference source is other useful communication signals.The algorithm model mainly consists of an encoder,a separator,and a decoder,which directly separates and processes the aliased signals composed of two source signals.Specifically,a multi scale structure is introduced into the encoder to extract different scale features of aliased signals;Introducing void convolution,residual learning,and attention mechanisms into the separator to increase the receptive field of the network,while fully mining the potential information of each source signal;A multi scale structure corresponding to the encoder is used in the decoder to reconstruct the signal waveform.The algorithm uses supervised training and uses scale invariant signal-to-noise ratio as a loss function to calculate the training error.In thesis,simulation experiments have been conducted on two source mixed signals of QPSK,8PSK,and 16 QAM,and the separation performance has been evaluated using bit error rate.Experimental results show that the algorithm has extensive and good separation performance for mixed signals of the same frequency,non-mixed signals of the same frequency,mixed signals of different modulation types,and mixed signals of higher order modulation.