Research On Deep Learning-based Blind Separation Algorithm For Modulation Signal

Blind signal separation technology has always been one of the main components in the field of modern signal processing.The traditional blind separation research mainly focuses on the processing of speech,image,biomedical and other signals.At present,most of the blind separation algorithms are implemented under the condition that the number of source signals is known,but in practice,the number of source signals is usually unknown,so it is necessary to estimate the number of sources before blind separation.With the rapid improvement of computer performance,deep learning technology has been successfully applied in communication,image,voice and other fields.Based on the ability of deep learning to automatically extract feature information from big data,the thesis applies the method of deep learning to the estimation of the number of sources and the blind separation of modulation signals.The research contents are summarized as follows:1)Because the current deep learning architecture and optimization algorithm can not effectively learn the complex baseband signal feature information,this thesis establishes a three-dimensional convolution neural network architecture to realize the complex baseband signal feature learning.In this architecture,the real part and the virtual part of the input multi-channel complex signal are synthetically learned by three-dimensional convolution layer,and the output is reconstructed according to the operation rules,and the multi-channel complex mapping is learned.Compared with other current solutions,this architecture is simple to implement,can effectively learn complex mapping,and avoids the problem that activation function,optimization algorithm and so on cannot be supported in the complex domain.2)The thesis analyzes the traditional methods of source number estimation,and designs a convolutional neural network composed of three-dimensional convolutional neural network architecture and full connection layer in series.The thesis studies and analyzes this method through simulation experiment,and compares this method with common source number estimation methods.The final results show that the proposed method can effectively estimate the number of sources of the mixed signal when it is disturbed by noise,and its performance is significantly improved compared with the traditional method.In the case of low SNR,it can also achieve high accuracy.3)The thesis proposes a blind separation method of modulation signals based on deep learning,and a convolution network is constructed by using a three-dimensional convolution neural network architecture.The thesis discusses the performance of proposed method for separating baseband or band-limited signals,and the influence of different carrier frequencies and modulation modes on the performance of the proposed method.The performance of common blind separation algorithms on blind separation tasks is compared.The experimental results show that the proposed method can realize blind separation for the modulation signals of the linear instantaneous mixed model and solve the problem of the uncertainty of the sequence of the separated signals.4)In order to demonstrate the feasibility of the proposed method in practical application,the thesis collects the required data through the establishment of experimental platform,and finally carries out the experimental verification.Experimental results show that the proposed method can meet the requirements of blind separation of modulation signals in practical applications.Because there are many interference factors in practical application,its separation performance is slightly lower than simulation experiment.