Research On Modulation Recognition Technology Of Digital Communication Signal Under Complex Electromagnetic Environment

In recent years,modulation recognition based on Deep Learning(DL)has attracted much attention due to its superiority in feature extraction and classification accuracy.Converting the constellation diagram features of communication signals into images and then using Deep Neural Networks(DNN)for classification,so as to complete the identification of modulated signals has been widely used.However,a serious problem in modulation recognition based on deep learning is that when the channel model of the modulation signal changes,the original trained model will be mismatched,resulting in a decline in the recognition rate.Today,with the rapid popularization of wireless communication applications,the electromagnetic environment is becoming more and more complex.In the complex electromagnetic environment,this problem is particularly prominent.This paper focuses on the problem of model mismatch in modulation recognition based on deep neural networks in complex electromagnetic environments,that is,when there is a difference in the distribution of test samples and training samples,the recognition rate will drop significantly.The article is divided into two steps to carry out the research on related issues.The first is to conduct research under the condition of only additive non-Gaussian impulse noise,and then to conduct research under the condition that multiplicative interference and additive non-Gaussian impulse noise exist simultaneously.The main work contents are as follows:Modulation recognition based on depth-domain adaptation under the condition of only non-Gaussian impulse noise has been intensively studied.Firstly,starting from the strengthening of the constellation diagram features,a real image bounded nonlinear function algorithm with better robustness to the characteristic index is proposed to solve the problems of the real image logarithmic method.In the part of IQ two-way signal matrix mapping algorithm,an improved graphic constellation projection algorithm is proposed,that is,the real image bounded nonlinear function algorithm is introduced into the graphic constellation projection algorithm,and the gray level is enhanced after matrix mapping to further strengthen the characteristics of the constellation diagram.Then the cross-domain recognition rate is further improved by a deep domain adaptive network.Since the domain adaptation loss of Deep Adaptation Networks(DAN)only considers the global situation,there will be local negative transfer phenomenon,so the local maximum average difference is used to replace the maximum average difference as the new domain adaptation loss,which improves the local Negative transfer phenomenon.Under the condition that multiplicative interference and additive non-Gaussian impulse noise exist simultaneously,the normal-mode blind equalization based on fractional low-order statistics and the norm-blind equalization algorithm based on the maximum correlation entropy criterion under non-Gaussian impulse noise are firstly studied.Due to the limited prior information of modulation recognition,the blind norm blind equalization algorithm based on the maximum correlation entropy criterion is determined to be used for modulation recognition.However,after the multipath channel model changes,even using the deep domain adaptive network for migration adaptation cannot achieve ideal recognition results.The analysis shows that this is related to the fact that the blind equalization algorithm used cannot correct the phase rotation,so a corrected normal modulus blind equalization algorithm based on the maximum correlation entropy criterion is proposed.Because the improved algorithm has the problem of non-convergence in the process of partial signal processing of the modulated signal set,the concept of constellation distribution equalization coefficient is proposed to solve this problem.Simulation results show that the proposed algorithm can effectively improve the problem of model mismatch.