Research On Digital Modulation Mode Recognition Based On Deep Learning

As an intermediate process of signal detection and demodulation,Automatic Modulation Classification(AMC)provides the modulation information of signals and is a key technology in the field of modern communication.Although the traditional AMC method performs well for signal modulation types with significantly different statistical characteristics,it is sensitive to channel environment and has high computational complexity,so it is not universal.Due to the great achievements of deep learning in image processing and speech recognition,how to combine deep learning with modulation recognition is the research direction of many scholars at present.Therefore,this thesis mainly studies the application of deep learning in signal modulation mode recognition,and the main research contents are as follows:(1)In order to solve the problem of poor recognition performance based on sequence feature modulation,this thesis mainly adopts the method of converting one-dimensional time series into two-dimensional images.After analyzing the statistical characteristics of the signal,it is transformed into two kinds of feature image.First,one-dimensional time series is transformed into two-dimensional time-frequency image by short-time Fourier Transform(STFT).Second,one-dimensional time series is converted into constellation diagram.In the process of signal transmission,the generated image is distorted due to the frequency shift,noise and so on.In this thesis,according to the different features of the image,the specific solution is proposed.In order to solve the problem of noise in STFT spectra,two methods of denoising using filter and wavelet packet are proposed.As for the interference points existing in the constellation map,the algorithm based on spatial density clustering is used to remove them.On this basis,the point density of different areas is mapped into different colors,so as to obtain the constellation map containing color information to enhance the features of the image.(2)In order to further improve the recognition performance of signal modulation mode under low signal-to-noise ratio,this thesis adopts deep residual network Res Net50 as the original network model and improves it on this basis.For the improved network model,a signal recognition algorithm based on local feature extraction and a multi-feature fusion algorithm are proposed.The recognition algorithm based on local feature extraction converts signals into time-frequency graphs through STFT and takes the time-frequency graphs as the input objects of the training model to prove the robustness of the improved network in this thesis.In order to solve the confusion problem in MQAM class with high signal-to-noise ratio,a signal modulation recognition algorithm based on multi-feature fusion is proposed in this thesis.Experiments show that compared with the recognition algorithm based on local feature extraction,the proposed algorithm not only effectively improves the in-class confusion of MQAM with a high signal-to-noise ratio,but also has an accuracy of more than 92% for all modulation types when the signal-to-noise ratio is greater than 0d B.It can be seen that the signal modulation recognition algorithm based on multi-feature fusion still has good classification and recognition effect under low signal-to-noise ratio.