Research On Modulation Pattern Recognition Based On Convolutional Neural Network

The automatic modulation recognition and parameter estimation of communication signals are the key technologies of modern communication technology,and are used in the fields of cooperative communication and non-cooperative communication.Because the electromagnetic environment is becoming more and more complex,further development of communication technology requires further research on modulation recognition and parameter estimation technology to adapt to today's complex environment.With the development of the times,the modulation pattern recognition of signals will also shift from empirical recognition to automatic modulation pattern recognition.This paper first studies the parameter estimation method of communication signals,and restores several traditional parameter estimation methods,such as frequency domain centering and cyclic spectrum analysis in carrier estimation,constellation restoration algorithm,frequency offset estimation algorithm and phase offset Estimation algorithm.In order to obtain the type of signal modulation,different parameter estimation algorithms are studied and the advantages and disadvantages are analyzed.It will pave the way for the research of modulation pattern recognition algorithm based on feature parameter extraction later.Secondly,this paper studies the modulation pattern recognition algorithm for feature parameter extraction,restores the representative constellation-based modulation pattern recognition algorithm and the high-order cumulant-based modulation pattern recognition algorithm,and discusses the feature parameters and classifiers.The relationship between selection and recognition range and anti-noise ability is found,and the shortcomings such as limited recognition range and poor anti-noise ability are found,so the following two new algorithms are proposed.Algorithm one is a modulation pattern recognition algorithm based on a block-shaped feature matrix.For modulation modes including PSK,FSK,ASK,QAM,etc.,the signal is framed to extract the characteristics of the signal's instantaneous phase,highest frequency,short-term energy,etc.in each frame Quantity,and finally constitute a feature vector.The signals of multiple frames form a feature matrix.This feature matrix is sent to a convolutional neural network for recognition to determine the signal modulation mode.Simulation experiments prove that when the recognition rate reaches more than 90%,the proposed algorithm reduces the signal-to-noise ratio requirement by about 1.5d B compared with the higher-order cumulant method,and reduces the signal-to-noise ratio requirement by about 3d B than the constellation method.Algorithm two is the modulation pattern recognition algorithm algorithm of amplitude and phase step recognition.Although algorithm one can recognize multiple signals,it has poor recognition performance for QAM signals,so this paper supplements algorithm two.This algorithm is only for QAM signals,and uses convolutional neural networks and subtraction clustering to identify the amplitude and phase of the signal step by step.This method does not depend on the elimination of frequency offset and phase offset of the signal,so it can play a good role in anti-frequency offset.Secondly,because there is no step of frequency offset elimination and phase offset elimination,the signal will not be lost during the preprocessing of frequency offset elimination and phase offset elimination.The simulation experiment proves that the recognition algorithm of step-by-step recognition of amplitude and phase has improved the anti-frequency deviation ability compared with the convolutional neural network algorithm based on constellation diagram.At the same time,when the recognition rate reaches more than 90%,the proposed algorithm reduces the signal-to-noise ratio requirement by about 3d B compared to the constellation method.