Research On The Application Of Machine Learning Based On Knowledge Embedding In Blind Communication Modulation Recognition

Modulation recognition is a key technology in wireless communication,especially blind modulation recognition in non-cooperative communication,which has important research significance.In recent years,machine learning algorithms have been widely used in the field of modulation recognition,among which intelligent recognition algorithms without prior information have become a research hotspot.However,machine learning algorithms face multiple problems when solving blind modulation recognition tasks.First,there are multiple influencing factors in the channel,such as Doppler shift,channel fading,and unpredictable noise,which can cause signal distortion and difficulty in identification.Secondly,the stability of widely used neural network algorithms needs to be solved urgently.Finally,more technical research is needed to enable the application of algorithms to real-world communication scenarios.This article proposes the following methods to solve the above problems:1.Aiming at the frequency offset caused by Doppler frequency shift and channel fading,this thesis proposes a frequency offset correction method based on the improved periodic graph method,which can correct the signal and obtain a more ideal constellation map,which has been verified on both the open source dataset and the actual collected data set.2.Aiming at the stability of neural networks,this thesis draws on the method of expert cyclic moment feature extraction to design a neural network structure with the same function but stronger extraction ability.At the same time,artificially extracted features are embedded in the neural network,which improves the modulation recognition accuracy to more than 90% when the signal-to-noise ratio exceeds 0d B,and ensures stability and convergence speed.3.Aiming at the problem of applying algorithms to actual communication scenarios,this thesis establishes a simple communication system,and establishes a software and hardware platform through GNURadio software and USRP(Universal Software Radio Peripheral)to collect and establish data sets.Finally,the effectiveness of the proposed algorithm is verified on this dataset.In this thesis,two algorithms are proposed to solve the modulation recognition task,which provide solutions from the perspectives of sequence features and image features.On the one hand,the neural network improvement scheme and framework are provided by manually designing feature embedding,which provides a reference design for stability improvement while improving accuracy.On the other hand,the frequency offset correction and kernel density information embedding in advance provide experimental support for the recognition based on constellation diagram.Finally,by verifying the data set obtained in the real scenario,the practical significance of blind modulation recognition in this study under non-cooperative conditions is shown.