Research On Digital Modulation Signal Demodulation Based On Machine Learning Methond

Digital signal modulation and demodulation based on machine learning methods can demodulate multiple digital modulation signals under the same network model without carrier recovery and filter processing,which helps to improve information transmission in communication systems High efficiency,alleviating the pressure of hardware circuit design in the increasingly complex communication environment.In addition,the digital signal modulation and demodulation based on the machine learning method can directly recover the symbol sequence from the fading signal without performing channel estimation,which is beneficial to improve the frequency utilization efficiency of the communication system.Therefore,the modulation and demodulation of digital signals based on machine learning methods has become one of the research hotspots in recent years.This article addresses the problem that the application of machine learning methods to demodulate multi-standard digital modulation signals requires a higher oversampling factor,and the problem of digital modulation signal demodulation methods based on machine learning in fading channels that have weak generalization capabilities for different modulation signals.Questions were studied.The main research contents are as follows:Firstly,the traditional demodulation method is compared with machine learning-based demodulation method,and the advantages of machine learning-based demodulation method are analyzed.Artificial Neural Network(ANN)and other Neural Network demodulator suitable for multi-mode digital modulation signal are analyzed.The above methods have some problems,such as high demand on over-sampling multiple of digital modulation signal,many Network parameters and great difficulty in training.Deep Belief Network(DBN)and other neural Network demodems suitable for digital modulation signals in fading channels are analyzed.The above methods have the problem of weak generalization ability for different digital modulation modes.Then,a multi-scale one-dimensional convolutional neural network demodulator is designed,which is suitable for multi-mode digital modulation signals under the condition of low over-sampling.The demodulator can demodulate four kinds of digital modulation signals,BPSK,4-QAM,8-QAM and 16-QAM,under the same over-sampling condition as the traditional demodulator,and can guarantee the same error performance as the traditional demodulation method.The simulation results show that the proposed digital modem can guarantee the demodulation error performance and reduce the requirement of sampling times and the complexity of the neural network structure under the Gaussian channel.Finally,A neural network cascading model is designed for demodulation of multi-mode digital modulation signals in fading channels.The model is composed of denoising autoencoder and demodulator.The influence of fading channel on the digital modulation signal is eliminated as far as possible by the encoding and decoding process of the denoising autoencoder,and the demodulator is used to demodulate the digital modulation signal.In Rayleigh fading channel was simulated and experimental results show that the cascade neural network model can be in the absence of channel estimation,realized the BFSK and BPSK modulation signal demodulation,show that the cascade neural network model,and in eliminating rely too much on the channel state information at the same time,to enhance the generalization ability of different modulation signals.