Research And Implementation Of Digital Signal Demodulation Algorithm Based On Neural Network

Nowadays,digital communication has become the mainstream communication method,and in the communication system,signal modulation and demodulation is a key part of information transmission.In different application scenarios,different communication protocols and modulation methods are often used.However,traditional digital radio systems use dedicated circuits,which results in its single-function,poor flexibility and nonprogrammable characteristics,and cannot be used in multimode communication.In this context,software-defined radio has gradually developed.A neural network is a mathematical model that can be implemented by software,and it has excellent performance when dealing with non-linear problems such as image and natural language processing.As the neural network has good learning ability and generalization ability,it can obtain the ability of demodulating digital signal by learning a large amount of relevant data.Therefore,it is feasible and advantageous to use neural network as the demodulation module of softwaredefined radio system.This thesis proposes a demodulation algorithm based on One-Dimensional Convolutional Recurrent Neural Network(1-D CRNN).The object of this algorithm is the intermediate frequency sampling sequence of the received signal,and the sampling point corresponding to a single symbol is used as the input of the neural network to classify it.Then it is converted into bit information according to the classification category,and finally demodulation is completed.The algorithm connects 1-D CNN and RNN in series,extracts the characteristic information such as amplitude and phase in the signal with the help of 1-D CNN,and hands these time-related features to RNN for processing.This gives full play to the respective characteristics of CNN and RNN,and improves the accuracy of classifying symbol sampling points.Besides,this thesis also proposes a symbol synchronization algorithm based on 1-D CNN,which judges the position of the symbol by detecting the amplitude and phase changes between different symbols,and groups the intermediate frequency sampling sequence into symbols according to the position information.Only after correctly grouped data can be demodulated by the 1-D CRNN.Therefore,the 1-D CNN symbol algorithm plays a vital role.This thesis builds the neural network models of 1-D CNN and 1-D CRNN respectively,generates QPSK and 16 QAM signals and builds a data set,uses the training set and the verification set to train the neural network,and uses the test set to test its performance.The experimental results indicate that the network trained with data with lower snr has better noise adaptability,and the overall performance on the test set will be better.For 1-D CNN,it can accurately identify symbol positions in QPSK and 16 QAM signals,and the higher the snr of the signal,the more accurate the identification.For 1-D CRNN,the demodulation performance under Gaussian channel approaches the theoretical bit error rate curve.In the multipath fading channel,the single-symbol input is changed to the dual-symbol input,which has achieved a certain performance improvement,and the performance is better than some equalizer algorithms.