Neural Network Demodulator Of MPPSK Signal Over HF Channels

With the advent of the information age,communication has become an indispensable tool for people's life and production.Ultra-wideband and ultra-narrowband technologies are becoming more and more widely used,and most of them are asymmetrical modulations.HF channel is an old,classic and time-varying multipath channel,which has attracted much attention and application in the military field.At present,the demodulation performances of asymmetrical modulation over HF channels are unsatisfactory.This thesis processes two methods to demodulate an asymmetrical modulation signal called multi-position phase shift keying(MPPSK).One innovatively combines particle swarm optimization with artificial neural networks.The other applies convolutional neural networks and its variant algorithms in deep learning.First,the MPPSK technique is studied,which includes the classical MPPSK modulation,the modified MPPSK modulation and bipolar pulse MPPSK modulation.Much work has been done on the analysis of MPPSK from time and frequency domain.Two methods for the MPPSK modulator are also given: look-up table and switching.Secondly,the paper introduces the system model of MPPSK signal over HF channel,which includes the design principle and output of the baseband shaping filter,the propagation mode,mathematical model,standard channels and output results of the HF channel,the correlative and non-correlative detection methods.Thirdly,the demodulator for MPPSK signal over HF channel is designed based on particle swarm optimization and artificial neural networks.The influence factors of PSO-NN are studied in detail.A large number of experimental results show that PSO-NN algorithm combines the advantages of PSO and ANN,and overcomes the disadvantages of neural network,such as slow convergence and easy to fall into local optimum.Therefore,PSO-NN is superior to traditional methods in demodulation.The performance of multi-symbol joint demodulation is better than single-symbol demodulation.The performance of PSO-NN decreases with decreasing transmitting bandwidth and increasing code rate.The selection of transfer function and error function also has a large influence on the performance.Leaky-ReLU transfer function and cross-entropy error function are more suitable for PSO-NN in this case.Finally,the convolutional neural network,convolutional autoencoder and convolution recurrent neural network are implemented to design the demodulator of MPPSK signal over HF channel.The results show that deep learning algorithm can discover more intrinsic characteristics and links of the data.Thus,the demodulation performance of deep learning neural networks is superior to other traditional methods.The performance of multi-symbol joint demodulation is better than single-symbol demodulation,and it decreases with the decreasing of the transmitting bandwidth and the improvement of the code rate.The number of network layers and the size of the convolution kernel also affect the decision performance.Experimental results show that when the number of layers is 8 and the size of convolution kernel is 49,the demodulation performance of MPPSK signal over HF channel is better.The larger difference between the frequency of the interference and the system modulation frequency is,the better the anti-interference capability of the CNN network is.