Research On MPPSK Demodulator With Deep Neural Network Over Band-limited Channels

Nowadays,the frequency spectrum resource is more and more valuable with the development of information technology.Efficient modulation technology that can take into account the high spectrum utilization and high energy utilization characteristics is an effective way to solve the shortage of spectrum resource,which has important value.MPPSK modulation has the advantages of high spectrum utilization rate,high information transmission rate,strong encryption,etc.It has a wide range of applications in military and civilian fields.The traditional MPPSK demodulator based on the impacting filter can convert the position of symbol phase impulse into the amplitude impact,which is advantageous for the symbol detection through the impact amplitude decision.When the M is large or the channel environment is terrible,performance of the MPPSK demodulator will deteriorate.With the development of artificial intelligence and the advent of big data,deep learning(DL)has been widely and successfully applied in the fields of computer vision,pattern recognition,natural language processing,image classification,etc.The Stacked Sparse Auto-Encoder(SSAE)network introduces parameter pre-initialization,unsupervised learning and other deep learning advances on the basis of the traditional feedforward fully connected network,so SSAE network can automatically extract the high dimension features from the dataset,and further improve the performance of the network via data transformation and feature reconstruction.The following researches are constructed in this thesis.Firstly,the MPPSK modulation is derived from the development of efficient modulations,and the shortcomings of the traditional demodulator scheme via impact filtering are introduced.The viewpoint of introducing deep neural network(DNN)into the demodulation research of the MPPSK communication system is proposed.Also,the paper introduces the theoretical knowledge of neural network,such as training algorithm,activation function,numerical optimization,etc.In addition,the latest several advances of DL are also expounded.Secondly,the application of MPPSK demodulator with DNNs over band-limited channels is studied.This paper also introduces the SSAE network into the MPPSK demodulator of the MPPSK communication system over band-limited AWGN channel as the application scenario.Additionally,this paper analyzes and demonstrates the influence of neural network parameters on MPPSK demodulator,such as activation function,neural network dimension,numerical optimization,network structure optimization,etc.Simulation results and analysis conclusion of SSAE network are also given.Thirdly,the "multi-symbol joint decision" scheme is proposed based on the characteristics of MPPSK signals,which can extract inter-symbol interface to help demodulating the current symbol.Also,the effectiveness of this scheme for demodulating MPPSK signal over band-limted AWGN channel is demonstrated.Finally,the anti inter-symbol inteference application of MPPSK demodulator with DNNs over-band limted multipath channels is studied.However,Recurrent Neural Networks(RNNs)can solve the complex classification related to time series,so it is used in study of MPPSK demodulator over multipath channel.“Ensemble Learning” is introduced to form an integrated RNN to demodulate MPPSK signals in multipath channels.Conclusions and analysis are provided in the last.