Research On Shortwave Time-Varying Channel Equalization Method Based On Deep Neural Networks

The multipath and Doppler effects in shortwave time-varying channels lead to severe inter symbol interference(ISI)in transmission signals.Adaptive equalization technology is one of the commonly used measures to overcome ISI.However,shortwave timevarying channels are nonlinear,and traditional adaptive equalization algorithms such as the Least Mean Square(LMS)algorithm and Recursive Least Squares(RLS)algorithm lack the ability to handle nonlinear problems.In recent years,deep learning technology has shone brightly in the field of wireless communication,and a series of equalizers based on deep methods have emerged.This thesis focuses on the problem of frequency selective fading caused by multipath effects and time selective fading caused by Doppler frequency shift in the previous channel settings of shortwave time-varying channels.Different outdoor single input single output models are studied,and the SUI(Stanford University Inter)channel of Stanford University,which comprehensively considers different terrains in practical applications,is simulated and used as the basic channel model.At the same time,aiming at the problem that the underlying model of the traditional equalization algorithm is not suitable for the depth equalization algorithm,the conditional probability model based on which the depth equalization algorithm is based is analyzed.Aiming at the problem that the traditional adaptive equalization algorithm has poor ability to deal with nonlinear problems and the general one-way recurrent neural network equalizer has weak ability to equalize,on the basis of the depth equalization model based on conditional probability,this paper studies the two-way correlation of the input sequence of the equalization end of the HF time-varying channel,and proposes a depth equalizer based on the Bidirectional Gate Recurrent Unit(Bi-GRU),We have achieved better noise resistance and ability to cope with time-varying channels than LMS and RLS equalization algorithms and general unidirectional recurrent neural network equalizers on simulated SUI channels.This thesis focuses on the slow convergence speed and long training time of recurrent neural network equalizers,as well as the insensitivity of convolutional neural networks to temporal signals.Temporal Convolutional Network(TCN),commonly used for low bit rate speech synthesis,is studied.However,the unimproved TCN still cannot effectively address the equalization problem of time-varying channels.This thesis combines the characteristics of the input sequence at the equalization end and extensive experimental analysis to improve its main modules,and selects appropriate parameter combinations to design a deep equalizer based on the improved TCN.The equalizer outperforms LMS and RLS in terms of noise resistance and ability to cope with timevarying channels,and its convergence performance and training cost are far superior to those of recurrent neural network equalizers.The Bi-GRU based deep equalizer and the improved TCN equalizer proposed in this thesis can greatly alleviate ISI in shortwave time-varying channels,and the latter brings smaller time overhead and better convergence performance compared to general deep equalizers,opening up a new path for deep methods to solve shortwave time-varying channel equalization problems.