Design Of An Adaptive Equalizer Based On Neural Network

The importance of communication technology in modern science and technology continues to increase.As a kind of technology which is capable of improve communication quality apparently,adaptive equalization occupies a very important position in digital communication technology.It is quite meaningful to design a kind of adaptive equalizer which has high performance and efficiency.In this thesis,based on neural network and machine learning technology,many kinds of neural network adaptive equalizers based on traditional structure are researched,neural network with decision feedback structure is adopted as the basic structure of adaptive equalizer.The performance of the equalizer is superior to traditional adaptive equalizer based on forward neural network structure.This kind of equalizer has less computational complexity compared to recurrent neural network and is suitable to wireless communication system.The adaptive equalizer is optimized from two parts including algorithm and structure in this thesis.In terms of traditional neural network training algorithms,a kind of improved algorithm is proposed.Compared to other traditional first order training algorithms,iteration number of the proposed algorithm decrease by more than 50%,the improved algorithm has less steady state mean square error and is suitable to training neural network adaptive equalizer.Ensemble learning method in machine learning is combined with decision feedback neural network adaptive equalizer in this thesis,a kind of integrated decision feedback neural network adaptive equalizer is designed.This kind of equalizer has stronger generation ability and less mean square error and the integrated structure equalizer is able to decrease mean square error by 2dB compared to single structure equalizer.Logic design,Modelsim simulation and FPGA verification of integrated decision feedback neural network adaptive equalizer are given in this thesis.The number of base learner is appropriate when designing the circuit of integrated decision feedback neural network adaptive equalizer.The result of simulation demonstrates that,despite of a certain degree of the accuracy loss,the mean square error of integrated decision feedback neural network adaptive equalizer still decrease by 1.7dB compared to single structure equalizer.Because of less base learner which costs less hardware utilization resource,a great tradeoff between equalization performance and hardware consumption is achieved.