Design Of A Blind Equalizer Based On Recurrent Neural Network

With the vast development of digital communications,the demand for better communication quality is growing higher.Channel equalization,which is capable of significantly improving communication quality,can be divided into adaptive equalization and blind equalization according to the existence of training data or not.Blind equalization has higher bandwidth utilization and can be applied in more areas for the absence of training data.Therefore,the design of a better blind equalizer is quite meaningful.In this thesis,recurrent neural network is used to construct a blind equalizer for its powerful information processing ability and learning ability after comparing various kinds of structures,and some improvements on its training algorithm and structure is completed.After analyzing some common first order gradient descent optimization algorithms,a new first order gradient descent optimization algorithm has been proposed based on NAG and RMSprop.Moreover,this new algorithm converges faster and produces less mean squared error compared with other first order gradient descent optimization algorithms.According to simulation results,new algorithm tends to converge after 50 iterations while other algorithms need more than 300 iterations,which means that it is extremely suitable for the training of blind equalizer based on recurrent neural network.In order to further improve the performance of blind equalizer,the combination of ensemble learning and blind equalization is proposed and has been proved to be able to enhance blind equalizer's robustness and generation ability.The MSE of an equalizer will decrease by at least 1dB when adds the number of base learners up to more than 3.According to Python simulation results,MSE of the improved blind equalizer is 8dB less than ordinary blind equalizer which is based on recurrent neural network without applying new first order gradient descent optimization algorithm and ensemble learning.Finally,circuit design and FPGA verification of the improved blind equalizer are completed in this thesis.In circuit design,ensemble learning is realized in a serial way,which costs less hardware resource and has achieved a great balance between performance of blind equalizer and resource consumption.Despite that some approximate calculations in ModelSim simulation degrade its accuracy,the improved blind equalizer is still 6.7dB better than the ordinary one.Moreover,this blind equalizer has been proved to accomplish expected requirement after FPGA verification.