Study On Online Blind Equalization Algorithm Of Extreme Learning Machine

There are two main types of equalization techniques that can effectively improve the effects of multipath effects and channel distortion on communication quality: adaptive equalization techniques and blind adaptive equalization technique.The former can effectively eliminate intersymbol interference caused by multipath effects,but it requires the constant transmission of training sequences,resulting in the loss of the effective spectrum.The blind equalization technique is able to achieve channel equalization without training sequences,only using the priori information of the receiving sequence.It is gradually attracting attention of researchers.Among them,neural network has become one of the important directions for solving the blind equalization problem due to its good nonlinear mapping ability.As a novel single-hidden layer feedforward neural network,extreme learning machine(ELM)not only inherits its good generalization ability,but also has the advantages of being simple and running fast.However,there is little research on blind equalization algorithms based on ELM,and the research results are batch training algorithms,which cannot meet the requirements of real-time communication system.So,this thesis proposes an constant modulus algorithm based on the online sequence ELM(OS-ELM-CMA)and on-line blind equalization algorithms based on the recursive least squares algorithm.The main contributions are as follows:(1)After further investigating of ELM and their related algorithms,a constant modulus algorithm based on online sequence ELM is proposed(OS-ELMCMA).The algorithm will construct the cost function by introducing the cost function of the Regularized ELM(RELM)to Online Sequential Extreme Learning Machine(OS-ELM),combining the error function of the constant modulus algorithm(CMA).And then,solve the output weights of the initial learning phase of OS-ELM by.the mehod of Iterative Re-Weighted Least Squares(IRWLS)In the learning pro-cess,OS-ELM-CMA can add data one by one or block by block,which meets the requirements of online blind equalization.Simulation results show that,compared with the traditional CMA and the RELM-CMA based on bathes,the proposed OS-ELM-CMA not only achieves online blind equalization,but also has a lower mean square error values.(2)To address the disadvantages of the OS-ELM needing a bulk of data to train the network in the initial learning phase,based on ELM,a recursive least squares constant modulus algorithm(ELM-RLS-CMA)and a recursive least squares multimode algorithm(ELM-RLS-CMA)are proposed for the constant signals and multimode signals,respectively.In addition,some experiments are designed for nonlinear channel represented by satellite channel to verify the effect of different parameters on these algorithms and the differences of performance between the two algorithms for QPSK normal-mode signals and 16 QAM multimode signals.The results show that ELM-RLS-MMA not only has comparable performance to ELM-RLS-CMA,but also can solve the phase rotation.(3)In order to comprehensively verify the ability of the proposed ELM-RLS,further based on the prediction principle,the ELM-RLS is used as a nonlinear error filter,training the output weights by the RLS,and the automatic gain device and the the rotation factor enables online blind equalization of 16 QAM multi-mode signals.The ELM-PEF and ELM-RLS-MMA are proved to have comparable performance in simulation experiments.All of them can solve the phase deflection problem with good equalization capability.