Study On Blind Equalization Algorithm Of Satellite Channel Based On Extreme Learning Machine

In satellite communication system,the non-linearity and group delay of satellite channel lead to distortion of transmission signal,spectrum leakage and interference to adjacent channels,which seriously affect the quality of communication.Receiver blind equalization technology can recover the distorted signal adaptively without obtaining channel prior information and training sequence.It can save spectrum resources and effectively overcome the impact of non-ideal characteristics of satellite channel.Extreme Learning Machine(ELM)is a new type of single hidden layer feedforward neural network.Because of its good non-linear mapping,fast global search ability and simple network topology,ELM is widely used to solve classification and regression problems.The fields include image classification,face recognition,disease diagnosis,time series analysis,channel equalization and so on.In this paper,the blind equalization algorithm of satellite channel based on extreme learning machine is studied.The specific work is as follows.(1)Aiming at blind equalization of high-order QAM signals over satellite channels,the basic theoretical knowledge of extreme learning machine(ELM)is studied in depth,including the random and optimal generation of hidden layer parameters(input weights and biases),the constraint solving of output weights,linear and non-linear output layers and the fully complex ELM.(2)The neural network blind equalization method based on prediction principle and traditional cost function is deeply studied.In the framework of complex extreme learning machine,the Prediction-based ELM Blind Equalization Algorithms(P-ELM-BEA)and the ELM-SVR-CMA is proposed,which is based on Constant Modulus Algorithms(CMA)and trained by Support Vector Regression(SVR).For 16 QAM signals in linear and non-linear channels,the Prediction-based ELM Blind Equalization algorithm(P-ELM-BEA)and ELM-SVR-CMA simulation experiments are completed.The experimental results show that compared with P-ELM-BEA,ELM-SVR-CMA has lower Mean Square Error(MSE)under the same experimental conditions.However,both algorithms have phase rotation problems.(3)In order to overcome the phase rotation problem of ELM-SVR-CMA,a multimodulus algorithm(MMA)is used to replace CMA in ELM-SVR-CMA in the framework of the fully complex extreme learning machine,aiming at blind equalization of high-order QAM signals in memory-based non-linear satellite channel.ELM-SVRMMA,a multi-mode blind equalization algorithm for extreme learning machine,is proposed.On this basis,a dual-mode blind equalization scheme ELM-SVR-MMA-DD is constructed by combining decision guidance(DD)algorithm.The simulation results show that ELM-SVR-MMA-DD can successfully equalize 16 QAM signal in memory non-linear satellite channel by selecting appropriate activation function and using nonlinear output function.The constellation is clear and phase rotation is eliminated.(4)In order to further improve the performance of ELM-SVR-MMA and ELMSVR-MMA-DD for blind equalization of satellite channel,Restricted Boltzmann Machine(RBM)is embedded in the fully complex extreme learning machine.The weights and biases obtained by RBM learning are used to replace the parameters randomly generated by traditional ELM.An improved blind equalization algorithm based on RBM,ELM-R-SVR-MMA,is proposed and constructed.The corresponding dual-mode blind equalization scheme is ELM-R-SVR-MMA-DD.The simulation results show that,compared with ELM-SVR-MMA and ELM-SVR-MMA-DD,ELM-R-SVR-MMA and ELM-R-SVR-MMA-DD can achieve lower MSE level under the same data packet size,and the constellations are clearer and more compact after equalization,and the number of hidden layer nodes is fewer,and the network topology is simpler.Compared with the classical blind equalization algorithm based on Volterra filtering,the proposed algorithms have better equalization performance with fewer data points.