| Massive MIMO can significantly improve the spectral efficiency of the system,and has become a key technology in 5G.However,with the increase of antennas number,the performance of traditional detection methods is limited.The complexity of the detection increases sharply with the number of antennas.In massive MIMO scenery,the complexity of message-passing algorithm is relatively low.It can achieve convergence and obtain better detection results.Cascading the detector and the channel decoder for Turbo iteration can also effectively enhance the BER performance of the receiver.In this paper,a low complexity expection propagation(EP)algorithm using the Turbo iterative framework is proposed,which has better bit error rate performance under various loaded conditions.Based on the framework of iterative detection and decoding(IDD),a low complexity successive over relaxation(SOR)expectation propagation(EP)is proposed for massive MIMO.When calculating the posterior distribution of the received signal in EP iteration,matrix inversion with high complexity can be replaced by a linear equation,which can be solved by SOR.In IDD structure,the feedback of the decoder is used as a priori information of EP iteration.The feedback symbols generated by the high reliability decoder have small variance and make the coefficient matrix of linear equation diagonally dominant.Therefore,SOR can converge quickly under this scheme.The above SOR method reduces the computational complexity from O(NT3)to O(NT2).Numerical results show that SOR-EP maintains almost as good performance as traditional EP.The proposed method converges in various loaded scenarios,especially in the heavily-loaded case.This is superior to other low-complexity EP methods.Applying the proposed low-complexity SOR-EP algorithm to DEP(for the soft information of the transmitted symbols obtained from the feedback of the decoder,the moment matching is also performed to speed up the convergence of the outer iteration)can still reduce its complexity without affecting its convergence speed.By using the optimization methods for SOR,A-SOR(adaptive relaxation factor SOR)and NA-SOR(fixed relaxation factor,optimized SOR method),better BER performance is further obtained.At the same time,the deep learning method is used to further optimize the SOR-EP without Turbo iteration to improve the BER performance.Due to the channel hardening characteristics,the estimated value of matrix inversion and the actual value will differ by a fixed multiple.A multiplicative factor is added to the estimated value of matrix inversion to make it closer to the real matrix inversion value;In addition,each iteration of EP involves the prior messages passing of the transmit symbols.In order to ensure the stability of the messages passing,a damping factor is added between each iteration.A model-driven EP detection network SOR-EPNet is constructed to optimize the above parameters and improve the performance of the low-complexity SOR-EP without Turbo iteration. |
