Nonlinear Distortion Compensation For OFDM-MIMO Systems Based On Machine Learning

Next-generation wireless networks must be well-prepared for ultra-high-speed data services to meet the growing demands of end users or applications.Orthogonal Frequency Division Multiplexing(OFDM)is a viable broadband technology that can be used to support high bit rate services.Multiple-Input-Multiple-Output(MIMO)systems deploy multiple antennas on both the transmitter and receiver ends of the wireless link,which can increase system capacity without increasing bandwidth or transmission power.Therefore,the combined system of MIMO and OFDM,namely OFDM-MIMO system,has been widely used in many public wireless systems due to its commercial feasibility.In this system,the OFDM signal has a large Peak-to-Average Power ratio(PAPR),so the performance of the transceiver is very sensitive to nonlinear distortion caused by the power amplifier when operating near the nonlinear saturation region.Due to the nonlinear memory effect of the power amplifier,intersymbol crosstalk will be introduced in the OFDM-MIMO system,resulting in the distortion of the output signal and reducing the spectral efficiency.Hybrid precoding technology can resist multipath fading,solve the problem of channel interference,and improve spectral efficiency.It is an important signal processing method in MIMO systems.Therefore,this paper uses memory polynomials to model nonlinear amplifiers,studies the crosstalk between symbols in the structure combined with hybrid precoding,and proposes a digital predistortion method based on Recursive Least Squares(RLS)combined with hybrid precoding.The method can extract the parameters of the digital predistortion module through adaptive learning in the training phase.The simulation results show that this method can reduce the Normalized Nean Square Error(NMSE)of the system by about 8dB compared with the digital predistortion method based on the Least Mean Squares(LMS)algorithm combined with hybrid precoding.At the same time,this paper uses different numbers of RF links to verify the pre-distortion algorithm when the SNR is equal to 20dB.The simulation results show that the algorithm can improve the spectral efficiency by about 16%compared with the LMS algorithm when the RF link number is equal to 8 and the OFDM signal can be demodulated correctly at the receiving end.In addition to the problem of inter-symbol crosstalk caused by nonlinear amplifiers,a part of the signal output by the amplifier will be coupled with each other when passing through parallel multiple RF channels,that is,crosstalk between different channels occurs.Aiming at this problem,this paper proposes a CNN-BiLSTM digital predistortion method based on attention mechanism.The method firstly models the predistorter of the power amplifier through the neural network structure,and finally realizes the compensation of the crosstalk effect between different channels and eliminates the nonlinear characteristics of the power amplifier.In this paper,the number of output channels and filter size of CNN network in the proposed algorithm are set to 128 and 3,respectively,and the number of hidden layer neurons and training rounds of BiLSTM network are set to 128 and 55,respectively,to verify the algorithm.The simulation results show that the proposed algorithm can reduce the NMSE of the signal by about 20dB compared with the traditional algorithm,and reduce it by about 5dB compared with other machine learning algorithms,and can obtain constellation diagrams that are closer to the linear output signal.Therefore,the proposed method can better compensate the crosstalk effect between different channels and the nonlinearity of the power amplifier.