Deep MIMO Detection Based On QAM Signal Decoupling

Multiple-input multiple-output（MIMO）technology is one of the important technologies for 4G/5G and even for future wireless communications.With the increasing number of antennas,the spectrum efficiency of the communication system have been greatly enhanced,and the quality of communication has been significantly improved.However,at the same time,the interference between different antennas also increases,which greatly increases the complexity of the detectionalgorithm.Therefore,it is of great significance to study MIMO signal detection algorithms with high accuracy and low complexity.This thesis studies the LISA deep detection algorithm based on Long Short-Term Memory（LSTM）network.The main work is as follows:（1）The performance of the traditional LISA deep detection algorithm degrades in the varying channel scenario,and deteriorates sharply with the increase of the modulation order of QAM signal.This thesis proposes to transform the high-order QAM signals into QPSK signals for detection,so as to transform the complex quantization order recognition of QAM signal into simple positive/negative polarity recognition of QPSK signal.This can fully exploit the advantages of neural network in binary classification.（2）In order to better learn the information of the system channel,this thesis uses the Bi-directional Long Short-Term Memory（Bi-LSTM）network to model the signal detection system.In order to adapt to the sequential characteristics of Bi-directional Long Short-Term Memory（Bi-LSTM）network,the detection order of signals is sorted according to the equivalent channel conditions,so as to reduce the error propagation caused by Bi-LSTM serial sequential processing.（3）Aiming at the problem of storage and calculation in the signal detection based on deep learning,this thesis compresses and accelerates the model of the improved LISA algorithm.This thesis firstly analyzes the complexity of each module of the proposed MIMO detection system,uses the low-rank decomposition to accelerate the compression of Bi-LSTM network in the improved LISA algorithm,and proposes a channel pruning method based on cell state mask to further improve the ratio of acceleration and compression.Simulation results show that the improved LISA algorithm significantly improves the performance of LISA algorithm in varying channel scenarios.In the case of 16QAM modulation and BER of 10^-2,the detection performance of the improved algorithm is about 2d B higher than that of the original LISA algorithm,achieving the quasi-optimal performance.At the same time,with the increase of modulation order,the detection performance of the improved LISA algorithm is still very excellent.Even if the storage of the compact model is reduced by 2/3（the amount of calculation is reduced by more than half）,the compact model still has higher detection accuracy than the original LISA algorithm.