Research On Key Technologies Of SCMA Based On Deep Learning

With the rapid development of various intelligent terminal access,virtual reality,augmented reality,Internet of Things and mobile Internet,more requirements are putting forward for fifth generation(5G)wireless communication,such as massive connection,high throughput,low latency,and better service experience and so on.In order to achieve the requirements,non-orthogonal multiple access allows multiple users to share on the same time,frequency,space,code or power domain,which can increase the number of user connections.Sparse code multiple access(SCMA)is a kind of non-orthogonal multiple access shared by code domain.Due to non-orthogonality,the detection algorithm of the SCMA receiver is more complicated than the orthogonal system.By combining SCMA with MIMO(Multiple Input Multiple Output)or OFDM(Orthogonal Frequency Division Multiplexing),it is possible to further use antennas or subcarriers to improve throughput.In addition,deep learning as a kind of machine learning,has been widely used in computer vision and natural language processing.Therefore,it is a feasible method to solve many existing communication problems by applying deep learning to future mobile communication networks.This thesis mainly studies the detection problem of SCMA by reducing the detection complexity and improving the detection reliability through sphere decoding and message passing algorithm based on deep learning,and solves the peak-to-average ratio(PAPR)problem of OFDM and SCMA-OFDM systems through deep learning.Firstly,improved sphere decoding algorithms are considered for SCMA detection.In the SCMA system,the complexity of optimal maximum likelihood detection is very high.Therefore,on the one hand,the receiver of SCMA uses the message passing algorithm(MPA)to realize low complexity detection by exploiting the sparse at the expense of some performance loss.On the other hand,the sphere decoding algorithm is a feasible method to achieve optimal detection with low complexity.Sphere decoding is used for detection.Firstly the original SCMA detection problem is converted into the integer least squares problem based on the structure of SCMA codebook.Then the Schnorr Euchner(SE)sphere decoding algorithm is used to detect uncoded SCMA system.In order to reduce the complexity in the search stage of sphere decoding,the matrix is reordered before the search stage to remove unnecessary nodes as early as possible,and the non-zero lower bound is calculated to reduce the number of visited nodes in the search stage.The improved SE(ISE)can achieve the optimal maximum likelihood performance with 1.1% complexity of MPA algorithm in the SCMA system without channel coding.In addition,in order to further reduce the complexity of ISE algorithm,ISE algorithm based on deep learning is proposed.By using the minimum radius of sphere decoding obtained by deep neural network,we can reduce the computational complexity of ISE at high signal-to-noise ratio region while ensuring the reliability of ISE.However,due to the hard decision of the ISE sphere decoding algorithm in the uncoded SCMA system,some information is discarded and SE is no longer suitable for coded SCMA system.Therefore channel ordering and non-zero lower bound are applied to the single tree search(STS)sphere decoding algorithm for coded SCMA detection.The modified STS(MSTS)can reduce the computational complexity of the MPA algorithm by about 76% in the coded SCMA system.Secondly,deep learning is proposed to optimize the improved message passing algorithm.The shuffled message passing algorithm(SMPA)is serial strategy of MPA for SCMA detection,which can accelerate the convergence rate of MPA.However,it still achieves the near-optimal performance due to effect of cycle in factor graph.Therefore,we propose SMPA based on deep learning(DNN-SMPA),which can be constructed by unfolding SMPA into deep neural network.We reduce the influence of the cycle by training the weights assigned to the edges of the factor graph.With the same computational complexity,DNN-SMPA has a better bit error rate(BER)than other message passing algorithms,specifically,DNN-SMPA with I = 3 iterations has respectively about 0.4 d B,0.8 d B and 1.0 d B gain over SMPA,DNNMPA and MPA in Rayleigh-fading channels.And DNN-SMPA has fewer training parameters than a fully connected neural network.Finally,deep learning is considered to improve the tone reservation(TR)algorithm to reduce the PAPR of OFDM and SCMA-OFDM systems.PAPR in OFDM and SCMA-OFDM systems will cause the high-power amplifier at the transmitter to work in the saturation region,resulting in in-band distortion and out-of-band distortion that interferes with other signals,reducing system reliability.In order to reduce the PAPR of OFDM and SCMA-OFDM systems,we propose tone reservation algorithm based on deep learning(DL-TR),which can be constructed by unrolling the TR algorithm into a deep neural network.The loss function contains PAPR and the power of reserved subcarrier signals.The peak ratios as the network parameters are optimized by training.Given the same computational complexity,the DL-TR algorithm has few training parameters.The PAPR of DL-TR has about 0.8 d B and 3.1 d B gain over TR and original OFDM under 64 sub-carrier numbers,respectively.The BER of DL-TR is closer to that of original and is about 0.02 d B better than TR.There are similar results in SCMAOFDM.DL-TR can achieve lower PAPR and BER than existing algorithms.