MIMO-NOMA Signal Detection Method Based On Model-driven Deep Transfer Learning

Non-orthogonal multiple access technology(NOMA),as a key technology that can support large-scale connections of users or devices and improve spectrum efficiency,is increasingly combined with multiple-input multiple-output(MIMO)technology,i.e.,MIMO-NOMA system.In wide-coverage multi-node dense access scenarios,MIMO-NOMA systems have obvious performance advantages in providing large system capacity and high spectrum efficiency.However,MIMO-NOMA systems use a combination of non-orthogonal transmission and multiple antennas,which undoubtedly increases the processing complexity of the transmitters and receivers,and poses challenges to the design and implementation of the transmitters and receivers.One of them is how to design the signal detection algorithm at the receiver.In recent years,Deep Learning(DL)and its branches have been widely and effectively applied in the field of communication to solve communication physical layer problems.In view of this,based on the DL method,this paper deeply studies the signal detection algorithm of MIMO-NOMA system.The main contents are as follows:First,for the downlink MIMO-NOMA system,a model-driven deep learning signal detector named LPCG-SIC detector is proposed.The method firstly rewrites the linear minimum mean square error equalizer in the conjugate gradient algorithm,and introduces an interference adaptive term with learnable parameters to eliminate the interference between users.Then the CG algorithm is improved by introducing a simple and effective preconditioning matrix transformation,which can ensure the convergence of iterative search and accelerate the convergence speed.Finally,the deep unfolding approach is used to combine the improved algorithm with the successive interference cancellation(SIC)structure in NOMA detection,and the search step-sizes and conjugate direction step-sizes of each iteration are extended from scalars to vectors as the learnable parameters of each layer.The detection network has fewer parameters,lower computational complexity,only needs to optimize a small number of trainable parameters.The simulation results show that the LPCG-SIC detector has better bit error rate performance than several other MIMO-NOMA detectors.Second,in order to overcome the shortcomings of most existing DL-based detectors that only focus on signal detection tasks in a given environment and cannot adapt to changes in feature space or distribution,a DTL-based detection algorithm framework and three model-driven DTL strategies are proposed,namely DTL-Det.Consisting of three stages: offline learning,transfer learning,and online detection,the framework is a general DTL workflow for signal detection,applicable to any model-driven deep unfolding network.In the proposed framework,a pre-trained deep network model for signal detection is first built in the source domain,and invariant common features are extracted through offline learning.Once the system detects changes in the communication environment,it will adopt the proposed DTL strategy to transfer the knowledge of the common features of the source domain and the target domain in the pre-training model to the target domain through online transfer learning,and then fine-tune the pre-trained network model.Finally,the well-trained network is used to decode the signal online.Simulation results show that the proposed DTL algorithm and transfer strategies can obtain significant performance gain in bit error rate compared to no-transfer methods,and only require relatively few training samples and parameters.