NOMA Signal Detection Method Based On Deep Learning

Non-orthogonal multiple access(NOMA),one of the most promising technologies responding to the growing demand for access to the mobile Internet and the massive Internet of Things,has received active attention in recent years.Compared with traditional orthogonal multiple access(OMA)technologies,NOMA systems have higher spectral efficiency,faster transmission rates,and larger system capacity.NOMA actively introduces interference between users,and its high spectral efficiency is achieved at the cost of increasing receiver complexity.Therefore,it is necessary to improve the signal detection accuracy of NOMA system and reduce the complexity of the receivers.Due to the vigorous development of artificial intelligence in recent years,deep learning technology has been increasingly widely applied in the field of communication.Therefore,this thesis studies the detection method of uplink NOMA system based on deep learning.Firstly,this thesis proposes a receiver scheme named DLSI for uplink NOMA system based on data-driven deep learning.In this scheme,the channel matrix is decomposed into QL to obtain the equivalent signal observation information and channel state information,and the solution of signal detection objective function is constructed iteratively.Then deep neural network(DNN)is used to realize the iterative steps and restore the symbols of the transmitted signals one by one.After all symbols of one user are detected,the next user’s symbols are detected using a successive interference cancellation(SIC)algorithm.A soft decision layer is added to the output of each DNN,and the soft information is input to the next DNN to improve the accuracy of training.Soft information contains more signal knowledge than hard decisions and is therefore also used in SIC steps,which reduces error propagation to some extent.Simulation results show that DLSI has better performance than traditional minimum mean square error(MMSE)detection methods and other deep learning methods.Although data-driven DLSI has good simulation performance,its performance is highly dependent on parameters,and the DLSI is easy to fall into the local optimal state during training,and a large number of parameters need to be trained to achieve good results.Therefore,this thesis also proposes a pilot assisted channel estimation and signal detection joint optimization receiver method,named PA-LSIC,for uplink NOMA systems with unknown channel information.This method uses model-driven neural network,combines deep learning with traditional detection methods,introduces trainable projection gradient descent algorithm and preserves the SIC detection structure.In this scheme,step size and projection softness are regarded as learnable parameters,and a learnable noise cancellation factor is added to channel estimation to solve the problem of noise interference in channel estimation,and a learnable interference cancellation factor is added to reduce noise interference in channel estimation to solve the problem of incomplete interference elimination and error propagation in SIC process.Learning parameters are added on the basis of traditional channel estimation and signal detection algorithm to ensure the minimum performance.The PA-LSIC reduces the complexity of training and implementation,reduces learnable parameters and improves the efficiency of learning and implementation compared with data-driven deep learning methods.