Research On Multi-user Detection For Non-orthogonal Multiple Access Based On Variational Learning

Non-orthogonal multiple access has received extensive attention as a promising technical scheme to realize future massive access in these years.However,there are some problems and challenges with this technology in the following aspects.Firstly,in the design of the transceiver of NOMA system,the codebook and multiuser detection of the existing transceiver are designed individually,which is difficult to be included in the unified optimization framework.This is also a limit to the further improvement of bit error rate performance.The joint design of transceivers,although achieved by deep learning techniques,fails to achieve a better compromise on reliability and spectrum resource utilization.There is still room to improve the corresponding bit error rate performance for different overload rate requirements.Secondly,NOMA system requires good channel estimation.Existing algorithms based on linear minimum mean square error estimation and compressive sensing highly rely on channel statistics as prior information.Though the deep learning-based channel estimation method compensates for the problem of obtaining prior information through training data,it requires a great number of training samples matching the actual data,and its robustness is insufficient.Finally,the multiuser detection algorithms of NOMA are mostly iterative ones with good bit error rate performance but high computational complexity.Multiuser detection based on deep learning can ensure that there is no great loss of performance while reducing computational complexity.However,network models are mostly used as black boxes and lack design basis and criteria,resulting in poor robustness.This paper studies the joint design methods of NOMA transceiver that are applicable to different overload rates,the channel estimation methods without extracting prior information manually,and the low complexity NOMA multiuser detection methods in the single-cell static environment of a massive access scenario aiming at the above three problems.The main work is as follows:Firstly,aiming at the problem that independent design of transceiver leads to limited performance and joint design methods based on deep learning cannot better match different overload rate requirements,a joint design method based on multiuser variational information bottleneck is proposed to achieve a better compromise on reliability and spectrum resource utilization.An optimization criterion is established in this joint design approach to balance compression and recovery of user-transmitted information.The upper bound analytical form of the criterion was derived by the variational inference method as the mutual information in the criterion was difficult to solve and the joint design objective function under the criterion was obtained.The transceiver is further parameterized through an autoencoder to achieve the joint design of NOMA transceiver.The bit error rate performance of the proposed joint design method and the adaptability to different overload rate requirements are verified by numerical simulation.Secondly,an uplink channel estimation method based on data-model dual-drive variational learning is proposed for the problem that the channel estimation of NOMA system depended on channel statistics as prior information.The posterior distribution mean of user channel gain is derived based on the variational inference method.The inference process is further analyzed to discover the calculation steps that require the involvement of prior information in order to elucidate how the inference process is influenced by prior information.This part is parameterized by the neural network so that this part could be realized by the neural network model and training data without the need for known prior information(channel gain distribution variance).Numerical simulations are carried out to verify the detection performance of the proposed algorithm under the conditions of no prior information,a small number of training samples,and inconsistent statistical distribution with the measured data.Finally,a low complexity NOMA multiuser detection algorithm based on variational learning is proposed for the high computational complexity of the existing NOMA iterative multiuser detection algorithm and the low computational complexity of the deep learning-based multiuser detection algorithm with insufficient design basis and criteria.NOMA multiuser signal superposition is modeled as a binary encoding model to reduce the computational complexity of the inference process.On the basis of this binary encoding model,the posterior probability solution process for each user transmitted bit is derived based on the variational inference methods.The posterior probability solution process is equal to the forward propagation calculation process of a recurrent neural network.In addition,there is a specific numerical relationship between the network model structure and the system parameter configuration.The weight parameters of each calculation unit of the recurrent neural network are set differently to further accelerate the calculation of the posterior probabilities of the user transmitted bits.The bit error rate performance and training convergence performance of the proposed method at different overload rates are verified by simulation.At last,the computational complexity of different algorithms is compared.