Research On MIMO Detection Algorithms Based On Variational Bayesian Inference

Massive Multiple-Input Multiple-Output(Massive MIMO)technology has been applied to The 5th Generation(5G)wireless communication system,which significantly improves spectrum efficiency.In the future communication systems,this technology will still attract the attention of academia and industry.In a typical up-link multi-user MIMO system,the base station serves multiple users using the same time-frequency resource,causing the received signal a superposition of multiple users.So detection algorithms are needed to decode the signals for each user.With the development of deep neural networks,researchers recently designed MIMO signal detection algorithms based on model-driven deep learning technique,to improve accuracy and efficiency.However,in the case of high rate millimeter wave MIMO transmission,the Analog-to-Digital Converters(ADCs)at the receiver,which are working at high sampling rate,can not meet the sampling precision required by those detection algorithms.Therefore,the detection performance of the receiver decreases significantly.Moreover,ADCs with capability of high sampling rate and high precision are too expensive to be commercially used in large-scale MIMO.Onebit quantization ADCs,also known as comparators,are easy to reach high sampling rate while cheap to deploy.But one-bit quantization loses a lot of information of the received signal,therefore a new type of detection algorithms need to be designed specifically for one-bit quantization.Based on the previous research of detection algorithms for non-quantized MIMO system,this thesis designed a one-bit MIMO detection algorithm named One VBN.And it is based on model-driven deep learning and variational Bayesian inference technique.Firstly,the posterior distribution of transmitting signal to be estimated is derived by variational Bayesian inference theory.Then,iteratively updating steps of the signal to be estimated and intermediate variables are derived by variational expectation maximization(VEM)technique.Finally,by adding learnable parameters into each iteration,the iteration steps are deep unrolled into multiple model-based neural networks.In the independent identically distributed Gaussian channels,compared with the state-of-the-art algorithm named Deep HOTML,One VBN has a better performance than it.In the Rural Macro-Non Line of Sight(RMa-NLOS)channel defined by 3rd Generation Partnership Project(3GPP),One VBN has a clear advantage over Deep HOTML.Because the number of learnable parameters is less than traditional neural networks,One VBN is easy to train and it can adjust parameters timely with the change of wireless channels.In the case where channel estimation error exists,One VBN still holds its advantages because of robustness.In addition,the proposed One VBN algorithm does not need prior information of noise variance,and has low computational complexity while converges faster than the others.