Research On MIMO Signal Detection Algorithm Based On Deep Learning In Complex Scenes

In the physical layer of a wireless network system,a fundamental scientific problem is to recover the transmitted information at the receiver as accurate and fast as possible.In practical scenarios,the receiver often suffers from complicated and dynamic noises or interferences.There are many reasons for this phenomenon,one of which is that excessive reuse of spectrum resources causes serious signal interferences in wireless network,and the interferences are often dynamically correlated in time or frequency domain.In addition,for some scenarios such as long wave,underwater communication and multi-access systems,the noise or interference is often non-Gaussian,somehow impulsive,and non-analytical.The diversity of practical environments significantly affects the performance of signal detection algorithms.Due to that the existing multi-input multi-output(MIMO)signal detection algorithms mainly focus on specific noise or interference models(such as white Gaussian noise),they may fail to work well in practice,which has become the bottleneck of the system.Therefore,it has been a challenge for us to design an intelligent signal detection scheme of MIMO system in the presence of complicated and dynamic noise or interference,so that the system is able to adapt to various practical scenarios.Firstly,in order to solve this problem,this thesis proposes a novel signal detection framework based on the principle of maximum an normalizing flow estimation(MANFE),which do not require any statistical knowledge on the noise.The proposed framework is a probabilistic model which can effectively approximate the unknown noise distribution via a normalizing flow.More importantly,the proposed framework is driven by an unsupervised learning approach that does not require any labeled data or statistical knowledge.Instead,it only requires noise samples.In order to reduce the computational complexity of the proposed framework,this thesis further proposes a low-complexity version of MANFE,which uses a signal detector for initial estimation and significantly reduces the search space accordingly.Simulation results show that the proposed framework outperforms other existing algorithms in terms of bit error rate(BER),and it can reach the optimal maximum likelihood estimation performance under Gaussian noise.Moreover,the performance of the lower complexity MANFE is also better than the Euclidean distance based maximum likelihood detector in impulsive noise environment.Secondly,this thesis further proposes an iterative framework which jointly uses a deep convolutional neural network(DCNN)and a conventional signal detection algorithm.The proposed framework can capture the local correlation of noises among different symbols.As a general framework,the proposed iterative framework can improve the performance of conventional signal detection algorithms in correlated noise environment.In this framework,the conventional signal detection algorithm is used for the initial estimation,and the DCNN is used to improve the effective signal-to-noise ratio(SNR)in the presence of correlated noise or interference.In addition,the simulation results are presented to verify the effectiveness of the proposed iterative detection framework.