Research And Optimization Of MIMO Communication Systems Based On Deep Neural Network

With the development of the information society,various mobile communication terminals have gradually become popular.People's lives are inseparable from wireless communication services,which have unknowingly covered every corner of people's daily lives.As the number of users grows,the corresponding data is also increasing rapidly as well.This trend raises higher performance requirements for wireless communication technology,but traditional wireless communication technology is difficult to meet such requirements.In order to meet people's increasing requirements for wireless communication technology,Multiple-Input Multiple-Output(MIMO)technology is proposed.MIMO technology can achieve higher system capacity,data transmission speed and spectrum utilization by using multiple antennas at the transmitter side and the receiver side of the communication system,which can meet people's increasing requirements for wireless communication and is considered as one of the most promising wireless communication technologies.But at the same time,MIMO technology also faces a series of challenges such as high system complexity and inter-channel interference.In recent years,deep learning technology has performed prominently in many fields by using the ability to extract potential features and laws from data.In some fields,it has gradually replaced many traditional methods,making neural networks return to people's field of vision in the form of deep learning,triggering the trend of further research and application.Therefore,it can be considered that deep learning is also worthy to be researched for solving related problems of MIMO systems.This paper researches the application of deep neural networks in deep learning to MIMO systems.The main research contents are as follows:1)Focusing on signal detection in MIMO systems,this paper proposes a parallel detection network based on deep neural networks for MIMO systems.Compared with the traditional MIMO detection network based on fully connected deep neural networks,the parallel detection network proposed in this paper is inspired by the random forest algorithm and residual learning,which modifies multiple traditional fully connected detection networks with the idea of residual learning and connects these modified detection networks in a parallel manner.Each modified detection network can be used as an independent detection network.By introducing some differences into these modified detection networks,the diversity of the detection results can be achieved.After that,the detection result closest to the real transmitted signal is selected by using the Maximum Likelihood(ML)method as the final detection result of the parallel detection network.Simulation results show that,compared to some traditional detection algorithms and the fully connected detection network,the parallel detection network proposed in this paper can achieve better detection performance in different MIMO systems.2)Focusing on the codebook optimization problem in Generalized Space Shift Keying(GSSK)systems developed based on MIMO systems,this paper proposes a codebook generation network for GSSK systems based on deep neural networks.GSSK systems can carry information bits by activating several transmitting antennas and using the spatial information of these antennas.When the number of antennas is fixed,the performance of the GSSK system can be improved by optimizing the codebook design.The network structure,loss function,activation function and training strategy of the codebook generation network proposed in this paper is specially designed,which can enable the codebook generation network to generate optimized codebooks for current channel conditions and noise.Simulation results show that,compared with the codebooks generated by the traditional codebook generation methods,the codebook generated by the codebook generation network proposed in this paper can help different GSSK systems achieve better performance.