Neural Network Based MIMO Communication Technologies Under Inter-cell Interference Of The Same Frequencey

With the rapid development of mobile Internet,Internet of Things,cloud com-puting and other technologies,the demand for mobile data traffic presents an explo-sive growth,and the boundary of mobile communication is also being continuously broadened.The rapid development of mobile communication service requires higher efficiency and reliability of the physical layer of a wireless network.Multiple input multiple output(MIMO)technology is a viable way to significantly improve spectral efficiency.However,in cellular network planning,the use of small frequency reuse factor will bring large co-frequency interference between cells,especially at the edge of the cell.Co-frequency interference has become one of the main factors that limit the performance improvement of the cellular network.Based on the neural networks,this paper mainly studies the downlink key technologies in the MIMO multi-cell co-frequency interference environment,and at the same time,the framework is extended to the complex-valued neural network for exploration and analysis.For the channel estimation task,the traditional channel estimation algorithm relies on the prior information and has strong assumptions,so it is difficult to achieve good performance in the actual channel estimation task.In the current wireless MIMO sys-tem,orthogonal frequency division multiplexing(OFDM)modulation is used to divide the communication resources into two dimensions in the time-frequency domain,and the channel parameter values can be estimated by borrowing the relevant technology in the field of deep learning and image processing.However,the current solutions,such as ChannelNet,do not optimize the neural network architecture for channel estimation problem,thus the computation complexity and estimation error performance need to be further improved.Therefore,based on the idea of image superresolution,multiple neural network modules composed of residuals and dense blocks are stacked together to form RD-CENet.The channel information in time-frequency domain is regarded as a low-quality two-dimensional image after linear interpolation based on least square(LS)estimation at pilot positions,and then is mapped to a high-quality channel estimate by RD-CENet.The simulation results show that,compared with ChannelNet,RD-CENET can reduce the computational complexity and improve the performance of channel es-timation in multi-path fading channel environment.At the same time,RD-CENET has good robustness and can still work effectively when the SIR or Doppler shift of the chan-nels does not match the training set.Furthermore,we extend the neural network into the corresponding complex-valued neural network,and the performance of the neural network is compared with that of the real-valued nueral network.After the estimated channel values are obtained,it is necessary to use a certain de-tection technology to restore the transmitted signal.In the design of detection algorithm,superresolution network is adopted at first.According to the simulation results,partial superresolution networks can not achieve a good performance for the signal detection of such a discrete value fitting problem.Therefore,we design the data organization,network structure and loss function based on the traditional interference suppression combining(IRC)algorithm,and propose the interference suppression receiving net-work based on IRC.Compared with the superresolution network,it has achieved more stable performance in multipath fading channel,and has lower complexity and faster training convergence.It suffices to use the data at the pilot frequency to complete real-time training and track the changes of channel.Compared with the traditional merge detection algorithm,the performance of the interference suppression receiving network based on IRC is better under the actual channel estimation with the estimation error.At the end of the paper,the superresolution scheme and IRC scheme are extended to the complex network,and the simulation performance is verified.Moreover,the IRC based receiver neural network is extended to the complex-valued neural network by two different methods and compared.