Research On Deep Learning-based Channel State Information Feedback For Massive MIMO

In the massive multi-input multi-output(MIMO)environment,effectively reducing the feedback overhead of channel state information(CSI)has been a hotspot in the field of signal processing.On the one hand,the feedback overhead increases sharply with the increase of the number of massive MIMO antennas;on the other hand,since channel reciprocity in frequency division duplex(FDD)mode is not available,CSI feedback becomes more difficult.To this end,this thesis focuses on the shortcomings of the existing CSI feedback algorithm,combined with deep learning to carry out in-depth research on the feedback reconstruction of CSI,that is to reduce the feedback overhead of CSI in the FDD system under massive MIMO environment and effectively reconstruct the CSI at the base station.The specific research contents are as follows:1.Aiming at the existing CSI feedback reconstruction method without considering the weight between features,a CSI feedback reconstruction method based on convolutional neural network(CNN),long short term memory(LSTM)network and attention mechanism is proposed.This method uses CNN to extract the features of CSI,compress it into codewords and feed it back to the base station,where CNN is used to decompress the codewords.On this basis,combined with the LSTM network to extract the time-dependent features of CSI,and use the attention mechanism to assign weights to the features.This method uses a single-level LSTM network to significantly reduce the number of training parameters and achieve a compromise between performance and complexity.Simulation results show that this method can effectively reduce the feedback overhead of CSI,and is superior to the existing methods in terms of CSI compression and recovery accuracy.2.Aiming at the existing CNN-based CSI feedback method that less considers the correlation between channels and the noise of the feedback link,a CSI feedback reconstruction method based on channel correlation,denoising network and bi-directional long and short memory(Bi LSTM)network is proposed.This method uses known uplink CSI to help recover the unknown downlink CSI,that is,the bidirectional correlation between amplitude and absolute value between the uplink and downlink CSI is used to compress and decompress the downlink CSI.On this basis,the Bi LSTM network is used to extract the time-correlation features of the CSI of the previous antenna and the latter antenna,and the denoising network is used to train the proposed feedback network to enhance the anti-interference ability of the network.Simulation results show that,compared with the method using only downlink CSI,this method has better robustness,can significantly reduce the feedback load of CSI,and can accurately reconstruct CSI at the base station.