New channel autoencoder and deep-learning based channel state information feedback schemes for massive MIMO systems

Abstract

In B5G communications, machine learning is envisioned to be a key enabling technology in developing the communication systems of the future. A major field in machine learning for communications is channel autoencoders and deep-learning-based channel state information feedback schemes which are a modification of channel autoencoders. Channel autoencoders and deep-learning-based channel state information feedback schemes strive to optimize a part of the communication system end-to-end by using artificial neural networks. It is postulated that the non-linearity of the neural networks would enable these machine learning for communication schemes to find new optimal points that no other conventional algorithms were able to achieve, at a lower complexity. In this thesis, we propose two new channel autoencoder structures as well as a new deep-learning based channel state information feedback structure. ConvAE is a channel autoencoder structure and uses residual blocks with convolutional layers. Residual blocks are made of two convolutional layers, two batch normalization layers, and a residual connection. Residual blocks solve the vanishing gradient problem that other conventional convolutional-layer-based networks suffer. This configuration increases performance while decreasing computational complexity at run-time compared to conventional channel autoencoders. To verify the performance of ConvAE, simulations were done under a 2-by-2 Rayleigh fading MIMO channel. These simulations showed that ConvAE was able to surpass the BER and achievable rate performance of the conventional fully-connected-layer based channel autoencoder. ConvAE-Advanced is a channel autoencoder that achieves adaptive transmission across multiple timeslots, using more resources when the communication channel is favorable. ConvAE-Advanced utilizes an unexploited input dimension in ConvAE for adaptive transmission. This unexploited input dimension was utilized by adding a convolutional layer with 1-by-1 kernels to the residual block structure. To verify the performance of ConvAE-Advanced, simulations were done under a 2-by-2 WINNER2 MIMO channel. Simulations showed that ConvAE-Advanced was able to achieve a better BER and rate performance compared to ConvAE. Moreover, ConvAE-Advanced can attain a better BER performance without additional resource usage compared to ConvAE. ChannelAttention is a deep-learning-based channel state information feedback structure and uses residual blocks with self-attention layers. ChannelAttention utilized self-attention layers as well as a larger number of channels compared to the conventional deep-learning-based channel state information feedback scheme, CsiNet. This enabled ChannelAttention to have a larger receptive field size as well as a larger model capacity. The configuration of ChannelAttention improves normalized mean squared error and cosine similarity performance while increasing complexity. To verify the performance of ChannelAttention, simulations were done under a 64-by-64 QuaDRiGa channel model based on the 3GPP 38.901 urban microcell scenario. Simulations show that ChannelAttention surpasses the normalized mean square error and cosine similarity performance of the conventional CsiNet scheme across all compression ratio ranges.