Compression of Channel Coefficients with Neural Networks for NR and LTE

We employ a resource block (RB1)-based compression method using neural networks for the channel coefficients under the specification of the third generation partnership project (3GPP). An autoencoder is trained to compress/decompress the channel coefficients, and the same compressor/decompressor is used for all RBs. Because the proposed method is RB-based, it universally applies to various combinations of resource configurations allowed by 3GPP. It is essential to compress the channel coefficients because they are stored in a buffer that takes a large memory when a large bandwidth is allocated. The buffer provides the channel coefficients to different blocks of the baseband modem. We investigate the compression considering two formats for the complex channel coefficients, Cartesian and polar. For each case, we train and test a separate autoencoder to maximize the compression performance. We reduce the buffer size by about 3 times without losing the performance by more than 0.1 dB, and the proposed algorithm is reliably applicable to both new radio (NR) and long-term evolution (LTE).