2D/3D ResNet Deep Neural Network for 4G and 5G NR Wireless Channel Estimation

For estimating the MIMO channels for both the uplink and the downlink, we used residual deep neural network. The proposed residual deep neural network (ResNET) channel estimation outperforms the MMSE channel estimation by more than 2 dB under EPA uplink channels and is capable of operating on untrained wireless channels. According to the 3GPP NR release 15 standard, the proposed 3D ResNET channel estimation approach exhibits a 0.5 to 1 dB gap to the ideal channel estimation and a 1 to 1.6 dB gain over MMSE channel estimation for 5G New Radio (NR) physical layer links, including demodulation reference symbols (DM-RS Type 2) and synchronization signal (SS) bursts periodicity (30 ms), DL-SCH transport channel coding modulation and coding scheme 13 (MCS), code rate 490 over 1024 under 15 iterations of Normalize Min-sum 5G eMBB LDPC decoder with scale 0.75, HARQ off (no retransmissions), PDSCH precoding using SVD (PDSCH Mapping Type A, 4×4 antennas), 20 resource blocks (240 subcariers), clustered delay line (CDL-C) channel with subcarier spacing 30kHZ, delay spread 300e-9, doppler shift 5Hz, central frequency 3.5GHz demonstrates a 0.5 to 1 dB gap to the perfect channel estimate and a gain of between 1 and 1.6 dB over MMSE channel estimation. The suggested solution enables a 25% boost in 5G NR throughput.