Bundle-based Channel Estimation for 5G NR with Supervised Learning

This paper explores to improve the quality of channel estimation when a bundling structure is imposed under 3GPP new radio (NR) specification. When bundling is configured, the precoding per bundle is determined solely by a gNB, which is transparent to a user equipment (UE). Due to the possibility that each bundle is precoded in a different manner, the power delay profile (PDP) needs to be estimated within a bundle in a frequency domain, not in a time domain. One simple approach is to assume that PDP is uniformly distributed up to the maximum delay spread. This method is acceptable when a channel is propagated within a short delay spread, but is not in general. Alternatively, this paper utilizes channel correlation even across bundles in a frequency domain and estimates PDP values by using all the correlation directly. A deep learning technique is applied with a neural network to generate correlation as if not to be precoded. To stabilize the performance, post-processing has been added to the neural network such as filtering, truncation and normalization. Simulation results show that the proposed method achieves the performance close to that of ideal PDP and outperforms the uniform PDP assumption by more than 2 dB.