Doppler Spread Estimation for 5G NR with Supervised Learning

We present a Doppler spread estimation algorithm based on supervised learning, which is compatible with 5G new radio (NR) specifications of 3rd Generation Partnership Project (3GPP). Compared to the conventional algorithm using a predefined function to characterize the relationship between Doppler spread and the channel correlation, the proposed algorithm learns such relationship from the training samples collected in practice. Specifically, a multi-layer perceptron (MLP) neural network is trained as the function approximator to estimate Doppler spread. We design the input features to the MLP, and propose a scheme for training data collection based on the performance metric, i.e., block error rate (BLER), and our domain knowledge of the system. Simulation results show the improved BLER performance, e.g., l.ldB gain at 10% BLER, is obtained in a scenario of high Doppler spread at frequency range 2 (FR2).