Deep-learning-based Bouncing-order Prediction for Propagation Channel Characterization Using Graph Modeling

Abstract

The propagation of electromagnetic waves can be traced along multiple paths where interactions between waves and objects occur resulting in bouncing effect. The number of bounces, called as bouncing-times, can significantly influence the fading characteristics, the delay and angular spreads of receiver signals. Therefore, it is meaningful and valuable to research on bouncing-times. Graph modeling is an efficient technique that can be applied to simulation of the channel characteristics. Bouncing-times is an important parameter for graph modeling of propagation characteristics with fidelity and accuracy. A proper setting of the bouncing-times can largely reduce the computational cost of graph modeling with no tradeoff of the performance. In this work, a novel method based on deep learning (DL) applied to prediction of bouncing-times accurately is proposed. We choose foliage scenario as sample environment and train a neural network that can predict bouncing-times directly from the power delay profile (PDP) of a propagation channel. Both experimental PDPs and measurement PDPs are used to validate the neural network being trained. The result shows that the neural network being trained can accurately predict the bouncing-times for foliage scenario. With appropriate modifications, the proposed DL method can be generalized to more scenarios of parametric channel characterization.