Predicting AAM Path Loss through Neural Networks and Statistical Modeling

Advanced Air Mobility (AAM) aircraft are expected to travel urban and suburban routes at an altitude not to exceed 3,000 feet above ground level (AGL). They will be required to transmit and receive data with ground systems managed by the Provider of Support for UAMs (PSUs) or by Supplemental Data Service Providers (SDSPs) and, in some cases, with FAA systems. The air-ground data link must be secure, reliable, and available. In this study, we focus on signal path loss computation as airborne UAM flights traverse their flight path. The strength of these signals (or, equivalently, the amount of signal path loss) is a weak function of atmospheric conditions (humidity, temperature, and atmospheric pressure) for currently planned aviation bands, but a strong function of multipath interference effects and Doppler shifts. We describe initial results from empirical statistical models, with which we plan to combine our machine-learning models. We train a Convolutional Neural Network (CNN) using physics-based ray tracing computations, after which the CNN can predict the path loss. The results show that, while the physics-based system requires two to twelve minutes to compute the path loss from a transmitter, after training, the CNN can predict the path loss in 1.045 seconds, a 100-fold improvement in performance.