Bayesian machine learning framework for time-domain prediction of multirotor vehicle noise

This work presents a Bayesian machine learning framework developed to predict aeroacoustic time-series signals generated by a quadrotor vehicle in forward flight at varying velocities. In this effort, a Gaussian process (GP) regression model is trained using a database of simulated signals produced by the Comprehensive Multi-rotor Noise Assessment framework. Unlike traditional frequency-domain models, the GP model directly predicts the time-domain signal, inherently capturing both amplitude and phase information of relevant frequency components. This capability is achieved by partitioning the tonal and broadband components during pre-processing, and capturing each component via a blade passage frequency-informed Fourier kernel and a Gaussian likelihood model, respectively. The resulting model is probabilistic in nature, inherently capturing the associated prediction uncertainty. Quantitative evaluations demonstrate strong agreement with ground truth signals in both time and frequency domains, with mean loudness errors of 1.11% in decibels and 5.55% in sones. The mean psychoacoustic annoyance error is found to be approximately 10%. The model is also computationally efficient compared to traditional physics-based solvers, requiring 0.1803 s to generate a time-series signal sampled at 44 100 Hz on a single NVIDIA A100 GPU (NVIDIA, Santa Clara, CA).