Bayesian machine learning framework for time-domain prediction of multirotor vehicle noisea).

Area of Science:

• Acoustics
• Machine Learning
• Computational Fluid Dynamics

Background:

• Quadrotor vehicles generate complex aeroacoustic noise during forward flight.
• Predicting this noise is crucial for vehicle design and operational efficiency.
• Traditional methods often struggle with the time-domain complexity of aeroacoustic signals.

Purpose of the Study:

• To develop a novel Bayesian machine learning framework for predicting quadrotor aeroacoustic time-series signals.
• To directly model the time-domain signal, capturing both amplitude and phase information.
• To provide a computationally efficient alternative to traditional physics-based noise prediction models.

Main Methods:

• Utilized a Gaussian process (GP) regression model trained on simulated aeroacoustic data.
• Employed a Comprehensive Multi-rotor Noise Assessment framework for signal generation.
• Partitioned tonal and broadband noise components using a blade passage frequency-informed Fourier kernel and a Gaussian likelihood model.

Main Results:

• The GP model demonstrated strong agreement with ground truth signals in both time and frequency domains.
• Achieved low mean errors: 1.11% in loudness (dB), 5.55% in sones, and approximately 10% in psychoacoustic annoyance.
• The model is probabilistic, inherently quantifying prediction uncertainty.

Conclusions:

• The developed Bayesian GP framework offers an accurate and efficient method for predicting quadrotor aeroacoustic noise.
• This approach captures essential signal characteristics, including amplitude and phase, directly in the time domain.
• The model's computational efficiency makes it suitable for real-time applications and design optimization.