Low-Frequency Acoustic Source Detection and Localization

Abstract

In aviation, clear-air turbulence (CAT) is a major cause of in-flight injuries. It occurs in cloudless skies and cannot be detected by the onboard weather radar. Studies have predicted the extent of CAT to increase substantially in the next few decades, thus necessitating a method for detecting CAT.

With CAT known to generate low-frequency and infrasonic acoustic emissions, acoustic-based methods can potentially be deployed for detection and localization. This thesis studies low-frequency acoustic source detection and localization in the context of CAT.

Localizing low-frequency acoustic sources is challenging for acoustic beamforming which suffers from poor resolution at low source frequencies. A deep learning-based method is adopted as an alternative. Deep learning models for two-dimensional and three-dimensional acoustic source localization (ASL) have been built using synthetic data and computationally inexpensive neural network architectures. These models are necessary to prove the viability of deep learning for low-frequency ASL. The thesis then explores the potential of a deep learning-enabled, acoustics-based method for CAT detection in the future through a large-scale, virtual flight case, set up for the detection of a representative CAT source. The flight case tries to predict what an in-flight microphone will detect around a CAT source through a technique known as auralization which simulates the acoustic field of a source by modeling the sound propagation and determining what a receiver would hear.

The deep learning models yield promising qualitative and quantitative results that prove the feasibility of using deep learning for low-frequency ASL. Combined with the results from auralization, it can be concluded that there exists considerable scope for deep learning-enabled, acoustics-based detection and localization of CAT. The future work involves expanding the current scale of research with deeper network architectures to process real, in-flight acoustic data.