Machine Learning–Based Detection of External Gunshot Sound Sources in Propeller Acoustic Near Field

Abstract:
This paper presents a study of gunshot acoustic signal detectability in the near field of propeller noise, with a focus on the isolation of external gunshot signatures masked by propeller-induced noise. Controlled measurements were conducted in a Recirculation Delayed Anechoic Chamber (RDAC), where acoustic data were collected across varying rotor speeds, source locations, and propagation distances. Propeller noise characteristics were verified using UCD-QuietFly. The recorded signals were analyzed for the acoustic pressure, sound pressure level, and overall sound pressure level directivity to quantify masking effects. Results show that RPM is the dominant factor governing signal detectability. At 3000 RPM, the gunshot signal remains clearly identifiable within the low frequency range of 200–2000 Hz. At 4000 RPM, the signal becomes partially masked, while at 5000 RPM, propeller noise fully dominates and the gunshot signal becomes undetectable. Detectability is further reduced with increasing propagation distance. In-plane microphone locations provide improved detectability. A machine learning-based spectral separation framework was developed to suppress propeller noise and enhance the visibility of impulsive gunshot signatures in multichannel spectrograms. Experimental results show that learning-based denoising is effective at lower RPMs where the signal-to-noise ratio remains favorable, but performance degrades as broadband masking intensifies at higher rotor speeds.

Did you attend Forum 82? Click the preview below to access the full paper.