Journal:Applied Acoustics
Key Words:Deep learning; Fine-tuning; Scattering; Rotating beamforming
Abstract:Acoustic imaging of low-altitude aircraft in complex environments such as cabins, buildings, and mountains typically relies on beamforming, in which Green’s functions are tailored to account for scattering, but at a high computational cost. To efficiently predict tailored Green’s functions, a U-shaped convolutional neural network is designed to take physical parameters as input: two images of the real and imaginary parts of the source information, six images of the scatterer in orthographic projection, and two images of the real and imaginary parts of the incident wave field at the imaging plane. To enhance generalizability, the pre-trained network can be fine-tuned using some numerical data from out-of-training configurations. The developed approach reduces the time for scattering prediction from minutes to seconds and demonstrates accuracy on various scatterer shapes, both inside and outside the training datasets. Finally, the deep-learning-tailored Green’s functions are applied to rotating beamforming in both numerical and experimental settings, yielding results comparable to those obtained with numerical Green’s functions, along with an apparent reduction in computational time. This work highlights the potential of deep learning to accelerate beamforming in acoustically challenging environments.
Co-author:Jiahua He, Wangqiao Chen, Shengwu Chen
First Author:Jiaming Zhou
Indexed by:Journal paper
Correspondence Author:Lican Wang
Discipline:Engineering
First-Level Discipline:Aerospace Science and Technology
Document Type:J
Volume:253
Issue:111385
Page Number:1-10
Translation or Not:no
Date of Publication:2026-05-23
Included Journals:SCI、EI
Links to published journals:https://www.sciencedirect.com/science/article/pii/S0003682X26001660#s0065
Click:
The Last Update Time:..
