Colloquia

Samenvatting

In this study, a deep learning architecture is proposed for acoustic field reconstruction. The approach utilizes a convolutional neural network that processes stereo beamforming images to generate a volumetric representation of an acoustic field. Through synthetic data, the architecture is trained to reconstruct fields with several monopoles or line sources from the synthetic and measured input.

The synthetic and experimental results show the potential for a computationally efficient complete reconstruction of an acoustic field in 3D. Arbitrary acoustic fields, composed of monopoles and line sources, can be reconstructed with low computational effort (\SI{0.482}{\gflop}). A localization error of less than \SI{50}{\mili\meter} can be expected. The reconstructed sound pressure level field shows acoustic patterns, noise source locations, and source intensity, which can be used for characterization, identification, or localization.  As a practical demonstration, the trained model was used for a UAV tracking scenario, where it achieved high-accuracy acoustic trajectory tracking.

These results suggest that deep learning-based methods offer a viable path toward real-time and high-resolution 3D acoustic field reconstruction. This will have a significant implication on the field of acoustics, providing potential for acoustic diagnostics and environmental noise monitoring. Furthermore, the ability to reconstruct novel arbitrary acoustic fields in three dimensions lays the foundation for future research into more dynamic and complex acoustic scenes.