is thesis addresses key challenges in acoustics applications like troubleshooting, product optimization, and sound source modeling, focusing on the localization and characterization of sound sources using Beamforming techniques and an improved regularization of inverse methods. Traditional Beamforming methods, while effective in identifying main radiating sources, are limited by the array size and density, which affects their accuracy and increases cost. To address the high computation time and incompatibility with real-time applications of existing deconvolution methods, this work introduces a beamforming pre-processing step employing neural networks. The approach is based on a grid compression of the beamforming map and can drastically accelerate the deconvolution process by 2 to 3 orders of magnitude, enabling real-time applications without losing accuracy.

Furthermore, the thesis proposes a solution to the Beamforming assumption of a monopole source, a major limitation especially in aeroacoustic applications where higher-order sources like dipoles and quadrupoles may be present. By expanding the DAMAS deconvolution method, this thesis offers a method capable of handling these higher-order sources efficiently in real-time, without relying on a sensitive (and often unreliable) hyperparameter tuning. This method also supports solving for sources based on multiple Beamforming results simultaneously, overcoming the limitations of current costly approaches.

Another significant contribution of this work is in estimating the radiation characteristics of sound sources, traditionally dependent on expensive setups of spherical arrays for a high-resolution characterization. The thesis introduces a method, HELS Flow, which utilizes empirical priors learned through normalizing flows parameterized by neural networks, allowing for a substantial reduction in the number of microphones needed in the array without significantly compromising the resolution. This work shows that, after relevant training, the method can outperform traditional regularization approaches even when their hyperparameters are optimally selected.

Melden Sie sich hier an um Einladungen zu den Kolloquium-Vorträgen per E-Mail zu erhalten.
Register here to receive the invitations to colloquium talks via e-mail.

Zoom-Meeting-ID: 954 4073 3814
Passwort: 450783