Improving the analysis of aeroacoustic measurements through machine learning
Doctoral thesis
Date of Examination:2023-11-06
Date of issue:2023-11-20
Advisor:Dr. Carsten Spehr
Referee:Prof. Dr. Steffen Herbold
Referee:Prof. Dr. Alexander Ecker
Files in this item
Name:2023_09_19_Improving_the_analysis_of_aeroaco...pdf
Size:18.1Mb
Format:PDF
Abstract
English
This thesis focuses on improving the analysis of aeroacoustic imaging methods using automated data processing and machine learning. Imaging methods result in beamforming maps that are challenging to explore manually since they comprise complex, high-dimensional data, including spatial coordinates, frequency, and flow properties. The manual, iterative analysis is time-consuming, biased, and typically based on 2D beamforming maps that only insufficiently capture the complex source distribution of airframe sources. Further, acoustic array imaging methods often assume monopole sources and, thus, suffer from mismatches between model assumptions and actual sources. This thesis addresses these issues by proposing novel broadband beamforming methods based on the observation that most airframe noise sources are spatially compact and parameters such as their location do not change over frequency. A broadband approach improves the ratio of known to unknowns in the mathematical formulation of the problem, which allows for the inclusion of advanced model assumptions, such as dipoles and distributed sources. The resulting novel methods are broadband Global Optimization, a gridless covariance matrix fitting method, Broadband-CLEAN-SC, an adaptation of CLEAN based on Source Coherence (CLEAN-SC), and gridless beamforming using artificial neuronal networks with a permutation invariant loss. All proposed methods share that a beamforming solution is obtained for multiple frequencies simultaneously. The broadband approach improves the resolution at low frequencies. It suppresses side- and grating lobes (aliasing), improves the identification and spectra extraction from the results, and outperforms the corresponding small-band methods. This thesis proposes two clustering methods that identify sources in the high-dimensional beamforming maps and extract their spectra to post-process the industrial gold standard conventional beamforming and CLEAN-SC methods. The proposed ``Source Identification based on Spatial Normal Distribution'' (SIND) method is a clustering algorithm similar to a Gaussian Mixture Model. It is tailored to the source identification problem, with spatial discretization, a large number of estimated sources, and statistical noise as its main challenges. To determine the number of clusters, SIND does not rely on a priori hyper-parameters but determines the unknown number of sources iteratively from the spatial distribution of the data. The proposed ``Source Identification based on Hierarchical Clustering'' (SIHC) method clusters the data directly in space and frequency using the established Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) algorithm. The introduced automatic source identification capabilities allow the precise source identification and spectra extraction of 3D beamforming maps with no added effort, compared to the standard manual process, typically based on planar 2D beamforming maps with insufficient spatial resolution. This thesis provides an overview and insight into the aeroacoustic theory and introduces numerical features that explicitly formulate physical properties, enabling the deduction of aeroacoustic source mechanisms. The proposed formulas are robust towards noisy and degenerate spectra typically resulting from deconvolution methods such as CLEAN-SC. They are independent of the measured object, the amount of measured Mach numbers, and the Mach numbers themselves. Thus, they offer a comparability of the properties across different measurements, which was previously impossible. Sources can be visualized utilizing their high-dimensional feature space through dimensional-reduction methods and effectively clustered with HDBSCAN, offering a manual classification and interpretation guideline. This process facilitates the creation of an Expert Decision Support System (EDSS). The clusters proposed by the EDSS strongly correlate with the manually determined categories so that the expert can interpret them. Clustering results from industrial wind tunnel experiments on a Dornier 728 and Airbus 320 models are presented. The clustering accuracy, determined from a confusion matrix and a manual selection of its correct entries, is 77.04% for the Dornier 728 and 61.52% for the Airbus 320 experiment. The thesis presents a detailed aeroacoustic analysis and manual classification of all occurring airframe sources. Novel aeroacoustic observations are described, such as that the flap side edge and strake are composed of two sources each with different mechanisms and that many sources depend on the Mach number weaker than the one of a true Strouhal number. Also, some sources, such as the strake and cavity noise, show a Mach number dependency, even at a constant Reynolds number, while most sources are self-similar within a large Reynolds number range. In summary, this thesis presents an improved workflow for beamforming, post-processing, interpretation, and knowledge generation from aeroacoustic experiments. The proposed EDSS enables a complete aeroacoustic analysis of wind tunnel experiments, offering detailed insights into the nature of the sources. Further, the EDSS has proven its capabilities to be employed in situ to detect and fix spurious noise sources during experiments, offering new perspectives and a practical tool for researchers and practitioners in the field.
Keywords: beamforming; acoustics; aeroacoustics; machine learning; unsupervised learning; clustering; expert system