Universally applicable data-driven approaches for microphone arrays in acoustic testing

• Applicant: Professor Dr.-Ing. Ennes Sarradj
• Subject Area: Acoustics
• Term: since 2020
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 439144410

Microphone array methods are an important tool for identifying noise sources. Established model-based methods for localizing and quantifying sound sources primarily rely on physical models, which can lead to unreliable results due to the necessary simplifications involved. Reduced spatial resolution at large wavelengths and significant computational requirements limit the applicability of model-based microphone array methods. A promising alternative are data-driven microphone array methods, where complex relationships between source characteristics and captured signals are learned from acoustic training data.The potential resulting from this approach has not been fully realized, as simulated data has predominantly been used for training, which impairs the reliability of results in challenging experimental applications. In simple cases and especially for large wavelengths, data-driven methods still yield faster and sometimes more accurate results than conventional methods. However, currently published data-driven methods have limitations regarding their applicability. Changes in the measurement setup and environment require extensive adaptation through additional training.The aim of this research proposal is to achieve universal applicability for the the data-driven microphone array methods developed by the research group of the applicant. The methods shall be applicable independently of the measurement setup and environment, and shall produce reliable characterization results. Approaches will be investigated that consider known properties of the measurement setup and the environment as additional context information during training and method application. Subsequently, research will be conducted on how trained models can be adapted using experimentally acquired data to improve characterization results in experimental situations, including those that were not part of the training. The goal is to make the methods not only universally applicable but also sufficiently robust against interference. Finally, research will also be conducted to determine the extent to which already trained models can be adapted and improved based on a single measurement. To assess the desired generalization of the models, an experiment will also be conducted. The acquired experimental data serves as a suitable benchmark and will be openly accessible to other researchers.

• DFG Programme: Research Grants