Deep Learning Approaches for Microphone Arrays in Acoustic Testing

麦克风阵列方法是声源定位和表征的既定方法。使用合适的信号处理，不同声源的贡献可以在空间上映射。然而，低动态范围使得通常难以将映射的源贡献分配给其原因并识别源机制。出于这个原因，已经开发了许多不同的方法，这些方法在精度、计算工作量和抗干扰的鲁棒性方面差异很大。根据测量任务的不同，并不总是能获得令人满意的结果。该项目的主要目标是建立深度神经网络（DNN），用于麦克风阵列声源的定量表征。要找到一种替代已经建立的基于模型的方法，它提供了精确的结果，相对较少的计算工作。该项目的工作旨在显著增加DNN在这些和类似声学测量任务中的适当使用知识。首先，要开发一种能够以可再现的方式生成大量合成测量数据的方法。在此基础上，可以预见的方法，估计声源的位置和强度没有一个给定的网格的发展。此外，一个神经网络是要训练的声源的特性的逆问题的解决方案。最后，而不是离散的，频率上的描述的贡献，个别来源，它的目的是找到一种方法，估计所需的参数描述的功率谱的数据。还应根据实验获得的数据对所有开发的方法进行评价。

Microphone array methods are an established approach for the localization and characterization of acoustic sources. Using suitable signal processing, the contributions of different sound sources can be mapped spatially. However, a low dynamic range makes it often difficult to assign the mapped source contributions to their causes and to identify the source mechanisms. For this reason, a number of different methods have been developed, which differ greatly in terms of accuracy, computational effort and robustness against interference. Depending on the measurement task, satisfactory results are not always achieved. Consequently, the development of improved methods is desirable.The main goal of the project is to establish deep neural networks (DNN) for the quantitative characterization of sound sources with microphone arrays. An alternative to already established model-based methods is to be found, which provides precise results with comparatively little computational effort. The work in the project aims at a significant increase in knowledge on the appropriate use of DNN for these and similar acoustic measurement tasks. First, a method is to be developed that can generate large quantities of synthetic measurement data in a reproducible manner. Based on this, the development of a method is foreseen that estimates the location and strength of sound sources without a given grid. In addition, a neural network is to be trained for the solution of an inverse problem for the characterization of sound sources. Finally, instead of the discrete, frequency-wise description of the contribution of individual sources, it is intended to find a method which estimates the data necessary for a parametric description of the power spectrum. All developed methods should also be evaluated on the basis of experimentally obtained data.