Aeroakustische Quelllokalisation und Klassifikation mit künstlichen neuronalen Netzwerken / Aeroacoustic source localisation and classification with artificial neural networks

Kurzfassung

In the field of aeroacoustics, insufficient computational power has for long been limiting precise descriptions of the sonic beam formation, due to computationally expensive beam forming algorithms [15]. A new dawn is rising with the increasing popularity to replace analytical computation algorithms with trained neuronal networks. These are typically not only much faster but also able to model properties of systems which are physically unknown for now since they are hidden in the empirically obtained data sets. With more and more upcoming architectures, the range of possible applications is intensively increasing and they have turned out to be useful in interdisciplinary fields of science. In this work, for the first time, a neuronal architecture is specially designed to face the problem of identifying, classifying and locating sonic emitters, using the recognition of patterns within a measurement object called the cross spectrum matrix (CSM). The CSM is an easily obtained experimental quantity representing the cross correlation of all measured input pressures and incorporates information from the system in its spectral space. In order to understand the generation of the CSM, one has to consider co-dependencies between every emitter and receiver as well as the environment in with the sound is propagating, making it impossible to describe the system analytically only. At this point, an essential advantage of neuronal networks becomes apparent because it is minimally dependent on assumptions of the system and is driven purely based on data. Since neuronal networks are based on statistical operations, which have to be performed on a huge population of sampled data, the quality of the results depends significantly on the training data of the neural network. Therefore, the first big milestone of this work is the creation of versatile training data, allowing for the network to recognize underlying patterns in CSM data sets. To interpret the quality of the output from the neuronal network, it is compared to reliable analytical solutions within the possible approximations of conventional frequency domain beamforming. As a result of the study we find that, with the appropriate neural network architecture, the precision of localizing, validating and characterizing surpasses the expectations based on analytical evaluations. In particular, the optimal choice of model parameters turns out to be crucial. The derivation presented within this thesis, thus, provide a frame to properly set up the neuronal network. The thesis will be structured as follows. At the beginning, we will give an introduction to the most important theoretical concepts used or related to this work, explaining the framework of aeroacoustics, their experimental realization as well as neural networks. We will then proceed with a chapter describing the methods used herein with detailed descriptions of the data sets used. Afterwards, we will present the evaluation and the corresponding results, introducing several approaches for the neural network and summarizing the key findings of this study which is finally compared to real data.