Distributed Sound Source Localization Methods Using a Coarse Grid–Based Convolutional Neural Network

Abstract

In the field of aerodynamic noise, quick and correct source localization is of interest. The recent developing machine learning–based source localization methods are known for high-resolution and high calculation speed. However, the assumption of several point sources hinders machine learning methods in practical aeroacoustics experiments, in which a large number of sound sources are commonly distributed. In this paper, a coarse grid–based convolutional neural network (CG-CNN) method is proposed to predict the source strength at any position within the region of the scanning grid in a grid-independent way, which is effective for handling distributed sources in a fine grid. Instead of learning and predicting samples of point sources in a grid-free way or on the same fine grid in a grid-based way, the proposed method trains the model with point sources on a coarse grid and predicts the strength of these sources at any position. In the training process, the convolutional neural network model with a source cross-power matrix as inputs learns samples of sources on a coarse grid with low computational cost. In the predicting process, given that the source could be located at any position on the translated coarse grid, the CG-CNN method predicts the source strength on a translated grid point by directly changing the input of model according to the location of the grid. Simulation results prove that the method localizes point sources and line sources better than traditional beamforming methods in terms of accuracy and dynamic ranges. The CG-CNN method was applied in a wind tunnel experiment with a high-lift device with a serrated slat, and the distinct locations of sources are identified correctly. In general, the proposed method has high efficiency in learning from sources on a coarse grid and predicting sources at any position, which is helpful for distributed sources in aerodynamic noise investigations.