基于多层感知机与无网格策略的三维空间声源识别

3D Sound Source Identification Based on Multi-Layer Perceptron and Grid-Free Strategy

摘要:

以往波束形成算法将潜在声源区域划分成若干网格，所有的声源被映射到一个个网格点上，会导致错误的声源定位与强度计算，并且计算精度与效率受网格间距大小的影响。采用多层感知机神经网络以及无网格策略，能够提高声源识别的空间分辨率与计算效率。通过使用单个平面麦克风阵列对三维等强度双点声源进行识别定位，发现相较于传统互谱算法，多层感知机能够改善平面阵列在深度方向上较差的空间分辨率性能。此外，在定位误差方面，多层感知机优于传统互谱算法，同时声源识别的强度误差有所降低。在低频时，多层感知机性能优于波束形成算法，可用来弥补波束形成算法空间分辨率性能不佳的局限性。

Abstract:

In the past， the potential sound source area was divided into mounts of grids according to the beamforming algorithm， and all the sound sources were mapped into the grid points， which would lead to an incorrect sound source localization and intensity calculation， and the calculation accuracy and efficiency were affected by the size of grid spacing. In this paper， the multi-layer perceptron neural network and grid-free strategy are used to improve the spatial resolution and computational efficiency of sound source identification. Compared with the conventional cross-spectrum algorithm， with the algorithm of multi-layer perception， the spatial resolution can be improved in the depth direction as a planar array was applied to identify and localize two- point sound sources with the same intensity. In addition， multi-layer perceptron is superior to the conventional cross-spectrum algorithm in positioning error. Meanwhile， the intensity error of sound source identification is reduced. Moreover， multi-layer perceptron is superior to the beamforming algorithm at a low frequency range， which can be used as compensation for the poor spatial resolution of beamforming algorithm at this range.