Comparison of deconvolution methods for the visualization of acoustic sources based on cross-spectral imaging function beamforming

摘要

DAMAS, DAMAS2, NNLS, Fourier-based NNLS. CLEAN and CLEAN-SC are typical deconvolution methods, which have been used in the visualization of acoustic sources based on beamforming to improve the spatial resolution and the dynamic range effectively, it is of great significance to demonstrate and compare properties of these methods comprehensively. In this paper, these methods are applied to cross-spectral imaging function (CSIF) beamforming with auto-spectra exclusion and their properties are demonstrated and compared with each other first by computational simulations consisting of a single source, two incoherent sources and two coherent sources. All the deconvolution methods can visualize single source or incoherent sources in the region where the assumption of shift invariant point spread function is valid accurately and clearly. Not only the spatial resolution is improved dramatically, but also the sidelobes are eliminated effectively. In addition, these methods rank in a diminishing sequence of sidelobe elimination ability from CLEAN-SC CLEAN, DAMAS, Fourier-based NNLS, NNLS to DAMAS2. When the sources are out of the valid region, only DAMAS, NNLS, CLEAN and CLEAN-SC succeed in visualizing the sources and CLEAN-SC and CLEAN acquire the cleanest source images, then DAMAS, finally NNLS, while DAMAS2 and Fourier-based NNLS fail to not only locate the sources but also capture the strengths. DAMAS, DAMAS2, NNLS and Fourier-based NNLS have good availability for coherent sources in the valid region. In contrast, CLEAN fails to remove sidelobes effectively and CLEAN-SC can only detect one source. DAMAS2 and Fourier-based NNLS also perform poorly for coherent sources out of the valid region. Additionally, DAMAS2 and Fourier-based NNLS consume a minimum of time to conduct a calculation, CLEAN and CLEAN-SC take the second place, whereas DAMAS and NNLS are the slowest Then a series of experiments are performed on small loudspeakers to validate simulations and compare robustness of these deconvolution methods in practical applications. Some practical factors such as the frequency response characteristic mismatch among the measurement devices have almost no influence on the results of CLEAN-SC, bring some change to the results of DAMAS, DAMAS2, NNLS and Fourier-based NNLS in terms of reconstructed maximum values, sidelobes, etc, and contribute plenty of extra sidelobe contaminations to the results of CLEAN. The conclusions play a guiding significance on the application of these deconvolution methods in practical engineering.