Research On Separation Methods Of Complex Sound Sources

With the noise pollution becoming more and more serious nowadays,noise control work is particularly important.Sound source identification and sound source separation in complex environments are the main premise of noise control work.This thesis is based on the large underwater structures,and conducts theoretical research,simulation analysis and experimental verification for the identification and separation methods of complex sound sources.For the separation of sound sources in complex environments,the basic premise is to be able to effectively identify the sound source.Aiming at the problems of low computational efficiency,low resolution and inability to identify coherent sound sources in conventional beamforming and DAMAS,this thesis has researched a deconvolutional sound source imaging algorithm which can be used to identify coherent sound sources on the basis of DAMAS2.By removing the cross-spectral process in the DAMAS algorithm,the algorithm no longer needs to make assumptions about the number of non-coherent sound sources.Therefore,the algorithm can identify the coherent sound source effectively after removing the cross-spectrum process.In addition,the denoising process of principal component analysis is added to the algorithm,which makes the algorithm have the same noise robustness as the traditional algorithm.The optimized algorithm not only has a significant improvement over the computational efficiency of traditional algorithms,but also has an advantage for the identification of coherent sound sources.In order to solve the problem that the FastICA algorithm has the uncertainty of independent component ordering in the field of sound source separation,as well as to solve the problems of sensitive initial value,slow descending speed and complex calculation of Newton iteration method in the FastICA algorithm,this thesis studies a sound source separation optimization algorithm combining K-means clustering algorithm with the FastICA algorithm with damping factor,and elaborates on the third-order convergence and the fifth-order convergence in Newton's iterative method respectively.K-means clustering algorithm can be introduced to determine the ordering of independent components in the field of sound source separation.The FastICA algorithm with the third or fifth order convergence of damping factor can reduce the sensitivity of the initial value of the Newton iteration method,reduce the number of iterations and accelerate the convergence speed.The optimized algorithm can guarantee the separation effect and stability of the original algorithm.Finally,this thesis verifies the feasibility and effectiveness of two optimization algorithms in the separation of sound sources by simulation analysis and experiments.