Research And Implementation Of Blind Source Separation Algorithm For Sound Source Localization

With the development of sound positioning technology and increasing application requirements,more and more fields need to locate different types of target sound sources.For many real application scenarios,in addition to sounds from target sound sources that need to be positioned,there also exist sounds from other sound sources.Since characteristics of these sounds are generally unknown,and their channel propagation models are complex and unknown either,it is difficult to achieve sound source localization using conventional signal processing methods.Researches have been conducted on the problem of sound signal processing when the source signal and the mixed channel model are unknown,i.e.,blind signal separation.Although sound source and sound propagation models are unknown,these sound signals are statistically independent,so the most commonly used independent component analysis is to solve the blind signal separation problem.This thesis combines independent component analysis with sound source localization to solve sound source location problems in complex environments.This thesis first introduces the basic theory and related work of independent component analysis,and then introduces two commonly used algorithms in detail.The first is a fast fixed point algorithm based on negative entropy maximization.It uses negative entropy as the source signal.The metrics for statistically independent properties are calculated using an approximate Newton iteration method.The second is the joint approximate diagonalization algorithm.It builds a joint matrix through the cumulative tensor of the signal,and then uses the approximate diagonalization method for the mixed signal.Then these two commonly used algorithms are simulated to verify the effectiveness and performance difference of the algorithms in dealing with blind signal separation problems.After the separation signal is obtained,the target sound signal that needs to be located is also separated from background sounds.The problem of sound signal recognition is studied in this thesis.First,the commonly used frequency domain feature parameters are introduced in the sound recognition problem,and the second part is the introduction of commonly used sound recognition methods.The recognition algorithm uses Mel cepstrum coefficient as the characteristic parameter of the sound signal.The dynamic time warping is used as the recognition method to identify the target sound that needs to be located.Finally,the entire system is tested and analyzed.An array of microphones are used to collect the ambient sound signal.Then the mixed sound signal passes separation process by FastICA.The separated signal is extracted by taking the Mel cepstrum coefficient as the characteristic parameter,and the dynamic time warping is used for the recognition.Then find the target sound and calculate the position information through TDOA algorithm.The test results show that the system can achieve the target sound location under complex mixed sound conditions.