Research On Sparse Bayesian Learning For Acoustic Source Localization

Underwater passive source localization is one of the hotpots in the field of underwater acoustics.Matched field processing(MFP)is one of the representative technologies of the underwater target localization.It makes great use of acoustic channel,array and narrow or bandwidth procesing methods,and matches the received array data to a dictionary of replica vectors to localize one or several sources.The performance of MFP depends on the accurateness of the physical model and the amount of the data available.One main cause of the inaccuracy of the physical model is the environmental mismatch which may seriously decrease the performance of the localization.For the amount of the data available,broadband methods can increase the amount of the available data to improve the performance.The broadband MPFs can be divided into coherent processors and incoherent ones,but since the phase of the acoustic field of the coherent processors cannot be accurately predicted,most broadband methods are incoherent causing that their localization performance is insufficient compared to the coherent ones.In this thesis,the environmental mismatch and the broadband signal processing is studied,and then two kinds of source localization algorithms based on sparse Bayesian learning are proposed to solve the problem of the environmental mismatch and the phase prediction of the broadband processors.The main research contents and results of this thesis are summarized as follows:(1)For the environmental mismatch,a sparse Bayesian learning based on the predictable normal modes is proposed after studying the principles of the normal modes.Because different normal modes are affected differently by the environmental mismatch in the shallow water waveguide,the normal modes are classified into predictable and unpredictable ones.The predictable modes retain more correlation in the presence of the mismatch,which means they contain more source information.Unpredictable normal modes contain fewer correlations.For example,higher-order normal modes are more sensitive to the boundary interference due to multiple boundary reflections and medium delays.The correlation of the normal modes in the exist of the environmental uncertainty can be evaluated by calculating the second-order statistics of the horizontal wave numbers of the normal modes.Firstly,the normal mode horizontal wave number is divided into three parts,such as the errors of the source amplitude and range and the errors causing distortion of the ambiguity/likelihood function shape.Because the first two will not cause serious attenuation of positioning performance,the latter is assumed to be caused by the unpredictable normal modes.Therefore,the predictable normal modes are gained by minimizing the trace of the covariance of the errors causing distortion of the ambiguity/likelihood function shape.With the reconstructed replica vectors based on the predictable normal modes,an improved sparse Bayesain learning is used to localized source.The performance of this algorithm is evaluated and analyzed through simulation experiments and sea trials.The results show that the improvement can increase the localizing accuracy of the sparse Bayesian learning to a certain extent.(2)For the broadband processing,a cross-frequency incoherent sparse Bayesian learning is proposed.The broadband methods can be divided into incoherent ones and coherent ones.Compared with the incoherent processors,the coherent processors make full use of the source information of the cross-frequency term and the suppression of noise to increase the accuracy of positioning.While these coherent estimators need the the spectrum of the source,the matched phase coherent processor that avoids this problem by adding the perturbation factor into the replica vectors needs traverse the phase of the perturbation factor,which causes its high calculation cost.In response to this problem,based on the original sparse Bayesian learning,this thesis uses the phase correction to avoid seeking the optimal phase,which means it can make full use of the cross-frequency term without high calculation cost to increase the accuracy of localization.The performance of this algorithm is evaluated and analyzed through simulation experiments and sea trials.It can be seen from the results that the proposed method significantly improves the sparse Bayesian learning and greatly suppresses the sidelobes.