Research On Blind Source Separation Of Underwater Acoustic Signal

Due to the temporal and spatial variability of underwater channel along withthe volatility of ambient noise sources, the signals picked up by hydrophone arraysuffer severe distortions and the target information is masked by a great number ofinterferences. Blind source separation can retrieve the latent sources with little priorinformation. Using this technique to reduce noise and extract useful information,thereby realizing the detection, location and recognition of passive sonar signal,hence the accuracy and reliability of acoustic systems and devices can be improved.As a result, it has great practical significance to perform research on blind sourceseparation of underwater acoustic signal.This dissertation is rooted in the important applications of blind sourceseparation in underwater acoustic field, and the research is carried out around theblind source separation approach of underwater acoustic signal. Firstly, shipradiated noise as the source of passive sonar system is introduced. A mathematicalmodel is established according to its characteristics and a computer simulation isdetailed. Simulation results can truly reflect the typical acoustic signatures of shipradiated noise, and provide the data for the follow-up research work.Then, blind source separation algorithms which are based on independencecriterion are discussed. We start with the problem of estimating the number ofunknown sources. To reduce overestimation probability, the improved AIC isintroduced and tested with simulation. Subsequently, a composite algorithm whichcombines the advantages of joint approximate diagonalisation of eigen-matricesalgorithm and fast independent component analysis algorithm is proposed. Threealgorithms are compared and analyzed with similarity coefficient and performanceindex as the evaluation indicators. The result shows that the composite algorithmhas better separation performance in presence of noise or not.Finally, non-negative matrix factorization is applied to blind separation ofunderwater acoustic signal. Since NMF algorithms are sensitive to the initialization,least square is considered to improve the algorithms. The results of random,principal component analysis and least square initialization procedures on NMFalgorithms demonstrate the superiority of least square initialization technique aswell as the feasibility and validity of NMF in separating underwater acoustic signal.