Study On Blind Source Separation Of Underwater Acoustic Signal

In order to improve the detection performance of the passive sonar,the blind source separation(BSS)problem of underwater acoustic signals are studied based on the theory and application of BSS.By combining with the existing classic BSS algorithms,several BSS algorithms,which achieve separation for the underwater acoustic signals in the time domain and frequency domain,are proposed in this thesis.Moreover,the feasibility and effectiveness of these algorithms are validated through a variety of simulation data and real data.The main contents and innovations are described as follows:1.Several linear instantaneous mixture model-based classic blind separation algorithms are studied.First of all,three kinds of mathematical model of blind source separation are introduced: linear instantaneous mixture model,linear convolutive mixture model and non-linear mixture model,in which the linear instantaneous BSS problem is the foundation of the study.Secondly,the basic assumptions and the inherent uncertainties in the BSS problem are discussed.Then,several classic blind separation algorithms are introduced in detail,and the evaluation criteria of the separation performance is given.Finally,the separation performance of the underwater acoustic signals are tested under different conditions(e.g.the signal-to-noise ratio,the signal-to-interference ratio,etc.)by computer simulation,which laid a foundation for the study of linear convolutive BSS problems.2.A separation method for the underwater acoustic signal,which combing the second-order statistics and beamforming,is proposed.First of all,the mathematical model of the narrowband underwater acoustic array is introduced,which indicating that the underwater acoustic signal satisfies the convolutive mixture model in BSS problem.Then,the general ideas,the advantages and disadvantages of which achieve separation in the time domain and frequency domain are discussed respectively.After that,based on the non-stationary characteristics of the underwater acoustic signals,a separation algorithm is proposed by combing the second-order statistics and beamforming.The proposed method could not only improve the separation performance,but also greatly reduce the number of iterations.Moreover,the direction-of-arrival(DOA)information which obtained by multi-target beamforming,are combined with the correlation coefficient of separated sub-signals,to sorting the recoverd signals in different frequency point,and thus the permutation problem in the frequency domain BSS can be solved.Finally,the separation performance under different signal-to-noise ratio(SNR)and the ability to extract the weak target were verified by computer simulation.3.An underdetermined blind separation method for the underwater acoustic signal is proposed.When the number of sensors is less than the number of source signals,which is the underdetermined case,the blind source separation algorithms which are based on the independent component analysis are failed.In general,the underdetermined BSS algorithms are based on the sparseness of the source in time-frequency(T-F)domain,and achieve separation by clustering the characteristics belonging to the same signal.According to the different clustering methods,there are two kinds of algorithm to recover the source signals,which are called binary time-frequency masking and probabilistic(soft)time-frequency masking,respectively.First of all,the existing underdetermined blind source separation algorithms are studied.Then,when the initial value is far away from the true value,the soft T-F masking method which based on the expectation-maximazation(EM)algorithm would cause the local minimum problem.To solve this problem,the simulated anneling(SA)algorithm which has the global optimization capability is introduced,and a modified BSS algorithm which based on the annealing expectation-maximization(A-EM)method is proposed.Lastly,the effectiveness and robustness of the algorithm is verified by the computer simulation.4.A blind separation algorithm for the underwater acoustic signal based on a single vector hydrophone is proposed.Different from the traditional pressure hydrophone,the vector hydrophone could collect the sound pressure and the prticle velocity information simultaneously,and thus has the exellent direction-of-arrival(DOA)estimation capability.Firstly,the DOA estimation methods which based on a single vector hydorphone are introduced,on this basis,an improved multi-target DOA estimation method is proposed.Then,since the separation performance of the BSS method which based on the DOA information decreases rapidly when the sources are close to each other,an improved separation method which combine the DOA information and the mixing vector information together is proposed,and used for the recovery of the underwater acoustic signal.Lastly,the effectiveness of the method is verified by the computer simulation,and it is proved that the proposed method is significantly better than the baseline method when the source signals are close to each other.5.The methods of estimating the number of source signals under overdetermined,determined,underdetermined situations and which based on a single acoustic vector hydrophone are studied.In practice,the number of source signals are not known in advance,thus,how to estimate the number of sources accurately from the mixed signals is an important prerequisite to ensure the separation performance of the BSS algorithms.In this thesis,a variety of the methods to estimate the number of sources under different situations are studied,and the correct estimation probability of each method is discussed through computer simulation under different signal-to-noise ratio(SNR),number of snapshots and the number of source signals.6.A series of acoustic pool experiments are desinged to verify the performance of the BSS algorithms of the underwater acoustic signals proposed in this dissertation,and the impact of several parameters,such as: the number of sensors,the length of FFT,the length of separated filter,etc.In the mean time,the performance of the source number estimation methods are verified in practice by using the recorded data directly.