Research On Deep Learning-based Underwater Acoustic Source Separation Technology

Source separation technology can not only separate multiple target signals,but also reduce the noise interference to the target signal,which improve the receiving accuracy and reliability of the passive sonar system.Therefore,it has great practical significance to perform research on underwater acoustic signal separation.In this paper,we firstly look into the research process of traditional underwater acoustic separation and the traditional underwater separation system models are verified by simulation,so as to choose the appropriate system model for the application of deep learning.The performance analysis and comparison of Fast Independent Component Analysis(FastICA)algorithm and Joint Approximate Diagonalization of Eigenmatrices(JADE)algorithm based on positive and overdetermined models are made by using ship radiated noise simulation signals.The simulation results show that the performance of the two algorithms are similar,and the similarity coefficients can reach 0.94 and 0.95 respectively in the case of no noise.Binary time-frequency mask method is analysed for underdetermined model using linear frequency modulation(LFM)and achieve the similarity coefficient of 0.9 in the case of no noise.Since the source numbers estimation is the basement of the source separation,three kinds of source number estimation: Akaike Information Criterion(AIC),Minimum Description Length(MDL),Gerschgorin Disk Estimation(GDE)are compared with simulation.The simulation result shows that under high Signal-to-Noise Ratio(SNR)condition,MDL and GDE can achieve consistent estimation of the number of sources and the error probability is almost 0 while AIC has the overestimation problem.Based on the analysis and comparison of simulation results,the binary time-frequency mask method is chosen as the system model in this paper.Since the binary time-frequency mask method depends strongly on the clustering of the feature vectors,deep neural network learns from samples according to the energy dominance so as to embed each time-frequency point into a high-dimensional space and generate data-driven features for clustering.Recurrent Neural Network(RNN),Long Short Term Memory(LSTM)and Bidirectional Long Short Term Memory(Bi-LSTM)are selected successively as the deep model,and three kinds of actual data are used for training.Experimental results show that using Bi-LSTM can efficiently separate mixed signals composed of different types of underwater acoustic signals,and achieve a Similarity coefficient of 0.97.Furthermore,Compared with traditional time frequency mask method under various SNR conditions,although the performance is improved after the application of deep learning technology,the separation results are poor affected by the noise.However,experiment founds that increasing the cluster number can solve the influence of unknown noise.In order to further improve the recognition ability of the passive sonar,deep learning technology is also applied to feature extraction in the recognition process.In this paper,Deep Boltzmann Machine(DBM)is used to extract features from the raw time-frequency features of the underwater acoustic target signals,Three experiments are set up to verify the performance of the method:(1)By adjusting different structural parameters for comparison,DBM can achieve the recognition rate of 91% when the hidden layer structure is 200-50-20;(2)Compared with Multi-Layer Perceptron(MLP)and Deep Belief Network(DBN)in the same structure,DBM performs better;(3)Comparing the learned features with the artificial features including line spectrum,wave structure and Mel-frequency cepstral coefficient under the same classifier,experimental results show that the features extracted by DBM can achieve the better recognition rate.