Underwater Source Localization Based On Convolutional Neural Network And Transfer Learning

Passive localization of acoustic source in shallow water waveguide is an important issue in underwater acoustic detection.In this thesis,deep learning is introduced to source localization in underwater waveguide to formulate the problem of extracting location related infromation from measured data.Unlike the model-based matching field processing(MFP)algorithm,deep learning is a data-driven method.The combination of parameters are used to model the relationship between input data and labels.This thesis studies three problems:single source localization in shallow water,multiple sources localization in shallow water and localization in domain-shift scenarios.The first two are classical tasks in source localization,the last one is the key problem when deep learning localization algorithm is used in actual scene.In order to localize a single acousitc source in shallow water waveguide,classification model and multi-task regression model based on Convolutional Neural Network(CNN)are designed to jointly estimate the range and depth of the acoustic source.Simulation results show that CNN-c achieves higher localization accuracy than the traditional MFP method in low SNR scenes,and it also shows higher tolerance for system mismatch and environment mismatch.Result on experimental data also shows that CNN-c can achieve higher localization accuracy than MFP.In order to solve the problem of multiple sources localization,a multi-label learning(MLL)based algorithm is proposed in the scene with certain number of sources.By turning multi-class classifiers into a combination of binary classifiers,this algorithm successfully locates multiple acoustic sources.Our algorithm also shows ability to distinguish close sources.A post-processing algorithm based on neighborhood relative energy is designed to deal with arbitrary number of sources.When applied to sound source localization in real ocean environment,deep learning algorithms often face the problem of lack of data,but training network with simulation data may cause"domain shift".A localization algorithm based on pre-training fine-tuning is designed to deal with this problem.A small number of experimental data with labels are used to adjust the pre-training network parameters,so as to alleviate the performance loss caused by domain shift.An algorithm based on domain adaptive learning is designed for the senario where all experimental data are without labels.By aligning the the feature distribution of simulation data and experimental data,the localization accuracy on the experimental data is improved.