A multi-task learning convolutional neural network for source localization in deep ocean

• DOI: 10.1121/10.0001762
• Publication Date: 2020-08-20T12:57:26Z
• Authors: Yining Liu; Haiqiang Niu; Zhenglin Li
• Keywords: 0103 physical sciences; 14. Life underwater; 01 natural sciences

A multi-task learning (MTL) method with adaptively weighted losses applied to a convolutional neural network (CNN) is proposed estimate the range and depth of an acoustic source in deep ocean. The input normalized sample covariance matrices broadband data received by vertical line array. To handle environmental uncertainty, both training validation are generated propagation model based on multiple possible sets parameters. sensitivity analysis investigated examine effect mismatched parameters localization performance South China Sea environment. Among parameters, array tilt found be most important factor localization. Simulation results demonstrate that, compared conventional matched field processing (MFP), CNN MTL performs better more robust deep-ocean Tests real from also validate method. In specific ranges where MFP fails, reliably estimates depths underwater source.