在深海环境中使用光谱变换和卷积神经网络进行源深度估计。

Source depth estimation using spectral transformations and convolutional neural network in a deep-sea environment.

作者信息

Wang Wenbo, Wang Zhen, Su Lin, Hu Tao, Ren Qunyan, Gerstoft Peter, Ma Li

机构信息

Key Laboratory of Underwater Acoustic Environment, Institute of Acoustics, Chinese Academy of Sciences, Beijing 100190, China.

NoiseLab, Scripps Institution of Oceanography, University of California San Diego, La Jolla, California 92093-0238, USA.

出版信息

J Acoust Soc Am. 2020 Dec;148(6):3633. doi: 10.1121/10.0002911.

DOI:10.1121/10.0002911

PMID:33379929

Abstract

Multiple approaches for depth estimation in deep-ocean environments are discussed. First, a multispectral transformation for depth estimation (MSTDE) method based on the low-spatial-frequency interference in a constant sound speed is derived to estimate the source depth directly. To overcome the limitation of real sound-speed profiles and source bandwidths on the accuracy of MSTDE, a method based on a convolution neural network (CNN) and conventional beamforming (CBF) preprocessing is proposed. Further, transfer learning is adapted to tackle the effect of noise on the estimation result. At-sea data are used to test the performance of these methods, and results suggest that (1) the MSTDE can estimate the depth; however, the error increases with distance; (2) MSTDE error can be moderately compensated through a calculated factor; (3) the performance of deep-learning approach using CBF preprocessing is much better than those of MSTDE and traditional CNN.

摘要

讨论了深海环境中深度估计的多种方法。首先，推导了一种基于恒定声速下低空间频率干涉的深度估计多光谱变换（MSTDE）方法，以直接估计源深度。为了克服实际声速剖面和源带宽对MSTDE精度的限制，提出了一种基于卷积神经网络（CNN）和传统波束形成（CBF）预处理的方法。此外，采用迁移学习来解决噪声对估计结果的影响。利用海上数据测试了这些方法的性能，结果表明：（1）MSTDE可以估计深度；然而，误差随距离增加；（2）MSTDE误差可以通过计算因子进行适度补偿；（3）使用CBF预处理的深度学习方法的性能远优于MSTDE和传统CNN。