循环神经网络在浅海声速-声源联合反演中的应用  

Application of the sound speed profile and sound source location in shallow waters

作　　者：王同[1,2,3] 苏林 任群言[1,2] 王文博[1,2,3] 马力 WANG Tong;SU Lin;REN Qunyan;WANG Wenbo;MA Li(Key Laboratory of Underwater Acoustics Environment, Chinese Academy of Sciences, Beijing 100190, China;Institute of Acoustics, Chinese Academy of Sciences, Beijing 100190, China;University of Chinese Academy of Sciences, Beijing 100049, China)

机构地区：[1]中国科学院水声环境特性重点实验室,北京100190 [2]中国科学院声学研究所,北京100190 [3]中国科学院大学,北京100049

出　　处：《哈尔滨工程大学学报》2021年第8期1133-1139,共7页Journal of Harbin Engineering University

基　　金：国家自然科学基金项目(11704396).

摘　　要：利用序贯滤波对时变声速剖面进行反演追踪,通常会将声速剖面的时变特性描述为一阶随机游走过程。为了使状态方程更好的预测状态变量的走向,本文利用循环神经网络学习历史水文数据,对浅海环境下的时变声速剖面进行建模,利用集合卡尔曼滤波进行对声速剖面的反演,并对声源进行定位。结果相较于使用一阶随机游走过程的联合反演结果误差更小,在声源深度上均方根误差有着80%的降低,声速剖面反演结果误差有着38.2%的降低。本文通过实测声速剖面的仿真声学数据验证了该方法的可行性。The time-variant characteristics of the sound speed profile are usually described as a first-order random walk process when the sequential filtering approach is utilized for the time-varying sound speed profile inversion.To make state quotations better predict state variables,this paper uses the recurrent neural network to analyze historical hydrographic data and models for the time-varying sound speed profile in shallow water,thus accurately describing the time-variant characteristics of the sound speed profile.Based on this,Ensemble Kalman Filtering is used for the sound speed profile inversion and sound source location.The method improved results are obtained when compared with the joint inversion of the first-order random walk process.The root mean square error of sound source depth is reduced by 80%,and the error of sound speed profile inversion result is reduced by 38.2%.Feasibility of the method is verified by simulated acoustic data of the measured sound speed profile in this paper.

关 键 词：机器学习 循环神经网络 序贯滤波 集合卡尔曼滤波 声速剖面 声速剖面反演 水下目标 水声定位 

分 类 号：TB566[交通运输工程—水声工程]