基于多域特征提取和深度学习的声源被动测距  

Source ranging based on deep learning and multi-domain feature extraction:synthetic results

作　　者：肖旭 王同[1,2,3] 王文博 苏林[1,2] 马力 任群言[1,2,3] XIAO Xu;WANG Tong;WANG Wenbo;SU Lin;MA Li;REN Qunyan(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：《应用声学》2021年第1期121-130,共10页Journal of Applied Acoustics

基　　金：国家自然科学基金项目(11704396)。

摘　　要：采用一种基于多域特征提取的深度学习方法来实现声源被动测距。首先从声信号中提取多域特征,包含时域波形结构特征、时域包络特征、频域谱特征和基于短时傅里叶变换的时频联合域特征;然后基于不同谱表达计算出一组声学参数构成特征空间,在此基础上采用最大相关-最小冗余准则选出特征空间中与声源位置相关性高的关键特征作为模型输入;最后通过一种改进的深度神经网络实现声源距离的估计,引入自适应矩估计优化算法进行模型训练,利用L2和Dropout正则化策略实现网络参数稀疏化。通过声速正梯度浅海环境仿真实例对方法进行验证,对比分析了波形参数对测距性能和模型收敛速度的影响。结果表明,此方法在模型训练过程中收敛速度较快,预测性能较稳定,在所定条件下测试集上声源信号的综合测距精确率达到95%以上。Ranging of underwater acoustic sources is an important task in many practical applications.Deep neural networks(DNNs)have shown outstanding performance on illustrating complex nonlinear relationships.In terms of localization,DNNs were demonstrated to have low range estimation error under low SNR conditions.In this study,source localization is achieved by a deep learning method based on a set of multi-domain features extracted from sound signals.In this study,a comprehensive set of acoustic parameters are extracted from sound signals.The waveform,energy envelope,short-term Fourier transform are first estimated from the sound signals,based on which,the acoustic parameters are computed and used as the training data set for the estimation of source in a DNN.Training the deep architecture is achieved by Adam optimizer and dropout for parameter regularization with mean squared error(MSE)loss functions.

关 键 词：多域特征提取 深度学习 声源被动测距 

分 类 号：TB566[交通运输工程—水声工程]