基于DEMON谱和LSTM网络的水下运动目标噪声基频检测  

Fundamental frequency detection of underwater target noises using DEMON spectrum and LSTM network

作　　者：卢佳敏 宋三明 景严 张瑶[1,2] 谷浪[1,2] 鲁帆 胡志强 李硕[1,2] LU Jiamin;SONG Sanming;JING Yan;ZHANG Yao;GU Lang;LU Fan;HU Zhiqiang;LI Shuo(State Key Laboratory of Robotics,Shenyang Institute of Automation,Chinese Academy of Sciences,Shenyang 110016,China;Institutes for Robotics and Intelligent Manufacturing,Chinese Academy of Sciences,Shenyang 110169,China;University of Chinese Academy of Sciences,Beijing 100049,China;Institute of Acoustics,Chinese Academy of Sciences,Beijing 100190,China)

机构地区：[1]中国科学院沈阳自动化研究所机器人学国家重点实验室,沈阳110016 [2]中国科学院机器人与智能制造创新研究院,沈阳110169 [3]中国科学院大学,北京100049 [4]中国科学院声学研究所,北京100190

出　　处：《应用声学》2021年第5期745-753,共9页Journal of Applied Acoustics

基　　金：国家自然科学基金项目(61973297);中国科学院先导专项子课题(XDC03060105);中国科学院青年创新促进会课题(2020209);机器人学国家重点实验室课题(2017-Z010)。

摘　　要：传统的船舶辐射噪声基频检测方法不仅依赖大量的先验知识,而且对背景噪声非常敏感。为了提高目标识别的稳定性和精确性,该文提出了一种基于深度神经网络的基频检测算法。首先从多通道水听器信号中提取DEMON谱,然后直接将二维谱特征矩阵输入由卷积神经网络和长短时记忆网络构成的级联网络,最后通过稠密层输出实现对基频的估计。仿真和外场实验结果初步表明:深度网络能够实现无先验知识和不同信噪比条件下的基频检测,具有良好的泛化性能。长短时记忆网络能够高效地从时序DEMON谱中提取统计特征,提高基频估计精度。输入信号的时间长短会影响网络的检测精度,更长时间的信号能够获得更好的检测结果。The traditional fundamental frequency(F0)detection methods not only rely on prior knowledge,but also are very sensitive to ambient noises.In this paper,a fundamental frequency detection algorithm based on deep neural network is proposed to improve the stability and accuracy of target recognition.The DEMON spectrum matrix,which is composed of spectral vector extracted from each single hydrophone signal,is directly fed into the cascaded network made up of convolutional neural networks(CNN)and long short-term memory(LSTM)networks.Then,with the one-hot vector from the final dense layer,the fundamental frequency is estimated.The following conclusions can be drawn from computer simulation and field experiments:The deep learning-based method works well when no prior knowledge is assumed or signal to noise ratio varies,having good generalization performance.LSTM network can effectively extract the statistical characteristics from the DEMON spectrum sequence and improve the accuracy of the F0 estimation.The detection precision depends on the input signal length,and a better detection result could be obtained when a longer signal is available.

关 键 词：基频 深度网络 长短时记忆网络 卷积神经网络 水听器阵列 水下目标噪声 

分 类 号：TB566[交通运输工程—水声工程]