特征降维与融合的水声目标识别方法  

Underwater acoustic target recognition method based on feature dimension reduction and fusion

作　　者：李昊鑫 肖长诗[1,2] 元海文 郭玉滨 刘加轩 LI Haoxin;XIAO Changshi;YUAN Haiwei;GUO Yubin;LIU Jiaxuan(School of Navigation,Wuhan University of Technology,Wuhan 430063,China;Weihai Institute of Marine Information Science and Technology,Shandong Jiaotong University,Weihai 264299,China;School of Electrical and Information Engineering,Wuhan Institute of Technology,Wuhan 430205,China)

机构地区：[1]武汉理工大学航运学院,湖北武汉430063 [2]山东交通学院威海海洋信息科学与技术研究院,山东威海264299 [3]武汉工程大学电气信息学院,湖北武汉430205

出　　处：《哈尔滨工程大学学报》2025年第1期102-110,共9页Journal of Harbin Engineering University

基　　金：国家自然科学基金项目(52001235);湖北省自然科学基金项目(2022CBF313);山东省自然科学基金项目(ZR2020KE029).

摘　　要：为解决水声目标在强噪声环境下识别困难以及特征高维问题,本文提出一种将水声信号进行离散小波变换并提取其低频系数与重组一维梅尔倒谱系数融合的方法,以减少特征维度并弥补信息损失。利用1D-CNN-LSTM神经网络在DeepShip和ShipsEar 2个数据集上进行实验,识别准确率均在99%以上。结果表明:该算法能够有效抑制噪声,具备较强的鲁棒性。将所提算法应用到单船识别,实验结果表明该算法能够有效区分同类型的不同船舶。This study addresses the challenge of underwater acoustic target recognition in strongly noisy environments and the high-dimensional feature problem in recognition tasks.Thus,a method that applies discrete wavelet transform to hydroacoustic signals to extract low-frequency coefficients is proposed.These coefficients are then combined with the reconstructed one-dimensional Mel-frequency cepstral coefficients(MFCC).This approach aims to reduce the dimensionality of the features and compensate for the loss of information.Two datasets,namely Deep-Ship and ShipsEar,are used to conduct experiments on a 1D-CNN-LSTM neural network,achieving recognition accuracies of above 99%.The results demonstrate that the algorithm effectively suppresses noise and possesses robust performance.The algorithm is also applied to single-ship recognition,and the experimental results indicate that it effectively distinguishes different ships of the same type.

关 键 词：水声目标识别 离散小波变换 梅尔倒谱系数 特征融合 联合神经网络 单船识别 深度学习 船舶辐射噪声 

分 类 号：U675.79[交通运输工程—船舶及航道工程] TP181[交通运输工程—船舶与海洋工程]