基于水声环境空间中多模态深度融合模型的目标识别方法研究  

Research on the Underwater Target Recognition Based on Multi-Model Information Fusion with Deep-Learning-Based Multimodal Learning Model in Three-Dimensional Acoustic Map of the Underwater Acoustical Environment

作　　者：李琦[1] 孙桂玲[2] 黄翠[1] 刘颉[1] 常哲[1] 于金花[1] 文洪涛 LI Qi;SUN Gui-ling;HUANG Cui;LIU Jie;CHANG Zhe;YU Jin-hua;WEN Hong-tao(National Ocean Technology Center,Tianjin 300112,China;Nankai University,Tianjin 300350,China;Third Institute of Oceanography,Ministry of Natural Resources,Xiamen 361005,Fujian Province,China)

机构地区：[1]国家海洋技术中心,天津300112 [2]南开大学,天津300350 [3]自然资源部第三海洋研究所,福建厦门361005

出　　处：《海洋技术学报》2019年第6期35-45,共11页Journal of Ocean Technology

基　　金：国家海洋技术中心创新基金重点项目(Z-17005)

摘　　要：随着对水下目标特性研究的深入和声学探测技术的发展,基于单模态的阵列式信息融合或基于空间信息的分布式信息融合的水下目标识别方法研究已有一定成果,但针对复杂海况导致单一物理场或单一融合层次的系统识别性能提高有限等方面影响的水下目标识别方法研究还有所不足,因此,开展基于多模态深度融合模型的水下目标识别方法研究可利用模态互补,共享信息而提升识别率。文中在国内外研究基础上,深入研究了基于到达时差法和多模态方法组合的检测方法,初步形成了基于水声环境空间中多模态深度融合模型的识别框架,开展了海洋中典型自然与人为事件的信号分析与特征提取,并在此基础上,设计新型基于海底基站的被动识别系统。该系统同步记录和由位置等组成的时间序列标记声、磁和压数据,可实现高精度、高分辨率的识别。本研究可满足未来海洋观测对高性能水下目标探测、定位和跟踪系统的迫切需要,为海洋安全监管、海洋突发事件应急响应等领域提供新的技术手段和科学参考。As the analytic study of underwater target characteristics becomes more intensive and the acoustic detection technology becomes more mature,several achievements have been attained in the field of detection and recognition modeling for underwater target based on single modal information fusion of sensor array and space information fusion of distributed sensor array.However,only a few studies on multi-modal or deep multimodal fusion model of underwater target detection and identification methods have been conducted,of which the environmental factors and many interference sources may lead to detection system performance.It presents practical significance to conduct a study of detection and recognition methods of underwater target based on the deep multimodal data fusion model and a design of the measuring technique based on the seabed-base monitoring platform.This research makes full use of the complementary relationship between acoustic modes,magnetics modes and water pressure modes,and shares information to improve the recognition rate.The multi-modal kernel method for activity detection of underwater targets and the recognition model based on the multi-model information hybrid fusion with deep-learning-based multimodal(deep multimodal)learning model in three-dimensional space in underwater acoustical environment are developed by referencing relevant study results at home and abroad.The algorithm and method will be optimized,and a passive measurement system will be designed according to the features of the model.The device obtains the best of the underwater acoustic and magnetic and pressure signals for analysis of the characteristic parameters of underwater targets.The device also tests the validity of the model and the method using offshore test data.The underwater targets are detected and recognized precisely in the experimental pool.The device can offer appropriate measurement data and a new measuring technology for the study on detecting,positioning and tracking of underwater targets,and provide new technical

关 键 词：水下目标识别 多模态 水声环境 深度模型 目标特性 

分 类 号：P714[天文地球—海洋科学]