【作者】 李博；

【导师】 李刚； 易卿武；

【作者基本信息】 哈尔滨工程大学 ， 电子信息（专业学位）， 2023， 硕士

【摘要】 海洋声学环境中声道的特征分析是海洋研究领域的重要组成部分,通过海洋声场分析所获取的声道特征信息,对于海洋防灾减灾、保护海洋生态环境、维护海洋安全以及开展海洋科学研究等具有重大意义。当前声道识别方法主要通过声速数据本身的数学计算来识别声道位置及其特征信息,因此声速数据的准确度会直接影响到声道识别及其特征信息的准确度。但是在数据采集预处理过程中以及声速计算过程中,都会出现误差积累,会造成声速数据准确度较低的问题。声速剖面表示海水中声速随深度变化情况,并非所有声速剖面都具有声道,可以通过声速剖面分类判断声道是否存在及其类型。声速剖面分类是对声速数据进行特征提取并聚类分析,声速数据的准确度同样会影响到声速剖面分类的准确度。因此将声速剖面输出成图像,从数据识别变为图像识别,降低声速数据本身对声速剖面分类和声道识别的影响。卷积神经网络利用计算机通过神经网络结构自主地从图像数据样本中学习数据特征,可以提取出更高维、更抽象的数据特征,在声道目标识别和声速剖面图像分类方面具备潜在优势。本文提出基于卷积神经网络的声速剖面分类和声道目标智能识别算法,通过对海洋三维温盐场数据的预处理将声速剖面输出成图像,为算法提供图像数据集,通过文献分析确定声速剖面及声道类型,并设计时空泛化性实验。本文主要研究内容如下:(1)为了提供算法图像数据集,分别设计制作基于卷积神经网络图像分类和目标检测的图像数据集。通过分析选择三维海洋温盐场的数据集及声速计算公式,针对离散的声速序列进行插值方法的对比分析,并对Akima插值法进行算法改进。通过对文献中声速剖面类型的分析,将声速剖面及声道类型分为8种类型,在此基础上将声速剖面输出成图像,为算法提供图像数据集。(2)针对声速剖面的分类受声速数据准确度影响的问题,通过卷积神经网络提取声速剖面的图像特征,提出了一种基于VGG16(Visual Geometry Group 16)的声速剖面分类方法。从提取数据特征变为提取图像特征,根据图像的数据特征进行分类确定声速剖面中的声道及其类型。对VGG16的网络结构进行改进,使其适用于声速剖面分类。对VGG16网络中的参数进行调整以及设计与模糊C-均值聚类算法、自组织特征映射神经网络聚类算法和BP神经网络的对比实验,基于VGG16的声速剖面分类方法可以提升分类的准确度。(3)针对声道识别方法中声道位置及其特征信息受声速数据准确度影响的问题,在声速剖面分类确定声道及其类型的基础上,通过卷积神经网络进行声道目标检测识别,提出了一种基于改进Faster-RCNN的声道智能识别方法。从数据识别变为图像识别,对有声道目标的声速剖面进行声道智能识别,确定声道特征信息。针对声道目标种类少、目标单一等情况对Faster-RCNN目标检测算法进行网络模型的改进,基于改进FasterRCNN的声道智能识别方法可以进行声道特征智能识别。针对声速剖面分类和声道目标识别设计时空泛化性实验,实验结果表明,算法对声速剖面分类和声道目标检测识别在时间和空间上具备泛化性。更多还原

【Abstract】 The characteristic analysis of sound channel in the marine acoustic environment is an important component of the field of ocean research.The sound channel feature information obtained through the analysis of ocean sound fields is of great significance for marine disaster prevention and reduction,protection of marine ecological environment,maintenance of marine safety,and conducting marine scientific research.The current sound channel recognition methods mainly identify the position and feature information of the sound channel through mathematical calculations of the sound speed data itself.The accuracy of sound speed data directly affects the accuracy of sound channel recognition and its feature information.However,in the process of data collection and preprocessing,as well as in the calculation of sound speed,error accumulation can occur,which can lead to low accuracy of sound speed data.The sound speed profile represents the variation of sound speed with depth in seawater,and not all sound speed profiles have sound channel.The existence and type of sound channel can be determined through classification of sound speed profile.Sound speed profile classification involves feature extraction and clustering analysis of sound speed data,and the accuracy of sound speed data also affects the accuracy of sound speed profile classification.Therefore,outputting the sound speed profile into image changes from data recognition to image recognition,reducing the impact of sound speed data itself on sound speed profile classification and sound channel recognition.Convolutional neural network uses the computer to autonomously learn data features from image data samples through neural network structure,which can extract higher dimensional and more abstract data features.It has potential advantages in sound channel target recognition and sound speed profile image classification.This article proposes a convolutional neural network-based algorithm for classification of sound speed profile and intelligent recognition of sound channel target.By preprocessing ocean 3D temperature and salinity data,the sound speed profile is output as an image,providing an image dataset for the algorithm.Through literature analysis,the types of sound speed profile and sound channel are determined,and spatiotemporal generalization experiments are designed.The main research content of this paper is as follows:(1)To provide algorithmic image datasets,image datasets based on convolutional neural network for image classification and target detection are designed and produced.By analyzing and selecting the dataset of 3D ocean temperature and salinity field and the calculation formulae for sound speed.A comparative analysis of interpolation methods is conducted for discrete sound speed sequences.Algorithm improvement are made to Akima interpolation method.By analyzing the types of sound speed profile in literature,the types of sound speed profile and sound channel are classified into 8 types.Based on this,the sound speed profile is output as images,providing image datasets for the algorithm.(2)To address the issue of the accuracy of sound speed data affecting the classification of sound speed profile,convolutional neural network is used to extract image features of sound speed profile.A sound speed profile classification method based on VGG16(Visual Geometry Group 16)is proposed.From extracting data features to extracting image features,classify and determine the sound channel and their types in the sound speed profile based on the data features of the image.The network structure of VGG16 is improved to make it suitable for sound speed profile classification.By adjusting the parameters in the VGG16 network and designing comparative experiments with fuzzy C-means clustering algorithm,the self-organizing feature map neural network clustering algorithm and BP neural network.The sound speed profile classification method based on VGG16 can improve the accuracy of classification.(3)To address the issue of the accuracy of sound speed data affecting the position and feature information of the sound channel in sound channel recognition method.Based on the classification of sound speed profile to determine sound channel and its type.A convolutional neural network is used for sound channel target recognition.A sound channel intelligent recognition method based on improved Faster RCNN is proposed.Transform from data recognition to image recognition,intelligently recognize the sound speed profile of a target with sound channel,and determine the characteristic information of the sound channel.The FasterRCNN target detection algorithm is improved in the network model for situations of few types of sound channel targets and single target.The improved Faster-RCNN based sound channel intelligent recognition method can perform sound channel feature intelligent recognition.Design spatiotemporal generalization experiments for sound speed profile classification and sound channel target recognition.The experimental results show that the algorithm has generalization in both time and space for sound speed profile classification and sound channel target detection and recognition.更多还原