| With the development of the "Blue Economy," countries around the world are vigorously promoting the development of their marine economies and the exploration of marine resources.In the process of marine development and safeguarding maritime sovereignty,China urgently needs advanced technologies in various marine fields and their application in marine development.At the same time,with the widespread application of deep learning and reinforcement learning,researchers have been attempting to apply these techniques to marine research.This article focuses on the research of deep learning-based methods for annotating underwater acoustic signals,aiming to promote the advancement of marine research.Based on the aforementioned background,this article combines neural network models,feature fusion,attention mechanisms,and other strategies in deep learning with research on underwater acoustic signal annotation,proposing an automatic annotation method for underwater acoustic signals based on deep learning.In the field of underwater acoustics,voiceprint recognition is commonly used for automatic annotation of underwater sound.The most crucial parts of voiceprint recognition technology are feature extraction and feature representation.However,existing methods suffer from insufficient feature extraction and incomplete feature representation for underwater sound signals.Based on this,this thesis proposes a model for underwater sound annotation based on a Multi-Scale Hybrid Network(MSHN).The model employs a network architecture combining attention and residual connections for feature extraction,and utilizes an entropy-based method for feature representation.Finally,a classifier is used for automatic annotation of underwater sound signals.The MSHN model consists of two main parts: the feature extraction network and the feature representation method.These are also the two key research components of this thesis.Firstly,to address the issue of insufficient data feature extraction by the feature extraction network,a dual convolutional network model based on attention-residual(DCN-AR)is proposed in this thesis by combining ideas of attention mechanism,residual structure,feature fusion and optimization,using a multi-core convolutional fusion network(MCF)and a multi-layer optimized convolutional network(MOC).The MCF network extracts features by setting two branches with different sizes of convolution kernels,and uses attention mechanism to fuse the features from the branches to obtain multi-scale feature information.The MOC network mainly utilizes channel attention and optimized convolution operations to extract distinctive feature information.Finally,the MCF network and MOC network are integrated into a hybrid architecture in the DCN-AR model,which captures feature information that is both multi-scale and distinctive,and of significant importance.Lastly,a method based on information entropy called attentional statistical pooling(IEAP)was proposed in this thesis to combine global and local feature information,addressing the issue that existing feature representation methods can’t fully represent feature information.In the IEAP,the thesis introduces information entropy(IE)to obtain the information of each frame data,which is concatenated with the weighted mean and weighted standard deviation in the attentional statistical pooling method to become a feature representation that combines global and local features,thereby improving the reliability and accuracy of the labeling model.Finally,this thesis conducts multiple experiments on two different types of datasets to verify the effectiveness and reliability of the proposed deep learning-based automatic labeling method for underwater acoustic signals.The results demonstrate that the method has practical significance and can achieve high accuracy in labeling underwater acoustic signals. |
