面向原始辐射噪声的水声目标识别研究

Research on Underwater Acoustic Target Recognition for Original Radiated Noise

水声目标识别是一项利用目标辐射噪声特性对目标属性进行判别的模式识别技术,具有十分重要的经济价值和军事价值。传统水声目标识别方法依靠信号处理技术对目标辐射噪声进行特征提取,进而通过人工识别或设计分类器识别的方法实现目标类别的判别。通过设计特征进行特征提取的过程中,不可避免会造成目标辐射信号中信息的损失,而深度神经网络依靠其多层网络结构,具备强大的特征提取能力。文章以一维卷积神经网络作为构建水声目标识别的基本模型,利用梅尔频率倒谱系数特征的提取思路,创造性地设计了MFCC1D卷积神经网络。结果显示,在ShipsEar数据集与水下目标信号构成的混合数据集上,设计方法的识别准确率达到96.4%。

Underwater acoustic target recognition is a use of the target radiated noise characteristics of target attribute to discriminate the pattern recognition technology,is of great importance to economic and military value.Traditional underwater acoustic target recognition methods rely on feature extraction of target radiated noise signal processing technology,and using artificial recognition or design classifier recognition goal categories of discrimination.By designed features in the process of feature extraction,would inevitably cause the loss of information in the target radiation signal,and the depth of the neural network,relying on its multi-layer network structure,strong ability in feature extraction,based on the one-dimensional convolutional neural networks as building basic model of underwater acoustic target recognition,using the extraction of Mel frequency cepstrum coefficient characteristic way of thinking,creative MFCC1D convolution neural network was designed,in ShipsEar data sets with mixed data sets composed of underwater target signal recognition accuracy 96.4%.