基于面积加权GWT-GFT的水声目标识别-《计算机技术与发展》

文章信息/Info

Title:: Underwater Acoustic Target Recognition Based on Area Weighted GWT-GFT

Author(s):: CHEN Xin; SHAO Jie; WANG Xing-xing; YANG Xin; YANG Shi-yi-lin; School of Electronic and Information Engineering,Nanjing University of Aeronautics and Astronautics,Nanjing 211106,China

Keywords:: underwater acoustic target recognition; graph wavelet transform-graph Fourier transform; feature extraction; graph signal pro-cessing; vertex triangle area weighting

摘要:: 由于海洋环境的复杂性,水声目标的识别具有很大的挑战性。为解决这类复杂环境下特征提取的问题,提出了一种基于面积加权的图小波变换-图傅里叶变换(GWT-GFT)的分析方法。在完成数据预处理后,为了能够凸显顶点之间的关系,提出了一种新的基于顶点三角形面积的加权方法来构建图信号;构建好的图信号通过 GWT 分解为多尺度图分量;然后,利用 GFT 将这些分量从图域变换到特征值谱域进行分析;在此基础上,提取各分量特征值谱的特征;最后,利用基于高斯核函数的支持向量机(SVM)对获取的特征向量进行分类。基于水声信号 ShipsEar 数据库,采用 5 折交叉验证方法进行验证。与现有的其它方法相比,所提的模型以 36 个特征在 376 656 个样本上取得了 97. 22% 的准确率,证明了该分析方法的有效性和鲁棒性。

Abstract:: Due to the complexity of the marine environment,the recognition of underwater acoustic targets poses significant challenges.To address the problem of feature extraction in such complex environments,an analysis method based on area weighted Graph Wavelet Transform - Graph Fourier Transform (GWT-GFT) is proposed. After completing data preprocessing,a novel weighted method based on the triangle area of vertices is proposed to construct the graph signal in order to highlight the relationships between vertices. The con-structed graph signal is decomposed into multiple-scale graph components using GWT. Then,these components are transformed from the graph domain to the eigenvalue spectrum domain for analysis using GFT. Based on this,the characteristic eigenvalue spectra of each com-ponent are extracted. Finally,the obtained feature vectors are classified using Support Vector Machine (SVM) based on the Gaussian kernel function. Based on the ShipsEar database of underwater acoustic signals, a 5 - fold cross - validation method is employed for verification. Compared with other existing methods, the proposed model achieves a recognition accuracy of 97. 22% on a dataset consisting of 376 656 samples using 36 features. This result demonstrates the effectiveness and robustness of the proposed analysis meth-od.