摘要
以某研究所采集的湖泊环境下的水声目标信号的实测数据为实验样本,在对其进行预处理的基础上,提取了能够表征水声目标特性的15个标志性特征,并通过机器学习算法对水声目标进行自动识别,以验证这些标志性特征的性能.这组标志性特征对静止目标的识别率高于98%,对移动目标的识别率约90%,且AUC面积均保持在0.95以上的水平.实验结果表明:本文所提取的标志性特征能够在特征数量少,不依赖于深度学习方法的条件下,以极低的计算复杂度实现对水声目标的准确检测.
Based on the experimental evolution of underwater acoustic data in lake environment collected by a research institute,15 signature features which can fully present the characteristics of underwater acoustic target signals are extracted,and the performance of the features are verified by automatic recognition with use of machine learning algorithms.The accuracies on fixed targets are higher than 98%,and those for moving targets are around 90%,with AUC areas above 0.95.The experimental results show that the signature features extracted in this work can accurately detect underwater acoustic targets with very low computational complexity and only very few features,and without deep learning methods.
作者
江均均
黄敏
肖仲喆
JIANG Junjun;HUANG Min;XIAO Zhongzhe(School of Optoelectronic Science and Engineering,Soochow University,Suzhou,Jiangsu 215006,China)
出处
《复旦学报（自然科学版）》
CAS
CSCD
北大核心
2020年第3期299-306,共8页
Journal of Fudan University：Natural Science
基金
国家自然科学基金(61906128)
江苏省自然科学基金(BK20180834)。
关键词
水声信号
特征提取
机器学习
underwater acoustic signals
feature extraction
machine learning
