自然科学版
陕西师范大学学报(自然科学版)
全国声学大会专题
基于深度学习的典型海洋哺乳动物click信号识别方法
PDF下载 ()
高德洋,高大治*,李小雷
(中国海洋大学 信息科学与工程学院, 山东 青岛 266100)
高德洋,男,硕士研究生,主要研究方向为海洋声学。E-mail:673783376@qq.com;高大治,男,副教授,研究方向为海洋声学。E-mail:dzgao@ouc.edu.cn
摘要:
利用深度神经网络对3种典型海洋哺乳动物的click信号和脉冲噪声进行分类识别。首先对采集到的海洋哺乳动物click信号进行特征分析,给出频谱能量算法;之后利用前馈全连接神经网络对时域信号进行识别,研究神经网络参数的改变对识别结果的影响;最后利用卷积神经网络对时频信号进行了识别。结果表明:频谱能量算法识别准确率为69.83%,前馈全连接网络通过调整参数准确率可以达到98.28%,卷积神经网络准确率达到100%。由于实验数据规模较小、信号信噪比较高,所以神经网络的识别效果较好。深度学习方法能够比频谱能量算法取得更好的识别效果,调节前馈全连接网络的隐藏层参数,可提高识别效果。
关键词:
click信号;海洋哺乳动物;前馈全连接神经网络;卷积神经网络
收稿日期:
2019-09-30
中图分类号:
TB565.1
文献标识码:
A
文章编号:
1672-4291(2019)06-0037-07
基金项目:
国家自然科学基金(11674294,11874331)
Doi:
Marine mammals′ click signals classification based on deep learning
GAO Deyang, GAO Dazhi*, LI Xiaolei
(College of Information Science and Engineering, Ocean University of China, Qingdao 266100, Shandong, China)
Abstract:
Deep neural networks were used to classify click signals of three typical marine mammals and impulse noise. First, the click signals of marine mammals were analyzed and a spectral energy algorithm was proposed. Then, the time domain signals were identified by the feedforward fully connected neural network and the influence of the parameters of the neural network on the recognition results was studied.Finally, the convolution neural network was used to identify the time-frequency signals. The recognition results show that the recognition accuracy of spectral energy algorithm is 69.83%. The accuracy of feedforward fully connected network can reach 98.28% by adjusting parameters, and the accuracy of convolutional neural network can reach 100%. Due to the small scale of experimental data and high signal-to-noise ratio of signals, the neural network has a good recognition effect.The research shows that the deep learning method can achieve better identification effect than the spectral energy algorithm, and the identification effect can be improved by adjusting the hidden layer parameters of the feedforward fully connected network.
KeyWords:
click signals; marine mammals; feedforward fully connected neural network; convolutional neural network