SIA OpenIR  > 海洋机器人前沿技术中心
基于深度学习的水中运动目标噪声轴频检测算法研究
Alternative TitleResearch on Fundamental Frequency Detection Method for Underwater Target Noises by Deep Learning
卢佳敏
Department海洋机器人前沿技术中心
Thesis Advisor宋三明
Keyword基频检测 长短时记忆网络 水听器阵列 舰船噪声信号仿真 DEMON谱
Pages56页
Degree Discipline模式识别与智能系统
Degree Name硕士
2021-05-21
Degree Grantor中国科学院沈阳自动化研究所
Place of Conferral沈阳
Abstract对目标航行过程中向水体辐射出的噪声加以分析能够获得关于目标运动状态及本体特征等信息,进而达到探测和识别目的。研究人员对水中机动目标噪声进行了大量的分析,对目标噪声的产生过程进行建模并探索从中提取出有效特征的方法。其中LOFAR(Low Frequency Analysis Recording)谱和DEMON(Detection of Envelope Modulation On Noise)谱是在目标识别和探测过程中最常用的两种噪声特征,能够反映目标的工况和种类。航行速度作为一种重要工况,被许多研究人员关注。螺旋桨转速能直观反映航行速度,因此分析目标螺旋桨转速具有重要的研究意义。螺旋桨主轴的转动频率被称为轴频(基频),关于目标辐射噪声的轴频检测算法已经有了大量工作,但传统的船舶辐射噪声基频检测方法不仅依赖大量的先验知识,而且对背景噪声比较敏感。为了提高目标识别的稳定性和精确性,本文提出了一种基于深度神经网络的基频检测算法。首先,设计了一种基于水听器阵列信号和深度学习网络的水中目标噪声轴频检测算法。从多通道水听器信号中提取DEMON谱,引入梳状滤波器对DEMON谱的谐波结构进行增强。然后将谱特征输入CNN和LSTM构成的级联网络,对特征进行降维并提取其时序特征。最终将特征流经稠密层进行分类,并实现对基频的估计。其次,开展了水下目标辐射噪声及水听器接收端信号仿真。深度学习方法往往对样本量有一定的要求,但水中目标辐射噪声实测数据一般存在获取困难的问题。因此,本文介绍了两种常见的舰船噪声信号仿真方法,并通过Bellhop仿真海洋信道在传播过程中对信号的影响。利用仿真噪声信号对所设计网络进行训练。同时,获取一定量的实测噪声数据对网络进行微调,以适应实际环境下的噪声信号。最后,开展消声水槽和外场试验,对所提方法的可行性和精确率进行了测试。得到了如下结论:(1)深度网络能够实现无先验知识和不同信噪比条件下的基频检测,具有良好的泛化性能。(2)LSTM网络能够高效地从时序DEMON谱中提取统计特征,提高基频估计精度。(3)输入信号的时间长短会影响网络的检测精度,更长时间的信号能够获得更好的检测结果。
Other AbstractThe analysis of the noise radiated to the surrounding water during the target navigation can acquire the target's motion state and characteristics, by which the purpose of detection and identification can be achieved. A lot of researchers have contributed to the radiation noise modeling and the subsequent feature extraction. Among them, the LOFAR(Low Frequency Analysis Recording) spectrum and the DEMON(Detection of Envelope Modulation On Noise) spectrum are two of the most commonly features in target recognition and detection, which can characterize the working conditions and the category of targets. The sailing speed, as an important feature, has drawn attention from many researchers. Propeller speed can reflect the sailing speed directly, so the analysis of the target propeller speed has important research significance. Rotation frequency of the main shaft of the propeller is called the shaft frequency (fundamental frequency). A lot of work about the shaft frequency detection algorithm has been studied. However, the traditional fundamental frequency detection methods of radiated noise not only depend on prior knowledge, but also are sensitive to ambient noises. In order to improve the stability and accuracy of target recognition, an algorithm based on deep neural network is proposed in this paper to detect the fundamental frequency. First of all, a detection algorithm for underwater target shaft frequency based on hydrophone array signal and deep learning network is proposed. The DEMON spectrum is extracted from the multi-channel hydrophone signal, and the comb filter is applied to enhance the harmonic structure of the DEMON spectrum. Then, the spectral features are fed into the cascaded network composed of CNN(Convolutional Neural Network) and LSTM(Long-Short Term memory) to reduce the dimension and extract the temporal features. Finally, the features propagate through the dense layer for classification, with the output class label corresponding tothe fundamental frequency. Secondly, the radiation noise of underwater target and the signal received by hydrophone array are simulated. Deep learning based methods often require large amounts of data, however, it’s difficult to obtain measured data of radiation noise from underwater targets. Therefore, this paper introduced two common radiated noise simulation models, and the Bellhop is adopted to imitate the influence of ocean channel to noise signal. The simulation signals are used to train the network. At the same time, a certain amount of measured noise data is used to fine-tune the network to adapt to the actual environment. Finally, the experiments of anechoic tank and the field trials were carried out to prove the feasibility and accuracy of the proposed method. It comes to these conclusion: (1) The deep neural network can achieve the fundamental frequency detection under the conditions of no prior knowledge and different signal-to-noise ratio, and has good generalization performance. (2) The LSTM network can efficiently extract statistical features from the time sequence DEMON spectrum and improve the accuracy of fundamental frequency estimation. (3) The length of the input signal will affect the detection accuracy of the network, and longer signals can obtain better detection results.
Language中文
Contribution Rank1
Document Type学位论文
Identifierhttp://ir.sia.cn/handle/173321/28949
Collection海洋机器人前沿技术中心
Affiliation中国科学院沈阳自动化研究所
Recommended Citation
GB/T 7714
卢佳敏. 基于深度学习的水中运动目标噪声轴频检测算法研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2021.
Files in This Item:
File Name/Size DocType Version Access License
基于深度学习的水中运动目标噪声轴频检测算(2162KB)学位论文 开放获取CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[卢佳敏]'s Articles
Baidu academic
Similar articles in Baidu academic
[卢佳敏]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[卢佳敏]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.