基于一维卷积神经网络的舰船尾流检测

杨长生; 苟文博; 梁红

doi:10.11993/j.issn.2096-3920.2022-0052

基于一维卷积神经网络的舰船尾流检测

doi: 10.11993/j.issn.2096-3920.2022-0052

西北工业大学航海学院, 陕西西安, 710072

基金项目: 国家自然科学基金项目资助(61771398, 61971354)

详细信息

作者简介:
杨长生(1978-), 男, 副教授, 研究方向为水下信号与信息处理、仿生智能感知

中图分类号: TJ630.6; U674
计量
- 文章访问数: 74
- HTML全文浏览量: 18
- PDF下载量: 26
- 被引次数: 0
出版历程
- 收稿日期: 2022-08-29
- 修回日期: 2022-09-22
- 录用日期: 2022-11-09
- 网络出版日期: 2023-11-22

Ship Wake Detection Based on One-Dimensional Convolutional Neural Network

School of Marine Science and Technology, Northwestern Polytechnical University, Xi’an 710072, China

摘要

摘要: 为了提高舰船尾流的检测准确率, 提出一种基于一维卷积神经网络(1DCNN)的舰船尾流检测方法。利用舰船尾流散射回波模型构建仿真数据集, 通过水箱仿真实验验证散射回波模型的可靠性并构建实验数据集, 最后综合考虑不同结构模型的检测准确率和参数量搭建了1DCNN并与传统检测算法(基于单分类支持向量机和反向传播神经网络)在数据集上进行对比分析。仿真结果表明, 相比与传统检测算法, 文中所提出的1DCNN提高了不同信噪比下舰船尾流的检测准确率和检测效率, 具有较好的工程应用价值。
- 舰船尾流检测 /
- 一维卷积神经网络 /
- 单分类支持向量机 /
- 反向传播神经网络
Abstract: In order to improve the detection accuracy of ship wake, this paper proposed a ship wake detection method based on a one-dimensional convolutional neural network (1DCNN). Firstly, the simulation data set was constructed by using the ship wake scattering echo model. Then the reliability of the scattering echo model was verified by the water tank simulation experiment, and the experimental data set was constructed. Finally, a 1DCNN was built by comprehensively considering the detection accuracy and parameter quantity of different structural models and compared with the traditional detection algorithm (based on a one-class support vector machine and back propagation neural network) on the data set. The simulation results show that compared with the traditional detection algorithm, the 1DCNN proposed in this paper improves the detection accuracy and detection efficiency of ship wake under different signal-to-noise ratios and has good engineering application value.
- ship wake detection /
- one-dimensional convolutional neural network /
- one-class support vector machine /
- back propagation neural network

HTML全文

图 1 水箱实验场景布置图

Figure 1. Layout of water tank experiment scene

下载: 全尺寸图片幻灯片

图 2 仿真和实验散射回波

Figure 2. Simulated and experimental scattering echoes

下载: 全尺寸图片幻灯片

图 3 回波信号统计特性

Figure 3. Statistical properties of echo signals

下载: 全尺寸图片幻灯片

图 4 仿真和实验散射回波时频特性

Figure 4. Time-frequency characteristics of simulated and experimental scattering echoes

下载: 全尺寸图片幻灯片

图 5 数据集构成

Figure 5. Composition of data set

下载: 全尺寸图片幻灯片

图 6 1DCNN结构示意图

Figure 6. 1DCNN structure

下载: 全尺寸图片幻灯片

图 7 “Valid”填充方式示意图

Figure 7. “Valid” filling method

下载: 全尺寸图片幻灯片

图 8 1DCNN训练流程图

Figure 8. Flow chart of 1DCNN training

下载: 全尺寸图片幻灯片

图 9 1DCNN测试流程图

Figure 9. Flow chart of 1DCNN test

下载: 全尺寸图片幻灯片

图 10 高斯滤波器幅频响应曲线

Figure 10. Amplitude frequency response curve of Gaussian filter

下载: 全尺寸图片幻灯片

图 11 1DCNN模型识别准确率曲线图

Figure 11. Recognition accuracy curves of 1DCNN model

下载: 全尺寸图片幻灯片

表 1 数据集A构成

Table 1. Composition of data set A

距水面距离/m	收发平台速度/(m/s)	声源发射信号角 /(°)	有尾流散射回波个数	无尾流散射回波个数
15	0	90	120	120
15	0	60	120	120
15	20	90	120	120
15	20	60	120	120
20	0	90	120	120
20	0	60	120	120
20	20	90	120	120
20	20	60	120	120
25	0	90	120	120
25	0	60	120	120
25	20	90	120	120
25	20	60	120	120

下载: 导出CSV

表 2 不同卷积层数模型参数表

Table 2. Parameters of models with different convolutional layers

卷积层	卷积层数
卷积层	4	5	6
第1层	1×125, 32	1×125, 32	1×125, 32
第2层	1×15, 32	1×15, 32	1×15, 32
第3层	1×3, 64	1×3, 64	1×3, 64
第4层	1×3, 32	1×3, 32	1×3, 32
第5层	—	1×3, 32	1×3, 32
第6层	—	—	1×3, 32

下载: 导出CSV

表 3 不同卷积层数模型检测准确率

Table 3. Model detection accuracy with different convolu tional layers 单位：%

信噪比/dB	卷积层数
信噪比/dB	4	5	6
−20	98.32	99.48	99.64
−21	95.15	99.25	99.51
−22	92.82	99.10	99.20
−23	90.10	96.60	94.50
−24	87.85	93.28	90.15

下载: 导出CSV

表 4 不同卷积层数模型参数量

Table 4. Model parameter quantity with different convolution layers

卷积层数	参数量
4	362 90
5	39 394
6	42 498

下载: 导出CSV

表 5 1DCNN结构参数表

Table 5. Parameters of 1DCNN structure

结构	层参数	步长	填充	输入输出大小
输入层	—	—	—	1×25 000
卷积层	1×125, 32	16	Valid	1×1 555×32
池化层	1×3	2	Valid	1×776×32
卷积层	1×15, 32	4	Valid	1×191×32
池化层	1×3	2	Valid	1×96×32
卷积层	1×3, 64	2	Valid	1×47×64
池化层	1×3	2	Valid	1×24×64
卷积层	1×3, 32	1	Valid	1×22×32
池化层	1×3	2	Valid	1×11×32
卷积层	1×3, 32	1	Valid	1×9×32
GMP	32	—	—	—
全连接层	128	—	—	—
全连接层	2	—	—	—
Softmax	—	—	—	—

下载: 导出CSV

表 6 不同算法在数据集A和B的检测准确率

Table 6. Detection accuracy of different algorithms on data sets A and B 单位：%

算法	数据集A			数据集B
算法	SNR=−5 dB	SNR=−10 dB	SNR=−15 dB	数据集B
1DCNN	100.00	99.98	99.85	100.00
OCSVM	98.75	96.60	75.00	99.20
BP神经网络	95.20	90.00	63.75	98.43

下载: 导出CSV

表 7 变数据集检测准确率

Table 7. Detection accuracy of change dataset 单位：%

试验过程	SNR(数据集A)/dB
试验过程	−10	−15	−20
A→B	97.89	97.85	95.43
B→A	96.8	95.67	52.10

下载: 导出CSV

表 8 不同网络模块配置下的检测准确率

Table 8. Detection accuracy under different network module configurations

网络结构类别	数据集A检测准确率/%			数据集B 检测准确率/%	参数量
网络结构类别	SNR=−22 dB	SNR=−23 dB	SNR=−24 dB	数据集B 检测准确率/%	参数量
1(小卷积核和MP)	53.50	52.30	51.26	96.8	43 682
2(大卷积核和MP)	96.18	83.30	50.00	100.0	47 568
3(小卷积核和GMP)	53.99	53.21	52.53	98.2	35 450
4(大卷积核和GMP)	99.10	96.60	93.28	100.0	39 394

下载: 导出CSV

参考文献(11)

[1]	王成, 吴岩, 杨廷飞. 利用改进单分类支持向量机提升舰船尾流目标的检测准确率[J]. 兵工学报, 2020, 41(9): 1887-1893. Wang Cheng, Wu Yan, Yang Tingfei. Improving the detection accuracy of wake targets by improved one-class support vector machine[J]. Acta Armamentarii, 2020, 41(9): 1887-1893.
[2]	樊书宏, 严冰, 刘昆仑, 等. 舰船尾流侧向声检测方法[J]. 声学技术, 2011, 30(6): 474-479. Fan Shuhong, Yan Bing, Liu Kunlun, et al. A study of ships wake detection by side acoustic scattering[J]. Technical Acoustics, 2011, 30(6): 474-479.
[3]	Jeong S, Ban S W, Choi S, et al. Surface ship-wake detection using active sonar and one-class support vector machine[J]. IEEE Journal of Oceanic Engineering, 2012, 37(3): 456-466. doi: 10.1109/JOE.2012.2192344
[4]	金磊磊, 梁红, 杨长生. 基于卷积神经网络的水下目标声呐图像识别方法[J]. 西北工业大学学报, 2021, 39(2): 285-291. doi: 10.1051/jnwpu/20213920285 Jin Leilei, Liang Hong, Yang Changsheng. Sonar image recognition of underwater target based on convolutional neural network[J]. Joumal of Nonhwestem Polytechnical University, 2021, 39(2): 285-291. doi: 10.1051/jnwpu/20213920285
[5]	Ferguson E L, Ramakrishnan R, Williams S B, et al. Convolutional neural networks for passive monitoring of a shallow water environment using a single sensor[C]//2017 IEEE International Conference on Acoustics, Speech and Signal Processing(ICASSP). New Orleans, LA, USA: IEEE, 2017: 2657-2661.
[6]	Razzak M I, Naz S, Zaib A. Deep learning for medical image processing: Overview, challenges and the future[EB/OL]. (2017-04-22)[2022-08-29]. https://arxiv.org/abs/1704.06825v1.
[7]	R J 尤立克. 水声学原理[M]. 洪申, 译. 哈尔滨: 哈尔滨船舶工程学院出版社, 1990.
[8]	Keller J B, Miksis M. Bubble oscillations of large amplitude[J]. The Journal of the Acoustical Society of America, 1980, 68(2): 628-633. doi: 10.1121/1.384720
[9]	王虹斌. 水中气泡幕的多体多次声散射模型分析[J]. 船舶工程, 2006, 28(3): 30-33. doi: 10.13788/j.cnki.cbgc.2006.03.008 Wang Hongbin. Analysis of multi-body multi-dispersion model for bubble screen[J]. Ship Engineering, 2006, 28(3): 30-33. doi: 10.13788/j.cnki.cbgc.2006.03.008
[10]	Leifer I, Patro R K. The bubble mechanism for methane transport from the shallow sea bed to the surface: A review and sensitivity study[J]. Continental Shelf Research, 2002, 22(16): 2409-2428. doi: 10.1016/S0278-4343(02)00065-1
[11]	陈维兴, 崔朝臣, 李小菁, 等. 基于多种小波变换的一维卷积循环神经网络的风电机组轴承故障诊断[J]. 计量学报, 2006, 28(3): 30-33. Chen Weixing, Cui Chaochen, Li Xiaojing, et al. Bearing fault diagnosis of wind turbine based on multi-wavelet-1 d convolutional LSTM[J]. Acta Metrologica Sinica, 2006, 28(3): 30-33.