﻿ 一种神经网络的舰船目标噪声提取和分类技术
 舰船科学技术  2023, Vol. 45 Issue (20): 194-197    DOI: 10.3404/j.issn.1672-7649.2023.20.037 PDF

A neural network technique for extracting and classifying ship target noise
SUN Hang, WU Jian-jun, LI Lin
Kaifeng University, Kaifeng 475000, China
Abstract: This article analyzes the structural principles and learning methods of convolutional neural networks, and provides low-frequency line spectrum waveforms of ship noise, ocean spectrum curves, and line spectrum signal waveform curves. Finally, the ship target noise was extracted and classified, and the ship target noise feature response curve and noise signal feature value distribution map were obtained. The research in this article contributes to the further development of ship target noise processing technology.
Key words: neural network     noise extraction     classification technology
0 引　言

1 神经网络技术 1.1 卷积神经网络技术

 图 1 卷积神经网络整体结构 Fig. 1 The overall structure of convolutional neural networks

 $x_j^l = f\left( {u_j^l} \right)\text{，}$ (1)
 $u_j^l = \sum\limits_{i \in {M_j}} {x_i^{l - 1}} \times k_{ij}^l + b_j^l\text{。}$ (2)

 ${u_j} = b_j^ldown\left( {x_j^{l - 1}} \right) + b_j^l\text{。}$ (3)

 ${u^l} = {w^l}{x^{l - 1}} + {b^l}\text{。}$ (4)

1.2 神经网络学习过程

 $y_h^k = f\left( {\sum\limits_{i = 1}^{{N_1}} {{w_{ik}}gx_i^k + {\theta _k}} } \right)\text{，}$ (5)
 $z_j^k = f\left( {\sum\limits_{h = 1}^{{N_2}} {{w_{hj}}gx_h^k + {\gamma _j}} } \right)\text{。}$ (6)

 $\Delta \omega = - \eta \frac{{\partial E}}{{\partial \omega }}\text{，}$ (7)
 $\omega ' = \omega - \eta \frac{{\partial E}}{{\partial \omega }}\text{。}$ (8)

 $\Delta \omega \left( {k + 1} \right) = \left( {1 - {m_c}} \right)\nabla f\left( {\omega \left( k \right)} \right) + {m_c}\omega \text{。}$ (9)

 ${W_{ji}}\left( {t + 1} \right) = {W_{ji}}\left( t \right) + \eta \frac{{\partial E\left( t \right)}}{{\partial {W_{ji}}\left( t \right)}}\text{。}$ (10)

 $\eta \left( {t + 1} \right) = \eta \left( t \right) - \beta \frac{{\partial E}}{E}\text{，}$ (11)
 $\frac{{\Delta E}}{E} = \frac{{E\left( t \right) - E\left( {t - 1} \right)}}{{E\left( t \right)}}\text{。}$ (12)
2 舰船目标噪声建模仿真分析

 ${f_m} = mns\text{。}$ (13)

 $L\left( t \right) = \sum {{a_i}\left( {2{\text π} {f_i} + {\phi _i}\left( t \right)} \right)} \text{。}$ (14)

 图 2 舰船噪声低频线谱波形 Fig. 2 Low frequency line spectrum waveform of ship noise

 图 3 滤波器频率响应曲线 Fig. 3 Filter frequency response curve
 $X\left( t \right) = l\left( t \right) + [1 + a\left( t \right)]c\left( t \right)\text{。}$ (15)

 $SNR=10\mathrm{log}\left(\frac{信号功率}{噪声功率}\right) \text{。}$ (16)

 图 4 海洋噪声功率谱 Fig. 4 Ocean noise power spectrum

 图 5 线谱信号波形曲线 Fig. 5 Line spectrum signal waveform curve
3 舰船目标噪声提取及分类 3.1 舰船目标噪声提取

 $Mel\left( f \right) = 2\;595\lg \left( {1 + f/700} \right)\text{。}$ (17)

 图 6 Mel频率和实际频率之间的曲线关系 Fig. 6 Curve relationship between Mel frequency and actual frequency

 图 7 舰船目标噪声特征响应度曲线 Fig. 7 Characteristic response curve of ship target noise

 图 8 舰船目标噪声特征值大小曲线 Fig. 8 Ship target noise characteristic value size curve
3.2 舰船目标噪声分类

 $J=\frac{\left|S_w+S_b\right|}{\left|S_w\right|}\text{。}$ (18)

 $q = C_D^d = \frac{{D!}}{{\left( {D - d} \right)!d!}}\text{。}$ (19)

 图 9 识别准确率和训练次数之间的关系 Fig. 9 Identify the relationship between accuracy and the number of training sessions
4 结　语

 [1] 宋爱国, 陆佶人. 基于极限环的舰船噪声信号非线性特征分析及提取[J]. 声学学报, 1999(24): 407-415. [2] 刘洁, 陈劼, 韩冰, 等. 基于改进小波阈值去噪的深度学习水下目标分类[J]. 声学技术, 2023(42): 25-33. [3] 魏松杰, 张泽栋, 徐臻, 等. 基于多尺寸特征叠加的SAR舰船目标检测方法[J]. 湖南大学学报, 2021(48): 80-89. [4] 曾庆军, 王菲, 黄国建. 基于连续谱特征提取的被动声纳目标识别技术[J]. 上海交通大学学报, 2002(36): 382-386. [5] 王升贵, 胡桥, 陈迎亮, 等. 基于深度学习的水下目标识别方法研究[J]. 舰船科学技术, 2020(42): 141-145. [6] 李海涛, 刘振, 陈喆, 等. 融合人耳听觉特性与SAE模型的船舶辐射噪声分类方法[J]. 舰船科学技术, 2020(42): 172-176.