Application of Convolutional Neural Networks to the Detection of Non-cooperative Underwater Acoustic Pulse Signals
-
摘要: 针对非合作水声脉冲信号的检测问题, 文中提出了一种基于轻量化CNN的自动检测方法。在将一维时域声信号映射为二维时频谱图并执行降维与标准化预处理后, 引入到融合通道注意力机制的轻量化神经网络中, 通过自主学习和提取信号在图谱上的连续能量特征, 自适应地提高与目标时频结构相关通道的响应, 并抑制噪声相关通道的干扰, 将预处理过的数据处理为特征明显的128维数据。同时采用辅助分类机制, 利用交叉熵损失的排他性区分信噪, 根据128维数据对信号进行分类, 最终实现非合作信号的有效检测。仿真与盲测实验结果表明, 该方法在不同信噪比条件下均表现出良好的检测准确率和抗干扰能力。即使面对模型未知的信号类型, 依然能够保持较高的截获率和较低的虚警率, 为复杂海洋环境下的水声目标探测提供了有价值的技术参考。Abstract: To address the detection problem of non-cooperative underwater acoustic pulse signals, this paper proposes an automatic detection method based on a lightweight convolutional neural network. After mapping one-dimensional time-domain acoustic signals into two-dimensional time-frequency spectrograms and performing dimensionality reduction and normalization, the preprocessed data is fed into a lightweight neural network incorporating a channel attention mechanism. By autonomously learning and extracting the continuous energy features on the spectrograms, the network adaptively assigns high weights to features containing real targets while suppressing noise-related channel responses, thereby transforming the data into highly discriminative 128-dimensional features. Meanwhile, an auxiliary classification mechanism is employed to distinguish between signal and noise utilizing the exclusivity of cross-entropy loss. By classifying the signals based on these 128-dimensional features, the effective detection of non-cooperative signals is ultimately achieved. Simulation and blind test results demonstrate that the proposed method exhibits excellent detection accuracy and robust anti-interference capabilities under various signal-to-noise ratio conditions. Even when confronted with signal types unknown to the model, it maintains a high interception rate and a low false alarm rate, providing a valuable technical reference for underwater acoustic target detection in complex marine environments.
-
表 1 网络结构参数
Table 1. Network Configuration
模块 具体设置 输出尺寸 输入 单通道时频特征图 256×256×1 卷积模块1 卷积: 1→8, 卷积核3×3
步长1, 填充1
BN→ReLU→SE(8→2→8)
最大池化2×2, 步长2128×128×8 卷积模块2 卷积: 8→16, 卷积核3×3
步长1, 填充1
BN→ReLU→SE(16→4→16)
最大池化2×2, 步长264×64×16 卷积模块3 卷积: 16→32, 卷积核3×3
步长1, 填充1
BN→ReLU→SE(32→8→32)
最大池化2×2, 步长232×32×32 双分支
自适应池化平均池化与最大池化并行,
输出空间尺寸均为4×4每支路4×4×32 通道拼接
与展平两支路沿通道维拼接, 再展平 1 024维 特征映射层 全连接: 1 024→128
层归一化→ReLU→随机失活128维 辅助分类层 全连接: 128→7
归一化指数函数计算类别概率7维 表 2 实测数据盲测检测结果
Table 2. Blind test results of actual measured data
实测数据类别 测试
样本数判为第0类
(无信号)判为1-6类
(截获信号)噪声 1 1 0 信号 54 3 51 总体统计 55 4 51 表 5 代表性方法在低SNR条件下的检测结果
Table 5. Detection results of representative methods under low SNR conditions
方法 −22 dB AUC −20 dB AUC −20 dB时FPR≤
1%的最高Pd(%)频带能量检测 0.534 8 0.540 1 1.00 无SE卷积网络 0.778 3±0.011 9 0.887 7± 0.0149 48.00±4.51 MobileNetV3-
Small0.790 6± 0.0183 0.9045 ±0.016 556.89±3.22 文中方法 0.799 1±0.014 2 0.913 2±0.013 5 54.11±3.51 表 3 SE压缩比与特征维度的对照结果
Table 3. Comparison results of SE compression ratio and feature dimension
对照设置 参数量 未见信道
7分类准确率(%)−20 dBAUC 压缩比2, 128维 139 787 86.48±1.75 0.910 6±0.006 1 压缩比4,
128维(本文中)139 101 88.38±2.29 0.913 2±0.013 5 压缩比8, 128维 138 758 87.24±1.19 0.919 6±0.015 5 压缩比4, 64维 72 925 86.71±1.25 0.900 3±0.011 5 压缩比4, 256维 271 453 87.43±2.49 0.912 9±0.006 9 表 4 不同分类框架与训练目标的对照结果
Table 4. Comparison results of different classification frameworks and training objectives
方法 7分类
准确率(%)−20 dB AUC 直接2分类 不适用 0.968 7±0.009 2 仅7分类交叉熵 87.67±2.37 0.903 7±0.009 0 文中方案 88.38±2.29 0.913 2±0.013 5 表 6 不同网络的检测性能与计算开销
Table 6. Detection performance and computational overhead of different networks
方法 −20 dB AUC Pd/%(Pfa≤1%) 参数量/万 计算量/百万次 CPU推理时间/ms GPU推理时间/ms 无SE卷积网络 0.887 7±0.014 9 48.00±4.51 13.835 9 42.60 4.86 0.95 MobileNetV3-Small 0.904 5±0.016 5 56.89±3.22 152.474 3 67.44 39.55 7.59 文中方法 0.913 2±0.013 5 54.11±3.51 13.910 1 42.60 5.23 1.74 -
[1] 翟羽飞, 闫成琳, 周天, 等. 微弱非合作瞬态信号自适应检测算法[J]. 声学技术, 2025, 46: 1-13. doi: 10.16300/j.cnki.1000-3630.25063001Zhai Y F, Yan C L, Zhou T, et al. Adaptive detection algorithm for weak non-cooperative transient signal[J]. Technical Acoustics, 2025, 46: 1-13. doi: 10.16300/j.cnki.1000-3630.25063001 [2] 王晓燕, 方世良, 朱志峰. 非合作水声脉冲信号的熵检测方法[C]//中国声学学会. 泛在信息社会中的声学——中国声学学会2010年全国会员代表大会暨学术会议论文集. 《声学技术》, 2010: 80-81. [3] 赵安琪. 未知参量水声脉冲信号检测技术研究[D]. 哈尔滨: 哈尔滨工程大学, 2018. [4] 王晓燕, 方世良, 朱志峰. 基于ST-FRFT的非合作水声脉冲信号检测方法[J]. 信号处理, 2011, 27(8): 1271-1278. doi: 10.3969/j.issn.1003-0530.2011.08.026Wang X Y, Fang S L, Zhu Z F. Detection of non-cooperative underwater acoustic pulse signal based on ST-FRFT[J]. Signal Processing, 2011, 27(8): 1271-1278. doi: 10.3969/j.issn.1003-0530.2011.08.026 [5] 邓红超, 张刚强, 刘本奇. 基于时频联合的非合作水声信号盲检测[J]. 舰船科学技术, 2023, 45(1): 157-162. doi: 10.3404/j.issn.1672-7649.2023.01.028Deng H C, Zhang G Q, Liu B Q. Blind detection of non-cooperative underwater acoustic signals based on time-frequency domain association[J]. Ship Science and Technology, 2023, 45(1): 157-162. doi: 10.3404/j.issn.1672-7649.2023.01.028 [6] 李萌逸, 李记龙, 冯海泓. 双模型联合优化的水声通信信号检测与识别算法[J]. 声学学报, 2025, 50(5): 1327-1337. doi: 10.12395/0371-0025.2024163Li M Y, Li J L, Feng H H. Detection and recognition algorithm for underwater acoustic communication signals based on dual model joint optimization[J]. Acta Acustica, 2025, 50(5): 1327-1337. doi: 10.12395/0371-0025.2024163 [7] 王志桢, 王晓燕, 安良, 等. 水下小目标主动声辐射信号截获处理技术研究综述[J]. 水下无人系统学报, 2026, 34(3): 454-467,499. doi: 10.11993/j.issn.2096-3920.2026-0068Wang Z Z, Wang X Y, An L, et al. A review on interception and processing techniques of active acoustic radiated signals from small underwater targets[J]. Journal of Unmanned Undersea Systems, 2026, 34(3): 454-467,499. doi: 10.11993/j.issn.2096-3920.2026-0068 [8] 易振通, 吴瑰, 官端正, 等. 轻量化CNN的研究综述[J]. 工业控制计算机, 2022, 35(10): 109-111,114.Yi Z T, Wu G, Guan D Z, et al. Survey of research on lightweight convolutional neural networks[J]. Industrial Control Computer, 2022, 35(10): 109-111,114. [9] 巩文静, 李宇, 丁飞龙, 等. 基于注意力残差网络的非合作水声通信信号自动调制识别[J]. 声学学报, 2026, 51(1): 298-309. doi: 10.12395/0371-0025.2025139Gong W J, Li Y, Ding F L, et al. Automatic modulation recognition of non-cooperative underwater acoustic communication signals based on attention residual networks[J]. Acta Acustica, 2026, 51(1): 298-309. doi: 10.12395/0371-0025.2025139 [10] Yatigul R, Sittiwanchai T, Blattler A, et al. Enhancing squeeze-and-excitation networks with mixed pooling for image classification[C]//Proceedings of the 31st International Conference on Artificial Life and Robotics, ICAROB 2026. 2026: 39-42. [11] 王莉娜, 杨咚咚, 刘明明, 等. 短时傅里叶变换的参数选择方法[J]. 火控雷达技术, 2026, 55(1): 73-76. doi: 10.19472/j.cnki.1008-8652.2026.01.013Wang L N, Yang D D, Liu M M, et al. A method for selecting short-time Fourier transform parameters[J]. Fire Control Radar Technology, 2026, 55(1): 73-76. doi: 10.19472/j.cnki.1008-8652.2026.01.013 [12] 黄海宁, 李宝奇, 刘纪元, 等. 声呐图像水下目标识别综述与展望[J]. 电子与信息学报, 2024, 46(5): 1742-1760. doi: 10.11999/jeit231207Huang H N, Li B Q, Liu J Y, et al. Sonar image underwater target recognition: A comprehensive overview and prospects[J]. Journal of Electronics & Information Technology, 2024, 46(5): 1742-1760. doi: 10.11999/jeit231207 [13] 孙伟, 彭华, 李天昀, 等. 基于时频图像处理的宽带特定信号检测方法[J]. 信息工程大学学报, 2020, 21(5): 545-551. doi: 10.3969/j.issn.1671-0673.2020.05.006Sun W, Peng H, Li T Y, et al. Broadband specific signal detection method based on time-frequency image processing[J]. Journal of Information Engineering University, 2020, 21(5): 545-551. doi: 10.3969/j.issn.1671-0673.2020.05.006 [14] 兰州恒达彩印包装有限责任公司. 一种双三次图像插值模型的参数优化方法: CN201911310859.1[P]. 2023-10-27. [15] 东南大学;网络通信与安全紫金山实验室. 一种基于自监督学习的雷达信号分选识别方法: CN202210103370.2[P]. 2024-04-16. [16] 刘峙江, 赵晴晴, 徐立军. 基于轻量化CNN的水声通信前导信号检测方法[J]. 信号处理, 2023, 39(10): 1831-1841. doi: 10.16798/j.issn.1003-0530.2023.10.010Liu Z J, Zhao Q Q, Xu L J. Preamble signal detection method of underwater acoustic communication based on lightweight convolutional neural network[J]. Journal of Signal Processing, 2023, 39(10): 1831-1841. doi: 10.16798/j.issn.1003-0530.2023.10.010 -

下载: