在海洋遥感领域，水声目标分类识别一直是声呐系统的一项困难而又极其重要的任务，为了进一步提高在不同信噪比下水下声目标的识别准确率，本文提出一种使用多维域融合特征分别输入双通道模型的水声目标识别方法。首先，通过梅尔频率倒谱系数（MFCC）和短时傅里叶变换（STFT）提取声信号在频域和时频上的特征进行融合；其次，构建密集卷积神经网络（DenseCNN）和长短期记忆网络（LSTM）2个通道，DenseCNN通道架构采用跳跃连接重用所有以前的特征映射，以优化各种受损条件下的分类率，并采用SE注意力机制使得动态调整特征权重。LSTM通道捕捉时间相关性，对模型进行长依赖关系处理能力的补充。实验结果表明，该方法在–20～10 dB信噪比下的分类准确率优于其他先进的神经网络模型。
In the field of Marine remote sensing, the classification and recognition of underwater acoustic targets has always been a difficult and extremely important task for sonar systems. In order to improve the accuracy of underwater acoustic targets under different signal-to-noise ratios, this paper proposes a method of underwater acoustic target recognition using multi-domain fusion features to input dual channel models respectively. First, the features of acoustic signal in frequency domain and time frequency are extracted by Mel-Frequency Cepstral Coefficients (MFCC) and short-time Fourier transform (STFT). Secondly, dense convolutional neural network (DenseCNN) and long short term memory network (LSTM) are constructed. The DenseCNN channel architecture uses skip connections to reuse all previous feature maps to optimize classification rates under various damaged conditions, and SE attention mechanism enables dynamic adjustment of feature weights. LSTM channels capture temporal dependencies and complement the model's ability to handle long dependencies. Experimental results show that the classification accuracy of the proposed method is better than other advanced neural network models at –20～10 db SNR.
2024,46(20): 142-147 收稿日期：2024-1-2
DOI：10.3404/j.issn.1672-7649.2024.20.026
分类号：TB566
基金项目：陕西自然科学青年基金资助项目（2021JQ693）
作者简介：张晨颖(2000-),男,硕士研究生,研究方向为声信号识别
参考文献：
[1] 张奇, 笪良龙, 王超, 等. 基于深度学习的水声被动目标识别研究综述[J]. 电子与信息学报, 2023, 45(11): 4190-4202.
[2] DOAN V S, HUYNH T, KIM D S. Underwater acoustic target classification based on dense convolutional neural network[J]. IEEE Geoscience and Remote Sensing Letters, 2020, 19: 1-5.
[3] ZHOU A, ZHANG W, LI X, et al. A Novel Noise-Aware Deep Learning Model for Underwater Acoustic Denoising[J]. IEEE Transactions on Geoscience and Remote Sensing, 2023, 61: 1-13.
[4] WANG B, ZHANG W, ZHU Y, et al. An underwater acoustic target recognition method based on AMNet[J]. IEEE Geoscience and Remote Sensing Letters, 2023, 20: 1-5.
[5] 刘承伟, 洪峰, 冯海泓, 等. 结合多尺度卷积网络和双端注意力机制的水声目标识别[J]. 声学技术, 2023, 42(2): 161-167.
[6] 杜柏润, 章博. 基于卷积和循环网络的水声信号联合特征表示和识别方法[J]. 舰船科学技术, 2023, 45(15): 107-110.
DU Borun, ZHANG Bo. Joint feature representation and recognition method of underwater acoustic signals based on convolutional and cyclic networks[J]. Ship Science and Technology, 2023, 45(15): 107-110.
[7] 赵乾坤, 刘峰, 梁秀兵, 等. 基于时延神经网络模型的舰船辐射噪声目标识别[J]. 应用声学, 2023, 42(5): 1033–1041.
[8] 李紫鹏, 纪永强, 郭兵勇, 等. 基于特征模式分解的水声目标特征提取方法[J]. 哈尔滨工程大学学报, 2023, 44(9): 1542-1548.
[9] 李林桦, 王学渊, 李小霞. 融合多维时频特征的环境声识别[J]. 制造业自动化, 2021, 43(6): 49–52+94.
[10] 姜子星. 基于特征融合与软注意力机制的环境声识别技术研究[D].广州: 广东工业大学, 2022.
