基于一维卷积和LSTM网络的端到端水声目标识别  

End-to-end acoustic target recognition based on 1D convolutional and LSTM networks

作　　者：杨康[1,2] YANG Kang(Zhenjiang College of Technology,Zhenjiang 212003,China;Jiangsu University of Science and Technology,Zhenjiang 212003,China)

机构地区：[1]镇江市高等专科学校,江苏镇江212003 [2]江苏科技大学,江苏镇江212003

出　　处：《无线互联科技》2024年第23期73-76,共4页Wireless Internet Science and Technology

基　　金：镇江市高等专科学校科研项目,项目名称:基于水声的水下目标识别研究,项目编号:GZYB202319;江苏省研究生科研创新项目,项目名称:基于MIMO-OTFS的水声信道估计与均衡技术研究,项目编号:KYCX23_3908。

摘　　要：水声目标识别在国防和海洋环境监测等领域具有重要应用。然而,传统的时频域特征提取方法由于信息损失和环境适应性不足,限制了识别性能的提升。为克服这些局限性,文章提出了一种基于一维卷积神经网络(One-dimensional Convolutional Neural Network,1D CNN)与长短时记忆网络(Long Short-term Memory Network,LSTM)相结合的端到端水声目标识别模型(One-dimensional Long Short-term Memory,1DLSTM)。该模型直接以原始时域信号为输入,利用1D CNN提取局部特征,通过LSTM捕捉长程依赖关系,有效保留了信号的全局信息。在ShipsEar数据集上的实验结果表明,该模型的识别准确率高达93.91%,为水声目标端到端识别领域提供了一种新思路。Acoustic target recognition plays a crucial role in defense and marine environment monitoring.However,traditional time-frequency domain feature extraction methods often suffer from information loss and inadequate adaptability to varying environments,limiting their recognition performance.To address these limitations,this paper presents an end-to-end acoustic target recognition model(1DLSTM)that combines a one-dimensional convolutional neural network(1D CNN)with a long short-term memory network(LSTM).This model directly processes raw time-domain signals,using the 1D CNN to extract local features and the LSTM to capture long-term dependencies,thereby effectively preserving the global information of the signal.Experimental results on the ShipsEar dataset demonstrate that this model achieves a recognition accuracy of 93.91%,offering a novel approach to end-to-end acoustic target recognition.

关 键 词：深度学习 水声目标识别 端到端 

分 类 号：TB566[交通运输工程—水声工程] TP183[理学—物理]