Page 29 - 《应用声学》2023年第3期
P. 29

第 42 卷 第 3 期                                                                       Vol. 42, No. 3
             2023 年 5 月                          Journal of Applied Acoustics                      May, 2023

             ⋄ 研究报告 ⋄



                              卷积神经网络主动目标方位估计






                            王珍珠     1,2   赵 猛    1,2    任群言     1,2†  肖 旭     1,2   马 力     1,2


                                              (1 中国科学院声学研究所       北京   100190)
                                                (2 中国科学院大学      北京  100049)

                摘要：复杂海洋环境中信道的传输特性、时空变化、频散效应等一定程度上制约了主动声呐目标方位估计的性
                能。该文引入卷积神经网络，提出了适用于主动声呐中目标方位的高精度估计方法。仿真声场环境为浅海负
                梯度，主动发射信号为具有多普勒不变性质的双曲调频信号，水平线列阵作为接收装置，目标按仿真路线运动。
                该文利用 Kraken 进行声场数据仿真，并对接收的信号在频域做均匀加权常规波束形成，进而进行卷积神经网
                络的模型训练和测试。数值仿真研究表明，该文所用方法可以有效估计目标波达方向，对信噪比具有一定的鲁
                棒性。
                关键词：主动声呐；卷积神经网络；常规波束形成
                中图法分类号: TN911.7           文献标识码: A          文章编号: 1000-310X(2023)03-0467-07
                DOI: 10.11684/j.issn.1000-310X.2023.03.003






                  Active target azimuth estimation based on convolutional neural network



                      WANG Zhenzhu    1,2  ZHAO Meng  1,2  REN Qunyan   1,2  XIAO Xu  1,2  MA Li 1,2

                                (1 Institute of Acoustics Chinese Academy of Sciences, Beijing 100190, China)
                                   (2 University of Chinese Academy of Sciences, Beijing 100049, China)

                 Abstract: The performance of active sonar target azimuth estimation is restricted to some extent by channel
                 transmission characteristics, temporal and spatial variation and dispersion effect in complex ocean environment.
                 In this paper, a high precision estimation method for target orientation in active sonar is proposed used
                 convolutional neural network (CNN). The simulated acoustic field environment is shallow sea negative gradient,
                 the active transmitting signal is hyperbolic frequency modulation signal with Doppler invariant property, the
                 horizontal line array is the receiving device, and the target moves according to the simulated route. In this
                 paper, Kraken is used for sound field data simulation, and the received signal is uniformly weighted conventional
                 beamforming in the frequency domain, and then the model training and testing of CNN network is carried
                 out. Numerical simulation results show that the proposed method can effectively estimate the target arrival
                 direction and is robust to the signal-to-noise ratio.
                 Keywords: Active sonar; Convolutional neural network; Conventional beamforming



             2022-02-17 收稿; 2022-04-22 定稿
             作者简介: 王珍珠 (1998– ), 女, 河南商丘人, 硕士研究生, 研究方向: 信号与信息处理。
             † 通信作者 E-mail: renqunyan@mail.ioa.ac.cn
   24   25   26   27   28   29   30   31   32   33   34