Page 34 - 《应用声学》2021年第4期
P. 34

第 40 卷 第 4 期                                                                       Vol. 40, No. 4
             2021 年 7 月                          Journal of Applied Acoustics                      July, 2021

             ⋄ 研究报告 ⋄



                     时频谱图和数据增强的水声信号深度学习


                                               目标识别方法                    ∗




                                        刘 峰     †   罗再磊 沈同圣 赵德鑫



                                           (军事科学院    国防科技创新研究院       北京   100071)
                摘要：水声目标识别一直是水声领域研究的重点问题之一，深度学习方法可以有效地解决目标识别问题，然
                而，水声样本的稀少限制了该方法的应用。该文提出一种基于数据增强的水声信号深度学习目标识别方法，该
                方法以 Mel 功率谱作为网络的输入特征，通过对原始信号在时域和时频域的拉伸和掩蔽等变换，实现数据扩
                展和增加泛化性能的目的，最后，利用改进的 VGG 网络模型实现目标分类。实验结果表明，该文方法得到的水
                下目标识别准确率 (95.2%) 要优于其他 4 种对比方法，证明了该文提出的网络模型和数据增强方法均有助于
                提高目标分类性能。
                关键词：水声目标识别；卷积神经网络；数据增强；Mel 功率谱
                中图法分类号: TP391.4           文献标识码: A         文章编号: 1000-310X(2021)04-0518-07
                DOI: 10.11684/j.issn.1000-310X.2021.04.004




             Deep learning target recognition method of underwater acoustic signal based on

                                data augmentation and time-frequency spectrum


                                  LIU Feng LUO Zailei SHEN Tongsheng ZHAO Dexin

                    (National Lnnovation Lnstitute of Defense Technology, Academy of Military Sciences, Beijing 100071, China)

                 Abstract: Underwater acoustic target recognition has always been one of the key issues in the field of underwa-
                 ter acoustic research. Deep learning methods can effectively solve the problem of target recognition. However,
                 the scarcity of underwater acoustic samples limits the application of this method. This paper proposes a deep
                 learning target recognition method for underwater acoustic signals based on data enhancement. This method
                 uses Mel power spectrum as the input feature of the network, and in order to increase the generalization per-
                 formance of the method, data augmentation is achieved by stretching and masking the original signal in the
                 time domain and time-frequency domain. Finally, using an improved VGG network model to achieve target
                 classification. The experimental results show that the underwater target recognition accuracy (95.2%) obtained
                 by this method is better than the other four comparison methods, which demonstrates that the network model
                 and data enhancement method proposed in this paper can help to improve the target classification performance.
                 Keywords: Underwater acoustic target recognition; Convolutional neural network; Data augmentation; Mel
                 spectrogram


             2020-09-27 收稿; 2021-01-10 定稿
             国家自然科学基金项目 (41906169)
             ∗
             作者简介: 刘峰 (1988– ), 男, 黑龙江哈尔滨人, 博士, 助理研究员, 研究方向: 声信号处理, 目标识别。
             † 通信作者 E-mail: liufeng_cv@126.com
   29   30   31   32   33   34   35   36   37   38   39