欢迎访问《兵工学报》官方网站,今天是 分享到:

兵工学报 ›› 2021, Vol. 42 ›› Issue (11): 2444-2452.doi: 10.3969/j.issn.1000-1093.2021.11.018

• 论文 • 上一篇    下一篇

基于生成对抗网络的水声目标识别算法

薛灵芝, 曾向阳, 杨爽   

  1. (西北工业大学 航海学院, 陕西 西安 710072)
  • 上线日期:2021-12-27
  • 通讯作者: 曾向阳(1974—),男,教授,博士生导师 E-mail:zenggxy@nwpu.edu.cn
  • 作者简介:薛灵芝(1987—),女,博士研究生。E-mail: 466698993@qq.com
  • 基金资助:
    国家自然科学基金项目(11774291)

Underwater Acoustic Target Recognition Algorithm Based on Generative Adversarial Networks

XUE Lingzhi, ZENG Xiangyang, YANG Shuang   

  1. (School of Marine Science and Technology, Northwestern Polytechnical University, Xi'an 710072, Shaanxi, China)
  • Online:2021-12-27

摘要: 目标识别是水声探测领域的难题,也是研究热点。在水声目标识别实际应用中,标记样本数量不足是制约识别结果的主要因素之一。针对水声目标噪声数据具有的小样本特点,基于深度学习理论提出一种基于生成对抗网络的识别模型。该模型从生成模型与对抗模型的相互博弈中,学习更多有效的识别特征信息,并与深度自编码网络和深度置信网络模型进行对比。仿真实验结果表明:在样本数量有限的情况下,生成对抗网络模型的识别效果优于深度置信网络与深度自编码网络;3种深度学习模型的识别性能均优于先提取梅尔倒谱系数特征,再用Softmax分类的方法。为进一步测试所建模型的性能,研究了3种深度学习模型在不同信噪比下的鲁棒性,仿真实验结果表明:生成对抗网络模型对噪声具有更强的鲁棒性。

关键词: 水声目标, 生成对抗网络, 目标识别, 深度学习, 小样本

Abstract: In the practical application of underwater acoustic target recognition,one of the main factors restricting the recognition results is the insufficient quantity of labeled samples. For the small sample properties of underwater acoustic target noise,a generative adversarial networks(GAN)-based recognition algorithm is proposed based on deep learning theory. It can be used to learn more effective features with more discriminative information from the game between generated model and adversarial model,and it is compared with deep auto-encoder(DAE) network and deep belief network(DBN) models. The experimental results illustrate that the recognition performance of GAN network model is higher than those of DBN network and DAE network models when the number of samples is limited,and the recognition performances of the three deep learning models are better than the conventional approach of extracting Mel frequency cepstrum coefficient(MFCC) features and then classifying by Softmax. In addition,GAN network model is superior to DBN network and DAE network models in recognition rate when using training samples and test samples with different SNRs. The smulation experimental results indicate that the GAN network model is more robust to noise.

Key words: underwateracoustic, generativeadversarialnetwork, targetrecognition, deeplearning, smallsample

中图分类号: