电子学报 ›› 2022, Vol. 50 ›› Issue (6): 1281-1290.DOI: 10.12263/DZXB.20210805

所属专题: 长摘要论文 电磁频谱智能+

• 电磁频谱智能+ •    下一篇

基于孪生网络的电磁目标跨模式识别算法

张伟1,2, 王沙飞3, 林静然1,4(), 利强1,4, 邵怀宗1,4   

  1. 1.电子科技大学信息与通信工程学院,四川 成都 611731
    2.电子信息控制重点实验室,四川 成都 610036
    3.北方电子设备研究所,北京 100191
    4.鹏城实验室,广东 深圳 518055
  • 收稿日期:2021-06-29 修回日期:2021-12-23 出版日期:2022-06-25
    • 通讯作者:
    • 林静然
    • 作者简介:
    • 张 伟 男,1985年出生,山东泰安人.高级工程师,博士研究生.主要研究方向为非合作通信信号处理、电子对抗.
      王沙飞 男,1964年出生,河北张家口人.研究员,博士生导师.主要研究方向为非合作通信信号处理、电子对抗.
      林静然(通讯作者) 男,1978年出生,重庆人.教授,博士生导师.主要研究方向为电磁频谱智能管控与优化.
      利 强 男,1982年出生,四川成都人.副教授,博士生导师.主要研究方向为无线通信信号处理与优化.
      邵怀宗 男,1969年出生,四川巴中人.教授,博士生导师.主要研究方向为通信和雷达信号处理.
    • 基金资助:
    • 国家自然科学基金 (U19B2028)

Cross-Modal Recognition Algorithm of Electromagnetic Targets via Siamese Network

ZHANG Wei1,2, WANG Sha-fei3, LIN Jing-ran1,4(), LI Qiang1,4, SHAO Huai-zong1,4   

  1. 1.School of Information and Communication Engineering, University of Electronic Science and Technology of China, Chengdu, Sichuan 611731, China
    2.Science and Technology on Electronic Information Control Laboratory, Chengdu, Sichuan 610036, China
    3.Northern Institute of Electronic Equipment of China, Beijing 100191, China
    4.Peng Cheng Laboratory, Shenzhen, Guangdong 518055, China
  • Received:2021-06-29 Revised:2021-12-23 Online:2022-06-25 Published:2022-06-25
    • Corresponding author:
    • LIN Jing-ran
    • Supported by:
    • National Natural Science Foundation of China (U19B2028)

摘要:

以深度学习为代表的人工智能技术是解决电磁目标识别问题的一种有效方法.然而,在识别多模式电磁目标时,目标内部不同模式间数据的差异可能掩盖目标个体间的差异,当某种模式训练样本缺失或稀少时,该模式下的目标识别性能会显著下降.为此,提出一种基于孪生网络的电磁目标跨模式识别算法,在度量学习框架下通过优化设计网络结构和损失函数,引导网络在分类学习过程中拉近同一目标各模式数据间的距离,拉远不同目标数据间的距离,并结合邻近判决准则实现多模式电磁目标在非均衡数据集上的跨模式识别.基于实际数据的测试结果表明,在相同数据集和网络规模条件下,所提方法的跨模式识别率较经典卷积神经网络方法和数据增强方法提升20%.

长摘要
以深度学习为代表的人工智能技术是解决电磁目标识别问题的一种有效方法。然而,现有电磁目标设计日益复杂,大多可以灵活工作在多种差异巨大的模式上,这给目标识别带来新挑战。在识别多模式电磁目标时,目标内部不同模式间数据的差异可能掩盖目标个体间的差异,当某种模式训练样本缺失或稀少时,该模式下的目标识别性能会显著下降,甚至会将工作在不同模式的同一目标误判为多个目标。为此,提出一种基于孪生网络的电磁目标跨模式识别算法,在度量学习框架下通过优化设计网络结构和损失函数,引导网络在分类学习过程中拉近同一目标各模式数据间的距离,拉远不同目标数据间的距离,已达到提取不同模式间共有特征的目的,并结合邻近判决准则实现多模式电磁目标在非均衡数据集上的跨模式识别。在此基础上,进一步设计了联合判决反馈机制,当罕见模式下的测试数据到来时,利用多样本进行联合判决,根据当前判决结果标记新到数据,并利用新标注的数据反馈和更新网络,使网络具备自增长能力,逐渐提升罕见模式下的识别性能。基于实际数据的测试结果表明,在相同数据集和网络规模条件下,所提方法的跨模式识别率较经典卷积神经网络方法和数据增强方法提升20%。

关键词: 电磁目标, 跨模式识别, 孪生网络, 度量学习, 非均衡数据集

Abstract:

The artificial intelligence technology(e.g., deep learning) is an effective approach to electromagnetic target(ET) recognition. However, in the recognition of multi-mode ETs, when the training samples with a certain mode are missing or rare, the recognition rate with this mode will be significantly degraded. The reason mainly lies in that the data distance between different modes of the same ET may exceed the data distance between different ETs. To remedy this, a cross-modal ET recognition approach via Siamese network is developed in this paper. Following the framework of metric learning, we design the network structure and the loss function carefully, so that the recognition training process intentionally drives the Siamese network to enlarge the data distance between different ETs while shorten the distance between different modes of the same ET. Consequently, the multi-mode ETs can still be successfully recognized by employing certain distance-based decision criterion, even with imbalanced training data sets for different modes. The numerical results based on realistic data show that with same data sets and network size, the cross-modal recognition rate of the proposed approach is 20% higher than that of the classical convolutional neural network approach, and that of the popular data-enhancement approaches.

Extended Abstract
The artificial intelligence technology (e.g., deep learning) is an effective approach to electromagnetic target (ET) recognition. However, the design of current ETs is increasingly complex, and most of them can flexibly work on a variety of modes with huge differences, which brings new challenges to ET recognition. Specifically, in the recognition of multi-mode ETs, when the training samples with a certain mode are missing or rare, the recognition rate with this mode will be significantly degraded, or the same target working in different modes may even be mistakenly recognized as multiple targets. The reason mainly lies in that the data distance between different modes of the same ET may exceed the data distance between different ETs. To remedy this, a cross-modal ET recognition approach via Siamese network is developed in this paper. Following the framework of metric learning, we design the network structure and the loss function carefully, so that the recognition training process intentionally drives the Siamese network to enlarge the data distance between different ETs while shorten the distance between different modes of the same ET. As a result, the common features among the signals with different modes of the same target can be extracted, and hence the multi-mode ETs can still be successfully recognized by employing certain distance-based decision criterion, even with imbalanced training data sets for different modes. On this basis, a joint-decision-based feedback mechanism is further designed. When the test data with the rare mode arrives, multiple samples are used for joint decision. Then, the new-coming data is marked according to the joint decision results, and the network is re-trained with the new-labeled data. Consequently, the network gradually improves the recognition performance with the rare mode. The numerical results based on realistic data show that with same data sets and network size, the cross-modal recognition rate of the proposed approach is 20% higher than that of the classical convolutional neural network approach, and that of the popular data-enhancement approaches.

Key words: electromagnetic targets, cross-modal recognition, Siamese network, metric learning, imbalanced data set

中图分类号: