电子学报 ›› 2022, Vol. 50 ›› Issue (6): 1359-1369.DOI: 10.12263/DZXB.20210916

所属专题: 电磁频谱智能+

• 电磁频谱智能+ • 上一篇    下一篇

融合深度特征的电磁频谱异常检测算法

彭闯, 王伦文(), 胡炜林   

  1. 国防科技大学电子对抗学院,安徽 合肥 230037
  • 收稿日期:2021-07-15 修回日期:2021-12-27 出版日期:2022-06-25
    • 通讯作者:
    • 王伦文
    • 作者简介:
    • 彭 闯 男,1994年出生,山东曲阜人.国防科技大学电子对抗学院博士研究生.主要研究方向为深度学习、电磁环境态势认知等.E-mail: pengchuang17@nudt.edu.cn
      王伦文(通讯作者) 男,1966年出生,安徽休宁人.国防科技大学电子对抗学院教授,博士生导师.主要研究方向为智能信息处理、数据挖掘、电磁环境态势认知等.
      胡炜林 男,1997年出生,四川泸州人.国防科技大学电子对抗学院硕士研究生.主要研究方向为电磁态势认知、机器学习等.E-mail: hwl550@nudt.edu.cn
    • 基金资助:
    • 国家自然科学基金 (11975307); 国防科技创新特区项目 (19-H863-01-ZT-003-003-12)

Spectrum Anomaly Detection Algorithm Based on the Fusion of Depth Feature

PENG Chuang, WANG Lun-wen(), HU Wei-lin   

  1. College of Electronic Engineering, National University of Defense Technology, Hefei, Anhui 230037, China
  • Received:2021-07-15 Revised:2021-12-27 Online:2022-06-25 Published:2022-06-25
    • Corresponding author:
    • WANG Lun-wen
    • Supported by:
    • National Natural Science Foundation of China (11975307); National Defense Technology Innovation Special Zone Project (19-H863-01-ZT-003-003-12)

摘要:

针对电磁频谱异常检测效率不高等问题,该文结合卷积神经网络(Convolutional Neural Networks,CNN)以及长短时记忆神经(Long Short-Term Memory,LSTM)网络,提出一种融合深度特征的电磁频谱异常检测算法.首先构建深度特征提取网络,该网络包含能够分级提取深度特征的两路多层CNN以及LSTM;其次通过池化、合并等操作将网络模型提取的各级深度特征进行融合,实现频谱数据预测;最后计算预测数据与真实数据的均方误差,判别频谱异常.该算法能在无监督学习的条件下,实现多种类异常状态检测的.在公开频谱数据的多个频段对算法性能进行验证,结果表明本文算法能够有效地实现电磁频谱异常检测.

关键词: 深度学习, 异常检测, 频谱预测, 特征融合

Abstract:

To solve the problem of low efficiency of electromagnetic spectrum anomaly detection, we propose a new method of spectrum anomaly detection based on depth feature fusion which combines convolutional neural networks(CNN) and long short-term memory(LSTM) Networks. Firstly, a deep feature extraction network is constructed, which includes a multi-level CNN and a LSTM. The network can extract depth features in a hierarchical manner. Then, pooling layer, concatenate layer and other operations are used to fuse the depth features to achieve high-precision prediction of spectrum data. Finally, the mean square error between the predicted data and the real data is calculated by discriminating the spectrum anomaly. The algorithm can detect multiple kinds of abnormal states under the condition of unsupervised learning. We verified our algorithm in frequency bands of public spectrum data. The results show that our algorithm can effectively realize electromagnetic spectrum anomaly detection.

Key words: deep learning, anomaly detection, spectrum prediction, fusion of depth feature

中图分类号: