数据与计算发展前沿 ›› 2025, Vol. 7 ›› Issue (4): 42-53.

CSTR: 32002.14.jfdc.CN10-1649/TP.2025.04.004

doi: 10.11871/jfdc.issn.2096-742X.2025.04.004

• 专刊：空间科学大数据智能算法模型与工具 • 上一篇    下一篇

基于Geoformer的地磁超低频波智能识别研究

方少峰1,2(),邹自明1,2,*()   

  1. 1.中国科学院国家空间科学中心， 北京 100190
    2.国家空间科学数据中心， 北京 100190
  • 收稿日期:2025-05-20 出版日期:2025-08-20 发布日期:2025-08-21
  • 通讯作者: 邹自明
  • 作者简介:方少峰，博士，高级工程师，中国科学院国家空间科学中心，中国科学院青促会成员，主要研究方向为数据挖掘与空间天气学的交叉融合研究。
    本文中负责论文撰写与实验论证。
    FANG Shaofeng, Ph.D, is a senior engineer at the National Space Science Center CAS, and a member of Youth Innovation Promotion Association CAS. His main research interests include the intersection of data mining and space weather science.
    In this paper, he is responsible for writing the manuscript and conducting the experimental validation.
    E-mail: fangsf@nssc.ac.cn|邹自明，博士，研究员，博士生导师，中国科学院国家空间科学中心，国家空间科学数据中心主任。主要研究方向为空间科学与数据科学交叉领域研究。
    本文中负责论文架构设计，提出修改意见。
    ZOU Ziming, Ph.D, is a professor and Ph.D supervisor at the National Space Science Center, CAS, and the director of the National Space Science Data Center. His main research interests include the intersection of space science and data science.
    In this paper, he is responsible for designing the article’s overall structure and making revision suggestions.
    E-mail: mzou@nssc.ac.cn
  • 基金资助:
    国家重点研发计划“空间科学大数据智能管理与分析挖掘关键技术及应用”项目(2022YFF0711400)

Automated Detection of Geomagnetic Ultra-Low Frequency Waves Based on Geoformer

FANG Shaofeng1,2(),ZOU Ziming1,2,*()   

  1. 1. National Space Science Center, Chinese Academy of Sciences, Beijing 100190, China
    2. National Space Science Data Center, Beijing 100190, China
  • Received:2025-05-20 Online:2025-08-20 Published:2025-08-21
  • Contact: ZOU Ziming

RichHTML

6

PDF

6

摘要：

【目的】针对传统方法在噪声环境下地磁超低频波识别精度不足、长时序依赖特征建模能力较弱的问题，本文提出一种基于深度可分离卷积和时序Transformer的地磁超低频波识别框架Geoformer，通过融合卷积神经网络与改进型Transformer架构，结合优化的位置编码策略，提升地磁超低频波的识别准确率与模型泛化能力。【方法】该方法利用卷积神经网络提取地磁信号的局部时域特征与多通道空间相关性，同时通过时间绝对位置编码与高效相对位置编码增强模型对信号序列绝对位置与相对距离的感知能力，最终借助多头自注意力机制捕捉长时序依赖与多维度交互特征，实现精准分类。【结果】实验结果表明，相较于传统CNN与基础Transformer模型，Geoformer在真实地磁数据集上分类准确率提升12.2%，且明显优于LSTM、GRU、Resnet等传统深度学习模型。【局限】模型计算复杂度随信号长度增长呈二次增长，在超长时间序列实时处理中需结合降采样技术，且依赖高质量标注数据，在小样本场景下需引入迁移学习或自监督预训练。【结论】本文通过卷积-Transformer架构创新与位置编码优化，为地磁超低频波信号的智能分析提供了高效解决方案，有望在地球物理监测、空间天气预警等领域发挥重要作用。

关键词: 地磁超低频波, Geoformer, 信号分类

Abstract:

[Objective] Considering the limitations of traditional methods, such as insufficient classification accuracy of geomagnetic ultra-low frequency (ULF) waves in noisy environments and inferior capability in modeling long-term temporal dependencies, we propose Geoformer, a novel framework integrating depthwise separable convolution and temporal Transformer. By fusing convolutional neural networks (CNNs) with an improved Transformer architecture and optimizing positional encoding strategies, Geoformer enhances the recognition accuracy and model generalization ability for ULF waves. [Methods] This method uses CNNs to extract the local time-domain features and multi-channel spatial correlations of geomagnetic signals. Meanwhile, through time absolute positional encoding (tAPE) and efficient relative positional encoding (eRPE), it strengthens the model’s perception of the absolute positions and relative distances of signal sequences. Finally, it employs the multi-head self-attention mechanism to capture long-term temporal dependencies and multi-dimensional interactive features, achieving accurate classification. [Results] Experimental results indicate that, compared with traditional CNNs and basic Transformer models, Geoformer improves the classification accuracy by 12.2% on real world geomagnetic datasets and is significantly superior to traditional deep learning models like LSTM, GRU, and ResNet. [Limitations] The model’s computational complexity grows quadratically with the increase of signal length, necessitating the incorporation of downsampling techniques in real-time processing of ultra-long time sequences. Moreover, it relies on high-quality labeled data, so transfer learning or self-supervised pre-training should be introduced in small sample scenarios. [Conclusions] Through the innovation of the CNN-Transformer architecture and the optimization of positional encoding, this paper provides an efficient solution for the intelligent analysis of geomagnetic ultra-low frequency wave signals, holding significant promise for playing an important role in fields such as geophysical monitoring and space weather early warning.

Key words: geomagnetic ultra-low frequency waves, geoformer, signal classification