深度学习方法在地震事件分类中的应用及可解释性研究
作者:
作者单位:

作者简介:

路晓辰(1995-),男,在读硕士,主要从事地震学和人工智能研究。E-mail:lxc73731515@163.com

通讯作者:

基金项目:

中国地震局地震预测研究所基本科研业务专项(2020IESLZ01);甘肃省科技计划(21JR7RA790,21YF5FA031)


Application and interpretability of deep learning methods in seismic event classification
Author:
Affiliation:

Fund Project:

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
    摘要:

    采用2016—2020年福建台网所记录的爆破和天然地震事件以及背景噪声数据集,使用CNN模型、Inception10模型、ResNet18模型和Vgg16模型4种深度学习网络模型进行分类研究。针对深度学习网络模型的“黑盒”问题,将梯度类激活映射(Gradient- weighted Class Activation Mapping,Grad- CAM)算法引入这4种分类模型中,得到每个模型的可视化图。通过可视化图可以直观地看出模型在做出分类决策时对于不同波形特征的依赖权重,为模型的可解释性提供依据,进而提高模型的可信度。通过对模型的可视化图分析得出,分类效果更好的CNN模型和Vgg16模型在做出决策时更依赖于地震波形的震相特征,对于震前和震后的波段关注较小;而ResNet18模型和Inception10模型对于震相特征的关注不够敏锐。通过Grad- CAM算法对模型进行可视化分析得到的结果能够很好地反映模型的分类效果,对于改进和选择合适的分类模型具有重要意义。

    Abstract:

    In this paper, four deep learning network models, i.e., the CNN, ResNet18, Vgg16, and Inception10 models, were used to classify blasting and seismic events, and the dataset used blasting events and natural seismic events recorded by the Fujian station network from 2016 to 2020. The gradient- weighted class activation mapping (Grad- CAM) algorithm was introduced into the four classification models to address the “black box” problem of deep learning network models, and a visualization of each model was obtained. The visualization diagram provides an intuitive view of the model's reliance weights for different waveform features when making classification decisions, thus providing a basis for the model’s interpretability and improving its credibility. Analysis of the visualization diagram shows that the CNN and Vgg16 models with better classification effects rely more on the seismic phase characteristics of seismic waveforms when making decisions and pay less attention to the pre- earthquake and post- earthquake bands. In contrast, the ResNet18 and Inception10 models are insufficiently sensitive to the seismic phase characteristics. The results obtained from a visual analysis of the models through the Grad- CAM algorithm well reflect the classification effect of the models, which is important for improving the models and selecting an appropriate classification model.

    参考文献
    相似文献
    引证文献
引用本文

路晓辰,杨立明,杨兴悦,王祖东,王维欢,高永国,尹欣欣.深度学习方法在地震事件分类中的应用及可解释性研究[J].地震工程学报,2023,(2):474-482. LU Xiaochen, YANG Liming, YANG Xingyue, WANG Zudong, WANG Weihuan, GAO Yongguo, YIN Xinxin. Application and interpretability of deep learning methods in seismic event classification[J]. China Earthquake Engineering Journal,2023,(2):474-482.

复制
分享
文章指标
  • 点击次数:
  • 下载次数:
历史
  • 收稿日期:
  • 最后修改日期:
  • 录用日期:
  • 在线发布日期: 2023-04-03