Scholarlink

<p>As the depth of coal mining increases, seismic events pose a serious threat to mine safety, and studying the rupture mechanism helps to better understand the occurrence of mine-induced earthquakes. In this paper, we propose a new model, Graph-Fully Convolutional Focal Mechanism Network (GFFMNet), based on Graph Convolutional Networks (GCN) and Fully Convolutional Networks (FCN), for predicting the focal mechanisms of mine-induced seismic events. The model combines the spatial feature aggregation capabilities of GCN with the feature extraction abilities of FCN. Using monitoring data from the Dongtan Coal Mine in Shandong, China, we generated 210,000 samples and trained the model by minimizing the mean squared error (MSE) loss function. Experimental results show that GFFMNet significantly improves inference speed while maintaining prediction accuracy, with an average prediction time of only 0.12&#xa0;s per event and strong generalization capabilities with low error. Compared to the P-wave first-arrival polarity inversion method, GFFMNet demonstrates advantages in both information utilization and accuracy, with potential applications in real-time seismic analysis. Future research will focus on further optimizing the model to enhance prediction reliability.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

GFFMNet: Prediction Method for Coal Mine Seismic Focal Mechanisms Based on GCN

  • 期刊论文

  • 发表日期：

    2025年1月8日

查看全文
  • Kai Zhan,
  • Xigui Zheng,
  • Rui Xu,
  • Cong Wang,
  • Ping Song

摘要

As the depth of coal mining increases, seismic events pose a serious threat to mine safety, and studying the rupture mechanism helps to better understand the occurrence of mine-induced earthquakes. In this paper, we propose a new model, Graph-Fully Convolutional Focal Mechanism Network (GFFMNet), based on Graph Convolutional Networks (GCN) and Fully Convolutional Networks (FCN), for predicting the focal mechanisms of mine-induced seismic events. The model combines the spatial feature aggregation capabilities of GCN with the feature extraction abilities of FCN. Using monitoring data from the Dongtan Coal Mine in Shandong, China, we generated 210,000 samples and trained the model by minimizing the mean squared error (MSE) loss function. Experimental results show that GFFMNet significantly improves inference speed while maintaining prediction accuracy, with an average prediction time of only 0.12 s per event and strong generalization capabilities with low error. Compared to the P-wave first-arrival polarity inversion method, GFFMNet demonstrates advantages in both information utilization and accuracy, with potential applications in real-time seismic analysis. Future research will focus on further optimizing the model to enhance prediction reliability.