Scholarlink

<p>Weak microseismic signals generated during hydraulic fracturing contain critical information on fracture initiation and propagation, serving as a key basis for evaluating fracturing effectiveness and ensuring mining safety. However, in deep coal mining environments, strong mechanical disturbances, electromagnetic interference, and complex non-stationary noise lead to low signal-to-noise ratios and weak signal energy, posing significant challenges to traditional identification methods. To address these issues, a self-supervised GNN–Transformer framework (SS-GNN-MSTF-Transformer) is proposed for robust weak microseismic signal recognition. First, a multi-scale spatiotemporal feature extraction module (MSTF) is developed by integrating Short-Time Fourier Transform (STFT) with multi-scale convolution, enabling joint modeling of global and local time–frequency characteristics. Then, a Transformer encoder incorporating Rotary Position Encoding (RoPE) and multi-scale attention is employed to capture both long- and short-range temporal dependencies. Meanwhile, an adaptive graph neural network (GNN) is introduced to learn spatial correlations among monitoring stations in a data-driven manner, facilitating effective multi-station information fusion. Furthermore, a self-supervised learning strategy combining contrastive learning, masked time–frequency prediction, and graph structure recovery is designed to enhance feature representation under limited labeled data conditions. Experimental results based on real microseismic monitoring datasets demonstrate that the proposed method achieves an average accuracy and F1-score of 0.9343 under five-fold cross-validation, outperforming conventional methods such as SVM, KNN, RF, and BPNN. In addition, engineering validation on two independent field datasets shows that the model achieves F1-scores of 0.9341 and 0.9216, respectively, with improvements of 7.17% and 6.79% compared with the best traditional method (RF), confirming its strong generalization ability under different geological conditions and noise environments. The model also maintains an inference latency of approximately 14.9 ms per sample, demonstrating its feasibility for real-time monitoring applications. Overall, the proposed framework significantly improves the accuracy, robustness, and generalization ability of weak microseismic signal identification, providing a reliable and efficient solution for hydraulic fracturing monitoring, fracture propagation analysis, and safety assessment in deep mining environments.</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

A self-supervised GNN–Transformer framework for weak microseismic signal identification

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月25日

查看全文
  • Mingwei Liu,
  • Zhigang Deng,
  • Yunpeng Li,
  • Zhijia Zhang,
  • Tianyu Fu

摘要

Weak microseismic signals generated during hydraulic fracturing contain critical information on fracture initiation and propagation, serving as a key basis for evaluating fracturing effectiveness and ensuring mining safety. However, in deep coal mining environments, strong mechanical disturbances, electromagnetic interference, and complex non-stationary noise lead to low signal-to-noise ratios and weak signal energy, posing significant challenges to traditional identification methods. To address these issues, a self-supervised GNN–Transformer framework (SS-GNN-MSTF-Transformer) is proposed for robust weak microseismic signal recognition. First, a multi-scale spatiotemporal feature extraction module (MSTF) is developed by integrating Short-Time Fourier Transform (STFT) with multi-scale convolution, enabling joint modeling of global and local time–frequency characteristics. Then, a Transformer encoder incorporating Rotary Position Encoding (RoPE) and multi-scale attention is employed to capture both long- and short-range temporal dependencies. Meanwhile, an adaptive graph neural network (GNN) is introduced to learn spatial correlations among monitoring stations in a data-driven manner, facilitating effective multi-station information fusion. Furthermore, a self-supervised learning strategy combining contrastive learning, masked time–frequency prediction, and graph structure recovery is designed to enhance feature representation under limited labeled data conditions. Experimental results based on real microseismic monitoring datasets demonstrate that the proposed method achieves an average accuracy and F1-score of 0.9343 under five-fold cross-validation, outperforming conventional methods such as SVM, KNN, RF, and BPNN. In addition, engineering validation on two independent field datasets shows that the model achieves F1-scores of 0.9341 and 0.9216, respectively, with improvements of 7.17% and 6.79% compared with the best traditional method (RF), confirming its strong generalization ability under different geological conditions and noise environments. The model also maintains an inference latency of approximately 14.9 ms per sample, demonstrating its feasibility for real-time monitoring applications. Overall, the proposed framework significantly improves the accuracy, robustness, and generalization ability of weak microseismic signal identification, providing a reliable and efficient solution for hydraulic fracturing monitoring, fracture propagation analysis, and safety assessment in deep mining environments.