Related Experiment Videos
A self-supervised GNN-Transformer framework for weak microseismic signal identification
Mingwei Liu1, Zhigang Deng2, Yunpeng Li3
1School of Emergency Management and Safety Engineering, China University of Mining and Technology (Beijing), Beijing, 100083, China.
Scientific Reports
|May 25, 2026
Summary
A new self-supervised GNN-Transformer framework (SS-GNN-MSTF-Transformer) accurately identifies weak microseismic signals in deep mining. This robust method enhances fracture monitoring and safety assessments by overcoming complex noise challenges.
Area of Science:
- Geophysics
- Data Science
- Mining Engineering
Background:
- Weak microseismic signals are crucial for evaluating hydraulic fracturing effectiveness and ensuring mining safety.
- Deep mining environments present challenges like strong disturbances, electromagnetic interference, and complex noise, reducing signal-to-noise ratios.
- Traditional identification methods struggle with weak signals in noisy, complex mining conditions.
Purpose of the Study:
- To propose a robust framework for identifying weak microseismic signals in challenging deep mining environments.
- To enhance the accuracy, robustness, and generalization ability of microseismic signal recognition.
- To provide a reliable solution for hydraulic fracturing monitoring, fracture propagation analysis, and safety assessment.
Main Methods:
- Developed a self-supervised GNN-Transformer framework (SS-GNN-MSTF-Transformer).
- Integrated a multi-scale spatiotemporal feature extraction module (MSTF) using STFT and multi-scale convolution.
- Employed a Transformer encoder with RoPE and multi-scale attention for temporal dependencies and an adaptive GNN for spatial correlations.
- Utilized a self-supervised learning strategy including contrastive learning, masked time-frequency prediction, and graph structure recovery.
Main Results:
- Achieved an average accuracy of 0.9343 and F1-score under five-fold cross-validation on real microseismic data.
- Demonstrated superior performance over conventional methods (SVM, KNN, RF, BPNN).
- Engineering validation on independent field datasets yielded F1-scores of 0.9341 and 0.9216, showing significant improvement over RF.
- Maintained a low inference latency of approximately 14.9 ms per sample, indicating feasibility for real-time applications.
Conclusions:
- The proposed SS-GNN-MSTF-Transformer framework significantly improves weak microseismic signal identification accuracy, robustness, and generalization.
- The method effectively handles complex noise and data limitations in deep mining environments.
- The framework offers a reliable and efficient solution for critical applications in hydraulic fracturing and mining safety.
