TL;DR
This paper introduces a novel approach to predict and reconstruct abnormal mining pressure in longwall faces, particularly when encountering residual coal pillars. By combining clustering analysis with a CEEMDAN-Transformer deep learning model, the method accurately identifies and categorizes pressure anomalies, offering crucial early warnings for enhanced mine safety.
What Problem Does This Paper Address?
In multi-seam coal mining, lower seam operations often pass over previously mined areas (goafs) and leftover coal pillars from upper seams. This geological complexity can lead to unpredictable and severe "abnormal mining pressure" events. Current monitoring systems struggle to accurately identify these pressures, posing significant risks to safety and production, including equipment damage and potential accidents. The paper aims to bridge this gap by developing a more reliable prediction and early warning system for these hazardous conditions.
Methods
The core of the proposed method involves several key steps:
- Mechanism Analysis: The study first analyzes the physical mechanisms behind abnormal pressure, specifically how residual coal pillars and overlying rock layer instability contribute to stress transfer and dynamic loads.
- Data Preprocessing: Raw hydraulic support resistance data, which reflects mine pressure, is preprocessed. This includes bilinear interpolation to fill in missing data points and CEEMDAN (Complementary Ensemble Empirical Mode Decomposition with Adaptive Noise) for denoising and decomposing the time series data into intrinsic mode functions (IMFs).
- Deep Learning Prediction: A Transformer model with a multi-head attention mechanism is employed to predict the mining pressure time series (MPTS) based on the denoised IMF components. This architecture is chosen for its ability to capture complex temporal dependencies.
- Clustering Analysis: The k-means++ clustering algorithm is used to classify different types of abnormal mining pressure based on the predicted pressure data. This helps in categorizing anomalies like "cutting eye," "residual coal pillar," and "square breaking."
- Reconstruction: Finally, the clustering results are used to reconstruct and visualize the abnormal pressure areas, providing a spatial understanding of the predicted risks.
Datasets
The study utilizes historical hydraulic support resistance data from the 31317 mining face at the Chahasu coal mine. This dataset, comprising 1483 working resistance data points, serves as the basis for training and evaluating the proposed prediction and clustering models.
Key Insights and Ideas
- Residual Coal Pillars as Stress Concentrators: The paper clearly demonstrates how residual coal pillars act as significant stress concentrators, transferring loads from upper seams and causing dynamic and static stress superposition that leads to abnormal pressure in lower seams.
- Dynamic vs. Static Pressure: A crucial distinction is made between static pressure (before entering a pillar area) and dynamic pressure (during passage through a pillar area), highlighting the increased danger associated with the latter.
- Clustering for Anomaly Classification: The k-means++ algorithm proves effective in categorizing diverse abnormal pressure types, moving beyond simple pressure thresholds to understand the nature of the anomaly.
- Transformer's Predictive Power: The Transformer model, enhanced by CEEMDAN denoising, shows strong capability in predicting complex, non-stationary mining pressure time series, outperforming simpler methods.
- Integrated Approach is Key: The paper emphasizes that a combination of physical mechanism understanding, advanced data processing (interpolation, decomposition), deep learning prediction, and clustering analysis is essential for robust abnormal pressure prediction.
Key Results
- The developed method successfully identified six distinct types of abnormal mine pressure during a specific mining advance range (3921.4–5050.4 m).
- The prediction model achieved Mean Absolute Error (MAE) values lower than 2000 kN for hydraulic support working resistance predictions, indicating high accuracy.
- The model predicted one abnormal pressure event between hydraulic supports 80#–129# in the 5050.4–5219.5 m advance range, corresponding to the 18th square breaking area, which was verified as accurate through actual production data.
- A critical disaster criterion was established: when predicted hydraulic support working resistance exceeds 19,000 KN, immediate measures (like increasing advance speed) are recommended to mitigate risks.
Limitations and Future Work
- Data Availability: The paper notes that the specific mining data used cannot be made publicly available due to its sensitive nature, which limits broader replication and benchmarking.
- Generalizability: While validated on one mine, further testing on diverse geological conditions and mining scenarios would enhance confidence in the model's generalizability.
- Real-time Implementation: The paper focuses on prediction and reconstruction. Future work could explore seamless integration into real-time monitoring and control systems for immediate response.
- Further Refinement of Clustering: While k-means++ was effective, exploring other advanced clustering techniques or ensemble methods could potentially further refine the classification of pressure types.
Summary updated