A Sensor Data Prediction and Early-Warning Method for Coal Mining Faces Based on the MTGNN-Bayesian-IF-DBSCAN...

Area of Science:

• Intelligent Systems
• Data Science
• Mining Engineering

Background:

• Coal mine safety monitoring requires accurate gas concentration prediction and anomaly detection.
• Existing methods struggle with complex spatiotemporal dependencies and adaptive anomaly identification.

Purpose of the Study:

• To propose an integrated prediction and early-warning method for intelligent coal mine safety.
• To enhance gas concentration prediction accuracy and anomaly detection reliability.

Main Methods:

• Utilized Multi-Task Graph Neural Network (MTGNN) to model spatiotemporal gas concentration and wind speed data.
• Employed Bayesian optimization to adaptively fuse Isolation Forest (IF) and Density-Based Spatial Clustering of Applications with Noise (DBSCAN) for anomaly detection.
• Integrated original concentration and residual features for robust anomaly identification.

Main Results:

• MTGNN achieved high-precision gas concentration prediction (MAE as low as 0.00237, RMSE < 0.0203).
• The fused anomaly detection method achieved an F1 score of 1.0 with optimal weights (w_if=0.43, w_dbscan=0.52).
• Gas anomaly detection rate reached 93-96% with a false alarm rate below 5%.

Conclusions:

• The MTGNN-Bayesian-IF-DBSCAN framework offers a novel technical pathway for intelligent gas warning in coal mines.
• The study demonstrates enhanced accuracy, reliability, and robustness in gas prediction and anomaly detection.
• The proposed method shows potential for generalization to other industrial sensor network applications.