Interpretable Graph-Temporal-Spectral Fusion for Precursor-Related Anomaly Detection in Underground Mine Sensor Networks.

Underground coal mine safety monitoring relies on multi-source sensor networks, but abnormal-state detection remains challenging because methane, airflow, dust, and equipment-operation signals are non-stationary, heterogeneous, and constrained by ventilation and mining disturbances. This study proposes GasNet, an interpretable engineering-prior graph-temporal-spectral framework for precursor-related anomaly detection. Variables are organized into methane-related core sensors and environmental-operational modulation sensors. A directed sensor graph is constructed using ventilation causality, sensor deployment, and shearer-coupling relationships. GasNet then integrates a graph convolutional network for spatial-topological modeling, TimesNet for temporal-spectral pattern extraction, and cross-attention for adaptive feature fusion. An unsupervised reconstruction strategy identifies intervals deviating from learned normal production patterns. Field validation was conducted on the 31002 fully mechanized working face of the Xinyuan Coal Mine, where eight precursor-related abnormal intervals were annotated from monitoring data and field records. GasNet achieved a Precision of 0.881, a Recall of 1.000, an F1-score of 0.937, a false-alarm rate of 0.0017, and zero missed detections, with the highest F1-score among seven time-series baselines. Interpretability analysis further provided feature-fusion and sensor-time evidence for warning review. These results support the feasibility of GasNet for interpretable anomaly detection in the investigated working face.