Research on multi-factor adaptive integrated early warning method for coal mine disaster risks based on multi-task learning

The reliable early warning of risks associated with gas, fire, dust, and roof hazards is crucial for the safe mining of coal mines. Traditional warning methods suffer from singular disaster risk warnings, low integration of risk information across different indicators, and insufficient perception of multi-hazard coupling relationships. To address these challenges, this paper proposes a method for adaptive integration of risk warnings that quantitatively learns the relationships between various indicators and warning tasks. Anomaly-transformer and E2GAN models are first employed to detect anomalies and impute missing values in time-series data. Subsequently, an improved MMoE model is used for multi-indicator fusion and prediction, allowing the simultaneous forecasting of future trends for all early-warning indicators. Finally, an adaptive multi-hazard risk integration warning model is developed, utilizing original and predicted data to calculate the current and future risk probabilities for various hazards. Comprehensive risk identification and warning are then performed using a multi-hazard grading identification. Experimental results show that the improved MMoE model outperforms LSTNet and TCN in prediction accuracy, and the integration model exceeds CNN and GRU in warning performance. Field validation confirms that this approach effectively identifies risks and enhances the reliability of intelligent early warning systems, ensuring coal mining safety.