Risk level early warning for coal gas compound dynamic disasters via multi-source information fusion and deep learning

Under deep mining conditions, the failure mechanisms of coal gas compound dynamic disasters are highly complex and involve strong multi-factor coupling, so accurate early warning is critical for ensuring safe mine production. This study proposes a deep learning-based early-warning approach driven by multi-source information fusion and develops an SCSSA-MSDA-TFT time-series intelligent early-warning model. In this framework, an Sine-Cosine and Cauchy-enhanced Sparrow Search Algorithm (SCSSA) is employed to adaptively optimize the model hyperparameters, Multi-Source Domain Adaptation (MSDA) is introduced to align the distributions of heterogeneous monitoring data and achieve unified feature representation, and a Temporal Fusion Transformer (TFT) is used to efficiently extract the dynamic evolutionary characteristics of multi-source time-series indicators, thereby enabling risk-level early warning. For multi-source information such as microseismic monitoring and gas-related parameters, a data-driven risk-level calibration procedure for coal gas compound dynamic disasters is constructed. Taking a Composite Risk Index (CRI) as the core, the CRI series is temporally smoothed, and the high-risk threshold is determined based on Receiver Operating Characteristic (ROC) curve analysis. Cluster validity evaluation is then used to assess the consistency between the calibrated risk levels and the intrinsic data structure. Furthermore, a compound dynamic disaster early-warning indicator system is established: an XGBoost multi-class baseline model is trained and global Shapley Additive Explanations (SHAP) importance is computed, which, combined with sliding-window robustness checks and subset selection criteria, yields a compact indicator subset that balances physical interpretability and discriminative efficiency. The results show that the proposed model achieves a macro-averaged F1-score of 0.965 and an accuracy of 0.961 on the test set, significantly outperforming the comparison and ablation models. The model can accurately capture multi-scale precursory signals of coal gas compound dynamic disasters and realize precise prediction and early warning of risk levels. The proposed deep learning fusion early-warning approach effectively integrates multi-source information and establishes a coherent risk-level calibration and indicator system, offering substantial engineering application value for improving the accuracy and reliability of risk-level early warning for coal gas compound dynamic disasters.