An explainable integrated early-warning method for coal and gas outburst risk based on multimodal transfer learning

Reliable, interpretable early warning of coal and gas outbursts remains challenging because monitoring data are noisy, temporal indicators and non-temporal parameters are heterogeneous, outburst samples are sparse, and geological conditions vary substantially. Existing statistical and deep-learning methods struggle to capture coupled precursor relationships among acoustic emission (AE), electromagnetic radiation (EMR), and gas signals under noisy, heterogeneous cross-scenario conditions, and usually provide probability-based warnings without mechanism-consistent explanations. To address these limitations, an integrated framework is developed. First, an adaptive denoising method that combines the discrete Fourier transform, the whale optimization algorithm, and hierarchical density-based spatial clustering of applications with noise (HDBSCAN) is proposed to suppress interference while preserving AE and EMR dynamics. Statistical analyses identify precursor response characteristics preceding gas-dynamic events. Second, a TimeCMA-based multimodal forecasting model using cross-modality alignment predicts the short- and medium-term evolution of AE, EMR, and gas signals. Using field monitoring data, it achieves mean absolute error values of 0.0312, 0.3775, and 0.4152 and mean squared error values of 0.0035, 1.7002, and 0.7448 at horizons of 32, 64, and 128 steps, respectively. Third, a transfer-learning-based early-warning model embeds multi-scale temporal statistical features into structured prompts. Forecasting representations are transferred to early-warning classification to mitigate the scarcity and imbalance of labeled outburst-risk samples. The model achieves an accuracy of 0.976, an F1 score of 0.972, and an area under the receiver operating characteristic curve of 0.968, and field validation in coal mines in Guizhou and Sichuan confirms generalization. Finally, a coal-gas-outburst knowledge graph is constructed from standards, literature, and case reports. By integrating Shapley additive explanations (SHAP)-derived temporal attributions with non-temporal seam parameters, a graph retrieval-augmented generation (GraphRAG)-enhanced lightweight large language model generates evidence-grounded and mechanism-aligned interpretations of risk evolution. Comparative evaluation shows that the knowledge-graph-enhanced model provides higher causal consistency, domain specificity, and scenario adaptability than the baseline model. Overall, the proposed framework provides a deployable paradigm for proactive coal and gas outburst warning under limited labeled samples, delivering accurate forecasting, robust early warning, and mechanism-consistent interpretation through precursor representation transfer, model attribution, and decision support informed by domain knowledge.