Abstract: To address the problems of strong nonlinear coupling and significant spatial heterogeneity in multivariate time series data from coal mines, a risk early warning model for underground coal mine personnel integrating multi-source time series data was proposed. A multimodal data synchronization method based on a co-directional dual-pointer sliding window was adopted. Combined with Kalman filtering, a delay compensation mechanism was introduced to improve interpolation accuracy, thereby achieving high-precision time alignment of signals with different sampling frequencies. A ten-dimensional feature vector was constructed, and the SHAP method was utilized for global and local importance analysis to eliminate redundant features, achieving efficient dimensionality reduction. This significantly improved the interpretability and robustness of model decision-making while maintaining prediction performance. The Mixture of Attention Heads (MOA) mechanism was incorporated to capture the nonlinear dependencies and potential coupling features of multi-source signals. The MOA-Transformer model was constructed, where the Transformer encoder structure was used for feature-engineering-based risk level classification. The MOA was then employed to construct feature representations for classification. Field test results showed that the proposed model outperformed models such as recurrent neural networks and convolutional neural networks in terms of accuracy, precision, recall, and F₁-score. It could achieve high detection rates and low false alarm rates under conditions of few abnormal events, providing a feasible technical approach for risk identification and graded early warning for underground coal mine personnel.