Abstract: This study proposes a deep learning-based risk assessment method to address challenges in data fusion from multiple heterogeneous sources when evaluating water inrush hazards in coal mines with high-pressure aquifers. The method implements a network architecture of feature extraction-temporal analysis-risk assessment. It employs attention mechanisms for efficient extraction of geological structure features, utilizes LSTM networks to capture temporal evolution patterns of hydrogeological parameters, and applies fully connected layers for risk level classification. Experimental results demonstrate a risk prediction accuracy of 93.2%, meeting real-time early warning requirements.