A frame-difference-aware rockburst early warning method based on the spatiotemporal distribution of microseismicity...

Under deep underground coal mining conditions, the spatiotemporal evolution of the microseismic energy density cloud map reflects the accumulation, instability, and release of stress in the surrounding rock, and thus provides a critical precursor for dynamic rockburst prediction and intelligent early warning. However, most existing approaches focus on single-point or low-dimensional microseismic parameters and fail to capture the complex spatiotemporal dynamics jointly driven by geological structures and mining-induced disturbances. To address this limitation, we propose a Frame-Difference-aware Spatiotemporal Fusion Network (FDSF-Net), a fully spatiotemporal framework for microseismic energy density cloud map prediction and rockburst early warning in underground coal mines. FDSF-Net encodes sequential energy density cloud maps to extract explicit spatiotemporal variation features, and incorporates a channel attention mechanism that injects dynamic disturbance information derived from adjacent frame differences into the original feature space, thereby enhancing sensitivity to abrupt energy changes and localized anomalies. A multiscale spatial convolution module is introduced to capture local aggregation patterns under different receptive fields, while attention-based feature enhancement strengthens the responses in critical regions. A hierarchical convolutional fusion structure integrates shallow-, middle-, and deep-level spatiotemporal representations through adaptive attention weighting. Subsequently, a temporal-channel mixing convolution enables global information interaction, and a decoder reconstructs the future energy density cloud map distribution. Experiments on long-term microseismic monitoring data from the 40302 working face of a coal mine in Shaanxi Province demonstrate that FDSF-Net achieves an MAE of 109.75 and an RMSE of 3.53. Feature visualization confirms that the model effectively learns multiscale spatial patterns, and cross-mine validation verifies its robustness and generalization capability, supporting the development of accurate and interpretable intelligent rockburst early-warning systems.