Abstract
Abstract
Early warning of mine seismic events is critical for the safety of underground mining operations, yet remains challenging due to the high dimensionality and inherent noise of monitoring data, which obscure the latent dynamics preceding instability. Here, we propose a spatiotemporal dynamic prediction model (SDPM) that integrates an auto-reservoir neural network (ARNN) with spatiotemporal information (STI) transformation grounded in Takens’ embedding theorem. The framework converts simultaneous multi-sensor observations into low-dimensional temporal trajectories describing system evolution over a short future window. By analyzing representations in both observable and latent spaces, SDPM jointly detects anomalies and fluctuations associated with critical slowing down, enabling early identification of tipping points preceding strong seismic events. Tests on synthetic systems and field observations from three coal-mining faces show that SDPM consistently detects precursory changes before strong seismic events and outperforms representative forecasting and dedicated warning benchmarks. The results further reveal a trade-off between temporal specificity and detection sensitivity: shorter warning-association windows provide more temporally localized warnings, whereas longer windows capture earlier and more dispersed precursors. Together, these findings establish a data-driven framework for extracting early-warning signals from high-dimensional monitoring data and support its use as a decision-support tool for anticipating critical transitions in complex engineering systems.