Mechanisms of strong mining pressure under high-intensity mining are complex, and precursors are often weak. To address this, an instability probability prediction method is proposed. This method is based on the coupled evolution mechanism of "unsteady energy accumulation" and "system structural ordering." First, Singular Spectrum Analysis (SSA) and Wigner-Ville Distribution (WVD) are used to capture time-frequency characteristics of unsteady energy release. Recurrence Plot (RP) analysis quantifies dynamic entropy changes during the rock system''s transition from disorder to order. A spatiotemporal-dynamic precursor indicator system is thus constructed. Second, Logistic regression validates the significant predictive capability of these physical indicators. Guided by these mechanisms, an improved 1D Convolutional Neural Network (1D-CNN) with a channel attention mechanism is developed. This architecture translates physical inductive bias into endogenous network constraints. By employing physics-aware attention to adaptively recalibrate feature weights, the model achieves manifold decoupling and precise extraction of critical frequency bands and dynamic modes. Results reveal distinct instability signatures: a negative shift in frequency centroid and a drop in recurrence entropy. Comparative tests against LSTM, GRU, and SVM confirm the improved 1D-CNN''s significant advantages in feature capture and efficiency. The model achieves an overall accuracy of 92.2%. It correctly identifies 100% of strong mining pressure events, providing an average warning time of 68.7±12.5 minutes. This work offers a novel theoretical and technical approach for real-time, graded early warning of dynamic mining hazards.