TSF-Mlstm: A Spatiotemporal Feature-Enhanced Monitoring Method for Coalmine Pressure Prediction

Accurate and reliable monitoring of coalmine pressure is critical for ensuring safe underground operations and preventing catastrophic roof collapse. However, coalmine pressure signals are inherently complex and highly variable, often affected by geological heterogeneity, mining disturbances, and sensor noise, leading to significant challenges in acquiring stable and interpretable measurements. In this study, we present a novel measurement enhancement framework for coalmine pressure monitoring, which augments traditional hydraulic support sensors with intelligent signal modeling and spatiotemporal feature alignment mechanisms. Unlike conventional approaches that focus solely on raw sensor outputs, our method acts as a data-driven post-measurement refinement module, designed to mitigate noise interference, model temporal-spatial dependencies, and provide real-time prediction feedback for early warning systems. The proposed framework integrates a multidimensional measurement modeling strategy, in which a Mamba-based long-sequence module is fused with a Long Short-Term Memory (LSTM) network to form a hybrid signal processing pipeline. This architecture enables the system to learn deep temporal-spatial patterns embedded in noisy and nonstationary measurement sequences, thus enhancing data stability, reliability, and interpretability. Rather than replacing traditional sensors, our method complements and extends their functions by enabling intelligent post-measurement analysis and predictive forecasting, thereby reinforcing the practical utility of existing instrumentation in harsh underground environments. Experimental validation on real-world coalmine hydraulic support monitoring datasets shows that the proposed framework consistently outperforms traditional measurementdriven forecasting baselines across various prediction horizons. Specifically, the proposed TSF-Mlstm achieves relative MAE improvements of 8.19%, 9.67%, 8.00%, 10.60%, 12.52%, and 12.05% over the baseline methods. In addition, compared with three recent hybrid Mamba–LSTM variants, it further reduces the average MAE by 8.22%, 11.55%, and 14.55%, demonstrating its stronger long-range forecasting stability. These quantitative improvements demonstrate the superior predictive stability of our system under complex and noisy underground conditions. The proposed work thus contributes a viable direction toward intelligent coalmine measurement enhancement and supports early risk detection through advanced sensor data modeling.